]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
818ab71a | 2 | Copyright (C) 2003-2016 Free Software Foundation, Inc. |
b8698a0f | 3 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
4 | and Ira Rosen <irar@il.ibm.com> |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 3, or (at your option) any later | |
11 | version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
c7131fb2 | 25 | #include "backend.h" |
957060b5 AM |
26 | #include "target.h" |
27 | #include "rtl.h" | |
ebfd146a | 28 | #include "tree.h" |
c7131fb2 | 29 | #include "gimple.h" |
c7131fb2 | 30 | #include "ssa.h" |
957060b5 AM |
31 | #include "optabs-tree.h" |
32 | #include "insn-config.h" | |
33 | #include "recog.h" /* FIXME: for insn_data */ | |
34 | #include "cgraph.h" | |
957060b5 | 35 | #include "dumpfile.h" |
c7131fb2 | 36 | #include "alias.h" |
40e23961 | 37 | #include "fold-const.h" |
d8a2d370 | 38 | #include "stor-layout.h" |
2fb9a547 | 39 | #include "tree-eh.h" |
45b0be94 | 40 | #include "gimplify.h" |
5be5c238 | 41 | #include "gimple-iterator.h" |
18f429e2 | 42 | #include "gimplify-me.h" |
442b4905 | 43 | #include "tree-cfg.h" |
e28030cf | 44 | #include "tree-ssa-loop-manip.h" |
ebfd146a | 45 | #include "cfgloop.h" |
0136f8f0 AH |
46 | #include "tree-ssa-loop.h" |
47 | #include "tree-scalar-evolution.h" | |
ebfd146a | 48 | #include "tree-vectorizer.h" |
9b2b7279 | 49 | #include "builtins.h" |
70439f0d | 50 | #include "internal-fn.h" |
ebfd146a | 51 | |
7ee2468b SB |
52 | /* For lang_hooks.types.type_for_mode. */ |
53 | #include "langhooks.h" | |
ebfd146a | 54 | |
c3e7ee41 BS |
55 | /* Return the vectorized type for the given statement. */ |
56 | ||
57 | tree | |
58 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
59 | { | |
60 | return STMT_VINFO_VECTYPE (stmt_info); | |
61 | } | |
62 | ||
63 | /* Return TRUE iff the given statement is in an inner loop relative to | |
64 | the loop being vectorized. */ | |
65 | bool | |
66 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
67 | { | |
355fe088 | 68 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
c3e7ee41 BS |
69 | basic_block bb = gimple_bb (stmt); |
70 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
71 | struct loop* loop; | |
72 | ||
73 | if (!loop_vinfo) | |
74 | return false; | |
75 | ||
76 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
77 | ||
78 | return (bb->loop_father == loop->inner); | |
79 | } | |
80 | ||
81 | /* Record the cost of a statement, either by directly informing the | |
82 | target model or by saving it in a vector for later processing. | |
83 | Return a preliminary estimate of the statement's cost. */ | |
84 | ||
85 | unsigned | |
92345349 | 86 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 87 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 88 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 89 | { |
92345349 | 90 | if (body_cost_vec) |
c3e7ee41 | 91 | { |
92345349 | 92 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
ddf56386 RB |
93 | stmt_info_for_cost si = { count, kind, |
94 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
95 | misalign }; | |
96 | body_cost_vec->safe_push (si); | |
c3e7ee41 | 97 | return (unsigned) |
92345349 | 98 | (builtin_vectorization_cost (kind, vectype, misalign) * count); |
c3e7ee41 BS |
99 | } |
100 | else | |
310213d4 RB |
101 | return add_stmt_cost (stmt_info->vinfo->target_cost_data, |
102 | count, kind, stmt_info, misalign, where); | |
c3e7ee41 BS |
103 | } |
104 | ||
272c6793 RS |
105 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
106 | ||
107 | static tree | |
108 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
109 | { | |
110 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
111 | "vect_array"); | |
112 | } | |
113 | ||
114 | /* ARRAY is an array of vectors created by create_vector_array. | |
115 | Return an SSA_NAME for the vector in index N. The reference | |
116 | is part of the vectorization of STMT and the vector is associated | |
117 | with scalar destination SCALAR_DEST. */ | |
118 | ||
119 | static tree | |
355fe088 | 120 | read_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree scalar_dest, |
272c6793 RS |
121 | tree array, unsigned HOST_WIDE_INT n) |
122 | { | |
123 | tree vect_type, vect, vect_name, array_ref; | |
355fe088 | 124 | gimple *new_stmt; |
272c6793 RS |
125 | |
126 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
127 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
128 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
129 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
130 | build_int_cst (size_type_node, n), | |
131 | NULL_TREE, NULL_TREE); | |
132 | ||
133 | new_stmt = gimple_build_assign (vect, array_ref); | |
134 | vect_name = make_ssa_name (vect, new_stmt); | |
135 | gimple_assign_set_lhs (new_stmt, vect_name); | |
136 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
137 | |
138 | return vect_name; | |
139 | } | |
140 | ||
141 | /* ARRAY is an array of vectors created by create_vector_array. | |
142 | Emit code to store SSA_NAME VECT in index N of the array. | |
143 | The store is part of the vectorization of STMT. */ | |
144 | ||
145 | static void | |
355fe088 | 146 | write_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree vect, |
272c6793 RS |
147 | tree array, unsigned HOST_WIDE_INT n) |
148 | { | |
149 | tree array_ref; | |
355fe088 | 150 | gimple *new_stmt; |
272c6793 RS |
151 | |
152 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
153 | build_int_cst (size_type_node, n), | |
154 | NULL_TREE, NULL_TREE); | |
155 | ||
156 | new_stmt = gimple_build_assign (array_ref, vect); | |
157 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
158 | } |
159 | ||
160 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
161 | of *PTR. The memory reference replaces those in FIRST_DR | |
162 | (and its group). */ | |
163 | ||
164 | static tree | |
165 | create_array_ref (tree type, tree ptr, struct data_reference *first_dr) | |
166 | { | |
272c6793 RS |
167 | tree mem_ref, alias_ptr_type; |
168 | ||
169 | alias_ptr_type = reference_alias_ptr_type (DR_REF (first_dr)); | |
170 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); | |
171 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 172 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
173 | return mem_ref; |
174 | } | |
175 | ||
ebfd146a IR |
176 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
177 | ||
178 | /* Function vect_mark_relevant. | |
179 | ||
180 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
181 | ||
182 | static void | |
355fe088 | 183 | vect_mark_relevant (vec<gimple *> *worklist, gimple *stmt, |
83197f37 IR |
184 | enum vect_relevant relevant, bool live_p, |
185 | bool used_in_pattern) | |
ebfd146a IR |
186 | { |
187 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
188 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
189 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
355fe088 | 190 | gimple *pattern_stmt; |
ebfd146a | 191 | |
73fbfcad | 192 | if (dump_enabled_p ()) |
78c60e3d | 193 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 194 | "mark relevant %d, live %d.\n", relevant, live_p); |
ebfd146a | 195 | |
83197f37 IR |
196 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
197 | related pattern stmt instead of the original stmt. However, such stmts | |
198 | may have their own uses that are not in any pattern, in such cases the | |
199 | stmt itself should be marked. */ | |
ebfd146a IR |
200 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
201 | { | |
83197f37 IR |
202 | bool found = false; |
203 | if (!used_in_pattern) | |
204 | { | |
205 | imm_use_iterator imm_iter; | |
206 | use_operand_p use_p; | |
355fe088 | 207 | gimple *use_stmt; |
83197f37 | 208 | tree lhs; |
13c931c9 JJ |
209 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
210 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a | 211 | |
83197f37 IR |
212 | if (is_gimple_assign (stmt)) |
213 | lhs = gimple_assign_lhs (stmt); | |
214 | else | |
215 | lhs = gimple_call_lhs (stmt); | |
ebfd146a | 216 | |
83197f37 IR |
217 | /* This use is out of pattern use, if LHS has other uses that are |
218 | pattern uses, we should mark the stmt itself, and not the pattern | |
219 | stmt. */ | |
5ce9450f | 220 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
ab0ef706 JJ |
221 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs) |
222 | { | |
223 | if (is_gimple_debug (USE_STMT (use_p))) | |
224 | continue; | |
225 | use_stmt = USE_STMT (use_p); | |
226 | ||
13c931c9 JJ |
227 | if (!flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) |
228 | continue; | |
229 | ||
ab0ef706 JJ |
230 | if (vinfo_for_stmt (use_stmt) |
231 | && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (use_stmt))) | |
232 | { | |
233 | found = true; | |
234 | break; | |
235 | } | |
236 | } | |
83197f37 IR |
237 | } |
238 | ||
239 | if (!found) | |
240 | { | |
241 | /* This is the last stmt in a sequence that was detected as a | |
242 | pattern that can potentially be vectorized. Don't mark the stmt | |
243 | as relevant/live because it's not going to be vectorized. | |
244 | Instead mark the pattern-stmt that replaces it. */ | |
245 | ||
246 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
247 | ||
73fbfcad | 248 | if (dump_enabled_p ()) |
78c60e3d SS |
249 | dump_printf_loc (MSG_NOTE, vect_location, |
250 | "last stmt in pattern. don't mark" | |
e645e942 | 251 | " relevant/live.\n"); |
83197f37 IR |
252 | stmt_info = vinfo_for_stmt (pattern_stmt); |
253 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
254 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
255 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
256 | stmt = pattern_stmt; | |
257 | } | |
ebfd146a IR |
258 | } |
259 | ||
260 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
261 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
262 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
263 | ||
264 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
265 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
266 | { | |
73fbfcad | 267 | if (dump_enabled_p ()) |
78c60e3d | 268 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 269 | "already marked relevant/live.\n"); |
ebfd146a IR |
270 | return; |
271 | } | |
272 | ||
9771b263 | 273 | worklist->safe_push (stmt); |
ebfd146a IR |
274 | } |
275 | ||
276 | ||
277 | /* Function vect_stmt_relevant_p. | |
278 | ||
279 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
280 | "relevant for vectorization". | |
281 | ||
282 | A stmt is considered "relevant for vectorization" if: | |
283 | - it has uses outside the loop. | |
284 | - it has vdefs (it alters memory). | |
285 | - control stmts in the loop (except for the exit condition). | |
286 | ||
287 | CHECKME: what other side effects would the vectorizer allow? */ | |
288 | ||
289 | static bool | |
355fe088 | 290 | vect_stmt_relevant_p (gimple *stmt, loop_vec_info loop_vinfo, |
ebfd146a IR |
291 | enum vect_relevant *relevant, bool *live_p) |
292 | { | |
293 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
294 | ssa_op_iter op_iter; | |
295 | imm_use_iterator imm_iter; | |
296 | use_operand_p use_p; | |
297 | def_operand_p def_p; | |
298 | ||
8644a673 | 299 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
300 | *live_p = false; |
301 | ||
302 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
303 | if (is_ctrl_stmt (stmt) |
304 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
305 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 306 | *relevant = vect_used_in_scope; |
ebfd146a IR |
307 | |
308 | /* changing memory. */ | |
309 | if (gimple_code (stmt) != GIMPLE_PHI) | |
ac6aeab4 RB |
310 | if (gimple_vdef (stmt) |
311 | && !gimple_clobber_p (stmt)) | |
ebfd146a | 312 | { |
73fbfcad | 313 | if (dump_enabled_p ()) |
78c60e3d | 314 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 315 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 316 | *relevant = vect_used_in_scope; |
ebfd146a IR |
317 | } |
318 | ||
319 | /* uses outside the loop. */ | |
320 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
321 | { | |
322 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
323 | { | |
324 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
325 | if (!flow_bb_inside_loop_p (loop, bb)) | |
326 | { | |
73fbfcad | 327 | if (dump_enabled_p ()) |
78c60e3d | 328 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 329 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 330 | |
3157b0c2 AO |
331 | if (is_gimple_debug (USE_STMT (use_p))) |
332 | continue; | |
333 | ||
ebfd146a IR |
334 | /* We expect all such uses to be in the loop exit phis |
335 | (because of loop closed form) */ | |
336 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
337 | gcc_assert (bb == single_exit (loop)->dest); | |
338 | ||
339 | *live_p = true; | |
340 | } | |
341 | } | |
342 | } | |
343 | ||
344 | return (*live_p || *relevant); | |
345 | } | |
346 | ||
347 | ||
b8698a0f | 348 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 349 | |
ff802fa1 | 350 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
351 | used in STMT for anything other than indexing an array. */ |
352 | ||
353 | static bool | |
355fe088 | 354 | exist_non_indexing_operands_for_use_p (tree use, gimple *stmt) |
ebfd146a IR |
355 | { |
356 | tree operand; | |
357 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 358 | |
ff802fa1 | 359 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
360 | reference in STMT, then any operand that corresponds to USE |
361 | is not indexing an array. */ | |
362 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
363 | return true; | |
59a05b0c | 364 | |
ebfd146a IR |
365 | /* STMT has a data_ref. FORNOW this means that its of one of |
366 | the following forms: | |
367 | -1- ARRAY_REF = var | |
368 | -2- var = ARRAY_REF | |
369 | (This should have been verified in analyze_data_refs). | |
370 | ||
371 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 372 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
373 | for array indexing. |
374 | ||
375 | Therefore, all we need to check is if STMT falls into the | |
376 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
377 | |
378 | if (!gimple_assign_copy_p (stmt)) | |
5ce9450f JJ |
379 | { |
380 | if (is_gimple_call (stmt) | |
381 | && gimple_call_internal_p (stmt)) | |
382 | switch (gimple_call_internal_fn (stmt)) | |
383 | { | |
384 | case IFN_MASK_STORE: | |
385 | operand = gimple_call_arg (stmt, 3); | |
386 | if (operand == use) | |
387 | return true; | |
388 | /* FALLTHRU */ | |
389 | case IFN_MASK_LOAD: | |
390 | operand = gimple_call_arg (stmt, 2); | |
391 | if (operand == use) | |
392 | return true; | |
393 | break; | |
394 | default: | |
395 | break; | |
396 | } | |
397 | return false; | |
398 | } | |
399 | ||
59a05b0c EB |
400 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
401 | return false; | |
ebfd146a | 402 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
403 | if (TREE_CODE (operand) != SSA_NAME) |
404 | return false; | |
405 | ||
406 | if (operand == use) | |
407 | return true; | |
408 | ||
409 | return false; | |
410 | } | |
411 | ||
412 | ||
b8698a0f | 413 | /* |
ebfd146a IR |
414 | Function process_use. |
415 | ||
416 | Inputs: | |
417 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b8698a0f | 418 | - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt |
ff802fa1 | 419 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 420 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
421 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
422 | be performed. | |
ebfd146a IR |
423 | |
424 | Outputs: | |
425 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
426 | relevance info of the DEF_STMT of this USE: | |
427 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
428 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
429 | Exceptions: | |
430 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 431 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 432 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
433 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
434 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
435 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
436 | be modified accordingly. | |
437 | ||
438 | Return true if everything is as expected. Return false otherwise. */ | |
439 | ||
440 | static bool | |
355fe088 TS |
441 | process_use (gimple *stmt, tree use, loop_vec_info loop_vinfo, bool live_p, |
442 | enum vect_relevant relevant, vec<gimple *> *worklist, | |
aec7ae7d | 443 | bool force) |
ebfd146a IR |
444 | { |
445 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
446 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
447 | stmt_vec_info dstmt_vinfo; | |
448 | basic_block bb, def_bb; | |
355fe088 | 449 | gimple *def_stmt; |
ebfd146a IR |
450 | enum vect_def_type dt; |
451 | ||
b8698a0f | 452 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 453 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 454 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
455 | return true; |
456 | ||
81c40241 | 457 | if (!vect_is_simple_use (use, loop_vinfo, &def_stmt, &dt)) |
b8698a0f | 458 | { |
73fbfcad | 459 | if (dump_enabled_p ()) |
78c60e3d | 460 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 461 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
462 | return false; |
463 | } | |
464 | ||
465 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
466 | return true; | |
467 | ||
468 | def_bb = gimple_bb (def_stmt); | |
469 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
470 | { | |
73fbfcad | 471 | if (dump_enabled_p ()) |
e645e942 | 472 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n"); |
ebfd146a IR |
473 | return true; |
474 | } | |
475 | ||
b8698a0f L |
476 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
477 | DEF_STMT must have already been processed, because this should be the | |
478 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
479 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
480 | check that everything is as expected, and we are done. */ |
481 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
482 | bb = gimple_bb (stmt); | |
483 | if (gimple_code (stmt) == GIMPLE_PHI | |
484 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
485 | && gimple_code (def_stmt) != GIMPLE_PHI | |
486 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
487 | && bb->loop_father == def_bb->loop_father) | |
488 | { | |
73fbfcad | 489 | if (dump_enabled_p ()) |
78c60e3d | 490 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 491 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a IR |
492 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
493 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
494 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 495 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 496 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
497 | return true; |
498 | } | |
499 | ||
500 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
501 | outer-loop-header-bb: | |
502 | d = def_stmt | |
503 | inner-loop: | |
504 | stmt # use (d) | |
505 | outer-loop-tail-bb: | |
506 | ... */ | |
507 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
508 | { | |
73fbfcad | 509 | if (dump_enabled_p ()) |
78c60e3d | 510 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 511 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 512 | |
ebfd146a IR |
513 | switch (relevant) |
514 | { | |
8644a673 | 515 | case vect_unused_in_scope: |
7c5222ff IR |
516 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
517 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 518 | break; |
7c5222ff | 519 | |
ebfd146a | 520 | case vect_used_in_outer_by_reduction: |
7c5222ff | 521 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
522 | relevant = vect_used_by_reduction; |
523 | break; | |
7c5222ff | 524 | |
ebfd146a | 525 | case vect_used_in_outer: |
7c5222ff | 526 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 527 | relevant = vect_used_in_scope; |
ebfd146a | 528 | break; |
7c5222ff | 529 | |
8644a673 | 530 | case vect_used_in_scope: |
ebfd146a IR |
531 | break; |
532 | ||
533 | default: | |
534 | gcc_unreachable (); | |
b8698a0f | 535 | } |
ebfd146a IR |
536 | } |
537 | ||
538 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
539 | outer-loop-header-bb: | |
540 | ... | |
541 | inner-loop: | |
542 | d = def_stmt | |
06066f92 | 543 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
544 | stmt # use (d) */ |
545 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
546 | { | |
73fbfcad | 547 | if (dump_enabled_p ()) |
78c60e3d | 548 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 549 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 550 | |
ebfd146a IR |
551 | switch (relevant) |
552 | { | |
8644a673 | 553 | case vect_unused_in_scope: |
b8698a0f | 554 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 555 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 556 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
557 | break; |
558 | ||
ebfd146a IR |
559 | case vect_used_by_reduction: |
560 | relevant = vect_used_in_outer_by_reduction; | |
561 | break; | |
562 | ||
8644a673 | 563 | case vect_used_in_scope: |
ebfd146a IR |
564 | relevant = vect_used_in_outer; |
565 | break; | |
566 | ||
567 | default: | |
568 | gcc_unreachable (); | |
569 | } | |
570 | } | |
571 | ||
83197f37 IR |
572 | vect_mark_relevant (worklist, def_stmt, relevant, live_p, |
573 | is_pattern_stmt_p (stmt_vinfo)); | |
ebfd146a IR |
574 | return true; |
575 | } | |
576 | ||
577 | ||
578 | /* Function vect_mark_stmts_to_be_vectorized. | |
579 | ||
580 | Not all stmts in the loop need to be vectorized. For example: | |
581 | ||
582 | for i... | |
583 | for j... | |
584 | 1. T0 = i + j | |
585 | 2. T1 = a[T0] | |
586 | ||
587 | 3. j = j + 1 | |
588 | ||
589 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
590 | addressing of vectorized data-refs are handled differently. | |
591 | ||
592 | This pass detects such stmts. */ | |
593 | ||
594 | bool | |
595 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
596 | { | |
ebfd146a IR |
597 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
598 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
599 | unsigned int nbbs = loop->num_nodes; | |
600 | gimple_stmt_iterator si; | |
355fe088 | 601 | gimple *stmt; |
ebfd146a IR |
602 | unsigned int i; |
603 | stmt_vec_info stmt_vinfo; | |
604 | basic_block bb; | |
355fe088 | 605 | gimple *phi; |
ebfd146a | 606 | bool live_p; |
06066f92 IR |
607 | enum vect_relevant relevant, tmp_relevant; |
608 | enum vect_def_type def_type; | |
ebfd146a | 609 | |
73fbfcad | 610 | if (dump_enabled_p ()) |
78c60e3d | 611 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 612 | "=== vect_mark_stmts_to_be_vectorized ===\n"); |
ebfd146a | 613 | |
355fe088 | 614 | auto_vec<gimple *, 64> worklist; |
ebfd146a IR |
615 | |
616 | /* 1. Init worklist. */ | |
617 | for (i = 0; i < nbbs; i++) | |
618 | { | |
619 | bb = bbs[i]; | |
620 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 621 | { |
ebfd146a | 622 | phi = gsi_stmt (si); |
73fbfcad | 623 | if (dump_enabled_p ()) |
ebfd146a | 624 | { |
78c60e3d SS |
625 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
626 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
627 | } |
628 | ||
629 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 630 | vect_mark_relevant (&worklist, phi, relevant, live_p, false); |
ebfd146a IR |
631 | } |
632 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
633 | { | |
634 | stmt = gsi_stmt (si); | |
73fbfcad | 635 | if (dump_enabled_p ()) |
ebfd146a | 636 | { |
78c60e3d SS |
637 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
638 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
b8698a0f | 639 | } |
ebfd146a IR |
640 | |
641 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 642 | vect_mark_relevant (&worklist, stmt, relevant, live_p, false); |
ebfd146a IR |
643 | } |
644 | } | |
645 | ||
646 | /* 2. Process_worklist */ | |
9771b263 | 647 | while (worklist.length () > 0) |
ebfd146a IR |
648 | { |
649 | use_operand_p use_p; | |
650 | ssa_op_iter iter; | |
651 | ||
9771b263 | 652 | stmt = worklist.pop (); |
73fbfcad | 653 | if (dump_enabled_p ()) |
ebfd146a | 654 | { |
78c60e3d SS |
655 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
656 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
ebfd146a IR |
657 | } |
658 | ||
b8698a0f L |
659 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
660 | (DEF_STMT) as relevant/irrelevant and live/dead according to the | |
ebfd146a IR |
661 | liveness and relevance properties of STMT. */ |
662 | stmt_vinfo = vinfo_for_stmt (stmt); | |
663 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
664 | live_p = STMT_VINFO_LIVE_P (stmt_vinfo); | |
665 | ||
666 | /* Generally, the liveness and relevance properties of STMT are | |
667 | propagated as is to the DEF_STMTs of its USEs: | |
668 | live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO) | |
669 | relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO) | |
670 | ||
671 | One exception is when STMT has been identified as defining a reduction | |
672 | variable; in this case we set the liveness/relevance as follows: | |
673 | live_p = false | |
674 | relevant = vect_used_by_reduction | |
675 | This is because we distinguish between two kinds of relevant stmts - | |
b8698a0f | 676 | those that are used by a reduction computation, and those that are |
ff802fa1 | 677 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 678 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 679 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 680 | |
06066f92 IR |
681 | def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
682 | tmp_relevant = relevant; | |
683 | switch (def_type) | |
ebfd146a | 684 | { |
06066f92 IR |
685 | case vect_reduction_def: |
686 | switch (tmp_relevant) | |
687 | { | |
688 | case vect_unused_in_scope: | |
689 | relevant = vect_used_by_reduction; | |
690 | break; | |
691 | ||
692 | case vect_used_by_reduction: | |
693 | if (gimple_code (stmt) == GIMPLE_PHI) | |
694 | break; | |
695 | /* fall through */ | |
696 | ||
697 | default: | |
73fbfcad | 698 | if (dump_enabled_p ()) |
78c60e3d | 699 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 700 | "unsupported use of reduction.\n"); |
06066f92 IR |
701 | return false; |
702 | } | |
703 | ||
b8698a0f | 704 | live_p = false; |
06066f92 | 705 | break; |
b8698a0f | 706 | |
06066f92 IR |
707 | case vect_nested_cycle: |
708 | if (tmp_relevant != vect_unused_in_scope | |
709 | && tmp_relevant != vect_used_in_outer_by_reduction | |
710 | && tmp_relevant != vect_used_in_outer) | |
711 | { | |
73fbfcad | 712 | if (dump_enabled_p ()) |
78c60e3d | 713 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 714 | "unsupported use of nested cycle.\n"); |
7c5222ff | 715 | |
06066f92 IR |
716 | return false; |
717 | } | |
7c5222ff | 718 | |
b8698a0f L |
719 | live_p = false; |
720 | break; | |
721 | ||
06066f92 IR |
722 | case vect_double_reduction_def: |
723 | if (tmp_relevant != vect_unused_in_scope | |
724 | && tmp_relevant != vect_used_by_reduction) | |
725 | { | |
73fbfcad | 726 | if (dump_enabled_p ()) |
78c60e3d | 727 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 728 | "unsupported use of double reduction.\n"); |
7c5222ff | 729 | |
7c5222ff | 730 | return false; |
06066f92 IR |
731 | } |
732 | ||
733 | live_p = false; | |
b8698a0f | 734 | break; |
7c5222ff | 735 | |
06066f92 IR |
736 | default: |
737 | break; | |
7c5222ff | 738 | } |
b8698a0f | 739 | |
aec7ae7d | 740 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
741 | { |
742 | /* Pattern statements are not inserted into the code, so | |
743 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
744 | have to scan the RHS or function arguments instead. */ | |
745 | if (is_gimple_assign (stmt)) | |
746 | { | |
69d2aade JJ |
747 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
748 | tree op = gimple_assign_rhs1 (stmt); | |
749 | ||
750 | i = 1; | |
751 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
752 | { | |
753 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
aec7ae7d | 754 | live_p, relevant, &worklist, false) |
69d2aade | 755 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
aec7ae7d | 756 | live_p, relevant, &worklist, false)) |
566d377a | 757 | return false; |
69d2aade JJ |
758 | i = 2; |
759 | } | |
760 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 761 | { |
69d2aade | 762 | op = gimple_op (stmt, i); |
afbe6325 RB |
763 | if (TREE_CODE (op) == SSA_NAME |
764 | && !process_use (stmt, op, loop_vinfo, live_p, relevant, | |
765 | &worklist, false)) | |
07687835 | 766 | return false; |
9d5e7640 IR |
767 | } |
768 | } | |
769 | else if (is_gimple_call (stmt)) | |
770 | { | |
771 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
772 | { | |
773 | tree arg = gimple_call_arg (stmt, i); | |
774 | if (!process_use (stmt, arg, loop_vinfo, live_p, relevant, | |
aec7ae7d | 775 | &worklist, false)) |
07687835 | 776 | return false; |
9d5e7640 IR |
777 | } |
778 | } | |
779 | } | |
780 | else | |
781 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
782 | { | |
783 | tree op = USE_FROM_PTR (use_p); | |
784 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, | |
aec7ae7d | 785 | &worklist, false)) |
07687835 | 786 | return false; |
9d5e7640 | 787 | } |
aec7ae7d | 788 | |
3bab6342 | 789 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo)) |
aec7ae7d JJ |
790 | { |
791 | tree off; | |
3bab6342 | 792 | tree decl = vect_check_gather_scatter (stmt, loop_vinfo, NULL, &off, NULL); |
aec7ae7d JJ |
793 | gcc_assert (decl); |
794 | if (!process_use (stmt, off, loop_vinfo, live_p, relevant, | |
795 | &worklist, true)) | |
566d377a | 796 | return false; |
aec7ae7d | 797 | } |
ebfd146a IR |
798 | } /* while worklist */ |
799 | ||
ebfd146a IR |
800 | return true; |
801 | } | |
802 | ||
803 | ||
b8698a0f | 804 | /* Function vect_model_simple_cost. |
ebfd146a | 805 | |
b8698a0f | 806 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
807 | single op. Right now, this does not account for multiple insns that could |
808 | be generated for the single vector op. We will handle that shortly. */ | |
809 | ||
810 | void | |
b8698a0f | 811 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 BS |
812 | enum vect_def_type *dt, |
813 | stmt_vector_for_cost *prologue_cost_vec, | |
814 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
815 | { |
816 | int i; | |
92345349 | 817 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
818 | |
819 | /* The SLP costs were already calculated during SLP tree build. */ | |
820 | if (PURE_SLP_STMT (stmt_info)) | |
821 | return; | |
822 | ||
ebfd146a IR |
823 | /* FORNOW: Assuming maximum 2 args per stmts. */ |
824 | for (i = 0; i < 2; i++) | |
92345349 BS |
825 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
826 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, vector_stmt, | |
827 | stmt_info, 0, vect_prologue); | |
c3e7ee41 BS |
828 | |
829 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
92345349 BS |
830 | inside_cost = record_stmt_cost (body_cost_vec, ncopies, vector_stmt, |
831 | stmt_info, 0, vect_body); | |
c3e7ee41 | 832 | |
73fbfcad | 833 | if (dump_enabled_p ()) |
78c60e3d SS |
834 | dump_printf_loc (MSG_NOTE, vect_location, |
835 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 836 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
837 | } |
838 | ||
839 | ||
8bd37302 BS |
840 | /* Model cost for type demotion and promotion operations. PWR is normally |
841 | zero for single-step promotions and demotions. It will be one if | |
842 | two-step promotion/demotion is required, and so on. Each additional | |
843 | step doubles the number of instructions required. */ | |
844 | ||
845 | static void | |
846 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
847 | enum vect_def_type *dt, int pwr) | |
848 | { | |
849 | int i, tmp; | |
92345349 | 850 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 BS |
851 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
852 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
853 | void *target_cost_data; | |
8bd37302 BS |
854 | |
855 | /* The SLP costs were already calculated during SLP tree build. */ | |
856 | if (PURE_SLP_STMT (stmt_info)) | |
857 | return; | |
858 | ||
c3e7ee41 BS |
859 | if (loop_vinfo) |
860 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
861 | else | |
862 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
863 | ||
8bd37302 BS |
864 | for (i = 0; i < pwr + 1; i++) |
865 | { | |
866 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
867 | (i + 1) : i; | |
c3e7ee41 | 868 | inside_cost += add_stmt_cost (target_cost_data, vect_pow2 (tmp), |
92345349 BS |
869 | vec_promote_demote, stmt_info, 0, |
870 | vect_body); | |
8bd37302 BS |
871 | } |
872 | ||
873 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
874 | for (i = 0; i < 2; i++) | |
92345349 BS |
875 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
876 | prologue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt, | |
877 | stmt_info, 0, vect_prologue); | |
8bd37302 | 878 | |
73fbfcad | 879 | if (dump_enabled_p ()) |
78c60e3d SS |
880 | dump_printf_loc (MSG_NOTE, vect_location, |
881 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 882 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
883 | } |
884 | ||
0d0293ac | 885 | /* Function vect_cost_group_size |
b8698a0f | 886 | |
0d0293ac | 887 | For grouped load or store, return the group_size only if it is the first |
ebfd146a IR |
888 | load or store of a group, else return 1. This ensures that group size is |
889 | only returned once per group. */ | |
890 | ||
891 | static int | |
0d0293ac | 892 | vect_cost_group_size (stmt_vec_info stmt_info) |
ebfd146a | 893 | { |
355fe088 | 894 | gimple *first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a IR |
895 | |
896 | if (first_stmt == STMT_VINFO_STMT (stmt_info)) | |
e14c1050 | 897 | return GROUP_SIZE (stmt_info); |
ebfd146a IR |
898 | |
899 | return 1; | |
900 | } | |
901 | ||
902 | ||
903 | /* Function vect_model_store_cost | |
904 | ||
0d0293ac MM |
905 | Models cost for stores. In the case of grouped accesses, one access |
906 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a IR |
907 | |
908 | void | |
b8698a0f | 909 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
272c6793 | 910 | bool store_lanes_p, enum vect_def_type dt, |
92345349 BS |
911 | slp_tree slp_node, |
912 | stmt_vector_for_cost *prologue_cost_vec, | |
913 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
914 | { |
915 | int group_size; | |
92345349 | 916 | unsigned int inside_cost = 0, prologue_cost = 0; |
720f5239 | 917 | struct data_reference *first_dr; |
355fe088 | 918 | gimple *first_stmt; |
ebfd146a | 919 | |
8644a673 | 920 | if (dt == vect_constant_def || dt == vect_external_def) |
92345349 BS |
921 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
922 | stmt_info, 0, vect_prologue); | |
ebfd146a | 923 | |
0d0293ac MM |
924 | /* Grouped access? */ |
925 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
720f5239 IR |
926 | { |
927 | if (slp_node) | |
928 | { | |
9771b263 | 929 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
720f5239 IR |
930 | group_size = 1; |
931 | } | |
932 | else | |
933 | { | |
e14c1050 | 934 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 935 | group_size = vect_cost_group_size (stmt_info); |
720f5239 IR |
936 | } |
937 | ||
938 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
939 | } | |
0d0293ac | 940 | /* Not a grouped access. */ |
ebfd146a | 941 | else |
720f5239 IR |
942 | { |
943 | group_size = 1; | |
944 | first_dr = STMT_VINFO_DATA_REF (stmt_info); | |
945 | } | |
ebfd146a | 946 | |
272c6793 | 947 | /* We assume that the cost of a single store-lanes instruction is |
0d0293ac | 948 | equivalent to the cost of GROUP_SIZE separate stores. If a grouped |
272c6793 RS |
949 | access is instead being provided by a permute-and-store operation, |
950 | include the cost of the permutes. */ | |
cee62fee MM |
951 | if (!store_lanes_p && group_size > 1 |
952 | && !STMT_VINFO_STRIDED_P (stmt_info)) | |
ebfd146a | 953 | { |
e1377713 ES |
954 | /* Uses a high and low interleave or shuffle operations for each |
955 | needed permute. */ | |
956 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; | |
92345349 BS |
957 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
958 | stmt_info, 0, vect_body); | |
ebfd146a | 959 | |
73fbfcad | 960 | if (dump_enabled_p ()) |
78c60e3d | 961 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 962 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 963 | group_size); |
ebfd146a IR |
964 | } |
965 | ||
cee62fee | 966 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a | 967 | /* Costs of the stores. */ |
cee62fee MM |
968 | if (STMT_VINFO_STRIDED_P (stmt_info) |
969 | && !STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
f2e2a985 MM |
970 | { |
971 | /* N scalar stores plus extracting the elements. */ | |
f2e2a985 MM |
972 | inside_cost += record_stmt_cost (body_cost_vec, |
973 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), | |
974 | scalar_store, stmt_info, 0, vect_body); | |
f2e2a985 MM |
975 | } |
976 | else | |
977 | vect_get_store_cost (first_dr, ncopies, &inside_cost, body_cost_vec); | |
ebfd146a | 978 | |
cee62fee MM |
979 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
980 | inside_cost += record_stmt_cost (body_cost_vec, | |
981 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), | |
982 | vec_to_scalar, stmt_info, 0, vect_body); | |
983 | ||
73fbfcad | 984 | if (dump_enabled_p ()) |
78c60e3d SS |
985 | dump_printf_loc (MSG_NOTE, vect_location, |
986 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 987 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
988 | } |
989 | ||
990 | ||
720f5239 IR |
991 | /* Calculate cost of DR's memory access. */ |
992 | void | |
993 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 994 | unsigned int *inside_cost, |
92345349 | 995 | stmt_vector_for_cost *body_cost_vec) |
720f5239 IR |
996 | { |
997 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 998 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 999 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
1000 | |
1001 | switch (alignment_support_scheme) | |
1002 | { | |
1003 | case dr_aligned: | |
1004 | { | |
92345349 BS |
1005 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
1006 | vector_store, stmt_info, 0, | |
1007 | vect_body); | |
720f5239 | 1008 | |
73fbfcad | 1009 | if (dump_enabled_p ()) |
78c60e3d | 1010 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1011 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
1012 | break; |
1013 | } | |
1014 | ||
1015 | case dr_unaligned_supported: | |
1016 | { | |
720f5239 | 1017 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1018 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1019 | unaligned_store, stmt_info, |
92345349 | 1020 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1021 | if (dump_enabled_p ()) |
78c60e3d SS |
1022 | dump_printf_loc (MSG_NOTE, vect_location, |
1023 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1024 | "hardware.\n"); |
720f5239 IR |
1025 | break; |
1026 | } | |
1027 | ||
38eec4c6 UW |
1028 | case dr_unaligned_unsupported: |
1029 | { | |
1030 | *inside_cost = VECT_MAX_COST; | |
1031 | ||
73fbfcad | 1032 | if (dump_enabled_p ()) |
78c60e3d | 1033 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1034 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1035 | break; |
1036 | } | |
1037 | ||
720f5239 IR |
1038 | default: |
1039 | gcc_unreachable (); | |
1040 | } | |
1041 | } | |
1042 | ||
1043 | ||
ebfd146a IR |
1044 | /* Function vect_model_load_cost |
1045 | ||
0d0293ac MM |
1046 | Models cost for loads. In the case of grouped accesses, the last access |
1047 | has the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1048 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1049 | access scheme chosen. */ |
1050 | ||
1051 | void | |
92345349 BS |
1052 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, |
1053 | bool load_lanes_p, slp_tree slp_node, | |
1054 | stmt_vector_for_cost *prologue_cost_vec, | |
1055 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
1056 | { |
1057 | int group_size; | |
355fe088 | 1058 | gimple *first_stmt; |
ebfd146a | 1059 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; |
92345349 | 1060 | unsigned int inside_cost = 0, prologue_cost = 0; |
ebfd146a | 1061 | |
0d0293ac | 1062 | /* Grouped accesses? */ |
e14c1050 | 1063 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 1064 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && first_stmt && !slp_node) |
ebfd146a | 1065 | { |
0d0293ac | 1066 | group_size = vect_cost_group_size (stmt_info); |
ebfd146a IR |
1067 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
1068 | } | |
0d0293ac | 1069 | /* Not a grouped access. */ |
ebfd146a IR |
1070 | else |
1071 | { | |
1072 | group_size = 1; | |
1073 | first_dr = dr; | |
1074 | } | |
1075 | ||
272c6793 | 1076 | /* We assume that the cost of a single load-lanes instruction is |
0d0293ac | 1077 | equivalent to the cost of GROUP_SIZE separate loads. If a grouped |
272c6793 RS |
1078 | access is instead being provided by a load-and-permute operation, |
1079 | include the cost of the permutes. */ | |
7b5fc413 | 1080 | if (!load_lanes_p && group_size > 1 |
f2e2a985 | 1081 | && !STMT_VINFO_STRIDED_P (stmt_info)) |
ebfd146a | 1082 | { |
2c23db6d ES |
1083 | /* Uses an even and odd extract operations or shuffle operations |
1084 | for each needed permute. */ | |
1085 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; | |
1086 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, | |
1087 | stmt_info, 0, vect_body); | |
ebfd146a | 1088 | |
73fbfcad | 1089 | if (dump_enabled_p ()) |
e645e942 TJ |
1090 | dump_printf_loc (MSG_NOTE, vect_location, |
1091 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1092 | group_size); |
ebfd146a IR |
1093 | } |
1094 | ||
1095 | /* The loads themselves. */ | |
f2e2a985 | 1096 | if (STMT_VINFO_STRIDED_P (stmt_info) |
7b5fc413 | 1097 | && !STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
a82960aa | 1098 | { |
a21892ad BS |
1099 | /* N scalar loads plus gathering them into a vector. */ |
1100 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
92345349 | 1101 | inside_cost += record_stmt_cost (body_cost_vec, |
c3e7ee41 | 1102 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), |
92345349 | 1103 | scalar_load, stmt_info, 0, vect_body); |
a82960aa RG |
1104 | } |
1105 | else | |
1106 | vect_get_load_cost (first_dr, ncopies, | |
1107 | ((!STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1108 | || group_size > 1 || slp_node), | |
92345349 BS |
1109 | &inside_cost, &prologue_cost, |
1110 | prologue_cost_vec, body_cost_vec, true); | |
f2e2a985 | 1111 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
7b5fc413 RB |
1112 | inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct, |
1113 | stmt_info, 0, vect_body); | |
720f5239 | 1114 | |
73fbfcad | 1115 | if (dump_enabled_p ()) |
78c60e3d SS |
1116 | dump_printf_loc (MSG_NOTE, vect_location, |
1117 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1118 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1119 | } |
1120 | ||
1121 | ||
1122 | /* Calculate cost of DR's memory access. */ | |
1123 | void | |
1124 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1125 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1126 | unsigned int *prologue_cost, |
1127 | stmt_vector_for_cost *prologue_cost_vec, | |
1128 | stmt_vector_for_cost *body_cost_vec, | |
1129 | bool record_prologue_costs) | |
720f5239 IR |
1130 | { |
1131 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 1132 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 1133 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
1134 | |
1135 | switch (alignment_support_scheme) | |
ebfd146a IR |
1136 | { |
1137 | case dr_aligned: | |
1138 | { | |
92345349 BS |
1139 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1140 | stmt_info, 0, vect_body); | |
ebfd146a | 1141 | |
73fbfcad | 1142 | if (dump_enabled_p ()) |
78c60e3d | 1143 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1144 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1145 | |
1146 | break; | |
1147 | } | |
1148 | case dr_unaligned_supported: | |
1149 | { | |
720f5239 | 1150 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1151 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1152 | unaligned_load, stmt_info, |
92345349 | 1153 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1154 | |
73fbfcad | 1155 | if (dump_enabled_p ()) |
78c60e3d SS |
1156 | dump_printf_loc (MSG_NOTE, vect_location, |
1157 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1158 | "hardware.\n"); |
ebfd146a IR |
1159 | |
1160 | break; | |
1161 | } | |
1162 | case dr_explicit_realign: | |
1163 | { | |
92345349 BS |
1164 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1165 | vector_load, stmt_info, 0, vect_body); | |
1166 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1167 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1168 | |
1169 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1170 | the containing loop, the following cost should be added to the | |
92345349 | 1171 | prologue costs. */ |
ebfd146a | 1172 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1173 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1174 | stmt_info, 0, vect_body); | |
ebfd146a | 1175 | |
73fbfcad | 1176 | if (dump_enabled_p ()) |
e645e942 TJ |
1177 | dump_printf_loc (MSG_NOTE, vect_location, |
1178 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1179 | |
ebfd146a IR |
1180 | break; |
1181 | } | |
1182 | case dr_explicit_realign_optimized: | |
1183 | { | |
73fbfcad | 1184 | if (dump_enabled_p ()) |
e645e942 | 1185 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1186 | "vect_model_load_cost: unaligned software " |
e645e942 | 1187 | "pipelined.\n"); |
ebfd146a IR |
1188 | |
1189 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1190 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1191 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1192 | access, then the above cost should only be considered for one |
ff802fa1 | 1193 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1194 | and a realignment op. */ |
1195 | ||
92345349 | 1196 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1197 | { |
92345349 BS |
1198 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1199 | vector_stmt, stmt_info, | |
1200 | 0, vect_prologue); | |
ebfd146a | 1201 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1202 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1203 | vector_stmt, stmt_info, | |
1204 | 0, vect_prologue); | |
ebfd146a IR |
1205 | } |
1206 | ||
92345349 BS |
1207 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1208 | stmt_info, 0, vect_body); | |
1209 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1210 | stmt_info, 0, vect_body); | |
8bd37302 | 1211 | |
73fbfcad | 1212 | if (dump_enabled_p ()) |
78c60e3d | 1213 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1214 | "vect_model_load_cost: explicit realign optimized" |
1215 | "\n"); | |
8bd37302 | 1216 | |
ebfd146a IR |
1217 | break; |
1218 | } | |
1219 | ||
38eec4c6 UW |
1220 | case dr_unaligned_unsupported: |
1221 | { | |
1222 | *inside_cost = VECT_MAX_COST; | |
1223 | ||
73fbfcad | 1224 | if (dump_enabled_p ()) |
78c60e3d | 1225 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1226 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1227 | break; |
1228 | } | |
1229 | ||
ebfd146a IR |
1230 | default: |
1231 | gcc_unreachable (); | |
1232 | } | |
ebfd146a IR |
1233 | } |
1234 | ||
418b7df3 RG |
1235 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1236 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1237 | |
418b7df3 | 1238 | static void |
355fe088 | 1239 | vect_init_vector_1 (gimple *stmt, gimple *new_stmt, gimple_stmt_iterator *gsi) |
ebfd146a | 1240 | { |
ebfd146a | 1241 | if (gsi) |
418b7df3 | 1242 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1243 | else |
1244 | { | |
418b7df3 | 1245 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1246 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1247 | |
a70d6342 IR |
1248 | if (loop_vinfo) |
1249 | { | |
1250 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1251 | basic_block new_bb; |
1252 | edge pe; | |
a70d6342 IR |
1253 | |
1254 | if (nested_in_vect_loop_p (loop, stmt)) | |
1255 | loop = loop->inner; | |
b8698a0f | 1256 | |
a70d6342 | 1257 | pe = loop_preheader_edge (loop); |
418b7df3 | 1258 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1259 | gcc_assert (!new_bb); |
1260 | } | |
1261 | else | |
1262 | { | |
1263 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1264 | basic_block bb; | |
1265 | gimple_stmt_iterator gsi_bb_start; | |
1266 | ||
1267 | gcc_assert (bb_vinfo); | |
1268 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1269 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1270 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1271 | } |
ebfd146a IR |
1272 | } |
1273 | ||
73fbfcad | 1274 | if (dump_enabled_p ()) |
ebfd146a | 1275 | { |
78c60e3d SS |
1276 | dump_printf_loc (MSG_NOTE, vect_location, |
1277 | "created new init_stmt: "); | |
1278 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
ebfd146a | 1279 | } |
418b7df3 RG |
1280 | } |
1281 | ||
1282 | /* Function vect_init_vector. | |
ebfd146a | 1283 | |
5467ee52 RG |
1284 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1285 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1286 | vector type a vector with all elements equal to VAL is created first. | |
1287 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1288 | initialization at the loop preheader. | |
418b7df3 RG |
1289 | Return the DEF of INIT_STMT. |
1290 | It will be used in the vectorization of STMT. */ | |
1291 | ||
1292 | tree | |
355fe088 | 1293 | vect_init_vector (gimple *stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 | 1294 | { |
355fe088 | 1295 | gimple *init_stmt; |
418b7df3 RG |
1296 | tree new_temp; |
1297 | ||
5467ee52 RG |
1298 | if (TREE_CODE (type) == VECTOR_TYPE |
1299 | && TREE_CODE (TREE_TYPE (val)) != VECTOR_TYPE) | |
418b7df3 | 1300 | { |
5467ee52 | 1301 | if (!types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) |
418b7df3 | 1302 | { |
5a308cf1 IE |
1303 | /* Scalar boolean value should be transformed into |
1304 | all zeros or all ones value before building a vector. */ | |
1305 | if (VECTOR_BOOLEAN_TYPE_P (type)) | |
1306 | { | |
b3d51f23 IE |
1307 | tree true_val = build_all_ones_cst (TREE_TYPE (type)); |
1308 | tree false_val = build_zero_cst (TREE_TYPE (type)); | |
5a308cf1 IE |
1309 | |
1310 | if (CONSTANT_CLASS_P (val)) | |
1311 | val = integer_zerop (val) ? false_val : true_val; | |
1312 | else | |
1313 | { | |
1314 | new_temp = make_ssa_name (TREE_TYPE (type)); | |
1315 | init_stmt = gimple_build_assign (new_temp, COND_EXPR, | |
1316 | val, true_val, false_val); | |
1317 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
1318 | val = new_temp; | |
1319 | } | |
1320 | } | |
1321 | else if (CONSTANT_CLASS_P (val)) | |
42fd8198 | 1322 | val = fold_convert (TREE_TYPE (type), val); |
418b7df3 RG |
1323 | else |
1324 | { | |
b731b390 | 1325 | new_temp = make_ssa_name (TREE_TYPE (type)); |
0d0e4a03 | 1326 | init_stmt = gimple_build_assign (new_temp, NOP_EXPR, val); |
418b7df3 | 1327 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1328 | val = new_temp; |
418b7df3 RG |
1329 | } |
1330 | } | |
5467ee52 | 1331 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1332 | } |
1333 | ||
0e22bb5a RB |
1334 | new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_"); |
1335 | init_stmt = gimple_build_assign (new_temp, val); | |
418b7df3 | 1336 | vect_init_vector_1 (stmt, init_stmt, gsi); |
0e22bb5a | 1337 | return new_temp; |
ebfd146a IR |
1338 | } |
1339 | ||
a70d6342 | 1340 | |
ebfd146a IR |
1341 | /* Function vect_get_vec_def_for_operand. |
1342 | ||
ff802fa1 | 1343 | OP is an operand in STMT. This function returns a (vector) def that will be |
ebfd146a IR |
1344 | used in the vectorized stmt for STMT. |
1345 | ||
1346 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1347 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1348 | ||
1349 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
42fd8198 IE |
1350 | needs to be introduced. VECTYPE may be used to specify a required type for |
1351 | vector invariant. */ | |
ebfd146a IR |
1352 | |
1353 | tree | |
42fd8198 | 1354 | vect_get_vec_def_for_operand (tree op, gimple *stmt, tree vectype) |
ebfd146a IR |
1355 | { |
1356 | tree vec_oprnd; | |
355fe088 TS |
1357 | gimple *vec_stmt; |
1358 | gimple *def_stmt; | |
ebfd146a IR |
1359 | stmt_vec_info def_stmt_info = NULL; |
1360 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
42fd8198 | 1361 | tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo); |
ebfd146a | 1362 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
ebfd146a IR |
1363 | enum vect_def_type dt; |
1364 | bool is_simple_use; | |
1365 | tree vector_type; | |
1366 | ||
73fbfcad | 1367 | if (dump_enabled_p ()) |
ebfd146a | 1368 | { |
78c60e3d SS |
1369 | dump_printf_loc (MSG_NOTE, vect_location, |
1370 | "vect_get_vec_def_for_operand: "); | |
1371 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
e645e942 | 1372 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1373 | } |
1374 | ||
81c40241 | 1375 | is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt); |
ebfd146a | 1376 | gcc_assert (is_simple_use); |
73fbfcad | 1377 | if (dump_enabled_p ()) |
ebfd146a | 1378 | { |
78c60e3d | 1379 | int loc_printed = 0; |
ebfd146a IR |
1380 | if (def_stmt) |
1381 | { | |
78c60e3d SS |
1382 | if (loc_printed) |
1383 | dump_printf (MSG_NOTE, " def_stmt = "); | |
1384 | else | |
1385 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1386 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
ebfd146a IR |
1387 | } |
1388 | } | |
1389 | ||
1390 | switch (dt) | |
1391 | { | |
81c40241 | 1392 | /* operand is a constant or a loop invariant. */ |
ebfd146a | 1393 | case vect_constant_def: |
81c40241 | 1394 | case vect_external_def: |
ebfd146a | 1395 | { |
42fd8198 IE |
1396 | if (vectype) |
1397 | vector_type = vectype; | |
1398 | else if (TREE_CODE (TREE_TYPE (op)) == BOOLEAN_TYPE | |
1399 | && VECTOR_BOOLEAN_TYPE_P (stmt_vectype)) | |
1400 | vector_type = build_same_sized_truth_vector_type (stmt_vectype); | |
1401 | else | |
1402 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
1403 | ||
7569a6cc | 1404 | gcc_assert (vector_type); |
418b7df3 | 1405 | return vect_init_vector (stmt, op, vector_type, NULL); |
ebfd146a IR |
1406 | } |
1407 | ||
81c40241 | 1408 | /* operand is defined inside the loop. */ |
8644a673 | 1409 | case vect_internal_def: |
ebfd146a | 1410 | { |
ebfd146a IR |
1411 | /* Get the def from the vectorized stmt. */ |
1412 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1413 | |
ebfd146a | 1414 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1415 | /* Get vectorized pattern statement. */ |
1416 | if (!vec_stmt | |
1417 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1418 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1419 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1420 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1421 | gcc_assert (vec_stmt); |
1422 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1423 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1424 | else if (is_gimple_call (vec_stmt)) | |
1425 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1426 | else | |
1427 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1428 | return vec_oprnd; | |
1429 | } | |
1430 | ||
81c40241 | 1431 | /* operand is defined by a loop header phi - reduction */ |
ebfd146a | 1432 | case vect_reduction_def: |
06066f92 | 1433 | case vect_double_reduction_def: |
7c5222ff | 1434 | case vect_nested_cycle: |
81c40241 RB |
1435 | /* Code should use get_initial_def_for_reduction. */ |
1436 | gcc_unreachable (); | |
ebfd146a | 1437 | |
81c40241 | 1438 | /* operand is defined by loop-header phi - induction. */ |
ebfd146a IR |
1439 | case vect_induction_def: |
1440 | { | |
1441 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1442 | ||
1443 | /* Get the def from the vectorized stmt. */ | |
1444 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1445 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1446 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1447 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1448 | else | |
1449 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1450 | return vec_oprnd; |
1451 | } | |
1452 | ||
1453 | default: | |
1454 | gcc_unreachable (); | |
1455 | } | |
1456 | } | |
1457 | ||
1458 | ||
1459 | /* Function vect_get_vec_def_for_stmt_copy | |
1460 | ||
ff802fa1 | 1461 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1462 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1463 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1464 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1465 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1466 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1467 | DT is the type of the vector def VEC_OPRND. |
1468 | ||
1469 | Context: | |
1470 | In case the vectorization factor (VF) is bigger than the number | |
1471 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1472 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1473 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1474 | smallest data-type determines the VF, and as a result, when vectorizing |
1475 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1476 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1477 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1478 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1479 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1480 | ||
1481 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1482 | |
ebfd146a IR |
1483 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1484 | VS1.1: vx.1 = memref1 VS1.2 | |
1485 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1486 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1487 | |
1488 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1489 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1490 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1491 | VSnew.3: vz3 = vx.3 + ... | |
1492 | ||
1493 | The vectorization of S1 is explained in vectorizable_load. | |
1494 | The vectorization of S2: | |
b8698a0f L |
1495 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1496 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1497 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1498 | returns the vector-def 'vx.0'. |
1499 | ||
b8698a0f L |
1500 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1501 | function is called to get the relevant vector-def for each operand. It is | |
1502 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1503 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1504 | ||
b8698a0f L |
1505 | For example, to obtain the vector-def 'vx.1' in order to create the |
1506 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1507 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1508 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1509 | and return its def ('vx.1'). | |
1510 | Overall, to create the above sequence this function will be called 3 times: | |
1511 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1512 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1513 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1514 | ||
1515 | tree | |
1516 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1517 | { | |
355fe088 | 1518 | gimple *vec_stmt_for_operand; |
ebfd146a IR |
1519 | stmt_vec_info def_stmt_info; |
1520 | ||
1521 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1522 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1523 | return vec_oprnd; |
1524 | ||
1525 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1526 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1527 | gcc_assert (def_stmt_info); | |
1528 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1529 | gcc_assert (vec_stmt_for_operand); | |
ebfd146a IR |
1530 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) |
1531 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1532 | else | |
1533 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1534 | return vec_oprnd; | |
1535 | } | |
1536 | ||
1537 | ||
1538 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1539 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a IR |
1540 | |
1541 | static void | |
b8698a0f | 1542 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1543 | vec<tree> *vec_oprnds0, |
1544 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1545 | { |
9771b263 | 1546 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1547 | |
1548 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1549 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1550 | |
9771b263 | 1551 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1552 | { |
9771b263 | 1553 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1554 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1555 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1556 | } |
1557 | } | |
1558 | ||
1559 | ||
d092494c IR |
1560 | /* Get vectorized definitions for OP0 and OP1. |
1561 | REDUC_INDEX is the index of reduction operand in case of reduction, | |
1562 | and -1 otherwise. */ | |
ebfd146a | 1563 | |
d092494c | 1564 | void |
355fe088 | 1565 | vect_get_vec_defs (tree op0, tree op1, gimple *stmt, |
9771b263 DN |
1566 | vec<tree> *vec_oprnds0, |
1567 | vec<tree> *vec_oprnds1, | |
d092494c | 1568 | slp_tree slp_node, int reduc_index) |
ebfd146a IR |
1569 | { |
1570 | if (slp_node) | |
d092494c IR |
1571 | { |
1572 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1573 | auto_vec<tree> ops (nops); |
1574 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1575 | |
9771b263 | 1576 | ops.quick_push (op0); |
d092494c | 1577 | if (op1) |
9771b263 | 1578 | ops.quick_push (op1); |
d092494c IR |
1579 | |
1580 | vect_get_slp_defs (ops, slp_node, &vec_defs, reduc_index); | |
1581 | ||
37b5ec8f | 1582 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1583 | if (op1) |
37b5ec8f | 1584 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1585 | } |
ebfd146a IR |
1586 | else |
1587 | { | |
1588 | tree vec_oprnd; | |
1589 | ||
9771b263 | 1590 | vec_oprnds0->create (1); |
81c40241 | 1591 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 1592 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1593 | |
1594 | if (op1) | |
1595 | { | |
9771b263 | 1596 | vec_oprnds1->create (1); |
81c40241 | 1597 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 1598 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1599 | } |
1600 | } | |
1601 | } | |
1602 | ||
1603 | ||
1604 | /* Function vect_finish_stmt_generation. | |
1605 | ||
1606 | Insert a new stmt. */ | |
1607 | ||
1608 | void | |
355fe088 | 1609 | vect_finish_stmt_generation (gimple *stmt, gimple *vec_stmt, |
ebfd146a IR |
1610 | gimple_stmt_iterator *gsi) |
1611 | { | |
1612 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
310213d4 | 1613 | vec_info *vinfo = stmt_info->vinfo; |
ebfd146a IR |
1614 | |
1615 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1616 | ||
54e8e2c3 RG |
1617 | if (!gsi_end_p (*gsi) |
1618 | && gimple_has_mem_ops (vec_stmt)) | |
1619 | { | |
355fe088 | 1620 | gimple *at_stmt = gsi_stmt (*gsi); |
54e8e2c3 RG |
1621 | tree vuse = gimple_vuse (at_stmt); |
1622 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1623 | { | |
1624 | tree vdef = gimple_vdef (at_stmt); | |
1625 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1626 | /* If we have an SSA vuse and insert a store, update virtual | |
1627 | SSA form to avoid triggering the renamer. Do so only | |
1628 | if we can easily see all uses - which is what almost always | |
1629 | happens with the way vectorized stmts are inserted. */ | |
1630 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1631 | && ((is_gimple_assign (vec_stmt) | |
1632 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1633 | || (is_gimple_call (vec_stmt) | |
1634 | && !(gimple_call_flags (vec_stmt) | |
1635 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1636 | { | |
1637 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1638 | gimple_set_vdef (vec_stmt, new_vdef); | |
1639 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1640 | } | |
1641 | } | |
1642 | } | |
ebfd146a IR |
1643 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1644 | ||
310213d4 | 1645 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, vinfo)); |
ebfd146a | 1646 | |
73fbfcad | 1647 | if (dump_enabled_p ()) |
ebfd146a | 1648 | { |
78c60e3d SS |
1649 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); |
1650 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
ebfd146a IR |
1651 | } |
1652 | ||
ad885386 | 1653 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
8e91d222 JJ |
1654 | |
1655 | /* While EH edges will generally prevent vectorization, stmt might | |
1656 | e.g. be in a must-not-throw region. Ensure newly created stmts | |
1657 | that could throw are part of the same region. */ | |
1658 | int lp_nr = lookup_stmt_eh_lp (stmt); | |
1659 | if (lp_nr != 0 && stmt_could_throw_p (vec_stmt)) | |
1660 | add_stmt_to_eh_lp (vec_stmt, lp_nr); | |
ebfd146a IR |
1661 | } |
1662 | ||
70439f0d RS |
1663 | /* We want to vectorize a call to combined function CFN with function |
1664 | decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN | |
1665 | as the types of all inputs. Check whether this is possible using | |
1666 | an internal function, returning its code if so or IFN_LAST if not. */ | |
ebfd146a | 1667 | |
70439f0d RS |
1668 | static internal_fn |
1669 | vectorizable_internal_function (combined_fn cfn, tree fndecl, | |
1670 | tree vectype_out, tree vectype_in) | |
ebfd146a | 1671 | { |
70439f0d RS |
1672 | internal_fn ifn; |
1673 | if (internal_fn_p (cfn)) | |
1674 | ifn = as_internal_fn (cfn); | |
1675 | else | |
1676 | ifn = associated_internal_fn (fndecl); | |
1677 | if (ifn != IFN_LAST && direct_internal_fn_p (ifn)) | |
1678 | { | |
1679 | const direct_internal_fn_info &info = direct_internal_fn (ifn); | |
1680 | if (info.vectorizable) | |
1681 | { | |
1682 | tree type0 = (info.type0 < 0 ? vectype_out : vectype_in); | |
1683 | tree type1 = (info.type1 < 0 ? vectype_out : vectype_in); | |
d95ab70a RS |
1684 | if (direct_internal_fn_supported_p (ifn, tree_pair (type0, type1), |
1685 | OPTIMIZE_FOR_SPEED)) | |
70439f0d RS |
1686 | return ifn; |
1687 | } | |
1688 | } | |
1689 | return IFN_LAST; | |
ebfd146a IR |
1690 | } |
1691 | ||
5ce9450f | 1692 | |
355fe088 | 1693 | static tree permute_vec_elements (tree, tree, tree, gimple *, |
5ce9450f JJ |
1694 | gimple_stmt_iterator *); |
1695 | ||
1696 | ||
1697 | /* Function vectorizable_mask_load_store. | |
1698 | ||
1699 | Check if STMT performs a conditional load or store that can be vectorized. | |
1700 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
1701 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. | |
1702 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1703 | ||
1704 | static bool | |
355fe088 TS |
1705 | vectorizable_mask_load_store (gimple *stmt, gimple_stmt_iterator *gsi, |
1706 | gimple **vec_stmt, slp_tree slp_node) | |
5ce9450f JJ |
1707 | { |
1708 | tree vec_dest = NULL; | |
1709 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1710 | stmt_vec_info prev_stmt_info; | |
1711 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1712 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1713 | bool nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
1714 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
1715 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
57e2f6ad | 1716 | tree rhs_vectype = NULL_TREE; |
045c1278 | 1717 | tree mask_vectype; |
5ce9450f | 1718 | tree elem_type; |
355fe088 | 1719 | gimple *new_stmt; |
5ce9450f JJ |
1720 | tree dummy; |
1721 | tree dataref_ptr = NULL_TREE; | |
355fe088 | 1722 | gimple *ptr_incr; |
5ce9450f JJ |
1723 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
1724 | int ncopies; | |
1725 | int i, j; | |
1726 | bool inv_p; | |
1727 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; | |
1728 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
1729 | int gather_scale = 1; | |
1730 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
1731 | bool is_store; | |
1732 | tree mask; | |
355fe088 | 1733 | gimple *def_stmt; |
5ce9450f JJ |
1734 | enum vect_def_type dt; |
1735 | ||
1736 | if (slp_node != NULL) | |
1737 | return false; | |
1738 | ||
1739 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
1740 | gcc_assert (ncopies >= 1); | |
1741 | ||
1742 | is_store = gimple_call_internal_fn (stmt) == IFN_MASK_STORE; | |
1743 | mask = gimple_call_arg (stmt, 2); | |
045c1278 IE |
1744 | |
1745 | if (TREE_CODE (TREE_TYPE (mask)) != BOOLEAN_TYPE) | |
5ce9450f JJ |
1746 | return false; |
1747 | ||
1748 | /* FORNOW. This restriction should be relaxed. */ | |
1749 | if (nested_in_vect_loop && ncopies > 1) | |
1750 | { | |
1751 | if (dump_enabled_p ()) | |
1752 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1753 | "multiple types in nested loop."); | |
1754 | return false; | |
1755 | } | |
1756 | ||
1757 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) | |
1758 | return false; | |
1759 | ||
1760 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
1761 | return false; | |
1762 | ||
1763 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
1764 | return false; | |
1765 | ||
1766 | elem_type = TREE_TYPE (vectype); | |
1767 | ||
1768 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1769 | return false; | |
1770 | ||
f2e2a985 | 1771 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
5ce9450f JJ |
1772 | return false; |
1773 | ||
045c1278 IE |
1774 | if (TREE_CODE (mask) != SSA_NAME) |
1775 | return false; | |
1776 | ||
1777 | if (!vect_is_simple_use (mask, loop_vinfo, &def_stmt, &dt, &mask_vectype)) | |
1778 | return false; | |
1779 | ||
1780 | if (!mask_vectype) | |
1781 | mask_vectype = get_mask_type_for_scalar_type (TREE_TYPE (vectype)); | |
1782 | ||
56e39820 | 1783 | if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype)) |
045c1278 IE |
1784 | return false; |
1785 | ||
57e2f6ad IE |
1786 | if (is_store) |
1787 | { | |
1788 | tree rhs = gimple_call_arg (stmt, 3); | |
1789 | if (!vect_is_simple_use (rhs, loop_vinfo, &def_stmt, &dt, &rhs_vectype)) | |
1790 | return false; | |
1791 | } | |
1792 | ||
3bab6342 | 1793 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5ce9450f | 1794 | { |
355fe088 | 1795 | gimple *def_stmt; |
3bab6342 | 1796 | gather_decl = vect_check_gather_scatter (stmt, loop_vinfo, &gather_base, |
5ce9450f JJ |
1797 | &gather_off, &gather_scale); |
1798 | gcc_assert (gather_decl); | |
81c40241 RB |
1799 | if (!vect_is_simple_use (gather_off, loop_vinfo, &def_stmt, &gather_dt, |
1800 | &gather_off_vectype)) | |
5ce9450f JJ |
1801 | { |
1802 | if (dump_enabled_p ()) | |
1803 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1804 | "gather index use not simple."); | |
1805 | return false; | |
1806 | } | |
03b9e8e4 JJ |
1807 | |
1808 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
1809 | tree masktype | |
1810 | = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (TREE_CHAIN (arglist)))); | |
1811 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
1812 | { | |
1813 | if (dump_enabled_p ()) | |
1814 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1815 | "masked gather with integer mask not supported."); | |
1816 | return false; | |
1817 | } | |
5ce9450f JJ |
1818 | } |
1819 | else if (tree_int_cst_compare (nested_in_vect_loop | |
1820 | ? STMT_VINFO_DR_STEP (stmt_info) | |
1821 | : DR_STEP (dr), size_zero_node) <= 0) | |
1822 | return false; | |
1823 | else if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
045c1278 IE |
1824 | || !can_vec_mask_load_store_p (TYPE_MODE (vectype), |
1825 | TYPE_MODE (mask_vectype), | |
57e2f6ad IE |
1826 | !is_store) |
1827 | || (rhs_vectype | |
1828 | && !useless_type_conversion_p (vectype, rhs_vectype))) | |
5ce9450f JJ |
1829 | return false; |
1830 | ||
5ce9450f JJ |
1831 | if (!vec_stmt) /* transformation not required. */ |
1832 | { | |
1833 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
1834 | if (is_store) | |
1835 | vect_model_store_cost (stmt_info, ncopies, false, dt, | |
1836 | NULL, NULL, NULL); | |
1837 | else | |
1838 | vect_model_load_cost (stmt_info, ncopies, false, NULL, NULL, NULL); | |
1839 | return true; | |
1840 | } | |
1841 | ||
1842 | /** Transform. **/ | |
1843 | ||
3bab6342 | 1844 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5ce9450f JJ |
1845 | { |
1846 | tree vec_oprnd0 = NULL_TREE, op; | |
1847 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
1848 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
acdcd61b | 1849 | tree ptr, vec_mask = NULL_TREE, mask_op = NULL_TREE, var, scale; |
5ce9450f | 1850 | tree perm_mask = NULL_TREE, prev_res = NULL_TREE; |
acdcd61b | 1851 | tree mask_perm_mask = NULL_TREE; |
5ce9450f JJ |
1852 | edge pe = loop_preheader_edge (loop); |
1853 | gimple_seq seq; | |
1854 | basic_block new_bb; | |
1855 | enum { NARROW, NONE, WIDEN } modifier; | |
1856 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
1857 | ||
acdcd61b JJ |
1858 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); |
1859 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1860 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1861 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1862 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1863 | scaletype = TREE_VALUE (arglist); | |
1864 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
1865 | && types_compatible_p (srctype, masktype)); | |
1866 | ||
5ce9450f JJ |
1867 | if (nunits == gather_off_nunits) |
1868 | modifier = NONE; | |
1869 | else if (nunits == gather_off_nunits / 2) | |
1870 | { | |
1871 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
1872 | modifier = WIDEN; | |
1873 | ||
1874 | for (i = 0; i < gather_off_nunits; ++i) | |
1875 | sel[i] = i | nunits; | |
1876 | ||
557be5a8 | 1877 | perm_mask = vect_gen_perm_mask_checked (gather_off_vectype, sel); |
5ce9450f JJ |
1878 | } |
1879 | else if (nunits == gather_off_nunits * 2) | |
1880 | { | |
1881 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
1882 | modifier = NARROW; | |
1883 | ||
1884 | for (i = 0; i < nunits; ++i) | |
1885 | sel[i] = i < gather_off_nunits | |
1886 | ? i : i + nunits - gather_off_nunits; | |
1887 | ||
557be5a8 | 1888 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); |
5ce9450f | 1889 | ncopies *= 2; |
acdcd61b JJ |
1890 | for (i = 0; i < nunits; ++i) |
1891 | sel[i] = i | gather_off_nunits; | |
557be5a8 | 1892 | mask_perm_mask = vect_gen_perm_mask_checked (masktype, sel); |
5ce9450f JJ |
1893 | } |
1894 | else | |
1895 | gcc_unreachable (); | |
1896 | ||
5ce9450f JJ |
1897 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); |
1898 | ||
1899 | ptr = fold_convert (ptrtype, gather_base); | |
1900 | if (!is_gimple_min_invariant (ptr)) | |
1901 | { | |
1902 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
1903 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
1904 | gcc_assert (!new_bb); | |
1905 | } | |
1906 | ||
1907 | scale = build_int_cst (scaletype, gather_scale); | |
1908 | ||
1909 | prev_stmt_info = NULL; | |
1910 | for (j = 0; j < ncopies; ++j) | |
1911 | { | |
1912 | if (modifier == WIDEN && (j & 1)) | |
1913 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
1914 | perm_mask, stmt, gsi); | |
1915 | else if (j == 0) | |
1916 | op = vec_oprnd0 | |
81c40241 | 1917 | = vect_get_vec_def_for_operand (gather_off, stmt); |
5ce9450f JJ |
1918 | else |
1919 | op = vec_oprnd0 | |
1920 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
1921 | ||
1922 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
1923 | { | |
1924 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
1925 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 1926 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
5ce9450f JJ |
1927 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
1928 | new_stmt | |
0d0e4a03 | 1929 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
1930 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1931 | op = var; | |
1932 | } | |
1933 | ||
acdcd61b JJ |
1934 | if (mask_perm_mask && (j & 1)) |
1935 | mask_op = permute_vec_elements (mask_op, mask_op, | |
1936 | mask_perm_mask, stmt, gsi); | |
5ce9450f JJ |
1937 | else |
1938 | { | |
acdcd61b | 1939 | if (j == 0) |
81c40241 | 1940 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); |
acdcd61b JJ |
1941 | else |
1942 | { | |
81c40241 | 1943 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
acdcd61b JJ |
1944 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
1945 | } | |
5ce9450f | 1946 | |
acdcd61b JJ |
1947 | mask_op = vec_mask; |
1948 | if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask))) | |
1949 | { | |
1950 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)) | |
1951 | == TYPE_VECTOR_SUBPARTS (masktype)); | |
0e22bb5a | 1952 | var = vect_get_new_ssa_name (masktype, vect_simple_var); |
acdcd61b JJ |
1953 | mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op); |
1954 | new_stmt | |
0d0e4a03 | 1955 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, mask_op); |
acdcd61b JJ |
1956 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1957 | mask_op = var; | |
1958 | } | |
5ce9450f JJ |
1959 | } |
1960 | ||
1961 | new_stmt | |
1962 | = gimple_build_call (gather_decl, 5, mask_op, ptr, op, mask_op, | |
1963 | scale); | |
1964 | ||
1965 | if (!useless_type_conversion_p (vectype, rettype)) | |
1966 | { | |
1967 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
1968 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 1969 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
5ce9450f JJ |
1970 | gimple_call_set_lhs (new_stmt, op); |
1971 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 1972 | var = make_ssa_name (vec_dest); |
5ce9450f | 1973 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
0d0e4a03 | 1974 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
1975 | } |
1976 | else | |
1977 | { | |
1978 | var = make_ssa_name (vec_dest, new_stmt); | |
1979 | gimple_call_set_lhs (new_stmt, var); | |
1980 | } | |
1981 | ||
1982 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1983 | ||
1984 | if (modifier == NARROW) | |
1985 | { | |
1986 | if ((j & 1) == 0) | |
1987 | { | |
1988 | prev_res = var; | |
1989 | continue; | |
1990 | } | |
1991 | var = permute_vec_elements (prev_res, var, | |
1992 | perm_mask, stmt, gsi); | |
1993 | new_stmt = SSA_NAME_DEF_STMT (var); | |
1994 | } | |
1995 | ||
1996 | if (prev_stmt_info == NULL) | |
1997 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1998 | else | |
1999 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2000 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2001 | } | |
3efe2e2c JJ |
2002 | |
2003 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
2004 | from the IL. */ | |
e6f5c25d IE |
2005 | if (STMT_VINFO_RELATED_STMT (stmt_info)) |
2006 | { | |
2007 | stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
2008 | stmt_info = vinfo_for_stmt (stmt); | |
2009 | } | |
3efe2e2c JJ |
2010 | tree lhs = gimple_call_lhs (stmt); |
2011 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
2012 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
2013 | set_vinfo_for_stmt (stmt, NULL); | |
2014 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2015 | gsi_replace (gsi, new_stmt, true); | |
5ce9450f JJ |
2016 | return true; |
2017 | } | |
2018 | else if (is_store) | |
2019 | { | |
2020 | tree vec_rhs = NULL_TREE, vec_mask = NULL_TREE; | |
2021 | prev_stmt_info = NULL; | |
2022 | for (i = 0; i < ncopies; i++) | |
2023 | { | |
2024 | unsigned align, misalign; | |
2025 | ||
2026 | if (i == 0) | |
2027 | { | |
2028 | tree rhs = gimple_call_arg (stmt, 3); | |
81c40241 RB |
2029 | vec_rhs = vect_get_vec_def_for_operand (rhs, stmt); |
2030 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); | |
5ce9450f JJ |
2031 | /* We should have catched mismatched types earlier. */ |
2032 | gcc_assert (useless_type_conversion_p (vectype, | |
2033 | TREE_TYPE (vec_rhs))); | |
2034 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, | |
2035 | NULL_TREE, &dummy, gsi, | |
2036 | &ptr_incr, false, &inv_p); | |
2037 | gcc_assert (!inv_p); | |
2038 | } | |
2039 | else | |
2040 | { | |
81c40241 | 2041 | vect_is_simple_use (vec_rhs, loop_vinfo, &def_stmt, &dt); |
5ce9450f | 2042 | vec_rhs = vect_get_vec_def_for_stmt_copy (dt, vec_rhs); |
81c40241 | 2043 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2044 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2045 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2046 | TYPE_SIZE_UNIT (vectype)); | |
2047 | } | |
2048 | ||
2049 | align = TYPE_ALIGN_UNIT (vectype); | |
2050 | if (aligned_access_p (dr)) | |
2051 | misalign = 0; | |
2052 | else if (DR_MISALIGNMENT (dr) == -1) | |
2053 | { | |
2054 | align = TYPE_ALIGN_UNIT (elem_type); | |
2055 | misalign = 0; | |
2056 | } | |
2057 | else | |
2058 | misalign = DR_MISALIGNMENT (dr); | |
2059 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2060 | misalign); | |
08554c26 JJ |
2061 | tree ptr = build_int_cst (TREE_TYPE (gimple_call_arg (stmt, 1)), |
2062 | misalign ? misalign & -misalign : align); | |
5ce9450f JJ |
2063 | new_stmt |
2064 | = gimple_build_call_internal (IFN_MASK_STORE, 4, dataref_ptr, | |
08554c26 | 2065 | ptr, vec_mask, vec_rhs); |
5ce9450f JJ |
2066 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2067 | if (i == 0) | |
2068 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2069 | else | |
2070 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2071 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2072 | } | |
2073 | } | |
2074 | else | |
2075 | { | |
2076 | tree vec_mask = NULL_TREE; | |
2077 | prev_stmt_info = NULL; | |
2078 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); | |
2079 | for (i = 0; i < ncopies; i++) | |
2080 | { | |
2081 | unsigned align, misalign; | |
2082 | ||
2083 | if (i == 0) | |
2084 | { | |
81c40241 | 2085 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); |
5ce9450f JJ |
2086 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, |
2087 | NULL_TREE, &dummy, gsi, | |
2088 | &ptr_incr, false, &inv_p); | |
2089 | gcc_assert (!inv_p); | |
2090 | } | |
2091 | else | |
2092 | { | |
81c40241 | 2093 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2094 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2095 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2096 | TYPE_SIZE_UNIT (vectype)); | |
2097 | } | |
2098 | ||
2099 | align = TYPE_ALIGN_UNIT (vectype); | |
2100 | if (aligned_access_p (dr)) | |
2101 | misalign = 0; | |
2102 | else if (DR_MISALIGNMENT (dr) == -1) | |
2103 | { | |
2104 | align = TYPE_ALIGN_UNIT (elem_type); | |
2105 | misalign = 0; | |
2106 | } | |
2107 | else | |
2108 | misalign = DR_MISALIGNMENT (dr); | |
2109 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2110 | misalign); | |
08554c26 JJ |
2111 | tree ptr = build_int_cst (TREE_TYPE (gimple_call_arg (stmt, 1)), |
2112 | misalign ? misalign & -misalign : align); | |
5ce9450f JJ |
2113 | new_stmt |
2114 | = gimple_build_call_internal (IFN_MASK_LOAD, 3, dataref_ptr, | |
08554c26 | 2115 | ptr, vec_mask); |
b731b390 | 2116 | gimple_call_set_lhs (new_stmt, make_ssa_name (vec_dest)); |
5ce9450f JJ |
2117 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2118 | if (i == 0) | |
2119 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2120 | else | |
2121 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2122 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2123 | } | |
2124 | } | |
2125 | ||
3efe2e2c JJ |
2126 | if (!is_store) |
2127 | { | |
2128 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
2129 | from the IL. */ | |
e6f5c25d IE |
2130 | if (STMT_VINFO_RELATED_STMT (stmt_info)) |
2131 | { | |
2132 | stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
2133 | stmt_info = vinfo_for_stmt (stmt); | |
2134 | } | |
3efe2e2c JJ |
2135 | tree lhs = gimple_call_lhs (stmt); |
2136 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
2137 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
2138 | set_vinfo_for_stmt (stmt, NULL); | |
2139 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2140 | gsi_replace (gsi, new_stmt, true); | |
2141 | } | |
2142 | ||
5ce9450f JJ |
2143 | return true; |
2144 | } | |
2145 | ||
b1b6836e RS |
2146 | /* Return true if vector types VECTYPE_IN and VECTYPE_OUT have |
2147 | integer elements and if we can narrow VECTYPE_IN to VECTYPE_OUT | |
2148 | in a single step. On success, store the binary pack code in | |
2149 | *CONVERT_CODE. */ | |
2150 | ||
2151 | static bool | |
2152 | simple_integer_narrowing (tree vectype_out, tree vectype_in, | |
2153 | tree_code *convert_code) | |
2154 | { | |
2155 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_out)) | |
2156 | || !INTEGRAL_TYPE_P (TREE_TYPE (vectype_in))) | |
2157 | return false; | |
2158 | ||
2159 | tree_code code; | |
2160 | int multi_step_cvt = 0; | |
2161 | auto_vec <tree, 8> interm_types; | |
2162 | if (!supportable_narrowing_operation (NOP_EXPR, vectype_out, vectype_in, | |
2163 | &code, &multi_step_cvt, | |
2164 | &interm_types) | |
2165 | || multi_step_cvt) | |
2166 | return false; | |
2167 | ||
2168 | *convert_code = code; | |
2169 | return true; | |
2170 | } | |
5ce9450f | 2171 | |
ebfd146a IR |
2172 | /* Function vectorizable_call. |
2173 | ||
538dd0b7 | 2174 | Check if GS performs a function call that can be vectorized. |
b8698a0f | 2175 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
2176 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2177 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2178 | ||
2179 | static bool | |
355fe088 | 2180 | vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
190c2236 | 2181 | slp_tree slp_node) |
ebfd146a | 2182 | { |
538dd0b7 | 2183 | gcall *stmt; |
ebfd146a IR |
2184 | tree vec_dest; |
2185 | tree scalar_dest; | |
2186 | tree op, type; | |
2187 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
538dd0b7 | 2188 | stmt_vec_info stmt_info = vinfo_for_stmt (gs), prev_stmt_info; |
ebfd146a IR |
2189 | tree vectype_out, vectype_in; |
2190 | int nunits_in; | |
2191 | int nunits_out; | |
2192 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
190c2236 | 2193 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 2194 | vec_info *vinfo = stmt_info->vinfo; |
81c40241 | 2195 | tree fndecl, new_temp, rhs_type; |
355fe088 | 2196 | gimple *def_stmt; |
0502fb85 UB |
2197 | enum vect_def_type dt[3] |
2198 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
355fe088 | 2199 | gimple *new_stmt = NULL; |
ebfd146a | 2200 | int ncopies, j; |
6e1aa848 | 2201 | vec<tree> vargs = vNULL; |
ebfd146a IR |
2202 | enum { NARROW, NONE, WIDEN } modifier; |
2203 | size_t i, nargs; | |
9d5e7640 | 2204 | tree lhs; |
ebfd146a | 2205 | |
190c2236 | 2206 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2207 | return false; |
2208 | ||
8644a673 | 2209 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2210 | return false; |
2211 | ||
538dd0b7 DM |
2212 | /* Is GS a vectorizable call? */ |
2213 | stmt = dyn_cast <gcall *> (gs); | |
2214 | if (!stmt) | |
ebfd146a IR |
2215 | return false; |
2216 | ||
5ce9450f JJ |
2217 | if (gimple_call_internal_p (stmt) |
2218 | && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD | |
2219 | || gimple_call_internal_fn (stmt) == IFN_MASK_STORE)) | |
2220 | return vectorizable_mask_load_store (stmt, gsi, vec_stmt, | |
2221 | slp_node); | |
2222 | ||
0136f8f0 AH |
2223 | if (gimple_call_lhs (stmt) == NULL_TREE |
2224 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
ebfd146a IR |
2225 | return false; |
2226 | ||
0136f8f0 | 2227 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); |
5a2c1986 | 2228 | |
b690cc0f RG |
2229 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
2230 | ||
ebfd146a IR |
2231 | /* Process function arguments. */ |
2232 | rhs_type = NULL_TREE; | |
b690cc0f | 2233 | vectype_in = NULL_TREE; |
ebfd146a IR |
2234 | nargs = gimple_call_num_args (stmt); |
2235 | ||
1b1562a5 MM |
2236 | /* Bail out if the function has more than three arguments, we do not have |
2237 | interesting builtin functions to vectorize with more than two arguments | |
2238 | except for fma. No arguments is also not good. */ | |
2239 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
2240 | return false; |
2241 | ||
74bf76ed JJ |
2242 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
2243 | if (gimple_call_internal_p (stmt) | |
2244 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2245 | { | |
2246 | nargs = 0; | |
2247 | rhs_type = unsigned_type_node; | |
2248 | } | |
2249 | ||
ebfd146a IR |
2250 | for (i = 0; i < nargs; i++) |
2251 | { | |
b690cc0f RG |
2252 | tree opvectype; |
2253 | ||
ebfd146a IR |
2254 | op = gimple_call_arg (stmt, i); |
2255 | ||
2256 | /* We can only handle calls with arguments of the same type. */ | |
2257 | if (rhs_type | |
8533c9d8 | 2258 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 2259 | { |
73fbfcad | 2260 | if (dump_enabled_p ()) |
78c60e3d | 2261 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2262 | "argument types differ.\n"); |
ebfd146a IR |
2263 | return false; |
2264 | } | |
b690cc0f RG |
2265 | if (!rhs_type) |
2266 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 2267 | |
81c40241 | 2268 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[i], &opvectype)) |
ebfd146a | 2269 | { |
73fbfcad | 2270 | if (dump_enabled_p ()) |
78c60e3d | 2271 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2272 | "use not simple.\n"); |
ebfd146a IR |
2273 | return false; |
2274 | } | |
ebfd146a | 2275 | |
b690cc0f RG |
2276 | if (!vectype_in) |
2277 | vectype_in = opvectype; | |
2278 | else if (opvectype | |
2279 | && opvectype != vectype_in) | |
2280 | { | |
73fbfcad | 2281 | if (dump_enabled_p ()) |
78c60e3d | 2282 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2283 | "argument vector types differ.\n"); |
b690cc0f RG |
2284 | return false; |
2285 | } | |
2286 | } | |
2287 | /* If all arguments are external or constant defs use a vector type with | |
2288 | the same size as the output vector type. */ | |
ebfd146a | 2289 | if (!vectype_in) |
b690cc0f | 2290 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
2291 | if (vec_stmt) |
2292 | gcc_assert (vectype_in); | |
2293 | if (!vectype_in) | |
2294 | { | |
73fbfcad | 2295 | if (dump_enabled_p ()) |
7d8930a0 | 2296 | { |
78c60e3d SS |
2297 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2298 | "no vectype for scalar type "); | |
2299 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 2300 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
2301 | } |
2302 | ||
2303 | return false; | |
2304 | } | |
ebfd146a IR |
2305 | |
2306 | /* FORNOW */ | |
b690cc0f RG |
2307 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
2308 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
2309 | if (nunits_in == nunits_out / 2) |
2310 | modifier = NARROW; | |
2311 | else if (nunits_out == nunits_in) | |
2312 | modifier = NONE; | |
2313 | else if (nunits_out == nunits_in / 2) | |
2314 | modifier = WIDEN; | |
2315 | else | |
2316 | return false; | |
2317 | ||
70439f0d RS |
2318 | /* We only handle functions that do not read or clobber memory. */ |
2319 | if (gimple_vuse (stmt)) | |
2320 | { | |
2321 | if (dump_enabled_p ()) | |
2322 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2323 | "function reads from or writes to memory.\n"); | |
2324 | return false; | |
2325 | } | |
2326 | ||
ebfd146a IR |
2327 | /* For now, we only vectorize functions if a target specific builtin |
2328 | is available. TODO -- in some cases, it might be profitable to | |
2329 | insert the calls for pieces of the vector, in order to be able | |
2330 | to vectorize other operations in the loop. */ | |
70439f0d RS |
2331 | fndecl = NULL_TREE; |
2332 | internal_fn ifn = IFN_LAST; | |
2333 | combined_fn cfn = gimple_call_combined_fn (stmt); | |
2334 | tree callee = gimple_call_fndecl (stmt); | |
2335 | ||
2336 | /* First try using an internal function. */ | |
b1b6836e RS |
2337 | tree_code convert_code = ERROR_MARK; |
2338 | if (cfn != CFN_LAST | |
2339 | && (modifier == NONE | |
2340 | || (modifier == NARROW | |
2341 | && simple_integer_narrowing (vectype_out, vectype_in, | |
2342 | &convert_code)))) | |
70439f0d RS |
2343 | ifn = vectorizable_internal_function (cfn, callee, vectype_out, |
2344 | vectype_in); | |
2345 | ||
2346 | /* If that fails, try asking for a target-specific built-in function. */ | |
2347 | if (ifn == IFN_LAST) | |
2348 | { | |
2349 | if (cfn != CFN_LAST) | |
2350 | fndecl = targetm.vectorize.builtin_vectorized_function | |
2351 | (cfn, vectype_out, vectype_in); | |
2352 | else | |
2353 | fndecl = targetm.vectorize.builtin_md_vectorized_function | |
2354 | (callee, vectype_out, vectype_in); | |
2355 | } | |
2356 | ||
2357 | if (ifn == IFN_LAST && !fndecl) | |
ebfd146a | 2358 | { |
70439f0d | 2359 | if (cfn == CFN_GOMP_SIMD_LANE |
74bf76ed JJ |
2360 | && !slp_node |
2361 | && loop_vinfo | |
2362 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2363 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
2364 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2365 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
2366 | { | |
2367 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
2368 | { 0, 1, 2, ... vf - 1 } vector. */ | |
2369 | gcc_assert (nargs == 0); | |
2370 | } | |
2371 | else | |
2372 | { | |
2373 | if (dump_enabled_p ()) | |
2374 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 2375 | "function is not vectorizable.\n"); |
74bf76ed JJ |
2376 | return false; |
2377 | } | |
ebfd146a IR |
2378 | } |
2379 | ||
190c2236 JJ |
2380 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
2381 | ncopies = 1; | |
b1b6836e | 2382 | else if (modifier == NARROW && ifn == IFN_LAST) |
ebfd146a IR |
2383 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; |
2384 | else | |
2385 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2386 | ||
2387 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
2388 | needs to be generated. */ | |
2389 | gcc_assert (ncopies >= 1); | |
2390 | ||
2391 | if (!vec_stmt) /* transformation not required. */ | |
2392 | { | |
2393 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 2394 | if (dump_enabled_p ()) |
e645e942 TJ |
2395 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ===" |
2396 | "\n"); | |
c3e7ee41 | 2397 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
b1b6836e RS |
2398 | if (ifn != IFN_LAST && modifier == NARROW && !slp_node) |
2399 | add_stmt_cost (stmt_info->vinfo->target_cost_data, ncopies / 2, | |
2400 | vec_promote_demote, stmt_info, 0, vect_body); | |
2401 | ||
ebfd146a IR |
2402 | return true; |
2403 | } | |
2404 | ||
2405 | /** Transform. **/ | |
2406 | ||
73fbfcad | 2407 | if (dump_enabled_p ()) |
e645e942 | 2408 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
2409 | |
2410 | /* Handle def. */ | |
2411 | scalar_dest = gimple_call_lhs (stmt); | |
2412 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
2413 | ||
2414 | prev_stmt_info = NULL; | |
b1b6836e | 2415 | if (modifier == NONE || ifn != IFN_LAST) |
ebfd146a | 2416 | { |
b1b6836e | 2417 | tree prev_res = NULL_TREE; |
ebfd146a IR |
2418 | for (j = 0; j < ncopies; ++j) |
2419 | { | |
2420 | /* Build argument list for the vectorized call. */ | |
2421 | if (j == 0) | |
9771b263 | 2422 | vargs.create (nargs); |
ebfd146a | 2423 | else |
9771b263 | 2424 | vargs.truncate (0); |
ebfd146a | 2425 | |
190c2236 JJ |
2426 | if (slp_node) |
2427 | { | |
ef062b13 | 2428 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2429 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2430 | |
2431 | for (i = 0; i < nargs; i++) | |
9771b263 | 2432 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2433 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2434 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2435 | |
2436 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2437 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
2438 | { |
2439 | size_t k; | |
2440 | for (k = 0; k < nargs; k++) | |
2441 | { | |
37b5ec8f | 2442 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 2443 | vargs[k] = vec_oprndsk[i]; |
190c2236 | 2444 | } |
b1b6836e RS |
2445 | if (modifier == NARROW) |
2446 | { | |
2447 | tree half_res = make_ssa_name (vectype_in); | |
2448 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2449 | gimple_call_set_lhs (new_stmt, half_res); | |
2450 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2451 | if ((i & 1) == 0) | |
2452 | { | |
2453 | prev_res = half_res; | |
2454 | continue; | |
2455 | } | |
2456 | new_temp = make_ssa_name (vec_dest); | |
2457 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
2458 | prev_res, half_res); | |
2459 | } | |
70439f0d | 2460 | else |
b1b6836e RS |
2461 | { |
2462 | if (ifn != IFN_LAST) | |
2463 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2464 | else | |
2465 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2466 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2467 | gimple_call_set_lhs (new_stmt, new_temp); | |
2468 | } | |
190c2236 | 2469 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 2470 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2471 | } |
2472 | ||
2473 | for (i = 0; i < nargs; i++) | |
2474 | { | |
37b5ec8f | 2475 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2476 | vec_oprndsi.release (); |
190c2236 | 2477 | } |
190c2236 JJ |
2478 | continue; |
2479 | } | |
2480 | ||
ebfd146a IR |
2481 | for (i = 0; i < nargs; i++) |
2482 | { | |
2483 | op = gimple_call_arg (stmt, i); | |
2484 | if (j == 0) | |
2485 | vec_oprnd0 | |
81c40241 | 2486 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 2487 | else |
63827fb8 IR |
2488 | { |
2489 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
2490 | vec_oprnd0 | |
2491 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
2492 | } | |
ebfd146a | 2493 | |
9771b263 | 2494 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
2495 | } |
2496 | ||
74bf76ed JJ |
2497 | if (gimple_call_internal_p (stmt) |
2498 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2499 | { | |
2500 | tree *v = XALLOCAVEC (tree, nunits_out); | |
2501 | int k; | |
2502 | for (k = 0; k < nunits_out; ++k) | |
2503 | v[k] = build_int_cst (unsigned_type_node, j * nunits_out + k); | |
2504 | tree cst = build_vector (vectype_out, v); | |
2505 | tree new_var | |
0e22bb5a | 2506 | = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_"); |
355fe088 | 2507 | gimple *init_stmt = gimple_build_assign (new_var, cst); |
74bf76ed | 2508 | vect_init_vector_1 (stmt, init_stmt, NULL); |
b731b390 | 2509 | new_temp = make_ssa_name (vec_dest); |
0e22bb5a | 2510 | new_stmt = gimple_build_assign (new_temp, new_var); |
74bf76ed | 2511 | } |
b1b6836e RS |
2512 | else if (modifier == NARROW) |
2513 | { | |
2514 | tree half_res = make_ssa_name (vectype_in); | |
2515 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2516 | gimple_call_set_lhs (new_stmt, half_res); | |
2517 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2518 | if ((j & 1) == 0) | |
2519 | { | |
2520 | prev_res = half_res; | |
2521 | continue; | |
2522 | } | |
2523 | new_temp = make_ssa_name (vec_dest); | |
2524 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
2525 | prev_res, half_res); | |
2526 | } | |
74bf76ed JJ |
2527 | else |
2528 | { | |
70439f0d RS |
2529 | if (ifn != IFN_LAST) |
2530 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2531 | else | |
2532 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
74bf76ed JJ |
2533 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2534 | gimple_call_set_lhs (new_stmt, new_temp); | |
2535 | } | |
ebfd146a IR |
2536 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2537 | ||
b1b6836e | 2538 | if (j == (modifier == NARROW ? 1 : 0)) |
ebfd146a IR |
2539 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; |
2540 | else | |
2541 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2542 | ||
2543 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2544 | } | |
b1b6836e RS |
2545 | } |
2546 | else if (modifier == NARROW) | |
2547 | { | |
ebfd146a IR |
2548 | for (j = 0; j < ncopies; ++j) |
2549 | { | |
2550 | /* Build argument list for the vectorized call. */ | |
2551 | if (j == 0) | |
9771b263 | 2552 | vargs.create (nargs * 2); |
ebfd146a | 2553 | else |
9771b263 | 2554 | vargs.truncate (0); |
ebfd146a | 2555 | |
190c2236 JJ |
2556 | if (slp_node) |
2557 | { | |
ef062b13 | 2558 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2559 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2560 | |
2561 | for (i = 0; i < nargs; i++) | |
9771b263 | 2562 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2563 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2564 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2565 | |
2566 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2567 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
2568 | { |
2569 | size_t k; | |
9771b263 | 2570 | vargs.truncate (0); |
190c2236 JJ |
2571 | for (k = 0; k < nargs; k++) |
2572 | { | |
37b5ec8f | 2573 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
2574 | vargs.quick_push (vec_oprndsk[i]); |
2575 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 | 2576 | } |
70439f0d RS |
2577 | if (ifn != IFN_LAST) |
2578 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2579 | else | |
2580 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
190c2236 JJ |
2581 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2582 | gimple_call_set_lhs (new_stmt, new_temp); | |
2583 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 2584 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2585 | } |
2586 | ||
2587 | for (i = 0; i < nargs; i++) | |
2588 | { | |
37b5ec8f | 2589 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2590 | vec_oprndsi.release (); |
190c2236 | 2591 | } |
190c2236 JJ |
2592 | continue; |
2593 | } | |
2594 | ||
ebfd146a IR |
2595 | for (i = 0; i < nargs; i++) |
2596 | { | |
2597 | op = gimple_call_arg (stmt, i); | |
2598 | if (j == 0) | |
2599 | { | |
2600 | vec_oprnd0 | |
81c40241 | 2601 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 2602 | vec_oprnd1 |
63827fb8 | 2603 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2604 | } |
2605 | else | |
2606 | { | |
336ecb65 | 2607 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 2608 | vec_oprnd0 |
63827fb8 | 2609 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 2610 | vec_oprnd1 |
63827fb8 | 2611 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2612 | } |
2613 | ||
9771b263 DN |
2614 | vargs.quick_push (vec_oprnd0); |
2615 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
2616 | } |
2617 | ||
b1b6836e | 2618 | new_stmt = gimple_build_call_vec (fndecl, vargs); |
ebfd146a IR |
2619 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2620 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
2621 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2622 | ||
2623 | if (j == 0) | |
2624 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2625 | else | |
2626 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2627 | ||
2628 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2629 | } | |
2630 | ||
2631 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a | 2632 | } |
b1b6836e RS |
2633 | else |
2634 | /* No current target implements this case. */ | |
2635 | return false; | |
ebfd146a | 2636 | |
9771b263 | 2637 | vargs.release (); |
ebfd146a | 2638 | |
ebfd146a IR |
2639 | /* The call in STMT might prevent it from being removed in dce. |
2640 | We however cannot remove it here, due to the way the ssa name | |
2641 | it defines is mapped to the new definition. So just replace | |
2642 | rhs of the statement with something harmless. */ | |
2643 | ||
dd34c087 JJ |
2644 | if (slp_node) |
2645 | return true; | |
2646 | ||
ebfd146a | 2647 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
2648 | if (is_pattern_stmt_p (stmt_info)) |
2649 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
2650 | else | |
2651 | lhs = gimple_call_lhs (stmt); | |
3cc2fa2a JJ |
2652 | |
2653 | if (gimple_call_internal_p (stmt) | |
2654 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2655 | { | |
2656 | /* Replace uses of the lhs of GOMP_SIMD_LANE call outside the loop | |
2657 | with vf - 1 rather than 0, that is the last iteration of the | |
2658 | vectorized loop. */ | |
2659 | imm_use_iterator iter; | |
2660 | use_operand_p use_p; | |
355fe088 | 2661 | gimple *use_stmt; |
3cc2fa2a JJ |
2662 | FOR_EACH_IMM_USE_STMT (use_stmt, iter, lhs) |
2663 | { | |
2664 | basic_block use_bb = gimple_bb (use_stmt); | |
2665 | if (use_bb | |
2666 | && !flow_bb_inside_loop_p (LOOP_VINFO_LOOP (loop_vinfo), use_bb)) | |
2667 | { | |
2668 | FOR_EACH_IMM_USE_ON_STMT (use_p, iter) | |
2669 | SET_USE (use_p, build_int_cst (TREE_TYPE (lhs), | |
2670 | ncopies * nunits_out - 1)); | |
2671 | update_stmt (use_stmt); | |
2672 | } | |
2673 | } | |
2674 | } | |
2675 | ||
9d5e7640 | 2676 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); |
ebfd146a | 2677 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 2678 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
2679 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
2680 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
2681 | |
2682 | return true; | |
2683 | } | |
2684 | ||
2685 | ||
0136f8f0 AH |
2686 | struct simd_call_arg_info |
2687 | { | |
2688 | tree vectype; | |
2689 | tree op; | |
2690 | enum vect_def_type dt; | |
2691 | HOST_WIDE_INT linear_step; | |
2692 | unsigned int align; | |
17b658af | 2693 | bool simd_lane_linear; |
0136f8f0 AH |
2694 | }; |
2695 | ||
17b658af JJ |
2696 | /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME, |
2697 | is linear within simd lane (but not within whole loop), note it in | |
2698 | *ARGINFO. */ | |
2699 | ||
2700 | static void | |
2701 | vect_simd_lane_linear (tree op, struct loop *loop, | |
2702 | struct simd_call_arg_info *arginfo) | |
2703 | { | |
355fe088 | 2704 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
17b658af JJ |
2705 | |
2706 | if (!is_gimple_assign (def_stmt) | |
2707 | || gimple_assign_rhs_code (def_stmt) != POINTER_PLUS_EXPR | |
2708 | || !is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))) | |
2709 | return; | |
2710 | ||
2711 | tree base = gimple_assign_rhs1 (def_stmt); | |
2712 | HOST_WIDE_INT linear_step = 0; | |
2713 | tree v = gimple_assign_rhs2 (def_stmt); | |
2714 | while (TREE_CODE (v) == SSA_NAME) | |
2715 | { | |
2716 | tree t; | |
2717 | def_stmt = SSA_NAME_DEF_STMT (v); | |
2718 | if (is_gimple_assign (def_stmt)) | |
2719 | switch (gimple_assign_rhs_code (def_stmt)) | |
2720 | { | |
2721 | case PLUS_EXPR: | |
2722 | t = gimple_assign_rhs2 (def_stmt); | |
2723 | if (linear_step || TREE_CODE (t) != INTEGER_CST) | |
2724 | return; | |
2725 | base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t); | |
2726 | v = gimple_assign_rhs1 (def_stmt); | |
2727 | continue; | |
2728 | case MULT_EXPR: | |
2729 | t = gimple_assign_rhs2 (def_stmt); | |
2730 | if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t)) | |
2731 | return; | |
2732 | linear_step = tree_to_shwi (t); | |
2733 | v = gimple_assign_rhs1 (def_stmt); | |
2734 | continue; | |
2735 | CASE_CONVERT: | |
2736 | t = gimple_assign_rhs1 (def_stmt); | |
2737 | if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE | |
2738 | || (TYPE_PRECISION (TREE_TYPE (v)) | |
2739 | < TYPE_PRECISION (TREE_TYPE (t)))) | |
2740 | return; | |
2741 | if (!linear_step) | |
2742 | linear_step = 1; | |
2743 | v = t; | |
2744 | continue; | |
2745 | default: | |
2746 | return; | |
2747 | } | |
2748 | else if (is_gimple_call (def_stmt) | |
2749 | && gimple_call_internal_p (def_stmt) | |
2750 | && gimple_call_internal_fn (def_stmt) == IFN_GOMP_SIMD_LANE | |
2751 | && loop->simduid | |
2752 | && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME | |
2753 | && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0)) | |
2754 | == loop->simduid)) | |
2755 | { | |
2756 | if (!linear_step) | |
2757 | linear_step = 1; | |
2758 | arginfo->linear_step = linear_step; | |
2759 | arginfo->op = base; | |
2760 | arginfo->simd_lane_linear = true; | |
2761 | return; | |
2762 | } | |
2763 | } | |
2764 | } | |
2765 | ||
0136f8f0 AH |
2766 | /* Function vectorizable_simd_clone_call. |
2767 | ||
2768 | Check if STMT performs a function call that can be vectorized | |
2769 | by calling a simd clone of the function. | |
2770 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2771 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2772 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2773 | ||
2774 | static bool | |
355fe088 TS |
2775 | vectorizable_simd_clone_call (gimple *stmt, gimple_stmt_iterator *gsi, |
2776 | gimple **vec_stmt, slp_tree slp_node) | |
0136f8f0 AH |
2777 | { |
2778 | tree vec_dest; | |
2779 | tree scalar_dest; | |
2780 | tree op, type; | |
2781 | tree vec_oprnd0 = NULL_TREE; | |
2782 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
2783 | tree vectype; | |
2784 | unsigned int nunits; | |
2785 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2786 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
310213d4 | 2787 | vec_info *vinfo = stmt_info->vinfo; |
0136f8f0 | 2788 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
81c40241 | 2789 | tree fndecl, new_temp; |
355fe088 TS |
2790 | gimple *def_stmt; |
2791 | gimple *new_stmt = NULL; | |
0136f8f0 AH |
2792 | int ncopies, j; |
2793 | vec<simd_call_arg_info> arginfo = vNULL; | |
2794 | vec<tree> vargs = vNULL; | |
2795 | size_t i, nargs; | |
2796 | tree lhs, rtype, ratype; | |
2797 | vec<constructor_elt, va_gc> *ret_ctor_elts; | |
2798 | ||
2799 | /* Is STMT a vectorizable call? */ | |
2800 | if (!is_gimple_call (stmt)) | |
2801 | return false; | |
2802 | ||
2803 | fndecl = gimple_call_fndecl (stmt); | |
2804 | if (fndecl == NULL_TREE) | |
2805 | return false; | |
2806 | ||
d52f5295 | 2807 | struct cgraph_node *node = cgraph_node::get (fndecl); |
0136f8f0 AH |
2808 | if (node == NULL || node->simd_clones == NULL) |
2809 | return false; | |
2810 | ||
2811 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
2812 | return false; | |
2813 | ||
2814 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
2815 | return false; | |
2816 | ||
2817 | if (gimple_call_lhs (stmt) | |
2818 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
2819 | return false; | |
2820 | ||
2821 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); | |
2822 | ||
2823 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2824 | ||
2825 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt)) | |
2826 | return false; | |
2827 | ||
2828 | /* FORNOW */ | |
2829 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
2830 | return false; | |
2831 | ||
2832 | /* Process function arguments. */ | |
2833 | nargs = gimple_call_num_args (stmt); | |
2834 | ||
2835 | /* Bail out if the function has zero arguments. */ | |
2836 | if (nargs == 0) | |
2837 | return false; | |
2838 | ||
2839 | arginfo.create (nargs); | |
2840 | ||
2841 | for (i = 0; i < nargs; i++) | |
2842 | { | |
2843 | simd_call_arg_info thisarginfo; | |
2844 | affine_iv iv; | |
2845 | ||
2846 | thisarginfo.linear_step = 0; | |
2847 | thisarginfo.align = 0; | |
2848 | thisarginfo.op = NULL_TREE; | |
17b658af | 2849 | thisarginfo.simd_lane_linear = false; |
0136f8f0 AH |
2850 | |
2851 | op = gimple_call_arg (stmt, i); | |
81c40241 RB |
2852 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &thisarginfo.dt, |
2853 | &thisarginfo.vectype) | |
0136f8f0 AH |
2854 | || thisarginfo.dt == vect_uninitialized_def) |
2855 | { | |
2856 | if (dump_enabled_p ()) | |
2857 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2858 | "use not simple.\n"); | |
2859 | arginfo.release (); | |
2860 | return false; | |
2861 | } | |
2862 | ||
2863 | if (thisarginfo.dt == vect_constant_def | |
2864 | || thisarginfo.dt == vect_external_def) | |
2865 | gcc_assert (thisarginfo.vectype == NULL_TREE); | |
2866 | else | |
2867 | gcc_assert (thisarginfo.vectype != NULL_TREE); | |
2868 | ||
6c9e85fb JJ |
2869 | /* For linear arguments, the analyze phase should have saved |
2870 | the base and step in STMT_VINFO_SIMD_CLONE_INFO. */ | |
17b658af JJ |
2871 | if (i * 3 + 4 <= STMT_VINFO_SIMD_CLONE_INFO (stmt_info).length () |
2872 | && STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]) | |
6c9e85fb JJ |
2873 | { |
2874 | gcc_assert (vec_stmt); | |
2875 | thisarginfo.linear_step | |
17b658af | 2876 | = tree_to_shwi (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]); |
6c9e85fb | 2877 | thisarginfo.op |
17b658af JJ |
2878 | = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 1]; |
2879 | thisarginfo.simd_lane_linear | |
2880 | = (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 3] | |
2881 | == boolean_true_node); | |
6c9e85fb JJ |
2882 | /* If loop has been peeled for alignment, we need to adjust it. */ |
2883 | tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo); | |
2884 | tree n2 = LOOP_VINFO_NITERS (loop_vinfo); | |
17b658af | 2885 | if (n1 != n2 && !thisarginfo.simd_lane_linear) |
6c9e85fb JJ |
2886 | { |
2887 | tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2); | |
17b658af | 2888 | tree step = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]; |
6c9e85fb JJ |
2889 | tree opt = TREE_TYPE (thisarginfo.op); |
2890 | bias = fold_convert (TREE_TYPE (step), bias); | |
2891 | bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step); | |
2892 | thisarginfo.op | |
2893 | = fold_build2 (POINTER_TYPE_P (opt) | |
2894 | ? POINTER_PLUS_EXPR : PLUS_EXPR, opt, | |
2895 | thisarginfo.op, bias); | |
2896 | } | |
2897 | } | |
2898 | else if (!vec_stmt | |
2899 | && thisarginfo.dt != vect_constant_def | |
2900 | && thisarginfo.dt != vect_external_def | |
2901 | && loop_vinfo | |
2902 | && TREE_CODE (op) == SSA_NAME | |
2903 | && simple_iv (loop, loop_containing_stmt (stmt), op, | |
2904 | &iv, false) | |
2905 | && tree_fits_shwi_p (iv.step)) | |
0136f8f0 AH |
2906 | { |
2907 | thisarginfo.linear_step = tree_to_shwi (iv.step); | |
2908 | thisarginfo.op = iv.base; | |
2909 | } | |
2910 | else if ((thisarginfo.dt == vect_constant_def | |
2911 | || thisarginfo.dt == vect_external_def) | |
2912 | && POINTER_TYPE_P (TREE_TYPE (op))) | |
2913 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; | |
17b658af JJ |
2914 | /* Addresses of array elements indexed by GOMP_SIMD_LANE are |
2915 | linear too. */ | |
2916 | if (POINTER_TYPE_P (TREE_TYPE (op)) | |
2917 | && !thisarginfo.linear_step | |
2918 | && !vec_stmt | |
2919 | && thisarginfo.dt != vect_constant_def | |
2920 | && thisarginfo.dt != vect_external_def | |
2921 | && loop_vinfo | |
2922 | && !slp_node | |
2923 | && TREE_CODE (op) == SSA_NAME) | |
2924 | vect_simd_lane_linear (op, loop, &thisarginfo); | |
0136f8f0 AH |
2925 | |
2926 | arginfo.quick_push (thisarginfo); | |
2927 | } | |
2928 | ||
2929 | unsigned int badness = 0; | |
2930 | struct cgraph_node *bestn = NULL; | |
6c9e85fb JJ |
2931 | if (STMT_VINFO_SIMD_CLONE_INFO (stmt_info).exists ()) |
2932 | bestn = cgraph_node::get (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[0]); | |
0136f8f0 AH |
2933 | else |
2934 | for (struct cgraph_node *n = node->simd_clones; n != NULL; | |
2935 | n = n->simdclone->next_clone) | |
2936 | { | |
2937 | unsigned int this_badness = 0; | |
2938 | if (n->simdclone->simdlen | |
2939 | > (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
2940 | || n->simdclone->nargs != nargs) | |
2941 | continue; | |
2942 | if (n->simdclone->simdlen | |
2943 | < (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
2944 | this_badness += (exact_log2 (LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
2945 | - exact_log2 (n->simdclone->simdlen)) * 1024; | |
2946 | if (n->simdclone->inbranch) | |
2947 | this_badness += 2048; | |
2948 | int target_badness = targetm.simd_clone.usable (n); | |
2949 | if (target_badness < 0) | |
2950 | continue; | |
2951 | this_badness += target_badness * 512; | |
2952 | /* FORNOW: Have to add code to add the mask argument. */ | |
2953 | if (n->simdclone->inbranch) | |
2954 | continue; | |
2955 | for (i = 0; i < nargs; i++) | |
2956 | { | |
2957 | switch (n->simdclone->args[i].arg_type) | |
2958 | { | |
2959 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
2960 | if (!useless_type_conversion_p | |
2961 | (n->simdclone->args[i].orig_type, | |
2962 | TREE_TYPE (gimple_call_arg (stmt, i)))) | |
2963 | i = -1; | |
2964 | else if (arginfo[i].dt == vect_constant_def | |
2965 | || arginfo[i].dt == vect_external_def | |
2966 | || arginfo[i].linear_step) | |
2967 | this_badness += 64; | |
2968 | break; | |
2969 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
2970 | if (arginfo[i].dt != vect_constant_def | |
2971 | && arginfo[i].dt != vect_external_def) | |
2972 | i = -1; | |
2973 | break; | |
2974 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
d9a6bd32 | 2975 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
2976 | if (arginfo[i].dt == vect_constant_def |
2977 | || arginfo[i].dt == vect_external_def | |
2978 | || (arginfo[i].linear_step | |
2979 | != n->simdclone->args[i].linear_step)) | |
2980 | i = -1; | |
2981 | break; | |
2982 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
d9a6bd32 JJ |
2983 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
2984 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
e01d41e5 JJ |
2985 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
2986 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
2987 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
2988 | /* FORNOW */ |
2989 | i = -1; | |
2990 | break; | |
2991 | case SIMD_CLONE_ARG_TYPE_MASK: | |
2992 | gcc_unreachable (); | |
2993 | } | |
2994 | if (i == (size_t) -1) | |
2995 | break; | |
2996 | if (n->simdclone->args[i].alignment > arginfo[i].align) | |
2997 | { | |
2998 | i = -1; | |
2999 | break; | |
3000 | } | |
3001 | if (arginfo[i].align) | |
3002 | this_badness += (exact_log2 (arginfo[i].align) | |
3003 | - exact_log2 (n->simdclone->args[i].alignment)); | |
3004 | } | |
3005 | if (i == (size_t) -1) | |
3006 | continue; | |
3007 | if (bestn == NULL || this_badness < badness) | |
3008 | { | |
3009 | bestn = n; | |
3010 | badness = this_badness; | |
3011 | } | |
3012 | } | |
3013 | ||
3014 | if (bestn == NULL) | |
3015 | { | |
3016 | arginfo.release (); | |
3017 | return false; | |
3018 | } | |
3019 | ||
3020 | for (i = 0; i < nargs; i++) | |
3021 | if ((arginfo[i].dt == vect_constant_def | |
3022 | || arginfo[i].dt == vect_external_def) | |
3023 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
3024 | { | |
3025 | arginfo[i].vectype | |
3026 | = get_vectype_for_scalar_type (TREE_TYPE (gimple_call_arg (stmt, | |
3027 | i))); | |
3028 | if (arginfo[i].vectype == NULL | |
3029 | || (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
3030 | > bestn->simdclone->simdlen)) | |
3031 | { | |
3032 | arginfo.release (); | |
3033 | return false; | |
3034 | } | |
3035 | } | |
3036 | ||
3037 | fndecl = bestn->decl; | |
3038 | nunits = bestn->simdclone->simdlen; | |
3039 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3040 | ||
3041 | /* If the function isn't const, only allow it in simd loops where user | |
3042 | has asserted that at least nunits consecutive iterations can be | |
3043 | performed using SIMD instructions. */ | |
3044 | if ((loop == NULL || (unsigned) loop->safelen < nunits) | |
3045 | && gimple_vuse (stmt)) | |
3046 | { | |
3047 | arginfo.release (); | |
3048 | return false; | |
3049 | } | |
3050 | ||
3051 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
3052 | needs to be generated. */ | |
3053 | gcc_assert (ncopies >= 1); | |
3054 | ||
3055 | if (!vec_stmt) /* transformation not required. */ | |
3056 | { | |
6c9e85fb JJ |
3057 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (bestn->decl); |
3058 | for (i = 0; i < nargs; i++) | |
3059 | if (bestn->simdclone->args[i].arg_type | |
3060 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
3061 | { | |
17b658af | 3062 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_grow_cleared (i * 3 |
6c9e85fb JJ |
3063 | + 1); |
3064 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (arginfo[i].op); | |
3065 | tree lst = POINTER_TYPE_P (TREE_TYPE (arginfo[i].op)) | |
3066 | ? size_type_node : TREE_TYPE (arginfo[i].op); | |
3067 | tree ls = build_int_cst (lst, arginfo[i].linear_step); | |
3068 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (ls); | |
17b658af JJ |
3069 | tree sll = arginfo[i].simd_lane_linear |
3070 | ? boolean_true_node : boolean_false_node; | |
3071 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (sll); | |
6c9e85fb | 3072 | } |
0136f8f0 AH |
3073 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; |
3074 | if (dump_enabled_p ()) | |
3075 | dump_printf_loc (MSG_NOTE, vect_location, | |
3076 | "=== vectorizable_simd_clone_call ===\n"); | |
3077 | /* vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); */ | |
3078 | arginfo.release (); | |
3079 | return true; | |
3080 | } | |
3081 | ||
3082 | /** Transform. **/ | |
3083 | ||
3084 | if (dump_enabled_p ()) | |
3085 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); | |
3086 | ||
3087 | /* Handle def. */ | |
3088 | scalar_dest = gimple_call_lhs (stmt); | |
3089 | vec_dest = NULL_TREE; | |
3090 | rtype = NULL_TREE; | |
3091 | ratype = NULL_TREE; | |
3092 | if (scalar_dest) | |
3093 | { | |
3094 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3095 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); | |
3096 | if (TREE_CODE (rtype) == ARRAY_TYPE) | |
3097 | { | |
3098 | ratype = rtype; | |
3099 | rtype = TREE_TYPE (ratype); | |
3100 | } | |
3101 | } | |
3102 | ||
3103 | prev_stmt_info = NULL; | |
3104 | for (j = 0; j < ncopies; ++j) | |
3105 | { | |
3106 | /* Build argument list for the vectorized call. */ | |
3107 | if (j == 0) | |
3108 | vargs.create (nargs); | |
3109 | else | |
3110 | vargs.truncate (0); | |
3111 | ||
3112 | for (i = 0; i < nargs; i++) | |
3113 | { | |
3114 | unsigned int k, l, m, o; | |
3115 | tree atype; | |
3116 | op = gimple_call_arg (stmt, i); | |
3117 | switch (bestn->simdclone->args[i].arg_type) | |
3118 | { | |
3119 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3120 | atype = bestn->simdclone->args[i].vector_type; | |
3121 | o = nunits / TYPE_VECTOR_SUBPARTS (atype); | |
3122 | for (m = j * o; m < (j + 1) * o; m++) | |
3123 | { | |
3124 | if (TYPE_VECTOR_SUBPARTS (atype) | |
3125 | < TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)) | |
3126 | { | |
3127 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); | |
3128 | k = (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
3129 | / TYPE_VECTOR_SUBPARTS (atype)); | |
3130 | gcc_assert ((k & (k - 1)) == 0); | |
3131 | if (m == 0) | |
3132 | vec_oprnd0 | |
81c40241 | 3133 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3134 | else |
3135 | { | |
3136 | vec_oprnd0 = arginfo[i].op; | |
3137 | if ((m & (k - 1)) == 0) | |
3138 | vec_oprnd0 | |
3139 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3140 | vec_oprnd0); | |
3141 | } | |
3142 | arginfo[i].op = vec_oprnd0; | |
3143 | vec_oprnd0 | |
3144 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, | |
3145 | size_int (prec), | |
3146 | bitsize_int ((m & (k - 1)) * prec)); | |
3147 | new_stmt | |
b731b390 | 3148 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3149 | vec_oprnd0); |
3150 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3151 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3152 | } | |
3153 | else | |
3154 | { | |
3155 | k = (TYPE_VECTOR_SUBPARTS (atype) | |
3156 | / TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)); | |
3157 | gcc_assert ((k & (k - 1)) == 0); | |
3158 | vec<constructor_elt, va_gc> *ctor_elts; | |
3159 | if (k != 1) | |
3160 | vec_alloc (ctor_elts, k); | |
3161 | else | |
3162 | ctor_elts = NULL; | |
3163 | for (l = 0; l < k; l++) | |
3164 | { | |
3165 | if (m == 0 && l == 0) | |
3166 | vec_oprnd0 | |
81c40241 | 3167 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3168 | else |
3169 | vec_oprnd0 | |
3170 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3171 | arginfo[i].op); | |
3172 | arginfo[i].op = vec_oprnd0; | |
3173 | if (k == 1) | |
3174 | break; | |
3175 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, | |
3176 | vec_oprnd0); | |
3177 | } | |
3178 | if (k == 1) | |
3179 | vargs.safe_push (vec_oprnd0); | |
3180 | else | |
3181 | { | |
3182 | vec_oprnd0 = build_constructor (atype, ctor_elts); | |
3183 | new_stmt | |
b731b390 | 3184 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3185 | vec_oprnd0); |
3186 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3187 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3188 | } | |
3189 | } | |
3190 | } | |
3191 | break; | |
3192 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3193 | vargs.safe_push (op); | |
3194 | break; | |
3195 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
3196 | if (j == 0) | |
3197 | { | |
3198 | gimple_seq stmts; | |
3199 | arginfo[i].op | |
3200 | = force_gimple_operand (arginfo[i].op, &stmts, true, | |
3201 | NULL_TREE); | |
3202 | if (stmts != NULL) | |
3203 | { | |
3204 | basic_block new_bb; | |
3205 | edge pe = loop_preheader_edge (loop); | |
3206 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
3207 | gcc_assert (!new_bb); | |
3208 | } | |
17b658af JJ |
3209 | if (arginfo[i].simd_lane_linear) |
3210 | { | |
3211 | vargs.safe_push (arginfo[i].op); | |
3212 | break; | |
3213 | } | |
b731b390 | 3214 | tree phi_res = copy_ssa_name (op); |
538dd0b7 | 3215 | gphi *new_phi = create_phi_node (phi_res, loop->header); |
0136f8f0 | 3216 | set_vinfo_for_stmt (new_phi, |
310213d4 | 3217 | new_stmt_vec_info (new_phi, loop_vinfo)); |
0136f8f0 AH |
3218 | add_phi_arg (new_phi, arginfo[i].op, |
3219 | loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
3220 | enum tree_code code | |
3221 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3222 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3223 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3224 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3225 | widest_int cst |
3226 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3227 | ncopies * nunits); | |
3228 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3229 | tree phi_arg = copy_ssa_name (op); |
0d0e4a03 JJ |
3230 | new_stmt |
3231 | = gimple_build_assign (phi_arg, code, phi_res, tcst); | |
0136f8f0 AH |
3232 | gimple_stmt_iterator si = gsi_after_labels (loop->header); |
3233 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); | |
3234 | set_vinfo_for_stmt (new_stmt, | |
310213d4 | 3235 | new_stmt_vec_info (new_stmt, loop_vinfo)); |
0136f8f0 AH |
3236 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), |
3237 | UNKNOWN_LOCATION); | |
3238 | arginfo[i].op = phi_res; | |
3239 | vargs.safe_push (phi_res); | |
3240 | } | |
3241 | else | |
3242 | { | |
3243 | enum tree_code code | |
3244 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3245 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3246 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3247 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3248 | widest_int cst |
3249 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3250 | j * nunits); | |
3251 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3252 | new_temp = make_ssa_name (TREE_TYPE (op)); |
0d0e4a03 JJ |
3253 | new_stmt = gimple_build_assign (new_temp, code, |
3254 | arginfo[i].op, tcst); | |
0136f8f0 AH |
3255 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3256 | vargs.safe_push (new_temp); | |
3257 | } | |
3258 | break; | |
3259 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
e01d41e5 JJ |
3260 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3261 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3262 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3263 | default: |
3264 | gcc_unreachable (); | |
3265 | } | |
3266 | } | |
3267 | ||
3268 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
3269 | if (vec_dest) | |
3270 | { | |
3271 | gcc_assert (ratype || TYPE_VECTOR_SUBPARTS (rtype) == nunits); | |
3272 | if (ratype) | |
b731b390 | 3273 | new_temp = create_tmp_var (ratype); |
0136f8f0 AH |
3274 | else if (TYPE_VECTOR_SUBPARTS (vectype) |
3275 | == TYPE_VECTOR_SUBPARTS (rtype)) | |
3276 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3277 | else | |
3278 | new_temp = make_ssa_name (rtype, new_stmt); | |
3279 | gimple_call_set_lhs (new_stmt, new_temp); | |
3280 | } | |
3281 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3282 | ||
3283 | if (vec_dest) | |
3284 | { | |
3285 | if (TYPE_VECTOR_SUBPARTS (vectype) < nunits) | |
3286 | { | |
3287 | unsigned int k, l; | |
3288 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); | |
3289 | k = nunits / TYPE_VECTOR_SUBPARTS (vectype); | |
3290 | gcc_assert ((k & (k - 1)) == 0); | |
3291 | for (l = 0; l < k; l++) | |
3292 | { | |
3293 | tree t; | |
3294 | if (ratype) | |
3295 | { | |
3296 | t = build_fold_addr_expr (new_temp); | |
3297 | t = build2 (MEM_REF, vectype, t, | |
3298 | build_int_cst (TREE_TYPE (t), | |
3299 | l * prec / BITS_PER_UNIT)); | |
3300 | } | |
3301 | else | |
3302 | t = build3 (BIT_FIELD_REF, vectype, new_temp, | |
3303 | size_int (prec), bitsize_int (l * prec)); | |
3304 | new_stmt | |
b731b390 | 3305 | = gimple_build_assign (make_ssa_name (vectype), t); |
0136f8f0 AH |
3306 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3307 | if (j == 0 && l == 0) | |
3308 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3309 | else | |
3310 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3311 | ||
3312 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3313 | } | |
3314 | ||
3315 | if (ratype) | |
3316 | { | |
3317 | tree clobber = build_constructor (ratype, NULL); | |
3318 | TREE_THIS_VOLATILE (clobber) = 1; | |
3319 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3320 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3321 | } | |
3322 | continue; | |
3323 | } | |
3324 | else if (TYPE_VECTOR_SUBPARTS (vectype) > nunits) | |
3325 | { | |
3326 | unsigned int k = (TYPE_VECTOR_SUBPARTS (vectype) | |
3327 | / TYPE_VECTOR_SUBPARTS (rtype)); | |
3328 | gcc_assert ((k & (k - 1)) == 0); | |
3329 | if ((j & (k - 1)) == 0) | |
3330 | vec_alloc (ret_ctor_elts, k); | |
3331 | if (ratype) | |
3332 | { | |
3333 | unsigned int m, o = nunits / TYPE_VECTOR_SUBPARTS (rtype); | |
3334 | for (m = 0; m < o; m++) | |
3335 | { | |
3336 | tree tem = build4 (ARRAY_REF, rtype, new_temp, | |
3337 | size_int (m), NULL_TREE, NULL_TREE); | |
3338 | new_stmt | |
b731b390 | 3339 | = gimple_build_assign (make_ssa_name (rtype), tem); |
0136f8f0 AH |
3340 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3341 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, | |
3342 | gimple_assign_lhs (new_stmt)); | |
3343 | } | |
3344 | tree clobber = build_constructor (ratype, NULL); | |
3345 | TREE_THIS_VOLATILE (clobber) = 1; | |
3346 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3347 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3348 | } | |
3349 | else | |
3350 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); | |
3351 | if ((j & (k - 1)) != k - 1) | |
3352 | continue; | |
3353 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); | |
3354 | new_stmt | |
b731b390 | 3355 | = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0); |
0136f8f0 AH |
3356 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3357 | ||
3358 | if ((unsigned) j == k - 1) | |
3359 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3360 | else | |
3361 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3362 | ||
3363 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3364 | continue; | |
3365 | } | |
3366 | else if (ratype) | |
3367 | { | |
3368 | tree t = build_fold_addr_expr (new_temp); | |
3369 | t = build2 (MEM_REF, vectype, t, | |
3370 | build_int_cst (TREE_TYPE (t), 0)); | |
3371 | new_stmt | |
b731b390 | 3372 | = gimple_build_assign (make_ssa_name (vec_dest), t); |
0136f8f0 AH |
3373 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3374 | tree clobber = build_constructor (ratype, NULL); | |
3375 | TREE_THIS_VOLATILE (clobber) = 1; | |
3376 | vect_finish_stmt_generation (stmt, | |
3377 | gimple_build_assign (new_temp, | |
3378 | clobber), gsi); | |
3379 | } | |
3380 | } | |
3381 | ||
3382 | if (j == 0) | |
3383 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3384 | else | |
3385 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3386 | ||
3387 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3388 | } | |
3389 | ||
3390 | vargs.release (); | |
3391 | ||
3392 | /* The call in STMT might prevent it from being removed in dce. | |
3393 | We however cannot remove it here, due to the way the ssa name | |
3394 | it defines is mapped to the new definition. So just replace | |
3395 | rhs of the statement with something harmless. */ | |
3396 | ||
3397 | if (slp_node) | |
3398 | return true; | |
3399 | ||
3400 | if (scalar_dest) | |
3401 | { | |
3402 | type = TREE_TYPE (scalar_dest); | |
3403 | if (is_pattern_stmt_p (stmt_info)) | |
3404 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
3405 | else | |
3406 | lhs = gimple_call_lhs (stmt); | |
3407 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
3408 | } | |
3409 | else | |
3410 | new_stmt = gimple_build_nop (); | |
3411 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
3412 | set_vinfo_for_stmt (stmt, NULL); | |
3413 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2865f32a | 3414 | gsi_replace (gsi, new_stmt, true); |
0136f8f0 AH |
3415 | unlink_stmt_vdef (stmt); |
3416 | ||
3417 | return true; | |
3418 | } | |
3419 | ||
3420 | ||
ebfd146a IR |
3421 | /* Function vect_gen_widened_results_half |
3422 | ||
3423 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 3424 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 3425 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
3426 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
3427 | needs to be created (DECL is a function-decl of a target-builtin). | |
3428 | STMT is the original scalar stmt that we are vectorizing. */ | |
3429 | ||
355fe088 | 3430 | static gimple * |
ebfd146a IR |
3431 | vect_gen_widened_results_half (enum tree_code code, |
3432 | tree decl, | |
3433 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
3434 | tree vec_dest, gimple_stmt_iterator *gsi, | |
355fe088 | 3435 | gimple *stmt) |
b8698a0f | 3436 | { |
355fe088 | 3437 | gimple *new_stmt; |
b8698a0f L |
3438 | tree new_temp; |
3439 | ||
3440 | /* Generate half of the widened result: */ | |
3441 | if (code == CALL_EXPR) | |
3442 | { | |
3443 | /* Target specific support */ | |
ebfd146a IR |
3444 | if (op_type == binary_op) |
3445 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
3446 | else | |
3447 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
3448 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3449 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
3450 | } |
3451 | else | |
ebfd146a | 3452 | { |
b8698a0f L |
3453 | /* Generic support */ |
3454 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
3455 | if (op_type != binary_op) |
3456 | vec_oprnd1 = NULL; | |
0d0e4a03 | 3457 | new_stmt = gimple_build_assign (vec_dest, code, vec_oprnd0, vec_oprnd1); |
ebfd146a IR |
3458 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3459 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 3460 | } |
ebfd146a IR |
3461 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3462 | ||
ebfd146a IR |
3463 | return new_stmt; |
3464 | } | |
3465 | ||
4a00c761 JJ |
3466 | |
3467 | /* Get vectorized definitions for loop-based vectorization. For the first | |
3468 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
3469 | scalar operand), and for the rest we get a copy with | |
3470 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
3471 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
3472 | The vectors are collected into VEC_OPRNDS. */ | |
3473 | ||
3474 | static void | |
355fe088 | 3475 | vect_get_loop_based_defs (tree *oprnd, gimple *stmt, enum vect_def_type dt, |
9771b263 | 3476 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
3477 | { |
3478 | tree vec_oprnd; | |
3479 | ||
3480 | /* Get first vector operand. */ | |
3481 | /* All the vector operands except the very first one (that is scalar oprnd) | |
3482 | are stmt copies. */ | |
3483 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
81c40241 | 3484 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt); |
4a00c761 JJ |
3485 | else |
3486 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
3487 | ||
9771b263 | 3488 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3489 | |
3490 | /* Get second vector operand. */ | |
3491 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 3492 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3493 | |
3494 | *oprnd = vec_oprnd; | |
3495 | ||
3496 | /* For conversion in multiple steps, continue to get operands | |
3497 | recursively. */ | |
3498 | if (multi_step_cvt) | |
3499 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
3500 | } | |
3501 | ||
3502 | ||
3503 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
3504 | For multi-step conversions store the resulting vectors and call the function | |
3505 | recursively. */ | |
3506 | ||
3507 | static void | |
9771b263 | 3508 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
355fe088 | 3509 | int multi_step_cvt, gimple *stmt, |
9771b263 | 3510 | vec<tree> vec_dsts, |
4a00c761 JJ |
3511 | gimple_stmt_iterator *gsi, |
3512 | slp_tree slp_node, enum tree_code code, | |
3513 | stmt_vec_info *prev_stmt_info) | |
3514 | { | |
3515 | unsigned int i; | |
3516 | tree vop0, vop1, new_tmp, vec_dest; | |
355fe088 | 3517 | gimple *new_stmt; |
4a00c761 JJ |
3518 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
3519 | ||
9771b263 | 3520 | vec_dest = vec_dsts.pop (); |
4a00c761 | 3521 | |
9771b263 | 3522 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
3523 | { |
3524 | /* Create demotion operation. */ | |
9771b263 DN |
3525 | vop0 = (*vec_oprnds)[i]; |
3526 | vop1 = (*vec_oprnds)[i + 1]; | |
0d0e4a03 | 3527 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
4a00c761 JJ |
3528 | new_tmp = make_ssa_name (vec_dest, new_stmt); |
3529 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
3530 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3531 | ||
3532 | if (multi_step_cvt) | |
3533 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 3534 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
3535 | else |
3536 | { | |
3537 | /* This is the last step of the conversion sequence. Store the | |
3538 | vectors in SLP_NODE or in vector info of the scalar statement | |
3539 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
3540 | if (slp_node) | |
9771b263 | 3541 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 3542 | else |
c689ce1e RB |
3543 | { |
3544 | if (!*prev_stmt_info) | |
3545 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3546 | else | |
3547 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
4a00c761 | 3548 | |
c689ce1e RB |
3549 | *prev_stmt_info = vinfo_for_stmt (new_stmt); |
3550 | } | |
4a00c761 JJ |
3551 | } |
3552 | } | |
3553 | ||
3554 | /* For multi-step demotion operations we first generate demotion operations | |
3555 | from the source type to the intermediate types, and then combine the | |
3556 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
3557 | type. */ | |
3558 | if (multi_step_cvt) | |
3559 | { | |
3560 | /* At each level of recursion we have half of the operands we had at the | |
3561 | previous level. */ | |
9771b263 | 3562 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
3563 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
3564 | stmt, vec_dsts, gsi, slp_node, | |
3565 | VEC_PACK_TRUNC_EXPR, | |
3566 | prev_stmt_info); | |
3567 | } | |
3568 | ||
9771b263 | 3569 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3570 | } |
3571 | ||
3572 | ||
3573 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
3574 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
3575 | the resulting vectors and call the function recursively. */ | |
3576 | ||
3577 | static void | |
9771b263 DN |
3578 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
3579 | vec<tree> *vec_oprnds1, | |
355fe088 | 3580 | gimple *stmt, tree vec_dest, |
4a00c761 JJ |
3581 | gimple_stmt_iterator *gsi, |
3582 | enum tree_code code1, | |
3583 | enum tree_code code2, tree decl1, | |
3584 | tree decl2, int op_type) | |
3585 | { | |
3586 | int i; | |
3587 | tree vop0, vop1, new_tmp1, new_tmp2; | |
355fe088 | 3588 | gimple *new_stmt1, *new_stmt2; |
6e1aa848 | 3589 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 3590 | |
9771b263 DN |
3591 | vec_tmp.create (vec_oprnds0->length () * 2); |
3592 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
3593 | { |
3594 | if (op_type == binary_op) | |
9771b263 | 3595 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
3596 | else |
3597 | vop1 = NULL_TREE; | |
3598 | ||
3599 | /* Generate the two halves of promotion operation. */ | |
3600 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
3601 | op_type, vec_dest, gsi, stmt); | |
3602 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
3603 | op_type, vec_dest, gsi, stmt); | |
3604 | if (is_gimple_call (new_stmt1)) | |
3605 | { | |
3606 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
3607 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
3608 | } | |
3609 | else | |
3610 | { | |
3611 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
3612 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
3613 | } | |
3614 | ||
3615 | /* Store the results for the next step. */ | |
9771b263 DN |
3616 | vec_tmp.quick_push (new_tmp1); |
3617 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
3618 | } |
3619 | ||
689eaba3 | 3620 | vec_oprnds0->release (); |
4a00c761 JJ |
3621 | *vec_oprnds0 = vec_tmp; |
3622 | } | |
3623 | ||
3624 | ||
b8698a0f L |
3625 | /* Check if STMT performs a conversion operation, that can be vectorized. |
3626 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 3627 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
3628 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
3629 | ||
3630 | static bool | |
355fe088 TS |
3631 | vectorizable_conversion (gimple *stmt, gimple_stmt_iterator *gsi, |
3632 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
3633 | { |
3634 | tree vec_dest; | |
3635 | tree scalar_dest; | |
4a00c761 | 3636 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
3637 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
3638 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3639 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3640 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 3641 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
3642 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
3643 | tree new_temp; | |
355fe088 | 3644 | gimple *def_stmt; |
ebfd146a | 3645 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
355fe088 | 3646 | gimple *new_stmt = NULL; |
ebfd146a IR |
3647 | stmt_vec_info prev_stmt_info; |
3648 | int nunits_in; | |
3649 | int nunits_out; | |
3650 | tree vectype_out, vectype_in; | |
4a00c761 JJ |
3651 | int ncopies, i, j; |
3652 | tree lhs_type, rhs_type; | |
ebfd146a | 3653 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
3654 | vec<tree> vec_oprnds0 = vNULL; |
3655 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 3656 | tree vop0; |
4a00c761 | 3657 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 3658 | vec_info *vinfo = stmt_info->vinfo; |
4a00c761 | 3659 | int multi_step_cvt = 0; |
6e1aa848 DN |
3660 | vec<tree> vec_dsts = vNULL; |
3661 | vec<tree> interm_types = vNULL; | |
4a00c761 JJ |
3662 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
3663 | int op_type; | |
ef4bddc2 | 3664 | machine_mode rhs_mode; |
4a00c761 | 3665 | unsigned short fltsz; |
ebfd146a IR |
3666 | |
3667 | /* Is STMT a vectorizable conversion? */ | |
3668 | ||
4a00c761 | 3669 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3670 | return false; |
3671 | ||
8644a673 | 3672 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3673 | return false; |
3674 | ||
3675 | if (!is_gimple_assign (stmt)) | |
3676 | return false; | |
3677 | ||
3678 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3679 | return false; | |
3680 | ||
3681 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
3682 | if (!CONVERT_EXPR_CODE_P (code) |
3683 | && code != FIX_TRUNC_EXPR | |
3684 | && code != FLOAT_EXPR | |
3685 | && code != WIDEN_MULT_EXPR | |
3686 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
3687 | return false; |
3688 | ||
4a00c761 JJ |
3689 | op_type = TREE_CODE_LENGTH (code); |
3690 | ||
ebfd146a | 3691 | /* Check types of lhs and rhs. */ |
b690cc0f | 3692 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 3693 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
3694 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
3695 | ||
ebfd146a IR |
3696 | op0 = gimple_assign_rhs1 (stmt); |
3697 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
3698 | |
3699 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
3700 | && !((INTEGRAL_TYPE_P (lhs_type) | |
3701 | && INTEGRAL_TYPE_P (rhs_type)) | |
3702 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
3703 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
3704 | return false; | |
3705 | ||
e6f5c25d IE |
3706 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
3707 | && ((INTEGRAL_TYPE_P (lhs_type) | |
3708 | && (TYPE_PRECISION (lhs_type) | |
3709 | != GET_MODE_PRECISION (TYPE_MODE (lhs_type)))) | |
3710 | || (INTEGRAL_TYPE_P (rhs_type) | |
3711 | && (TYPE_PRECISION (rhs_type) | |
3712 | != GET_MODE_PRECISION (TYPE_MODE (rhs_type)))))) | |
4a00c761 | 3713 | { |
73fbfcad | 3714 | if (dump_enabled_p ()) |
78c60e3d | 3715 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
3716 | "type conversion to/from bit-precision unsupported." |
3717 | "\n"); | |
4a00c761 JJ |
3718 | return false; |
3719 | } | |
3720 | ||
b690cc0f | 3721 | /* Check the operands of the operation. */ |
81c40241 | 3722 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype_in)) |
b690cc0f | 3723 | { |
73fbfcad | 3724 | if (dump_enabled_p ()) |
78c60e3d | 3725 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3726 | "use not simple.\n"); |
b690cc0f RG |
3727 | return false; |
3728 | } | |
4a00c761 JJ |
3729 | if (op_type == binary_op) |
3730 | { | |
3731 | bool ok; | |
3732 | ||
3733 | op1 = gimple_assign_rhs2 (stmt); | |
3734 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
3735 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
3736 | OP1. */ | |
3737 | if (CONSTANT_CLASS_P (op0)) | |
81c40241 | 3738 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &vectype_in); |
4a00c761 | 3739 | else |
81c40241 | 3740 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1]); |
4a00c761 JJ |
3741 | |
3742 | if (!ok) | |
3743 | { | |
73fbfcad | 3744 | if (dump_enabled_p ()) |
78c60e3d | 3745 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3746 | "use not simple.\n"); |
4a00c761 JJ |
3747 | return false; |
3748 | } | |
3749 | } | |
3750 | ||
b690cc0f RG |
3751 | /* If op0 is an external or constant defs use a vector type of |
3752 | the same size as the output vector type. */ | |
ebfd146a | 3753 | if (!vectype_in) |
b690cc0f | 3754 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
3755 | if (vec_stmt) |
3756 | gcc_assert (vectype_in); | |
3757 | if (!vectype_in) | |
3758 | { | |
73fbfcad | 3759 | if (dump_enabled_p ()) |
4a00c761 | 3760 | { |
78c60e3d SS |
3761 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3762 | "no vectype for scalar type "); | |
3763 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 3764 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 3765 | } |
7d8930a0 IR |
3766 | |
3767 | return false; | |
3768 | } | |
ebfd146a | 3769 | |
e6f5c25d IE |
3770 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
3771 | && !VECTOR_BOOLEAN_TYPE_P (vectype_in)) | |
3772 | { | |
3773 | if (dump_enabled_p ()) | |
3774 | { | |
3775 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3776 | "can't convert between boolean and non " | |
3777 | "boolean vectors"); | |
3778 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
3779 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
3780 | } | |
3781 | ||
3782 | return false; | |
3783 | } | |
3784 | ||
b690cc0f RG |
3785 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
3786 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4a00c761 | 3787 | if (nunits_in < nunits_out) |
ebfd146a IR |
3788 | modifier = NARROW; |
3789 | else if (nunits_out == nunits_in) | |
3790 | modifier = NONE; | |
ebfd146a | 3791 | else |
4a00c761 | 3792 | modifier = WIDEN; |
ebfd146a | 3793 | |
ff802fa1 IR |
3794 | /* Multiple types in SLP are handled by creating the appropriate number of |
3795 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3796 | case of SLP. */ | |
437f4a00 | 3797 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a | 3798 | ncopies = 1; |
4a00c761 JJ |
3799 | else if (modifier == NARROW) |
3800 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
3801 | else | |
3802 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
b8698a0f | 3803 | |
ebfd146a IR |
3804 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
3805 | needs to be generated. */ | |
3806 | gcc_assert (ncopies >= 1); | |
3807 | ||
ebfd146a | 3808 | /* Supportable by target? */ |
4a00c761 | 3809 | switch (modifier) |
ebfd146a | 3810 | { |
4a00c761 JJ |
3811 | case NONE: |
3812 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
3813 | return false; | |
3814 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
3815 | &decl1, &code1)) | |
3816 | break; | |
3817 | /* FALLTHRU */ | |
3818 | unsupported: | |
73fbfcad | 3819 | if (dump_enabled_p ()) |
78c60e3d | 3820 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3821 | "conversion not supported by target.\n"); |
ebfd146a | 3822 | return false; |
ebfd146a | 3823 | |
4a00c761 JJ |
3824 | case WIDEN: |
3825 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
3826 | &code1, &code2, &multi_step_cvt, |
3827 | &interm_types)) | |
4a00c761 JJ |
3828 | { |
3829 | /* Binary widening operation can only be supported directly by the | |
3830 | architecture. */ | |
3831 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
3832 | break; | |
3833 | } | |
3834 | ||
3835 | if (code != FLOAT_EXPR | |
3836 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
3837 | <= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
3838 | goto unsupported; | |
3839 | ||
3840 | rhs_mode = TYPE_MODE (rhs_type); | |
3841 | fltsz = GET_MODE_SIZE (TYPE_MODE (lhs_type)); | |
3842 | for (rhs_mode = GET_MODE_2XWIDER_MODE (TYPE_MODE (rhs_type)); | |
3843 | rhs_mode != VOIDmode && GET_MODE_SIZE (rhs_mode) <= fltsz; | |
3844 | rhs_mode = GET_MODE_2XWIDER_MODE (rhs_mode)) | |
3845 | { | |
3846 | cvt_type | |
3847 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
3848 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
3849 | if (cvt_type == NULL_TREE) | |
3850 | goto unsupported; | |
3851 | ||
3852 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
3853 | { | |
3854 | if (!supportable_convert_operation (code, vectype_out, | |
3855 | cvt_type, &decl1, &codecvt1)) | |
3856 | goto unsupported; | |
3857 | } | |
3858 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
3859 | cvt_type, &codecvt1, |
3860 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
3861 | &interm_types)) |
3862 | continue; | |
3863 | else | |
3864 | gcc_assert (multi_step_cvt == 0); | |
3865 | ||
3866 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
3867 | vectype_in, &code1, &code2, |
3868 | &multi_step_cvt, &interm_types)) | |
4a00c761 JJ |
3869 | break; |
3870 | } | |
3871 | ||
3872 | if (rhs_mode == VOIDmode || GET_MODE_SIZE (rhs_mode) > fltsz) | |
3873 | goto unsupported; | |
3874 | ||
3875 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
3876 | codecvt2 = ERROR_MARK; | |
3877 | else | |
3878 | { | |
3879 | multi_step_cvt++; | |
9771b263 | 3880 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
3881 | cvt_type = NULL_TREE; |
3882 | } | |
3883 | break; | |
3884 | ||
3885 | case NARROW: | |
3886 | gcc_assert (op_type == unary_op); | |
3887 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
3888 | &code1, &multi_step_cvt, | |
3889 | &interm_types)) | |
3890 | break; | |
3891 | ||
3892 | if (code != FIX_TRUNC_EXPR | |
3893 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
3894 | >= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
3895 | goto unsupported; | |
3896 | ||
3897 | rhs_mode = TYPE_MODE (rhs_type); | |
3898 | cvt_type | |
3899 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
3900 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
3901 | if (cvt_type == NULL_TREE) | |
3902 | goto unsupported; | |
3903 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
3904 | &decl1, &codecvt1)) | |
3905 | goto unsupported; | |
3906 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
3907 | &code1, &multi_step_cvt, | |
3908 | &interm_types)) | |
3909 | break; | |
3910 | goto unsupported; | |
3911 | ||
3912 | default: | |
3913 | gcc_unreachable (); | |
ebfd146a IR |
3914 | } |
3915 | ||
3916 | if (!vec_stmt) /* transformation not required. */ | |
3917 | { | |
73fbfcad | 3918 | if (dump_enabled_p ()) |
78c60e3d | 3919 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3920 | "=== vectorizable_conversion ===\n"); |
4a00c761 | 3921 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
3922 | { |
3923 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
c3e7ee41 | 3924 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
8bd37302 | 3925 | } |
4a00c761 JJ |
3926 | else if (modifier == NARROW) |
3927 | { | |
3928 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
8bd37302 | 3929 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 JJ |
3930 | } |
3931 | else | |
3932 | { | |
3933 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
8bd37302 | 3934 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 | 3935 | } |
9771b263 | 3936 | interm_types.release (); |
ebfd146a IR |
3937 | return true; |
3938 | } | |
3939 | ||
3940 | /** Transform. **/ | |
73fbfcad | 3941 | if (dump_enabled_p ()) |
78c60e3d | 3942 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3943 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 3944 | |
4a00c761 JJ |
3945 | if (op_type == binary_op) |
3946 | { | |
3947 | if (CONSTANT_CLASS_P (op0)) | |
3948 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
3949 | else if (CONSTANT_CLASS_P (op1)) | |
3950 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
3951 | } | |
3952 | ||
3953 | /* In case of multi-step conversion, we first generate conversion operations | |
3954 | to the intermediate types, and then from that types to the final one. | |
3955 | We create vector destinations for the intermediate type (TYPES) received | |
3956 | from supportable_*_operation, and store them in the correct order | |
3957 | for future use in vect_create_vectorized_*_stmts (). */ | |
9771b263 | 3958 | vec_dsts.create (multi_step_cvt + 1); |
82294ec1 JJ |
3959 | vec_dest = vect_create_destination_var (scalar_dest, |
3960 | (cvt_type && modifier == WIDEN) | |
3961 | ? cvt_type : vectype_out); | |
9771b263 | 3962 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3963 | |
3964 | if (multi_step_cvt) | |
3965 | { | |
9771b263 DN |
3966 | for (i = interm_types.length () - 1; |
3967 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
3968 | { |
3969 | vec_dest = vect_create_destination_var (scalar_dest, | |
3970 | intermediate_type); | |
9771b263 | 3971 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3972 | } |
3973 | } | |
ebfd146a | 3974 | |
4a00c761 | 3975 | if (cvt_type) |
82294ec1 JJ |
3976 | vec_dest = vect_create_destination_var (scalar_dest, |
3977 | modifier == WIDEN | |
3978 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
3979 | |
3980 | if (!slp_node) | |
3981 | { | |
30862efc | 3982 | if (modifier == WIDEN) |
4a00c761 | 3983 | { |
c3284718 | 3984 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 3985 | if (op_type == binary_op) |
9771b263 | 3986 | vec_oprnds1.create (1); |
4a00c761 | 3987 | } |
30862efc | 3988 | else if (modifier == NARROW) |
9771b263 DN |
3989 | vec_oprnds0.create ( |
3990 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
3991 | } |
3992 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 3993 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 3994 | |
4a00c761 | 3995 | last_oprnd = op0; |
ebfd146a IR |
3996 | prev_stmt_info = NULL; |
3997 | switch (modifier) | |
3998 | { | |
3999 | case NONE: | |
4000 | for (j = 0; j < ncopies; j++) | |
4001 | { | |
ebfd146a | 4002 | if (j == 0) |
d092494c IR |
4003 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node, |
4004 | -1); | |
ebfd146a IR |
4005 | else |
4006 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
4007 | ||
9771b263 | 4008 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4009 | { |
4010 | /* Arguments are ready, create the new vector stmt. */ | |
4011 | if (code1 == CALL_EXPR) | |
4012 | { | |
4013 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4014 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4015 | gimple_call_set_lhs (new_stmt, new_temp); | |
4016 | } | |
4017 | else | |
4018 | { | |
4019 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
0d0e4a03 | 4020 | new_stmt = gimple_build_assign (vec_dest, code1, vop0); |
4a00c761 JJ |
4021 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4022 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4023 | } | |
4024 | ||
4025 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4026 | if (slp_node) | |
9771b263 | 4027 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
225ce44b RB |
4028 | else |
4029 | { | |
4030 | if (!prev_stmt_info) | |
4031 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4032 | else | |
4033 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4034 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4035 | } | |
4a00c761 | 4036 | } |
ebfd146a IR |
4037 | } |
4038 | break; | |
4039 | ||
4040 | case WIDEN: | |
4041 | /* In case the vectorization factor (VF) is bigger than the number | |
4042 | of elements that we can fit in a vectype (nunits), we have to | |
4043 | generate more than one vector stmt - i.e - we need to "unroll" | |
4044 | the vector stmt by a factor VF/nunits. */ | |
4045 | for (j = 0; j < ncopies; j++) | |
4046 | { | |
4a00c761 | 4047 | /* Handle uses. */ |
ebfd146a | 4048 | if (j == 0) |
4a00c761 JJ |
4049 | { |
4050 | if (slp_node) | |
4051 | { | |
4052 | if (code == WIDEN_LSHIFT_EXPR) | |
4053 | { | |
4054 | unsigned int k; | |
ebfd146a | 4055 | |
4a00c761 JJ |
4056 | vec_oprnd1 = op1; |
4057 | /* Store vec_oprnd1 for every vector stmt to be created | |
4058 | for SLP_NODE. We check during the analysis that all | |
4059 | the shift arguments are the same. */ | |
4060 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4061 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4062 | |
4063 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
4064 | slp_node, -1); | |
4065 | } | |
4066 | else | |
4067 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
4068 | &vec_oprnds1, slp_node, -1); | |
4069 | } | |
4070 | else | |
4071 | { | |
81c40241 | 4072 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 4073 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
4074 | if (op_type == binary_op) |
4075 | { | |
4076 | if (code == WIDEN_LSHIFT_EXPR) | |
4077 | vec_oprnd1 = op1; | |
4078 | else | |
81c40241 | 4079 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 4080 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4081 | } |
4082 | } | |
4083 | } | |
ebfd146a | 4084 | else |
4a00c761 JJ |
4085 | { |
4086 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
4087 | vec_oprnds0.truncate (0); |
4088 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
4089 | if (op_type == binary_op) |
4090 | { | |
4091 | if (code == WIDEN_LSHIFT_EXPR) | |
4092 | vec_oprnd1 = op1; | |
4093 | else | |
4094 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
4095 | vec_oprnd1); | |
9771b263 DN |
4096 | vec_oprnds1.truncate (0); |
4097 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
4098 | } |
4099 | } | |
ebfd146a | 4100 | |
4a00c761 JJ |
4101 | /* Arguments are ready. Create the new vector stmts. */ |
4102 | for (i = multi_step_cvt; i >= 0; i--) | |
4103 | { | |
9771b263 | 4104 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
4105 | enum tree_code c1 = code1, c2 = code2; |
4106 | if (i == 0 && codecvt2 != ERROR_MARK) | |
4107 | { | |
4108 | c1 = codecvt1; | |
4109 | c2 = codecvt2; | |
4110 | } | |
4111 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
4112 | &vec_oprnds1, | |
4113 | stmt, this_dest, gsi, | |
4114 | c1, c2, decl1, decl2, | |
4115 | op_type); | |
4116 | } | |
4117 | ||
9771b263 | 4118 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4119 | { |
4120 | if (cvt_type) | |
4121 | { | |
4122 | if (codecvt1 == CALL_EXPR) | |
4123 | { | |
4124 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4125 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4126 | gimple_call_set_lhs (new_stmt, new_temp); | |
4127 | } | |
4128 | else | |
4129 | { | |
4130 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4131 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4132 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4133 | vop0); | |
4a00c761 JJ |
4134 | } |
4135 | ||
4136 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4137 | } | |
4138 | else | |
4139 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
4140 | ||
4141 | if (slp_node) | |
9771b263 | 4142 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 4143 | else |
c689ce1e RB |
4144 | { |
4145 | if (!prev_stmt_info) | |
4146 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
4147 | else | |
4148 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4149 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4150 | } | |
4a00c761 | 4151 | } |
ebfd146a | 4152 | } |
4a00c761 JJ |
4153 | |
4154 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
4155 | break; |
4156 | ||
4157 | case NARROW: | |
4158 | /* In case the vectorization factor (VF) is bigger than the number | |
4159 | of elements that we can fit in a vectype (nunits), we have to | |
4160 | generate more than one vector stmt - i.e - we need to "unroll" | |
4161 | the vector stmt by a factor VF/nunits. */ | |
4162 | for (j = 0; j < ncopies; j++) | |
4163 | { | |
4164 | /* Handle uses. */ | |
4a00c761 JJ |
4165 | if (slp_node) |
4166 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
4167 | slp_node, -1); | |
ebfd146a IR |
4168 | else |
4169 | { | |
9771b263 | 4170 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
4171 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
4172 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
4173 | } |
4174 | ||
4a00c761 JJ |
4175 | /* Arguments are ready. Create the new vector stmts. */ |
4176 | if (cvt_type) | |
9771b263 | 4177 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4178 | { |
4179 | if (codecvt1 == CALL_EXPR) | |
4180 | { | |
4181 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4182 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4183 | gimple_call_set_lhs (new_stmt, new_temp); | |
4184 | } | |
4185 | else | |
4186 | { | |
4187 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4188 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4189 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4190 | vop0); | |
4a00c761 | 4191 | } |
ebfd146a | 4192 | |
4a00c761 | 4193 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 4194 | vec_oprnds0[i] = new_temp; |
4a00c761 | 4195 | } |
ebfd146a | 4196 | |
4a00c761 JJ |
4197 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
4198 | stmt, vec_dsts, gsi, | |
4199 | slp_node, code1, | |
4200 | &prev_stmt_info); | |
ebfd146a IR |
4201 | } |
4202 | ||
4203 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 4204 | break; |
ebfd146a IR |
4205 | } |
4206 | ||
9771b263 DN |
4207 | vec_oprnds0.release (); |
4208 | vec_oprnds1.release (); | |
4209 | vec_dsts.release (); | |
4210 | interm_types.release (); | |
ebfd146a IR |
4211 | |
4212 | return true; | |
4213 | } | |
ff802fa1 IR |
4214 | |
4215 | ||
ebfd146a IR |
4216 | /* Function vectorizable_assignment. |
4217 | ||
b8698a0f L |
4218 | Check if STMT performs an assignment (copy) that can be vectorized. |
4219 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4220 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4221 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4222 | ||
4223 | static bool | |
355fe088 TS |
4224 | vectorizable_assignment (gimple *stmt, gimple_stmt_iterator *gsi, |
4225 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
4226 | { |
4227 | tree vec_dest; | |
4228 | tree scalar_dest; | |
4229 | tree op; | |
4230 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a IR |
4231 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4232 | tree new_temp; | |
355fe088 | 4233 | gimple *def_stmt; |
ebfd146a | 4234 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
ebfd146a | 4235 | int ncopies; |
f18b55bd | 4236 | int i, j; |
6e1aa848 | 4237 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 4238 | tree vop; |
a70d6342 | 4239 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4240 | vec_info *vinfo = stmt_info->vinfo; |
355fe088 | 4241 | gimple *new_stmt = NULL; |
f18b55bd | 4242 | stmt_vec_info prev_stmt_info = NULL; |
fde9c428 RG |
4243 | enum tree_code code; |
4244 | tree vectype_in; | |
ebfd146a | 4245 | |
a70d6342 | 4246 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4247 | return false; |
4248 | ||
8644a673 | 4249 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
4250 | return false; |
4251 | ||
4252 | /* Is vectorizable assignment? */ | |
4253 | if (!is_gimple_assign (stmt)) | |
4254 | return false; | |
4255 | ||
4256 | scalar_dest = gimple_assign_lhs (stmt); | |
4257 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4258 | return false; | |
4259 | ||
fde9c428 | 4260 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 4261 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
4262 | || code == PAREN_EXPR |
4263 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
4264 | op = gimple_assign_rhs1 (stmt); |
4265 | else | |
4266 | return false; | |
4267 | ||
7b7ec6c5 RG |
4268 | if (code == VIEW_CONVERT_EXPR) |
4269 | op = TREE_OPERAND (op, 0); | |
4270 | ||
465c8c19 JJ |
4271 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
4272 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4273 | ||
4274 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4275 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4276 | case of SLP. */ | |
4277 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
4278 | ncopies = 1; | |
4279 | else | |
4280 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4281 | ||
4282 | gcc_assert (ncopies >= 1); | |
4283 | ||
81c40241 | 4284 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[0], &vectype_in)) |
ebfd146a | 4285 | { |
73fbfcad | 4286 | if (dump_enabled_p ()) |
78c60e3d | 4287 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4288 | "use not simple.\n"); |
ebfd146a IR |
4289 | return false; |
4290 | } | |
4291 | ||
fde9c428 RG |
4292 | /* We can handle NOP_EXPR conversions that do not change the number |
4293 | of elements or the vector size. */ | |
7b7ec6c5 RG |
4294 | if ((CONVERT_EXPR_CODE_P (code) |
4295 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 RG |
4296 | && (!vectype_in |
4297 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
4298 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
4299 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
4300 | return false; | |
4301 | ||
7b7b1813 RG |
4302 | /* We do not handle bit-precision changes. */ |
4303 | if ((CONVERT_EXPR_CODE_P (code) | |
4304 | || code == VIEW_CONVERT_EXPR) | |
4305 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
4306 | && ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4307 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4308 | || ((TYPE_PRECISION (TREE_TYPE (op)) | |
4309 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))) | |
4310 | /* But a conversion that does not change the bit-pattern is ok. */ | |
4311 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4312 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2dab46d5 IE |
4313 | && TYPE_UNSIGNED (TREE_TYPE (op))) |
4314 | /* Conversion between boolean types of different sizes is | |
4315 | a simple assignment in case their vectypes are same | |
4316 | boolean vectors. */ | |
4317 | && (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
4318 | || !VECTOR_BOOLEAN_TYPE_P (vectype_in))) | |
7b7b1813 | 4319 | { |
73fbfcad | 4320 | if (dump_enabled_p ()) |
78c60e3d SS |
4321 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4322 | "type conversion to/from bit-precision " | |
e645e942 | 4323 | "unsupported.\n"); |
7b7b1813 RG |
4324 | return false; |
4325 | } | |
4326 | ||
ebfd146a IR |
4327 | if (!vec_stmt) /* transformation not required. */ |
4328 | { | |
4329 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 4330 | if (dump_enabled_p ()) |
78c60e3d | 4331 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4332 | "=== vectorizable_assignment ===\n"); |
c3e7ee41 | 4333 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
4334 | return true; |
4335 | } | |
4336 | ||
4337 | /** Transform. **/ | |
73fbfcad | 4338 | if (dump_enabled_p ()) |
e645e942 | 4339 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
4340 | |
4341 | /* Handle def. */ | |
4342 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4343 | ||
4344 | /* Handle use. */ | |
f18b55bd | 4345 | for (j = 0; j < ncopies; j++) |
ebfd146a | 4346 | { |
f18b55bd IR |
4347 | /* Handle uses. */ |
4348 | if (j == 0) | |
d092494c | 4349 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node, -1); |
f18b55bd IR |
4350 | else |
4351 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
4352 | ||
4353 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 4354 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 4355 | { |
7b7ec6c5 RG |
4356 | if (CONVERT_EXPR_CODE_P (code) |
4357 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 4358 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
4359 | new_stmt = gimple_build_assign (vec_dest, vop); |
4360 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4361 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4362 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4363 | if (slp_node) | |
9771b263 | 4364 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 4365 | } |
ebfd146a IR |
4366 | |
4367 | if (slp_node) | |
f18b55bd IR |
4368 | continue; |
4369 | ||
4370 | if (j == 0) | |
4371 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4372 | else | |
4373 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4374 | ||
4375 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4376 | } | |
b8698a0f | 4377 | |
9771b263 | 4378 | vec_oprnds.release (); |
ebfd146a IR |
4379 | return true; |
4380 | } | |
4381 | ||
9dc3f7de | 4382 | |
1107f3ae IR |
4383 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
4384 | either as shift by a scalar or by a vector. */ | |
4385 | ||
4386 | bool | |
4387 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
4388 | { | |
4389 | ||
ef4bddc2 | 4390 | machine_mode vec_mode; |
1107f3ae IR |
4391 | optab optab; |
4392 | int icode; | |
4393 | tree vectype; | |
4394 | ||
4395 | vectype = get_vectype_for_scalar_type (scalar_type); | |
4396 | if (!vectype) | |
4397 | return false; | |
4398 | ||
4399 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4400 | if (!optab | |
4401 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
4402 | { | |
4403 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4404 | if (!optab | |
4405 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
4406 | == CODE_FOR_nothing)) | |
4407 | return false; | |
4408 | } | |
4409 | ||
4410 | vec_mode = TYPE_MODE (vectype); | |
4411 | icode = (int) optab_handler (optab, vec_mode); | |
4412 | if (icode == CODE_FOR_nothing) | |
4413 | return false; | |
4414 | ||
4415 | return true; | |
4416 | } | |
4417 | ||
4418 | ||
9dc3f7de IR |
4419 | /* Function vectorizable_shift. |
4420 | ||
4421 | Check if STMT performs a shift operation that can be vectorized. | |
4422 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4423 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
4424 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4425 | ||
4426 | static bool | |
355fe088 TS |
4427 | vectorizable_shift (gimple *stmt, gimple_stmt_iterator *gsi, |
4428 | gimple **vec_stmt, slp_tree slp_node) | |
9dc3f7de IR |
4429 | { |
4430 | tree vec_dest; | |
4431 | tree scalar_dest; | |
4432 | tree op0, op1 = NULL; | |
4433 | tree vec_oprnd1 = NULL_TREE; | |
4434 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4435 | tree vectype; | |
4436 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4437 | enum tree_code code; | |
ef4bddc2 | 4438 | machine_mode vec_mode; |
9dc3f7de IR |
4439 | tree new_temp; |
4440 | optab optab; | |
4441 | int icode; | |
ef4bddc2 | 4442 | machine_mode optab_op2_mode; |
355fe088 | 4443 | gimple *def_stmt; |
9dc3f7de | 4444 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
355fe088 | 4445 | gimple *new_stmt = NULL; |
9dc3f7de IR |
4446 | stmt_vec_info prev_stmt_info; |
4447 | int nunits_in; | |
4448 | int nunits_out; | |
4449 | tree vectype_out; | |
cede2577 | 4450 | tree op1_vectype; |
9dc3f7de IR |
4451 | int ncopies; |
4452 | int j, i; | |
6e1aa848 DN |
4453 | vec<tree> vec_oprnds0 = vNULL; |
4454 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
4455 | tree vop0, vop1; |
4456 | unsigned int k; | |
49eab32e | 4457 | bool scalar_shift_arg = true; |
9dc3f7de | 4458 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4459 | vec_info *vinfo = stmt_info->vinfo; |
9dc3f7de IR |
4460 | int vf; |
4461 | ||
4462 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
4463 | return false; | |
4464 | ||
4465 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
4466 | return false; | |
4467 | ||
4468 | /* Is STMT a vectorizable binary/unary operation? */ | |
4469 | if (!is_gimple_assign (stmt)) | |
4470 | return false; | |
4471 | ||
4472 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4473 | return false; | |
4474 | ||
4475 | code = gimple_assign_rhs_code (stmt); | |
4476 | ||
4477 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
4478 | || code == RROTATE_EXPR)) | |
4479 | return false; | |
4480 | ||
4481 | scalar_dest = gimple_assign_lhs (stmt); | |
4482 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
7b7b1813 RG |
4483 | if (TYPE_PRECISION (TREE_TYPE (scalar_dest)) |
4484 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4485 | { | |
73fbfcad | 4486 | if (dump_enabled_p ()) |
78c60e3d | 4487 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4488 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
4489 | return false; |
4490 | } | |
9dc3f7de IR |
4491 | |
4492 | op0 = gimple_assign_rhs1 (stmt); | |
81c40241 | 4493 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
9dc3f7de | 4494 | { |
73fbfcad | 4495 | if (dump_enabled_p ()) |
78c60e3d | 4496 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4497 | "use not simple.\n"); |
9dc3f7de IR |
4498 | return false; |
4499 | } | |
4500 | /* If op0 is an external or constant def use a vector type with | |
4501 | the same size as the output vector type. */ | |
4502 | if (!vectype) | |
4503 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
4504 | if (vec_stmt) | |
4505 | gcc_assert (vectype); | |
4506 | if (!vectype) | |
4507 | { | |
73fbfcad | 4508 | if (dump_enabled_p ()) |
78c60e3d | 4509 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4510 | "no vectype for scalar type\n"); |
9dc3f7de IR |
4511 | return false; |
4512 | } | |
4513 | ||
4514 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4515 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4516 | if (nunits_out != nunits_in) | |
4517 | return false; | |
4518 | ||
4519 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 4520 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &op1_vectype)) |
9dc3f7de | 4521 | { |
73fbfcad | 4522 | if (dump_enabled_p ()) |
78c60e3d | 4523 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4524 | "use not simple.\n"); |
9dc3f7de IR |
4525 | return false; |
4526 | } | |
4527 | ||
4528 | if (loop_vinfo) | |
4529 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4530 | else | |
4531 | vf = 1; | |
4532 | ||
4533 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4534 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4535 | case of SLP. */ | |
437f4a00 | 4536 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
9dc3f7de IR |
4537 | ncopies = 1; |
4538 | else | |
4539 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
4540 | ||
4541 | gcc_assert (ncopies >= 1); | |
4542 | ||
4543 | /* Determine whether the shift amount is a vector, or scalar. If the | |
4544 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
4545 | ||
dbfa87aa YR |
4546 | if ((dt[1] == vect_internal_def |
4547 | || dt[1] == vect_induction_def) | |
4548 | && !slp_node) | |
49eab32e JJ |
4549 | scalar_shift_arg = false; |
4550 | else if (dt[1] == vect_constant_def | |
4551 | || dt[1] == vect_external_def | |
4552 | || dt[1] == vect_internal_def) | |
4553 | { | |
4554 | /* In SLP, need to check whether the shift count is the same, | |
4555 | in loops if it is a constant or invariant, it is always | |
4556 | a scalar shift. */ | |
4557 | if (slp_node) | |
4558 | { | |
355fe088 TS |
4559 | vec<gimple *> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
4560 | gimple *slpstmt; | |
49eab32e | 4561 | |
9771b263 | 4562 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
4563 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
4564 | scalar_shift_arg = false; | |
4565 | } | |
4566 | } | |
4567 | else | |
4568 | { | |
73fbfcad | 4569 | if (dump_enabled_p ()) |
78c60e3d | 4570 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4571 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
4572 | return false; |
4573 | } | |
4574 | ||
9dc3f7de | 4575 | /* Vector shifted by vector. */ |
49eab32e | 4576 | if (!scalar_shift_arg) |
9dc3f7de IR |
4577 | { |
4578 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 4579 | if (dump_enabled_p ()) |
78c60e3d | 4580 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4581 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 4582 | |
aa948027 JJ |
4583 | if (!op1_vectype) |
4584 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
4585 | if (op1_vectype == NULL_TREE | |
4586 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 4587 | { |
73fbfcad | 4588 | if (dump_enabled_p ()) |
78c60e3d SS |
4589 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4590 | "unusable type for last operand in" | |
e645e942 | 4591 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
4592 | return false; |
4593 | } | |
9dc3f7de IR |
4594 | } |
4595 | /* See if the machine has a vector shifted by scalar insn and if not | |
4596 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 4597 | else |
9dc3f7de IR |
4598 | { |
4599 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4600 | if (optab | |
4601 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
4602 | { | |
73fbfcad | 4603 | if (dump_enabled_p ()) |
78c60e3d | 4604 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4605 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
4606 | } |
4607 | else | |
4608 | { | |
4609 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4610 | if (optab | |
4611 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
4612 | != CODE_FOR_nothing)) | |
4613 | { | |
49eab32e JJ |
4614 | scalar_shift_arg = false; |
4615 | ||
73fbfcad | 4616 | if (dump_enabled_p ()) |
78c60e3d | 4617 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4618 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
4619 | |
4620 | /* Unlike the other binary operators, shifts/rotates have | |
4621 | the rhs being int, instead of the same type as the lhs, | |
4622 | so make sure the scalar is the right type if we are | |
aa948027 | 4623 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
4624 | if (dt[1] == vect_constant_def) |
4625 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
4626 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
4627 | TREE_TYPE (op1))) | |
4628 | { | |
4629 | if (slp_node | |
4630 | && TYPE_MODE (TREE_TYPE (vectype)) | |
4631 | != TYPE_MODE (TREE_TYPE (op1))) | |
4632 | { | |
73fbfcad | 4633 | if (dump_enabled_p ()) |
78c60e3d SS |
4634 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4635 | "unusable type for last operand in" | |
e645e942 | 4636 | " vector/vector shift/rotate.\n"); |
21c0a521 | 4637 | return false; |
aa948027 JJ |
4638 | } |
4639 | if (vec_stmt && !slp_node) | |
4640 | { | |
4641 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
4642 | op1 = vect_init_vector (stmt, op1, | |
4643 | TREE_TYPE (vectype), NULL); | |
4644 | } | |
4645 | } | |
9dc3f7de IR |
4646 | } |
4647 | } | |
4648 | } | |
9dc3f7de IR |
4649 | |
4650 | /* Supportable by target? */ | |
4651 | if (!optab) | |
4652 | { | |
73fbfcad | 4653 | if (dump_enabled_p ()) |
78c60e3d | 4654 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4655 | "no optab.\n"); |
9dc3f7de IR |
4656 | return false; |
4657 | } | |
4658 | vec_mode = TYPE_MODE (vectype); | |
4659 | icode = (int) optab_handler (optab, vec_mode); | |
4660 | if (icode == CODE_FOR_nothing) | |
4661 | { | |
73fbfcad | 4662 | if (dump_enabled_p ()) |
78c60e3d | 4663 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4664 | "op not supported by target.\n"); |
9dc3f7de IR |
4665 | /* Check only during analysis. */ |
4666 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
4667 | || (vf < vect_min_worthwhile_factor (code) | |
4668 | && !vec_stmt)) | |
4669 | return false; | |
73fbfcad | 4670 | if (dump_enabled_p ()) |
e645e942 TJ |
4671 | dump_printf_loc (MSG_NOTE, vect_location, |
4672 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
4673 | } |
4674 | ||
4675 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
4676 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
4677 | && vf < vect_min_worthwhile_factor (code) | |
4678 | && !vec_stmt) | |
4679 | { | |
73fbfcad | 4680 | if (dump_enabled_p ()) |
78c60e3d | 4681 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4682 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
4683 | return false; |
4684 | } | |
4685 | ||
4686 | if (!vec_stmt) /* transformation not required. */ | |
4687 | { | |
4688 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 4689 | if (dump_enabled_p ()) |
e645e942 TJ |
4690 | dump_printf_loc (MSG_NOTE, vect_location, |
4691 | "=== vectorizable_shift ===\n"); | |
c3e7ee41 | 4692 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
9dc3f7de IR |
4693 | return true; |
4694 | } | |
4695 | ||
4696 | /** Transform. **/ | |
4697 | ||
73fbfcad | 4698 | if (dump_enabled_p ()) |
78c60e3d | 4699 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4700 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
4701 | |
4702 | /* Handle def. */ | |
4703 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4704 | ||
9dc3f7de IR |
4705 | prev_stmt_info = NULL; |
4706 | for (j = 0; j < ncopies; j++) | |
4707 | { | |
4708 | /* Handle uses. */ | |
4709 | if (j == 0) | |
4710 | { | |
4711 | if (scalar_shift_arg) | |
4712 | { | |
4713 | /* Vector shl and shr insn patterns can be defined with scalar | |
4714 | operand 2 (shift operand). In this case, use constant or loop | |
4715 | invariant op1 directly, without extending it to vector mode | |
4716 | first. */ | |
4717 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
4718 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
4719 | { | |
73fbfcad | 4720 | if (dump_enabled_p ()) |
78c60e3d | 4721 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4722 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 4723 | vec_oprnd1 = op1; |
8930f723 | 4724 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 4725 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
4726 | if (slp_node) |
4727 | { | |
4728 | /* Store vec_oprnd1 for every vector stmt to be created | |
4729 | for SLP_NODE. We check during the analysis that all | |
4730 | the shift arguments are the same. | |
4731 | TODO: Allow different constants for different vector | |
4732 | stmts generated for an SLP instance. */ | |
4733 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4734 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
4735 | } |
4736 | } | |
4737 | } | |
4738 | ||
4739 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
4740 | (a special case for certain kind of vector shifts); otherwise, | |
4741 | operand 1 should be of a vector type (the usual case). */ | |
4742 | if (vec_oprnd1) | |
4743 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
d092494c | 4744 | slp_node, -1); |
9dc3f7de IR |
4745 | else |
4746 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
d092494c | 4747 | slp_node, -1); |
9dc3f7de IR |
4748 | } |
4749 | else | |
4750 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
4751 | ||
4752 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 4753 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 4754 | { |
9771b263 | 4755 | vop1 = vec_oprnds1[i]; |
0d0e4a03 | 4756 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
9dc3f7de IR |
4757 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4758 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4759 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4760 | if (slp_node) | |
9771b263 | 4761 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
4762 | } |
4763 | ||
4764 | if (slp_node) | |
4765 | continue; | |
4766 | ||
4767 | if (j == 0) | |
4768 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4769 | else | |
4770 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4771 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4772 | } | |
4773 | ||
9771b263 DN |
4774 | vec_oprnds0.release (); |
4775 | vec_oprnds1.release (); | |
9dc3f7de IR |
4776 | |
4777 | return true; | |
4778 | } | |
4779 | ||
4780 | ||
ebfd146a IR |
4781 | /* Function vectorizable_operation. |
4782 | ||
16949072 RG |
4783 | Check if STMT performs a binary, unary or ternary operation that can |
4784 | be vectorized. | |
b8698a0f | 4785 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
4786 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4787 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4788 | ||
4789 | static bool | |
355fe088 TS |
4790 | vectorizable_operation (gimple *stmt, gimple_stmt_iterator *gsi, |
4791 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a | 4792 | { |
00f07b86 | 4793 | tree vec_dest; |
ebfd146a | 4794 | tree scalar_dest; |
16949072 | 4795 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 4796 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 4797 | tree vectype; |
ebfd146a IR |
4798 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4799 | enum tree_code code; | |
ef4bddc2 | 4800 | machine_mode vec_mode; |
ebfd146a IR |
4801 | tree new_temp; |
4802 | int op_type; | |
00f07b86 | 4803 | optab optab; |
523ba738 | 4804 | bool target_support_p; |
355fe088 | 4805 | gimple *def_stmt; |
16949072 RG |
4806 | enum vect_def_type dt[3] |
4807 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
355fe088 | 4808 | gimple *new_stmt = NULL; |
ebfd146a | 4809 | stmt_vec_info prev_stmt_info; |
b690cc0f | 4810 | int nunits_in; |
ebfd146a IR |
4811 | int nunits_out; |
4812 | tree vectype_out; | |
4813 | int ncopies; | |
4814 | int j, i; | |
6e1aa848 DN |
4815 | vec<tree> vec_oprnds0 = vNULL; |
4816 | vec<tree> vec_oprnds1 = vNULL; | |
4817 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 4818 | tree vop0, vop1, vop2; |
a70d6342 | 4819 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4820 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 IR |
4821 | int vf; |
4822 | ||
a70d6342 | 4823 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4824 | return false; |
4825 | ||
8644a673 | 4826 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
4827 | return false; |
4828 | ||
4829 | /* Is STMT a vectorizable binary/unary operation? */ | |
4830 | if (!is_gimple_assign (stmt)) | |
4831 | return false; | |
4832 | ||
4833 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4834 | return false; | |
4835 | ||
ebfd146a IR |
4836 | code = gimple_assign_rhs_code (stmt); |
4837 | ||
4838 | /* For pointer addition, we should use the normal plus for | |
4839 | the vector addition. */ | |
4840 | if (code == POINTER_PLUS_EXPR) | |
4841 | code = PLUS_EXPR; | |
4842 | ||
4843 | /* Support only unary or binary operations. */ | |
4844 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 4845 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 4846 | { |
73fbfcad | 4847 | if (dump_enabled_p ()) |
78c60e3d | 4848 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4849 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 4850 | op_type); |
ebfd146a IR |
4851 | return false; |
4852 | } | |
4853 | ||
b690cc0f RG |
4854 | scalar_dest = gimple_assign_lhs (stmt); |
4855 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
4856 | ||
7b7b1813 RG |
4857 | /* Most operations cannot handle bit-precision types without extra |
4858 | truncations. */ | |
045c1278 IE |
4859 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4860 | && (TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4861 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
7b7b1813 RG |
4862 | /* Exception are bitwise binary operations. */ |
4863 | && code != BIT_IOR_EXPR | |
4864 | && code != BIT_XOR_EXPR | |
4865 | && code != BIT_AND_EXPR) | |
4866 | { | |
73fbfcad | 4867 | if (dump_enabled_p ()) |
78c60e3d | 4868 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4869 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
4870 | return false; |
4871 | } | |
4872 | ||
ebfd146a | 4873 | op0 = gimple_assign_rhs1 (stmt); |
81c40241 | 4874 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
ebfd146a | 4875 | { |
73fbfcad | 4876 | if (dump_enabled_p ()) |
78c60e3d | 4877 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4878 | "use not simple.\n"); |
ebfd146a IR |
4879 | return false; |
4880 | } | |
b690cc0f RG |
4881 | /* If op0 is an external or constant def use a vector type with |
4882 | the same size as the output vector type. */ | |
4883 | if (!vectype) | |
b036c6c5 IE |
4884 | { |
4885 | /* For boolean type we cannot determine vectype by | |
4886 | invariant value (don't know whether it is a vector | |
4887 | of booleans or vector of integers). We use output | |
4888 | vectype because operations on boolean don't change | |
4889 | type. */ | |
4890 | if (TREE_CODE (TREE_TYPE (op0)) == BOOLEAN_TYPE) | |
4891 | { | |
4892 | if (TREE_CODE (TREE_TYPE (scalar_dest)) != BOOLEAN_TYPE) | |
4893 | { | |
4894 | if (dump_enabled_p ()) | |
4895 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
4896 | "not supported operation on bool value.\n"); | |
4897 | return false; | |
4898 | } | |
4899 | vectype = vectype_out; | |
4900 | } | |
4901 | else | |
4902 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
4903 | } | |
7d8930a0 IR |
4904 | if (vec_stmt) |
4905 | gcc_assert (vectype); | |
4906 | if (!vectype) | |
4907 | { | |
73fbfcad | 4908 | if (dump_enabled_p ()) |
7d8930a0 | 4909 | { |
78c60e3d SS |
4910 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4911 | "no vectype for scalar type "); | |
4912 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
4913 | TREE_TYPE (op0)); | |
e645e942 | 4914 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
4915 | } |
4916 | ||
4917 | return false; | |
4918 | } | |
b690cc0f RG |
4919 | |
4920 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4921 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4922 | if (nunits_out != nunits_in) | |
4923 | return false; | |
ebfd146a | 4924 | |
16949072 | 4925 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
4926 | { |
4927 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 4928 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1])) |
ebfd146a | 4929 | { |
73fbfcad | 4930 | if (dump_enabled_p ()) |
78c60e3d | 4931 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4932 | "use not simple.\n"); |
ebfd146a IR |
4933 | return false; |
4934 | } | |
4935 | } | |
16949072 RG |
4936 | if (op_type == ternary_op) |
4937 | { | |
4938 | op2 = gimple_assign_rhs3 (stmt); | |
81c40241 | 4939 | if (!vect_is_simple_use (op2, vinfo, &def_stmt, &dt[2])) |
16949072 | 4940 | { |
73fbfcad | 4941 | if (dump_enabled_p ()) |
78c60e3d | 4942 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4943 | "use not simple.\n"); |
16949072 RG |
4944 | return false; |
4945 | } | |
4946 | } | |
ebfd146a | 4947 | |
b690cc0f RG |
4948 | if (loop_vinfo) |
4949 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4950 | else | |
4951 | vf = 1; | |
4952 | ||
4953 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 4954 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 4955 | case of SLP. */ |
437f4a00 | 4956 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
b690cc0f RG |
4957 | ncopies = 1; |
4958 | else | |
4959 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
4960 | ||
4961 | gcc_assert (ncopies >= 1); | |
4962 | ||
9dc3f7de | 4963 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
4964 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
4965 | || code == RROTATE_EXPR) | |
9dc3f7de | 4966 | return false; |
ebfd146a | 4967 | |
ebfd146a | 4968 | /* Supportable by target? */ |
00f07b86 RH |
4969 | |
4970 | vec_mode = TYPE_MODE (vectype); | |
4971 | if (code == MULT_HIGHPART_EXPR) | |
523ba738 | 4972 | target_support_p = can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype)); |
00f07b86 RH |
4973 | else |
4974 | { | |
4975 | optab = optab_for_tree_code (code, vectype, optab_default); | |
4976 | if (!optab) | |
5deb57cb | 4977 | { |
73fbfcad | 4978 | if (dump_enabled_p ()) |
78c60e3d | 4979 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4980 | "no optab.\n"); |
00f07b86 | 4981 | return false; |
5deb57cb | 4982 | } |
523ba738 RS |
4983 | target_support_p = (optab_handler (optab, vec_mode) |
4984 | != CODE_FOR_nothing); | |
5deb57cb JJ |
4985 | } |
4986 | ||
523ba738 | 4987 | if (!target_support_p) |
ebfd146a | 4988 | { |
73fbfcad | 4989 | if (dump_enabled_p ()) |
78c60e3d | 4990 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4991 | "op not supported by target.\n"); |
ebfd146a IR |
4992 | /* Check only during analysis. */ |
4993 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
5deb57cb | 4994 | || (!vec_stmt && vf < vect_min_worthwhile_factor (code))) |
ebfd146a | 4995 | return false; |
73fbfcad | 4996 | if (dump_enabled_p ()) |
e645e942 TJ |
4997 | dump_printf_loc (MSG_NOTE, vect_location, |
4998 | "proceeding using word mode.\n"); | |
383d9c83 IR |
4999 | } |
5000 | ||
4a00c761 | 5001 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
5002 | if (!VECTOR_MODE_P (vec_mode) |
5003 | && !vec_stmt | |
5004 | && vf < vect_min_worthwhile_factor (code)) | |
7d8930a0 | 5005 | { |
73fbfcad | 5006 | if (dump_enabled_p ()) |
78c60e3d | 5007 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5008 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 5009 | return false; |
7d8930a0 | 5010 | } |
ebfd146a | 5011 | |
ebfd146a IR |
5012 | if (!vec_stmt) /* transformation not required. */ |
5013 | { | |
4a00c761 | 5014 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 5015 | if (dump_enabled_p ()) |
78c60e3d | 5016 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5017 | "=== vectorizable_operation ===\n"); |
c3e7ee41 | 5018 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
5019 | return true; |
5020 | } | |
5021 | ||
5022 | /** Transform. **/ | |
5023 | ||
73fbfcad | 5024 | if (dump_enabled_p ()) |
78c60e3d | 5025 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5026 | "transform binary/unary operation.\n"); |
383d9c83 | 5027 | |
ebfd146a | 5028 | /* Handle def. */ |
00f07b86 | 5029 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 5030 | |
ebfd146a IR |
5031 | /* In case the vectorization factor (VF) is bigger than the number |
5032 | of elements that we can fit in a vectype (nunits), we have to generate | |
5033 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
5034 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
5035 | from one copy of the vector stmt to the next, in the field | |
5036 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
5037 | stages to find the correct vector defs to be used when vectorizing | |
5038 | stmts that use the defs of the current stmt. The example below | |
5039 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
5040 | we need to create 4 vectorized stmts): | |
5041 | ||
5042 | before vectorization: | |
5043 | RELATED_STMT VEC_STMT | |
5044 | S1: x = memref - - | |
5045 | S2: z = x + 1 - - | |
5046 | ||
5047 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
5048 | there): | |
5049 | RELATED_STMT VEC_STMT | |
5050 | VS1_0: vx0 = memref0 VS1_1 - | |
5051 | VS1_1: vx1 = memref1 VS1_2 - | |
5052 | VS1_2: vx2 = memref2 VS1_3 - | |
5053 | VS1_3: vx3 = memref3 - - | |
5054 | S1: x = load - VS1_0 | |
5055 | S2: z = x + 1 - - | |
5056 | ||
5057 | step2: vectorize stmt S2 (done here): | |
5058 | To vectorize stmt S2 we first need to find the relevant vector | |
5059 | def for the first operand 'x'. This is, as usual, obtained from | |
5060 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
5061 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
5062 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
5063 | the vector stmt VS2_0, and as usual, record it in the | |
5064 | STMT_VINFO_VEC_STMT of stmt S2. | |
5065 | When creating the second copy (VS2_1), we obtain the relevant vector | |
5066 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
5067 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
5068 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
5069 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
5070 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
5071 | chain of stmts and pointers: | |
5072 | RELATED_STMT VEC_STMT | |
5073 | VS1_0: vx0 = memref0 VS1_1 - | |
5074 | VS1_1: vx1 = memref1 VS1_2 - | |
5075 | VS1_2: vx2 = memref2 VS1_3 - | |
5076 | VS1_3: vx3 = memref3 - - | |
5077 | S1: x = load - VS1_0 | |
5078 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
5079 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
5080 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
5081 | VS2_3: vz3 = vx3 + v1 - - | |
5082 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
5083 | |
5084 | prev_stmt_info = NULL; | |
5085 | for (j = 0; j < ncopies; j++) | |
5086 | { | |
5087 | /* Handle uses. */ | |
5088 | if (j == 0) | |
4a00c761 JJ |
5089 | { |
5090 | if (op_type == binary_op || op_type == ternary_op) | |
5091 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
5092 | slp_node, -1); | |
5093 | else | |
5094 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
5095 | slp_node, -1); | |
5096 | if (op_type == ternary_op) | |
36ba4aae | 5097 | { |
9771b263 DN |
5098 | vec_oprnds2.create (1); |
5099 | vec_oprnds2.quick_push (vect_get_vec_def_for_operand (op2, | |
81c40241 | 5100 | stmt)); |
36ba4aae | 5101 | } |
4a00c761 | 5102 | } |
ebfd146a | 5103 | else |
4a00c761 JJ |
5104 | { |
5105 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5106 | if (op_type == ternary_op) | |
5107 | { | |
9771b263 DN |
5108 | tree vec_oprnd = vec_oprnds2.pop (); |
5109 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
5110 | vec_oprnd)); | |
4a00c761 JJ |
5111 | } |
5112 | } | |
5113 | ||
5114 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 5115 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 5116 | { |
4a00c761 | 5117 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 5118 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 5119 | vop2 = ((op_type == ternary_op) |
9771b263 | 5120 | ? vec_oprnds2[i] : NULL_TREE); |
0d0e4a03 | 5121 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1, vop2); |
4a00c761 JJ |
5122 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5123 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5124 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5125 | if (slp_node) | |
9771b263 | 5126 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
5127 | } |
5128 | ||
4a00c761 JJ |
5129 | if (slp_node) |
5130 | continue; | |
5131 | ||
5132 | if (j == 0) | |
5133 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5134 | else | |
5135 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5136 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
5137 | } |
5138 | ||
9771b263 DN |
5139 | vec_oprnds0.release (); |
5140 | vec_oprnds1.release (); | |
5141 | vec_oprnds2.release (); | |
ebfd146a | 5142 | |
ebfd146a IR |
5143 | return true; |
5144 | } | |
5145 | ||
c716e67f XDL |
5146 | /* A helper function to ensure data reference DR's base alignment |
5147 | for STMT_INFO. */ | |
5148 | ||
5149 | static void | |
5150 | ensure_base_align (stmt_vec_info stmt_info, struct data_reference *dr) | |
5151 | { | |
5152 | if (!dr->aux) | |
5153 | return; | |
5154 | ||
52639a61 | 5155 | if (DR_VECT_AUX (dr)->base_misaligned) |
c716e67f XDL |
5156 | { |
5157 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
52639a61 | 5158 | tree base_decl = DR_VECT_AUX (dr)->base_decl; |
c716e67f | 5159 | |
428f0c67 JH |
5160 | if (decl_in_symtab_p (base_decl)) |
5161 | symtab_node::get (base_decl)->increase_alignment (TYPE_ALIGN (vectype)); | |
5162 | else | |
5163 | { | |
5164 | DECL_ALIGN (base_decl) = TYPE_ALIGN (vectype); | |
5165 | DECL_USER_ALIGN (base_decl) = 1; | |
5166 | } | |
52639a61 | 5167 | DR_VECT_AUX (dr)->base_misaligned = false; |
c716e67f XDL |
5168 | } |
5169 | } | |
5170 | ||
ebfd146a | 5171 | |
09dfa495 BM |
5172 | /* Given a vector type VECTYPE returns the VECTOR_CST mask that implements |
5173 | reversal of the vector elements. If that is impossible to do, | |
5174 | returns NULL. */ | |
5175 | ||
5176 | static tree | |
5177 | perm_mask_for_reverse (tree vectype) | |
5178 | { | |
5179 | int i, nunits; | |
5180 | unsigned char *sel; | |
5181 | ||
5182 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
5183 | sel = XALLOCAVEC (unsigned char, nunits); | |
5184 | ||
5185 | for (i = 0; i < nunits; ++i) | |
5186 | sel[i] = nunits - 1 - i; | |
5187 | ||
557be5a8 AL |
5188 | if (!can_vec_perm_p (TYPE_MODE (vectype), false, sel)) |
5189 | return NULL_TREE; | |
5190 | return vect_gen_perm_mask_checked (vectype, sel); | |
09dfa495 BM |
5191 | } |
5192 | ||
ebfd146a IR |
5193 | /* Function vectorizable_store. |
5194 | ||
b8698a0f L |
5195 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
5196 | can be vectorized. | |
5197 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
5198 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5199 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5200 | ||
5201 | static bool | |
355fe088 | 5202 | vectorizable_store (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 5203 | slp_tree slp_node) |
ebfd146a IR |
5204 | { |
5205 | tree scalar_dest; | |
5206 | tree data_ref; | |
5207 | tree op; | |
5208 | tree vec_oprnd = NULL_TREE; | |
5209 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5210 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
272c6793 | 5211 | tree elem_type; |
ebfd146a | 5212 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 5213 | struct loop *loop = NULL; |
ef4bddc2 | 5214 | machine_mode vec_mode; |
ebfd146a IR |
5215 | tree dummy; |
5216 | enum dr_alignment_support alignment_support_scheme; | |
355fe088 | 5217 | gimple *def_stmt; |
ebfd146a IR |
5218 | enum vect_def_type dt; |
5219 | stmt_vec_info prev_stmt_info = NULL; | |
5220 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 5221 | tree dataref_offset = NULL_TREE; |
355fe088 | 5222 | gimple *ptr_incr = NULL; |
ebfd146a IR |
5223 | int ncopies; |
5224 | int j; | |
355fe088 | 5225 | gimple *next_stmt, *first_stmt = NULL; |
0d0293ac | 5226 | bool grouped_store = false; |
272c6793 | 5227 | bool store_lanes_p = false; |
ebfd146a | 5228 | unsigned int group_size, i; |
6e1aa848 DN |
5229 | vec<tree> dr_chain = vNULL; |
5230 | vec<tree> oprnds = vNULL; | |
5231 | vec<tree> result_chain = vNULL; | |
ebfd146a | 5232 | bool inv_p; |
09dfa495 BM |
5233 | bool negative = false; |
5234 | tree offset = NULL_TREE; | |
6e1aa848 | 5235 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 5236 | bool slp = (slp_node != NULL); |
ebfd146a | 5237 | unsigned int vec_num; |
a70d6342 | 5238 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5239 | vec_info *vinfo = stmt_info->vinfo; |
272c6793 | 5240 | tree aggr_type; |
3bab6342 AT |
5241 | tree scatter_base = NULL_TREE, scatter_off = NULL_TREE; |
5242 | tree scatter_off_vectype = NULL_TREE, scatter_decl = NULL_TREE; | |
5243 | int scatter_scale = 1; | |
5244 | enum vect_def_type scatter_idx_dt = vect_unknown_def_type; | |
5245 | enum vect_def_type scatter_src_dt = vect_unknown_def_type; | |
355fe088 | 5246 | gimple *new_stmt; |
a70d6342 | 5247 | |
a70d6342 | 5248 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5249 | return false; |
5250 | ||
8644a673 | 5251 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
5252 | return false; |
5253 | ||
5254 | /* Is vectorizable store? */ | |
5255 | ||
5256 | if (!is_gimple_assign (stmt)) | |
5257 | return false; | |
5258 | ||
5259 | scalar_dest = gimple_assign_lhs (stmt); | |
ab0ef706 JJ |
5260 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
5261 | && is_pattern_stmt_p (stmt_info)) | |
5262 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
ebfd146a | 5263 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
38000232 | 5264 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF |
ebfd146a | 5265 | && TREE_CODE (scalar_dest) != INDIRECT_REF |
e9dbe7bb IR |
5266 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
5267 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
5268 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
5269 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
5270 | return false; |
5271 | ||
5272 | gcc_assert (gimple_assign_single_p (stmt)); | |
465c8c19 JJ |
5273 | |
5274 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
5275 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
5276 | ||
5277 | if (loop_vinfo) | |
5278 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
5279 | ||
5280 | /* Multiple types in SLP are handled by creating the appropriate number of | |
5281 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5282 | case of SLP. */ | |
5283 | if (slp || PURE_SLP_STMT (stmt_info)) | |
5284 | ncopies = 1; | |
5285 | else | |
5286 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
5287 | ||
5288 | gcc_assert (ncopies >= 1); | |
5289 | ||
5290 | /* FORNOW. This restriction should be relaxed. */ | |
5291 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) | |
5292 | { | |
5293 | if (dump_enabled_p ()) | |
5294 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5295 | "multiple types in nested loop.\n"); | |
5296 | return false; | |
5297 | } | |
5298 | ||
ebfd146a | 5299 | op = gimple_assign_rhs1 (stmt); |
81c40241 | 5300 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt)) |
ebfd146a | 5301 | { |
73fbfcad | 5302 | if (dump_enabled_p ()) |
78c60e3d | 5303 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5304 | "use not simple.\n"); |
ebfd146a IR |
5305 | return false; |
5306 | } | |
5307 | ||
272c6793 | 5308 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 5309 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 5310 | |
ebfd146a IR |
5311 | /* FORNOW. In some cases can vectorize even if data-type not supported |
5312 | (e.g. - array initialization with 0). */ | |
947131ba | 5313 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
5314 | return false; |
5315 | ||
5316 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
5317 | return false; | |
5318 | ||
f2e2a985 | 5319 | if (!STMT_VINFO_STRIDED_P (stmt_info)) |
09dfa495 | 5320 | { |
f2e2a985 MM |
5321 | negative = |
5322 | tree_int_cst_compare (loop && nested_in_vect_loop_p (loop, stmt) | |
5323 | ? STMT_VINFO_DR_STEP (stmt_info) : DR_STEP (dr), | |
5324 | size_zero_node) < 0; | |
5325 | if (negative && ncopies > 1) | |
09dfa495 BM |
5326 | { |
5327 | if (dump_enabled_p ()) | |
5328 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
f2e2a985 | 5329 | "multiple types with negative step.\n"); |
09dfa495 BM |
5330 | return false; |
5331 | } | |
f2e2a985 | 5332 | if (negative) |
09dfa495 | 5333 | { |
f2e2a985 MM |
5334 | gcc_assert (!grouped_store); |
5335 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
5336 | if (alignment_support_scheme != dr_aligned | |
5337 | && alignment_support_scheme != dr_unaligned_supported) | |
5338 | { | |
5339 | if (dump_enabled_p ()) | |
5340 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5341 | "negative step but alignment required.\n"); | |
5342 | return false; | |
5343 | } | |
5344 | if (dt != vect_constant_def | |
5345 | && dt != vect_external_def | |
5346 | && !perm_mask_for_reverse (vectype)) | |
5347 | { | |
5348 | if (dump_enabled_p ()) | |
5349 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5350 | "negative step and reversing not supported.\n"); | |
5351 | return false; | |
5352 | } | |
09dfa495 BM |
5353 | } |
5354 | } | |
5355 | ||
0d0293ac | 5356 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 5357 | { |
0d0293ac | 5358 | grouped_store = true; |
e14c1050 | 5359 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
cee62fee MM |
5360 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
5361 | if (!slp | |
5362 | && !PURE_SLP_STMT (stmt_info) | |
5363 | && !STMT_VINFO_STRIDED_P (stmt_info)) | |
b602d918 | 5364 | { |
272c6793 RS |
5365 | if (vect_store_lanes_supported (vectype, group_size)) |
5366 | store_lanes_p = true; | |
0d0293ac | 5367 | else if (!vect_grouped_store_supported (vectype, group_size)) |
b602d918 RS |
5368 | return false; |
5369 | } | |
b8698a0f | 5370 | |
cee62fee MM |
5371 | if (STMT_VINFO_STRIDED_P (stmt_info) |
5372 | && (slp || PURE_SLP_STMT (stmt_info)) | |
5373 | && (group_size > nunits | |
5374 | || nunits % group_size != 0)) | |
5375 | { | |
5376 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5377 | "unhandled strided group store\n"); | |
5378 | return false; | |
5379 | } | |
5380 | ||
ebfd146a IR |
5381 | if (first_stmt == stmt) |
5382 | { | |
5383 | /* STMT is the leader of the group. Check the operands of all the | |
5384 | stmts of the group. */ | |
e14c1050 | 5385 | next_stmt = GROUP_NEXT_ELEMENT (stmt_info); |
ebfd146a IR |
5386 | while (next_stmt) |
5387 | { | |
5388 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
5389 | op = gimple_assign_rhs1 (next_stmt); | |
81c40241 | 5390 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt)) |
ebfd146a | 5391 | { |
73fbfcad | 5392 | if (dump_enabled_p ()) |
78c60e3d | 5393 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5394 | "use not simple.\n"); |
ebfd146a IR |
5395 | return false; |
5396 | } | |
e14c1050 | 5397 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
5398 | } |
5399 | } | |
5400 | } | |
5401 | ||
3bab6342 AT |
5402 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5403 | { | |
355fe088 | 5404 | gimple *def_stmt; |
3bab6342 AT |
5405 | scatter_decl = vect_check_gather_scatter (stmt, loop_vinfo, &scatter_base, |
5406 | &scatter_off, &scatter_scale); | |
5407 | gcc_assert (scatter_decl); | |
81c40241 RB |
5408 | if (!vect_is_simple_use (scatter_off, vinfo, &def_stmt, &scatter_idx_dt, |
5409 | &scatter_off_vectype)) | |
3bab6342 AT |
5410 | { |
5411 | if (dump_enabled_p ()) | |
5412 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5413 | "scatter index use not simple."); | |
5414 | return false; | |
5415 | } | |
5416 | } | |
5417 | ||
ebfd146a IR |
5418 | if (!vec_stmt) /* transformation not required. */ |
5419 | { | |
5420 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
2e8ab70c RB |
5421 | /* The SLP costs are calculated during SLP analysis. */ |
5422 | if (!PURE_SLP_STMT (stmt_info)) | |
5423 | vect_model_store_cost (stmt_info, ncopies, store_lanes_p, dt, | |
5424 | NULL, NULL, NULL); | |
ebfd146a IR |
5425 | return true; |
5426 | } | |
5427 | ||
5428 | /** Transform. **/ | |
5429 | ||
c716e67f XDL |
5430 | ensure_base_align (stmt_info, dr); |
5431 | ||
3bab6342 AT |
5432 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5433 | { | |
5434 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE, op, src; | |
5435 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (scatter_decl)); | |
5436 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
5437 | tree ptr, mask, var, scale, perm_mask = NULL_TREE; | |
5438 | edge pe = loop_preheader_edge (loop); | |
5439 | gimple_seq seq; | |
5440 | basic_block new_bb; | |
5441 | enum { NARROW, NONE, WIDEN } modifier; | |
5442 | int scatter_off_nunits = TYPE_VECTOR_SUBPARTS (scatter_off_vectype); | |
5443 | ||
5444 | if (nunits == (unsigned int) scatter_off_nunits) | |
5445 | modifier = NONE; | |
5446 | else if (nunits == (unsigned int) scatter_off_nunits / 2) | |
5447 | { | |
5448 | unsigned char *sel = XALLOCAVEC (unsigned char, scatter_off_nunits); | |
5449 | modifier = WIDEN; | |
5450 | ||
5451 | for (i = 0; i < (unsigned int) scatter_off_nunits; ++i) | |
5452 | sel[i] = i | nunits; | |
5453 | ||
5454 | perm_mask = vect_gen_perm_mask_checked (scatter_off_vectype, sel); | |
5455 | gcc_assert (perm_mask != NULL_TREE); | |
5456 | } | |
5457 | else if (nunits == (unsigned int) scatter_off_nunits * 2) | |
5458 | { | |
5459 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
5460 | modifier = NARROW; | |
5461 | ||
5462 | for (i = 0; i < (unsigned int) nunits; ++i) | |
5463 | sel[i] = i | scatter_off_nunits; | |
5464 | ||
5465 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); | |
5466 | gcc_assert (perm_mask != NULL_TREE); | |
5467 | ncopies *= 2; | |
5468 | } | |
5469 | else | |
5470 | gcc_unreachable (); | |
5471 | ||
5472 | rettype = TREE_TYPE (TREE_TYPE (scatter_decl)); | |
5473 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5474 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5475 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5476 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5477 | scaletype = TREE_VALUE (arglist); | |
5478 | ||
5479 | gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE | |
5480 | && TREE_CODE (rettype) == VOID_TYPE); | |
5481 | ||
5482 | ptr = fold_convert (ptrtype, scatter_base); | |
5483 | if (!is_gimple_min_invariant (ptr)) | |
5484 | { | |
5485 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
5486 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
5487 | gcc_assert (!new_bb); | |
5488 | } | |
5489 | ||
5490 | /* Currently we support only unconditional scatter stores, | |
5491 | so mask should be all ones. */ | |
5492 | mask = build_int_cst (masktype, -1); | |
5493 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
5494 | ||
5495 | scale = build_int_cst (scaletype, scatter_scale); | |
5496 | ||
5497 | prev_stmt_info = NULL; | |
5498 | for (j = 0; j < ncopies; ++j) | |
5499 | { | |
5500 | if (j == 0) | |
5501 | { | |
5502 | src = vec_oprnd1 | |
81c40241 | 5503 | = vect_get_vec_def_for_operand (gimple_assign_rhs1 (stmt), stmt); |
3bab6342 | 5504 | op = vec_oprnd0 |
81c40241 | 5505 | = vect_get_vec_def_for_operand (scatter_off, stmt); |
3bab6342 AT |
5506 | } |
5507 | else if (modifier != NONE && (j & 1)) | |
5508 | { | |
5509 | if (modifier == WIDEN) | |
5510 | { | |
5511 | src = vec_oprnd1 | |
5512 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5513 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, perm_mask, | |
5514 | stmt, gsi); | |
5515 | } | |
5516 | else if (modifier == NARROW) | |
5517 | { | |
5518 | src = permute_vec_elements (vec_oprnd1, vec_oprnd1, perm_mask, | |
5519 | stmt, gsi); | |
5520 | op = vec_oprnd0 | |
5521 | = vect_get_vec_def_for_stmt_copy (scatter_idx_dt, vec_oprnd0); | |
5522 | } | |
5523 | else | |
5524 | gcc_unreachable (); | |
5525 | } | |
5526 | else | |
5527 | { | |
5528 | src = vec_oprnd1 | |
5529 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5530 | op = vec_oprnd0 | |
5531 | = vect_get_vec_def_for_stmt_copy (scatter_idx_dt, vec_oprnd0); | |
5532 | } | |
5533 | ||
5534 | if (!useless_type_conversion_p (srctype, TREE_TYPE (src))) | |
5535 | { | |
5536 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)) | |
5537 | == TYPE_VECTOR_SUBPARTS (srctype)); | |
0e22bb5a | 5538 | var = vect_get_new_ssa_name (srctype, vect_simple_var); |
3bab6342 AT |
5539 | src = build1 (VIEW_CONVERT_EXPR, srctype, src); |
5540 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, src); | |
5541 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5542 | src = var; | |
5543 | } | |
5544 | ||
5545 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
5546 | { | |
5547 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
5548 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 5549 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
3bab6342 AT |
5550 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
5551 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
5552 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5553 | op = var; | |
5554 | } | |
5555 | ||
5556 | new_stmt | |
5557 | = gimple_build_call (scatter_decl, 5, ptr, mask, op, src, scale); | |
5558 | ||
5559 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5560 | ||
5561 | if (prev_stmt_info == NULL) | |
5562 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5563 | else | |
5564 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5565 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5566 | } | |
5567 | return true; | |
5568 | } | |
5569 | ||
0d0293ac | 5570 | if (grouped_store) |
ebfd146a IR |
5571 | { |
5572 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 5573 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 5574 | |
e14c1050 | 5575 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
5576 | |
5577 | /* FORNOW */ | |
a70d6342 | 5578 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
5579 | |
5580 | /* We vectorize all the stmts of the interleaving group when we | |
5581 | reach the last stmt in the group. */ | |
e14c1050 IR |
5582 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
5583 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
5584 | && !slp) |
5585 | { | |
5586 | *vec_stmt = NULL; | |
5587 | return true; | |
5588 | } | |
5589 | ||
5590 | if (slp) | |
4b5caab7 | 5591 | { |
0d0293ac | 5592 | grouped_store = false; |
4b5caab7 IR |
5593 | /* VEC_NUM is the number of vect stmts to be created for this |
5594 | group. */ | |
5595 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 5596 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
52eab378 | 5597 | gcc_assert (GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_stmt)) == first_stmt); |
4b5caab7 | 5598 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
d092494c | 5599 | op = gimple_assign_rhs1 (first_stmt); |
4b5caab7 | 5600 | } |
ebfd146a | 5601 | else |
4b5caab7 IR |
5602 | /* VEC_NUM is the number of vect stmts to be created for this |
5603 | group. */ | |
ebfd146a IR |
5604 | vec_num = group_size; |
5605 | } | |
b8698a0f | 5606 | else |
ebfd146a IR |
5607 | { |
5608 | first_stmt = stmt; | |
5609 | first_dr = dr; | |
5610 | group_size = vec_num = 1; | |
ebfd146a | 5611 | } |
b8698a0f | 5612 | |
73fbfcad | 5613 | if (dump_enabled_p ()) |
78c60e3d | 5614 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5615 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 5616 | |
f2e2a985 MM |
5617 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
5618 | { | |
5619 | gimple_stmt_iterator incr_gsi; | |
5620 | bool insert_after; | |
355fe088 | 5621 | gimple *incr; |
f2e2a985 MM |
5622 | tree offvar; |
5623 | tree ivstep; | |
5624 | tree running_off; | |
5625 | gimple_seq stmts = NULL; | |
5626 | tree stride_base, stride_step, alias_off; | |
5627 | tree vec_oprnd; | |
f502d50e | 5628 | unsigned int g; |
f2e2a985 MM |
5629 | |
5630 | gcc_assert (!nested_in_vect_loop_p (loop, stmt)); | |
5631 | ||
5632 | stride_base | |
5633 | = fold_build_pointer_plus | |
f502d50e | 5634 | (unshare_expr (DR_BASE_ADDRESS (first_dr)), |
f2e2a985 | 5635 | size_binop (PLUS_EXPR, |
f502d50e MM |
5636 | convert_to_ptrofftype (unshare_expr (DR_OFFSET (first_dr))), |
5637 | convert_to_ptrofftype (DR_INIT(first_dr)))); | |
5638 | stride_step = fold_convert (sizetype, unshare_expr (DR_STEP (first_dr))); | |
f2e2a985 MM |
5639 | |
5640 | /* For a store with loop-invariant (but other than power-of-2) | |
5641 | stride (i.e. not a grouped access) like so: | |
5642 | ||
5643 | for (i = 0; i < n; i += stride) | |
5644 | array[i] = ...; | |
5645 | ||
5646 | we generate a new induction variable and new stores from | |
5647 | the components of the (vectorized) rhs: | |
5648 | ||
5649 | for (j = 0; ; j += VF*stride) | |
5650 | vectemp = ...; | |
5651 | tmp1 = vectemp[0]; | |
5652 | array[j] = tmp1; | |
5653 | tmp2 = vectemp[1]; | |
5654 | array[j + stride] = tmp2; | |
5655 | ... | |
5656 | */ | |
5657 | ||
cee62fee MM |
5658 | unsigned nstores = nunits; |
5659 | tree ltype = elem_type; | |
5660 | if (slp) | |
5661 | { | |
5662 | nstores = nunits / group_size; | |
5663 | if (group_size < nunits) | |
5664 | ltype = build_vector_type (elem_type, group_size); | |
5665 | else | |
5666 | ltype = vectype; | |
5667 | ltype = build_aligned_type (ltype, TYPE_ALIGN (elem_type)); | |
5668 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
f502d50e | 5669 | group_size = 1; |
cee62fee MM |
5670 | } |
5671 | ||
f2e2a985 MM |
5672 | ivstep = stride_step; |
5673 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
5674 | build_int_cst (TREE_TYPE (ivstep), | |
cee62fee | 5675 | ncopies * nstores)); |
f2e2a985 MM |
5676 | |
5677 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
5678 | ||
5679 | create_iv (stride_base, ivstep, NULL, | |
5680 | loop, &incr_gsi, insert_after, | |
5681 | &offvar, NULL); | |
5682 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 5683 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
f2e2a985 MM |
5684 | |
5685 | stride_step = force_gimple_operand (stride_step, &stmts, true, NULL_TREE); | |
5686 | if (stmts) | |
5687 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
5688 | ||
5689 | prev_stmt_info = NULL; | |
f502d50e MM |
5690 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (first_dr)), 0); |
5691 | next_stmt = first_stmt; | |
5692 | for (g = 0; g < group_size; g++) | |
f2e2a985 | 5693 | { |
f502d50e MM |
5694 | running_off = offvar; |
5695 | if (g) | |
f2e2a985 | 5696 | { |
f502d50e MM |
5697 | tree size = TYPE_SIZE_UNIT (ltype); |
5698 | tree pos = fold_build2 (MULT_EXPR, sizetype, size_int (g), | |
f2e2a985 | 5699 | size); |
f502d50e | 5700 | tree newoff = copy_ssa_name (running_off, NULL); |
f2e2a985 | 5701 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
f502d50e | 5702 | running_off, pos); |
f2e2a985 | 5703 | vect_finish_stmt_generation (stmt, incr, gsi); |
f2e2a985 | 5704 | running_off = newoff; |
f502d50e MM |
5705 | } |
5706 | for (j = 0; j < ncopies; j++) | |
5707 | { | |
5708 | /* We've set op and dt above, from gimple_assign_rhs1(stmt), | |
5709 | and first_stmt == stmt. */ | |
5710 | if (j == 0) | |
5711 | { | |
5712 | if (slp) | |
5713 | { | |
5714 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, NULL, | |
5715 | slp_node, -1); | |
5716 | vec_oprnd = vec_oprnds[0]; | |
5717 | } | |
5718 | else | |
5719 | { | |
5720 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
5721 | op = gimple_assign_rhs1 (next_stmt); | |
81c40241 | 5722 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
f502d50e MM |
5723 | } |
5724 | } | |
f2e2a985 | 5725 | else |
f502d50e MM |
5726 | { |
5727 | if (slp) | |
5728 | vec_oprnd = vec_oprnds[j]; | |
5729 | else | |
c079cbac | 5730 | { |
81c40241 | 5731 | vect_is_simple_use (vec_oprnd, vinfo, &def_stmt, &dt); |
c079cbac RB |
5732 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); |
5733 | } | |
f502d50e MM |
5734 | } |
5735 | ||
5736 | for (i = 0; i < nstores; i++) | |
5737 | { | |
5738 | tree newref, newoff; | |
355fe088 | 5739 | gimple *incr, *assign; |
f502d50e MM |
5740 | tree size = TYPE_SIZE (ltype); |
5741 | /* Extract the i'th component. */ | |
5742 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, | |
5743 | bitsize_int (i), size); | |
5744 | tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd, | |
5745 | size, pos); | |
5746 | ||
5747 | elem = force_gimple_operand_gsi (gsi, elem, true, | |
5748 | NULL_TREE, true, | |
5749 | GSI_SAME_STMT); | |
5750 | ||
5751 | newref = build2 (MEM_REF, ltype, | |
5752 | running_off, alias_off); | |
5753 | ||
5754 | /* And store it to *running_off. */ | |
5755 | assign = gimple_build_assign (newref, elem); | |
5756 | vect_finish_stmt_generation (stmt, assign, gsi); | |
5757 | ||
5758 | newoff = copy_ssa_name (running_off, NULL); | |
5759 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
5760 | running_off, stride_step); | |
5761 | vect_finish_stmt_generation (stmt, incr, gsi); | |
5762 | ||
5763 | running_off = newoff; | |
225ce44b RB |
5764 | if (g == group_size - 1 |
5765 | && !slp) | |
f502d50e MM |
5766 | { |
5767 | if (j == 0 && i == 0) | |
225ce44b RB |
5768 | STMT_VINFO_VEC_STMT (stmt_info) |
5769 | = *vec_stmt = assign; | |
f502d50e MM |
5770 | else |
5771 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = assign; | |
5772 | prev_stmt_info = vinfo_for_stmt (assign); | |
5773 | } | |
5774 | } | |
f2e2a985 | 5775 | } |
f502d50e | 5776 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
f2e2a985 MM |
5777 | } |
5778 | return true; | |
5779 | } | |
5780 | ||
9771b263 DN |
5781 | dr_chain.create (group_size); |
5782 | oprnds.create (group_size); | |
ebfd146a | 5783 | |
720f5239 | 5784 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 5785 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
5786 | /* Targets with store-lane instructions must not require explicit |
5787 | realignment. */ | |
5788 | gcc_assert (!store_lanes_p | |
5789 | || alignment_support_scheme == dr_aligned | |
5790 | || alignment_support_scheme == dr_unaligned_supported); | |
5791 | ||
09dfa495 BM |
5792 | if (negative) |
5793 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
5794 | ||
272c6793 RS |
5795 | if (store_lanes_p) |
5796 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
5797 | else | |
5798 | aggr_type = vectype; | |
ebfd146a IR |
5799 | |
5800 | /* In case the vectorization factor (VF) is bigger than the number | |
5801 | of elements that we can fit in a vectype (nunits), we have to generate | |
5802 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 5803 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
5804 | vect_get_vec_def_for_copy_stmt. */ |
5805 | ||
0d0293ac | 5806 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
5807 | |
5808 | S1: &base + 2 = x2 | |
5809 | S2: &base = x0 | |
5810 | S3: &base + 1 = x1 | |
5811 | S4: &base + 3 = x3 | |
5812 | ||
5813 | We create vectorized stores starting from base address (the access of the | |
5814 | first stmt in the chain (S2 in the above example), when the last store stmt | |
5815 | of the chain (S4) is reached: | |
5816 | ||
5817 | VS1: &base = vx2 | |
5818 | VS2: &base + vec_size*1 = vx0 | |
5819 | VS3: &base + vec_size*2 = vx1 | |
5820 | VS4: &base + vec_size*3 = vx3 | |
5821 | ||
5822 | Then permutation statements are generated: | |
5823 | ||
3fcc1b55 JJ |
5824 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
5825 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 5826 | ... |
b8698a0f | 5827 | |
ebfd146a IR |
5828 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
5829 | (the order of the data-refs in the output of vect_permute_store_chain | |
5830 | corresponds to the order of scalar stmts in the interleaving chain - see | |
5831 | the documentation of vect_permute_store_chain()). | |
5832 | ||
5833 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 5834 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 5835 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 5836 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
5837 | */ |
5838 | ||
5839 | prev_stmt_info = NULL; | |
5840 | for (j = 0; j < ncopies; j++) | |
5841 | { | |
ebfd146a IR |
5842 | |
5843 | if (j == 0) | |
5844 | { | |
5845 | if (slp) | |
5846 | { | |
5847 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c IR |
5848 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
5849 | NULL, slp_node, -1); | |
ebfd146a | 5850 | |
9771b263 | 5851 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
5852 | } |
5853 | else | |
5854 | { | |
b8698a0f L |
5855 | /* For interleaved stores we collect vectorized defs for all the |
5856 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
5857 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
5858 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
5859 | ||
0d0293ac | 5860 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 5861 | OPRNDS are of size 1. */ |
b8698a0f | 5862 | next_stmt = first_stmt; |
ebfd146a IR |
5863 | for (i = 0; i < group_size; i++) |
5864 | { | |
b8698a0f L |
5865 | /* Since gaps are not supported for interleaved stores, |
5866 | GROUP_SIZE is the exact number of stmts in the chain. | |
5867 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
5868 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
5869 | iteration of the loop will be executed. */ |
5870 | gcc_assert (next_stmt | |
5871 | && gimple_assign_single_p (next_stmt)); | |
5872 | op = gimple_assign_rhs1 (next_stmt); | |
5873 | ||
81c40241 | 5874 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
9771b263 DN |
5875 | dr_chain.quick_push (vec_oprnd); |
5876 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 5877 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
5878 | } |
5879 | } | |
5880 | ||
5881 | /* We should have catched mismatched types earlier. */ | |
5882 | gcc_assert (useless_type_conversion_p (vectype, | |
5883 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
5884 | bool simd_lane_access_p |
5885 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
5886 | if (simd_lane_access_p | |
5887 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
5888 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
5889 | && integer_zerop (DR_OFFSET (first_dr)) | |
5890 | && integer_zerop (DR_INIT (first_dr)) | |
5891 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
5892 | get_alias_set (DR_REF (first_dr)))) | |
5893 | { | |
5894 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
5895 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
5896 | (DR_REF (first_dr)), 0); | |
8928eff3 | 5897 | inv_p = false; |
74bf76ed JJ |
5898 | } |
5899 | else | |
5900 | dataref_ptr | |
5901 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
5902 | simd_lane_access_p ? loop : NULL, | |
09dfa495 | 5903 | offset, &dummy, gsi, &ptr_incr, |
74bf76ed | 5904 | simd_lane_access_p, &inv_p); |
a70d6342 | 5905 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 5906 | } |
b8698a0f | 5907 | else |
ebfd146a | 5908 | { |
b8698a0f L |
5909 | /* For interleaved stores we created vectorized defs for all the |
5910 | defs stored in OPRNDS in the previous iteration (previous copy). | |
5911 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
5912 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
5913 | next copy. | |
0d0293ac | 5914 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
5915 | OPRNDS are of size 1. */ |
5916 | for (i = 0; i < group_size; i++) | |
5917 | { | |
9771b263 | 5918 | op = oprnds[i]; |
81c40241 | 5919 | vect_is_simple_use (op, vinfo, &def_stmt, &dt); |
b8698a0f | 5920 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
9771b263 DN |
5921 | dr_chain[i] = vec_oprnd; |
5922 | oprnds[i] = vec_oprnd; | |
ebfd146a | 5923 | } |
74bf76ed JJ |
5924 | if (dataref_offset) |
5925 | dataref_offset | |
5926 | = int_const_binop (PLUS_EXPR, dataref_offset, | |
5927 | TYPE_SIZE_UNIT (aggr_type)); | |
5928 | else | |
5929 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
5930 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
5931 | } |
5932 | ||
272c6793 | 5933 | if (store_lanes_p) |
ebfd146a | 5934 | { |
272c6793 | 5935 | tree vec_array; |
267d3070 | 5936 | |
272c6793 RS |
5937 | /* Combine all the vectors into an array. */ |
5938 | vec_array = create_vector_array (vectype, vec_num); | |
5939 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 5940 | { |
9771b263 | 5941 | vec_oprnd = dr_chain[i]; |
272c6793 | 5942 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 5943 | } |
b8698a0f | 5944 | |
272c6793 RS |
5945 | /* Emit: |
5946 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
5947 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
5948 | new_stmt = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array); | |
5949 | gimple_call_set_lhs (new_stmt, data_ref); | |
267d3070 | 5950 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
5951 | } |
5952 | else | |
5953 | { | |
5954 | new_stmt = NULL; | |
0d0293ac | 5955 | if (grouped_store) |
272c6793 | 5956 | { |
b6b9227d JJ |
5957 | if (j == 0) |
5958 | result_chain.create (group_size); | |
272c6793 RS |
5959 | /* Permute. */ |
5960 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
5961 | &result_chain); | |
5962 | } | |
c2d7ab2a | 5963 | |
272c6793 RS |
5964 | next_stmt = first_stmt; |
5965 | for (i = 0; i < vec_num; i++) | |
5966 | { | |
644ffefd | 5967 | unsigned align, misalign; |
272c6793 RS |
5968 | |
5969 | if (i > 0) | |
5970 | /* Bump the vector pointer. */ | |
5971 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
5972 | stmt, NULL_TREE); | |
5973 | ||
5974 | if (slp) | |
9771b263 | 5975 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
5976 | else if (grouped_store) |
5977 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 5978 | vect_permute_store_chain(). */ |
9771b263 | 5979 | vec_oprnd = result_chain[i]; |
272c6793 | 5980 | |
aed93b23 RB |
5981 | data_ref = fold_build2 (MEM_REF, TREE_TYPE (vec_oprnd), |
5982 | dataref_ptr, | |
5983 | dataref_offset | |
5984 | ? dataref_offset | |
5985 | : build_int_cst (reference_alias_ptr_type | |
5986 | (DR_REF (first_dr)), 0)); | |
644ffefd | 5987 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 | 5988 | if (aligned_access_p (first_dr)) |
644ffefd | 5989 | misalign = 0; |
272c6793 RS |
5990 | else if (DR_MISALIGNMENT (first_dr) == -1) |
5991 | { | |
52639a61 RB |
5992 | if (DR_VECT_AUX (first_dr)->base_element_aligned) |
5993 | align = TYPE_ALIGN_UNIT (elem_type); | |
5994 | else | |
5995 | align = get_object_alignment (DR_REF (first_dr)) | |
5996 | / BITS_PER_UNIT; | |
5997 | misalign = 0; | |
272c6793 RS |
5998 | TREE_TYPE (data_ref) |
5999 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 6000 | align * BITS_PER_UNIT); |
272c6793 RS |
6001 | } |
6002 | else | |
6003 | { | |
6004 | TREE_TYPE (data_ref) | |
6005 | = build_aligned_type (TREE_TYPE (data_ref), | |
6006 | TYPE_ALIGN (elem_type)); | |
644ffefd | 6007 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 6008 | } |
aed93b23 RB |
6009 | if (dataref_offset == NULL_TREE |
6010 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
6011 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
6012 | misalign); | |
c2d7ab2a | 6013 | |
f234d260 BM |
6014 | if (negative |
6015 | && dt != vect_constant_def | |
6016 | && dt != vect_external_def) | |
09dfa495 BM |
6017 | { |
6018 | tree perm_mask = perm_mask_for_reverse (vectype); | |
6019 | tree perm_dest | |
6020 | = vect_create_destination_var (gimple_assign_rhs1 (stmt), | |
6021 | vectype); | |
b731b390 | 6022 | tree new_temp = make_ssa_name (perm_dest); |
09dfa495 BM |
6023 | |
6024 | /* Generate the permute statement. */ | |
355fe088 | 6025 | gimple *perm_stmt |
0d0e4a03 JJ |
6026 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
6027 | vec_oprnd, perm_mask); | |
09dfa495 BM |
6028 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
6029 | ||
6030 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); | |
6031 | vec_oprnd = new_temp; | |
6032 | } | |
6033 | ||
272c6793 RS |
6034 | /* Arguments are ready. Create the new vector stmt. */ |
6035 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
6036 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
6037 | |
6038 | if (slp) | |
6039 | continue; | |
6040 | ||
e14c1050 | 6041 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
6042 | if (!next_stmt) |
6043 | break; | |
6044 | } | |
ebfd146a | 6045 | } |
1da0876c RS |
6046 | if (!slp) |
6047 | { | |
6048 | if (j == 0) | |
6049 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6050 | else | |
6051 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6052 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6053 | } | |
ebfd146a IR |
6054 | } |
6055 | ||
9771b263 DN |
6056 | dr_chain.release (); |
6057 | oprnds.release (); | |
6058 | result_chain.release (); | |
6059 | vec_oprnds.release (); | |
ebfd146a IR |
6060 | |
6061 | return true; | |
6062 | } | |
6063 | ||
557be5a8 AL |
6064 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
6065 | VECTOR_CST mask. No checks are made that the target platform supports the | |
6066 | mask, so callers may wish to test can_vec_perm_p separately, or use | |
6067 | vect_gen_perm_mask_checked. */ | |
a1e53f3f | 6068 | |
3fcc1b55 | 6069 | tree |
557be5a8 | 6070 | vect_gen_perm_mask_any (tree vectype, const unsigned char *sel) |
a1e53f3f | 6071 | { |
d2a12ae7 | 6072 | tree mask_elt_type, mask_type, mask_vec, *mask_elts; |
2635892a | 6073 | int i, nunits; |
a1e53f3f | 6074 | |
22e4dee7 | 6075 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
22e4dee7 | 6076 | |
96f9265a RG |
6077 | mask_elt_type = lang_hooks.types.type_for_mode |
6078 | (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype))), 1); | |
22e4dee7 | 6079 | mask_type = get_vectype_for_scalar_type (mask_elt_type); |
a1e53f3f | 6080 | |
d2a12ae7 | 6081 | mask_elts = XALLOCAVEC (tree, nunits); |
aec7ae7d | 6082 | for (i = nunits - 1; i >= 0; i--) |
d2a12ae7 RG |
6083 | mask_elts[i] = build_int_cst (mask_elt_type, sel[i]); |
6084 | mask_vec = build_vector (mask_type, mask_elts); | |
a1e53f3f | 6085 | |
2635892a | 6086 | return mask_vec; |
a1e53f3f L |
6087 | } |
6088 | ||
cf7aa6a3 AL |
6089 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_p, |
6090 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ | |
557be5a8 AL |
6091 | |
6092 | tree | |
6093 | vect_gen_perm_mask_checked (tree vectype, const unsigned char *sel) | |
6094 | { | |
6095 | gcc_assert (can_vec_perm_p (TYPE_MODE (vectype), false, sel)); | |
6096 | return vect_gen_perm_mask_any (vectype, sel); | |
6097 | } | |
6098 | ||
aec7ae7d JJ |
6099 | /* Given a vector variable X and Y, that was generated for the scalar |
6100 | STMT, generate instructions to permute the vector elements of X and Y | |
6101 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
6102 | permuted vector variable. */ | |
a1e53f3f L |
6103 | |
6104 | static tree | |
355fe088 | 6105 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple *stmt, |
aec7ae7d | 6106 | gimple_stmt_iterator *gsi) |
a1e53f3f L |
6107 | { |
6108 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 6109 | tree perm_dest, data_ref; |
355fe088 | 6110 | gimple *perm_stmt; |
a1e53f3f | 6111 | |
acdcd61b | 6112 | perm_dest = vect_create_destination_var (gimple_get_lhs (stmt), vectype); |
b731b390 | 6113 | data_ref = make_ssa_name (perm_dest); |
a1e53f3f L |
6114 | |
6115 | /* Generate the permute statement. */ | |
0d0e4a03 | 6116 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
a1e53f3f L |
6117 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
6118 | ||
6119 | return data_ref; | |
6120 | } | |
6121 | ||
6b916b36 RB |
6122 | /* Hoist the definitions of all SSA uses on STMT out of the loop LOOP, |
6123 | inserting them on the loops preheader edge. Returns true if we | |
6124 | were successful in doing so (and thus STMT can be moved then), | |
6125 | otherwise returns false. */ | |
6126 | ||
6127 | static bool | |
355fe088 | 6128 | hoist_defs_of_uses (gimple *stmt, struct loop *loop) |
6b916b36 RB |
6129 | { |
6130 | ssa_op_iter i; | |
6131 | tree op; | |
6132 | bool any = false; | |
6133 | ||
6134 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6135 | { | |
355fe088 | 6136 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6137 | if (!gimple_nop_p (def_stmt) |
6138 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6139 | { | |
6140 | /* Make sure we don't need to recurse. While we could do | |
6141 | so in simple cases when there are more complex use webs | |
6142 | we don't have an easy way to preserve stmt order to fulfil | |
6143 | dependencies within them. */ | |
6144 | tree op2; | |
6145 | ssa_op_iter i2; | |
d1417442 JJ |
6146 | if (gimple_code (def_stmt) == GIMPLE_PHI) |
6147 | return false; | |
6b916b36 RB |
6148 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
6149 | { | |
355fe088 | 6150 | gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2); |
6b916b36 RB |
6151 | if (!gimple_nop_p (def_stmt2) |
6152 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2))) | |
6153 | return false; | |
6154 | } | |
6155 | any = true; | |
6156 | } | |
6157 | } | |
6158 | ||
6159 | if (!any) | |
6160 | return true; | |
6161 | ||
6162 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6163 | { | |
355fe088 | 6164 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6165 | if (!gimple_nop_p (def_stmt) |
6166 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6167 | { | |
6168 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); | |
6169 | gsi_remove (&gsi, false); | |
6170 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); | |
6171 | } | |
6172 | } | |
6173 | ||
6174 | return true; | |
6175 | } | |
6176 | ||
ebfd146a IR |
6177 | /* vectorizable_load. |
6178 | ||
b8698a0f L |
6179 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
6180 | can be vectorized. | |
6181 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
6182 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
6183 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
6184 | ||
6185 | static bool | |
355fe088 | 6186 | vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 6187 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a IR |
6188 | { |
6189 | tree scalar_dest; | |
6190 | tree vec_dest = NULL; | |
6191 | tree data_ref = NULL; | |
6192 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 6193 | stmt_vec_info prev_stmt_info; |
ebfd146a | 6194 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 6195 | struct loop *loop = NULL; |
ebfd146a | 6196 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 6197 | bool nested_in_vect_loop = false; |
c716e67f | 6198 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
272c6793 | 6199 | tree elem_type; |
ebfd146a | 6200 | tree new_temp; |
ef4bddc2 | 6201 | machine_mode mode; |
355fe088 | 6202 | gimple *new_stmt = NULL; |
ebfd146a IR |
6203 | tree dummy; |
6204 | enum dr_alignment_support alignment_support_scheme; | |
6205 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 6206 | tree dataref_offset = NULL_TREE; |
355fe088 | 6207 | gimple *ptr_incr = NULL; |
ebfd146a | 6208 | int ncopies; |
9b999e8c | 6209 | int i, j, group_size = -1, group_gap_adj; |
ebfd146a IR |
6210 | tree msq = NULL_TREE, lsq; |
6211 | tree offset = NULL_TREE; | |
356bbc4c | 6212 | tree byte_offset = NULL_TREE; |
ebfd146a | 6213 | tree realignment_token = NULL_TREE; |
538dd0b7 | 6214 | gphi *phi = NULL; |
6e1aa848 | 6215 | vec<tree> dr_chain = vNULL; |
0d0293ac | 6216 | bool grouped_load = false; |
272c6793 | 6217 | bool load_lanes_p = false; |
355fe088 | 6218 | gimple *first_stmt; |
4f0a0218 | 6219 | gimple *first_stmt_for_drptr = NULL; |
ebfd146a | 6220 | bool inv_p; |
319e6439 | 6221 | bool negative = false; |
ebfd146a IR |
6222 | bool compute_in_loop = false; |
6223 | struct loop *at_loop; | |
6224 | int vec_num; | |
6225 | bool slp = (slp_node != NULL); | |
6226 | bool slp_perm = false; | |
6227 | enum tree_code code; | |
a70d6342 IR |
6228 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
6229 | int vf; | |
272c6793 | 6230 | tree aggr_type; |
aec7ae7d JJ |
6231 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; |
6232 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
6233 | int gather_scale = 1; | |
6234 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
310213d4 | 6235 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 6236 | |
465c8c19 JJ |
6237 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
6238 | return false; | |
6239 | ||
6240 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
6241 | return false; | |
6242 | ||
6243 | /* Is vectorizable load? */ | |
6244 | if (!is_gimple_assign (stmt)) | |
6245 | return false; | |
6246 | ||
6247 | scalar_dest = gimple_assign_lhs (stmt); | |
6248 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
6249 | return false; | |
6250 | ||
6251 | code = gimple_assign_rhs_code (stmt); | |
6252 | if (code != ARRAY_REF | |
6253 | && code != BIT_FIELD_REF | |
6254 | && code != INDIRECT_REF | |
6255 | && code != COMPONENT_REF | |
6256 | && code != IMAGPART_EXPR | |
6257 | && code != REALPART_EXPR | |
6258 | && code != MEM_REF | |
6259 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
6260 | return false; | |
6261 | ||
6262 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
6263 | return false; | |
6264 | ||
6265 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
6266 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
6267 | ||
a70d6342 IR |
6268 | if (loop_vinfo) |
6269 | { | |
6270 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
6271 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
6272 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
6273 | } | |
6274 | else | |
3533e503 | 6275 | vf = 1; |
ebfd146a IR |
6276 | |
6277 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 6278 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 6279 | case of SLP. */ |
437f4a00 | 6280 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
6281 | ncopies = 1; |
6282 | else | |
6283 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
6284 | ||
6285 | gcc_assert (ncopies >= 1); | |
6286 | ||
6287 | /* FORNOW. This restriction should be relaxed. */ | |
6288 | if (nested_in_vect_loop && ncopies > 1) | |
6289 | { | |
73fbfcad | 6290 | if (dump_enabled_p ()) |
78c60e3d | 6291 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6292 | "multiple types in nested loop.\n"); |
ebfd146a IR |
6293 | return false; |
6294 | } | |
6295 | ||
f2556b68 RB |
6296 | /* Invalidate assumptions made by dependence analysis when vectorization |
6297 | on the unrolled body effectively re-orders stmts. */ | |
6298 | if (ncopies > 1 | |
6299 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
6300 | && ((unsigned)LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
6301 | > STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
6302 | { | |
6303 | if (dump_enabled_p ()) | |
6304 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6305 | "cannot perform implicit CSE when unrolling " | |
6306 | "with negative dependence distance\n"); | |
6307 | return false; | |
6308 | } | |
6309 | ||
7b7b1813 | 6310 | elem_type = TREE_TYPE (vectype); |
947131ba | 6311 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
6312 | |
6313 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
6314 | (e.g. - data copies). */ | |
947131ba | 6315 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 6316 | { |
73fbfcad | 6317 | if (dump_enabled_p ()) |
78c60e3d | 6318 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6319 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
6320 | return false; |
6321 | } | |
6322 | ||
ebfd146a | 6323 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 6324 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 6325 | { |
0d0293ac | 6326 | grouped_load = true; |
ebfd146a | 6327 | /* FORNOW */ |
3bab6342 | 6328 | gcc_assert (!nested_in_vect_loop && !STMT_VINFO_GATHER_SCATTER_P (stmt_info)); |
ebfd146a | 6329 | |
e14c1050 | 6330 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
d5f035ea RB |
6331 | |
6332 | /* If this is single-element interleaving with an element distance | |
6333 | that leaves unused vector loads around punt - we at least create | |
6334 | very sub-optimal code in that case (and blow up memory, | |
6335 | see PR65518). */ | |
72c0f643 | 6336 | bool force_peeling = false; |
d5f035ea | 6337 | if (first_stmt == stmt |
72c0f643 RB |
6338 | && !GROUP_NEXT_ELEMENT (stmt_info)) |
6339 | { | |
6340 | if (GROUP_SIZE (stmt_info) > TYPE_VECTOR_SUBPARTS (vectype)) | |
6341 | { | |
6342 | if (dump_enabled_p ()) | |
6343 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6344 | "single-element interleaving not supported " | |
6345 | "for not adjacent vector loads\n"); | |
6346 | return false; | |
6347 | } | |
6348 | ||
6349 | /* Single-element interleaving requires peeling for gaps. */ | |
6350 | force_peeling = true; | |
6351 | } | |
6352 | ||
6353 | /* If there is a gap in the end of the group or the group size cannot | |
6354 | be made a multiple of the vector element count then we access excess | |
6355 | elements in the last iteration and thus need to peel that off. */ | |
6356 | if (loop_vinfo | |
6357 | && ! STMT_VINFO_STRIDED_P (stmt_info) | |
6358 | && (force_peeling | |
6359 | || GROUP_GAP (vinfo_for_stmt (first_stmt)) != 0 | |
6360 | || (!slp && vf % GROUP_SIZE (vinfo_for_stmt (first_stmt)) != 0))) | |
d5f035ea RB |
6361 | { |
6362 | if (dump_enabled_p ()) | |
6363 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
72c0f643 RB |
6364 | "Data access with gaps requires scalar " |
6365 | "epilogue loop\n"); | |
6366 | if (loop->inner) | |
6367 | { | |
6368 | if (dump_enabled_p ()) | |
6369 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6370 | "Peeling for outer loop is not supported\n"); | |
6371 | return false; | |
6372 | } | |
6373 | ||
6374 | LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true; | |
d5f035ea RB |
6375 | } |
6376 | ||
b1af7da6 RB |
6377 | if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
6378 | slp_perm = true; | |
6379 | ||
7b5fc413 RB |
6380 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
6381 | if (!slp | |
6382 | && !PURE_SLP_STMT (stmt_info) | |
f2e2a985 | 6383 | && !STMT_VINFO_STRIDED_P (stmt_info)) |
b602d918 | 6384 | { |
272c6793 RS |
6385 | if (vect_load_lanes_supported (vectype, group_size)) |
6386 | load_lanes_p = true; | |
0d0293ac | 6387 | else if (!vect_grouped_load_supported (vectype, group_size)) |
b602d918 RS |
6388 | return false; |
6389 | } | |
f2556b68 RB |
6390 | |
6391 | /* Invalidate assumptions made by dependence analysis when vectorization | |
6392 | on the unrolled body effectively re-orders stmts. */ | |
6393 | if (!PURE_SLP_STMT (stmt_info) | |
6394 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
6395 | && ((unsigned)LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
6396 | > STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
6397 | { | |
6398 | if (dump_enabled_p ()) | |
6399 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6400 | "cannot perform implicit CSE when performing " | |
6401 | "group loads with negative dependence distance\n"); | |
6402 | return false; | |
6403 | } | |
96bb56b2 RB |
6404 | |
6405 | /* Similarly when the stmt is a load that is both part of a SLP | |
6406 | instance and a loop vectorized stmt via the same-dr mechanism | |
6407 | we have to give up. */ | |
6408 | if (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info) | |
6409 | && (STMT_SLP_TYPE (stmt_info) | |
6410 | != STMT_SLP_TYPE (vinfo_for_stmt | |
6411 | (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info))))) | |
6412 | { | |
6413 | if (dump_enabled_p ()) | |
6414 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6415 | "conflicting SLP types for CSEd load\n"); | |
6416 | return false; | |
6417 | } | |
ebfd146a IR |
6418 | } |
6419 | ||
a1e53f3f | 6420 | |
3bab6342 | 6421 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
aec7ae7d | 6422 | { |
355fe088 | 6423 | gimple *def_stmt; |
3bab6342 AT |
6424 | gather_decl = vect_check_gather_scatter (stmt, loop_vinfo, &gather_base, |
6425 | &gather_off, &gather_scale); | |
aec7ae7d | 6426 | gcc_assert (gather_decl); |
81c40241 RB |
6427 | if (!vect_is_simple_use (gather_off, vinfo, &def_stmt, &gather_dt, |
6428 | &gather_off_vectype)) | |
aec7ae7d | 6429 | { |
73fbfcad | 6430 | if (dump_enabled_p ()) |
78c60e3d | 6431 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6432 | "gather index use not simple.\n"); |
aec7ae7d JJ |
6433 | return false; |
6434 | } | |
6435 | } | |
f2e2a985 | 6436 | else if (STMT_VINFO_STRIDED_P (stmt_info)) |
7b5fc413 RB |
6437 | { |
6438 | if ((grouped_load | |
6439 | && (slp || PURE_SLP_STMT (stmt_info))) | |
6440 | && (group_size > nunits | |
b266b968 | 6441 | || nunits % group_size != 0)) |
7b5fc413 RB |
6442 | { |
6443 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6444 | "unhandled strided group load\n"); | |
6445 | return false; | |
6446 | } | |
6447 | } | |
319e6439 RG |
6448 | else |
6449 | { | |
6450 | negative = tree_int_cst_compare (nested_in_vect_loop | |
6451 | ? STMT_VINFO_DR_STEP (stmt_info) | |
6452 | : DR_STEP (dr), | |
6453 | size_zero_node) < 0; | |
6454 | if (negative && ncopies > 1) | |
6455 | { | |
73fbfcad | 6456 | if (dump_enabled_p ()) |
78c60e3d | 6457 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6458 | "multiple types with negative step.\n"); |
319e6439 RG |
6459 | return false; |
6460 | } | |
6461 | ||
6462 | if (negative) | |
6463 | { | |
08940f33 RB |
6464 | if (grouped_load) |
6465 | { | |
6466 | if (dump_enabled_p ()) | |
6467 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 TJ |
6468 | "negative step for group load not supported" |
6469 | "\n"); | |
08940f33 RB |
6470 | return false; |
6471 | } | |
319e6439 RG |
6472 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
6473 | if (alignment_support_scheme != dr_aligned | |
6474 | && alignment_support_scheme != dr_unaligned_supported) | |
6475 | { | |
73fbfcad | 6476 | if (dump_enabled_p ()) |
78c60e3d | 6477 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6478 | "negative step but alignment required.\n"); |
319e6439 RG |
6479 | return false; |
6480 | } | |
6481 | if (!perm_mask_for_reverse (vectype)) | |
6482 | { | |
73fbfcad | 6483 | if (dump_enabled_p ()) |
78c60e3d | 6484 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
6485 | "negative step and reversing not supported." |
6486 | "\n"); | |
319e6439 RG |
6487 | return false; |
6488 | } | |
6489 | } | |
7d75abc8 | 6490 | } |
aec7ae7d | 6491 | |
ebfd146a IR |
6492 | if (!vec_stmt) /* transformation not required. */ |
6493 | { | |
6494 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
2e8ab70c RB |
6495 | /* The SLP costs are calculated during SLP analysis. */ |
6496 | if (!PURE_SLP_STMT (stmt_info)) | |
6497 | vect_model_load_cost (stmt_info, ncopies, load_lanes_p, | |
6498 | NULL, NULL, NULL); | |
ebfd146a IR |
6499 | return true; |
6500 | } | |
6501 | ||
73fbfcad | 6502 | if (dump_enabled_p ()) |
78c60e3d | 6503 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6504 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a IR |
6505 | |
6506 | /** Transform. **/ | |
6507 | ||
c716e67f XDL |
6508 | ensure_base_align (stmt_info, dr); |
6509 | ||
3bab6342 | 6510 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
aec7ae7d JJ |
6511 | { |
6512 | tree vec_oprnd0 = NULL_TREE, op; | |
6513 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
6514 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
d3c2fee0 | 6515 | tree ptr, mask, var, scale, merge, perm_mask = NULL_TREE, prev_res = NULL_TREE; |
aec7ae7d JJ |
6516 | edge pe = loop_preheader_edge (loop); |
6517 | gimple_seq seq; | |
6518 | basic_block new_bb; | |
6519 | enum { NARROW, NONE, WIDEN } modifier; | |
6520 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
6521 | ||
6522 | if (nunits == gather_off_nunits) | |
6523 | modifier = NONE; | |
6524 | else if (nunits == gather_off_nunits / 2) | |
6525 | { | |
6526 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
6527 | modifier = WIDEN; | |
6528 | ||
6529 | for (i = 0; i < gather_off_nunits; ++i) | |
6530 | sel[i] = i | nunits; | |
6531 | ||
557be5a8 | 6532 | perm_mask = vect_gen_perm_mask_checked (gather_off_vectype, sel); |
aec7ae7d JJ |
6533 | } |
6534 | else if (nunits == gather_off_nunits * 2) | |
6535 | { | |
6536 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
6537 | modifier = NARROW; | |
6538 | ||
6539 | for (i = 0; i < nunits; ++i) | |
6540 | sel[i] = i < gather_off_nunits | |
6541 | ? i : i + nunits - gather_off_nunits; | |
6542 | ||
557be5a8 | 6543 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); |
aec7ae7d JJ |
6544 | ncopies *= 2; |
6545 | } | |
6546 | else | |
6547 | gcc_unreachable (); | |
6548 | ||
6549 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); | |
6550 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6551 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6552 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6553 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6554 | scaletype = TREE_VALUE (arglist); | |
d3c2fee0 | 6555 | gcc_checking_assert (types_compatible_p (srctype, rettype)); |
aec7ae7d JJ |
6556 | |
6557 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
6558 | ||
6559 | ptr = fold_convert (ptrtype, gather_base); | |
6560 | if (!is_gimple_min_invariant (ptr)) | |
6561 | { | |
6562 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
6563 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
6564 | gcc_assert (!new_bb); | |
6565 | } | |
6566 | ||
6567 | /* Currently we support only unconditional gather loads, | |
6568 | so mask should be all ones. */ | |
d3c2fee0 AI |
6569 | if (TREE_CODE (masktype) == INTEGER_TYPE) |
6570 | mask = build_int_cst (masktype, -1); | |
6571 | else if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
6572 | { | |
6573 | mask = build_int_cst (TREE_TYPE (masktype), -1); | |
6574 | mask = build_vector_from_val (masktype, mask); | |
03b9e8e4 | 6575 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
d3c2fee0 | 6576 | } |
aec7ae7d JJ |
6577 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) |
6578 | { | |
6579 | REAL_VALUE_TYPE r; | |
6580 | long tmp[6]; | |
6581 | for (j = 0; j < 6; ++j) | |
6582 | tmp[j] = -1; | |
6583 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
6584 | mask = build_real (TREE_TYPE (masktype), r); | |
d3c2fee0 | 6585 | mask = build_vector_from_val (masktype, mask); |
03b9e8e4 | 6586 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
aec7ae7d JJ |
6587 | } |
6588 | else | |
6589 | gcc_unreachable (); | |
aec7ae7d JJ |
6590 | |
6591 | scale = build_int_cst (scaletype, gather_scale); | |
6592 | ||
d3c2fee0 AI |
6593 | if (TREE_CODE (TREE_TYPE (rettype)) == INTEGER_TYPE) |
6594 | merge = build_int_cst (TREE_TYPE (rettype), 0); | |
6595 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (rettype))) | |
6596 | { | |
6597 | REAL_VALUE_TYPE r; | |
6598 | long tmp[6]; | |
6599 | for (j = 0; j < 6; ++j) | |
6600 | tmp[j] = 0; | |
6601 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (rettype))); | |
6602 | merge = build_real (TREE_TYPE (rettype), r); | |
6603 | } | |
6604 | else | |
6605 | gcc_unreachable (); | |
6606 | merge = build_vector_from_val (rettype, merge); | |
6607 | merge = vect_init_vector (stmt, merge, rettype, NULL); | |
6608 | ||
aec7ae7d JJ |
6609 | prev_stmt_info = NULL; |
6610 | for (j = 0; j < ncopies; ++j) | |
6611 | { | |
6612 | if (modifier == WIDEN && (j & 1)) | |
6613 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
6614 | perm_mask, stmt, gsi); | |
6615 | else if (j == 0) | |
6616 | op = vec_oprnd0 | |
81c40241 | 6617 | = vect_get_vec_def_for_operand (gather_off, stmt); |
aec7ae7d JJ |
6618 | else |
6619 | op = vec_oprnd0 | |
6620 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
6621 | ||
6622 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
6623 | { | |
6624 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
6625 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 6626 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
aec7ae7d JJ |
6627 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
6628 | new_stmt | |
0d0e4a03 | 6629 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
6630 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6631 | op = var; | |
6632 | } | |
6633 | ||
6634 | new_stmt | |
d3c2fee0 | 6635 | = gimple_build_call (gather_decl, 5, merge, ptr, op, mask, scale); |
aec7ae7d JJ |
6636 | |
6637 | if (!useless_type_conversion_p (vectype, rettype)) | |
6638 | { | |
6639 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
6640 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 6641 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
aec7ae7d JJ |
6642 | gimple_call_set_lhs (new_stmt, op); |
6643 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 6644 | var = make_ssa_name (vec_dest); |
aec7ae7d JJ |
6645 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
6646 | new_stmt | |
0d0e4a03 | 6647 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
6648 | } |
6649 | else | |
6650 | { | |
6651 | var = make_ssa_name (vec_dest, new_stmt); | |
6652 | gimple_call_set_lhs (new_stmt, var); | |
6653 | } | |
6654 | ||
6655 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6656 | ||
6657 | if (modifier == NARROW) | |
6658 | { | |
6659 | if ((j & 1) == 0) | |
6660 | { | |
6661 | prev_res = var; | |
6662 | continue; | |
6663 | } | |
6664 | var = permute_vec_elements (prev_res, var, | |
6665 | perm_mask, stmt, gsi); | |
6666 | new_stmt = SSA_NAME_DEF_STMT (var); | |
6667 | } | |
6668 | ||
6669 | if (prev_stmt_info == NULL) | |
6670 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6671 | else | |
6672 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6673 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6674 | } | |
6675 | return true; | |
6676 | } | |
f2e2a985 | 6677 | else if (STMT_VINFO_STRIDED_P (stmt_info)) |
7d75abc8 MM |
6678 | { |
6679 | gimple_stmt_iterator incr_gsi; | |
6680 | bool insert_after; | |
355fe088 | 6681 | gimple *incr; |
7d75abc8 | 6682 | tree offvar; |
7d75abc8 MM |
6683 | tree ivstep; |
6684 | tree running_off; | |
9771b263 | 6685 | vec<constructor_elt, va_gc> *v = NULL; |
7d75abc8 | 6686 | gimple_seq stmts = NULL; |
14ac6aa2 RB |
6687 | tree stride_base, stride_step, alias_off; |
6688 | ||
6689 | gcc_assert (!nested_in_vect_loop); | |
7d75abc8 | 6690 | |
f502d50e | 6691 | if (slp && grouped_load) |
ab313a8c RB |
6692 | first_dr = STMT_VINFO_DATA_REF |
6693 | (vinfo_for_stmt (GROUP_FIRST_ELEMENT (stmt_info))); | |
6694 | else | |
6695 | first_dr = dr; | |
6696 | ||
14ac6aa2 RB |
6697 | stride_base |
6698 | = fold_build_pointer_plus | |
ab313a8c | 6699 | (DR_BASE_ADDRESS (first_dr), |
14ac6aa2 | 6700 | size_binop (PLUS_EXPR, |
ab313a8c RB |
6701 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
6702 | convert_to_ptrofftype (DR_INIT (first_dr)))); | |
6703 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); | |
7d75abc8 MM |
6704 | |
6705 | /* For a load with loop-invariant (but other than power-of-2) | |
6706 | stride (i.e. not a grouped access) like so: | |
6707 | ||
6708 | for (i = 0; i < n; i += stride) | |
6709 | ... = array[i]; | |
6710 | ||
6711 | we generate a new induction variable and new accesses to | |
6712 | form a new vector (or vectors, depending on ncopies): | |
6713 | ||
6714 | for (j = 0; ; j += VF*stride) | |
6715 | tmp1 = array[j]; | |
6716 | tmp2 = array[j + stride]; | |
6717 | ... | |
6718 | vectemp = {tmp1, tmp2, ...} | |
6719 | */ | |
6720 | ||
ab313a8c RB |
6721 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step, |
6722 | build_int_cst (TREE_TYPE (stride_step), vf)); | |
7d75abc8 MM |
6723 | |
6724 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
6725 | ||
ab313a8c | 6726 | create_iv (unshare_expr (stride_base), unshare_expr (ivstep), NULL, |
7d75abc8 MM |
6727 | loop, &incr_gsi, insert_after, |
6728 | &offvar, NULL); | |
6729 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 6730 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
7d75abc8 | 6731 | |
ab313a8c RB |
6732 | stride_step = force_gimple_operand (unshare_expr (stride_step), |
6733 | &stmts, true, NULL_TREE); | |
7d75abc8 MM |
6734 | if (stmts) |
6735 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
6736 | ||
6737 | prev_stmt_info = NULL; | |
6738 | running_off = offvar; | |
ab313a8c | 6739 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (first_dr)), 0); |
7b5fc413 RB |
6740 | int nloads = nunits; |
6741 | tree ltype = TREE_TYPE (vectype); | |
b266b968 | 6742 | auto_vec<tree> dr_chain; |
7b5fc413 RB |
6743 | if (slp) |
6744 | { | |
6745 | nloads = nunits / group_size; | |
6746 | if (group_size < nunits) | |
6747 | ltype = build_vector_type (TREE_TYPE (vectype), group_size); | |
6748 | else | |
6749 | ltype = vectype; | |
6750 | ltype = build_aligned_type (ltype, TYPE_ALIGN (TREE_TYPE (vectype))); | |
6751 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
b266b968 RB |
6752 | if (slp_perm) |
6753 | dr_chain.create (ncopies); | |
7b5fc413 | 6754 | } |
7d75abc8 MM |
6755 | for (j = 0; j < ncopies; j++) |
6756 | { | |
6757 | tree vec_inv; | |
6758 | ||
7b5fc413 RB |
6759 | if (nloads > 1) |
6760 | { | |
6761 | vec_alloc (v, nloads); | |
6762 | for (i = 0; i < nloads; i++) | |
6763 | { | |
6764 | tree newref, newoff; | |
355fe088 | 6765 | gimple *incr; |
7b5fc413 RB |
6766 | newref = build2 (MEM_REF, ltype, running_off, alias_off); |
6767 | ||
6768 | newref = force_gimple_operand_gsi (gsi, newref, true, | |
6769 | NULL_TREE, true, | |
6770 | GSI_SAME_STMT); | |
6771 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, newref); | |
6772 | newoff = copy_ssa_name (running_off); | |
6773 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
6774 | running_off, stride_step); | |
6775 | vect_finish_stmt_generation (stmt, incr, gsi); | |
6776 | ||
6777 | running_off = newoff; | |
6778 | } | |
6779 | ||
6780 | vec_inv = build_constructor (vectype, v); | |
6781 | new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi); | |
6782 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
6783 | } | |
6784 | else | |
7d75abc8 | 6785 | { |
7b5fc413 RB |
6786 | new_stmt = gimple_build_assign (make_ssa_name (ltype), |
6787 | build2 (MEM_REF, ltype, | |
6788 | running_off, alias_off)); | |
6789 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6790 | ||
6791 | tree newoff = copy_ssa_name (running_off); | |
355fe088 | 6792 | gimple *incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
0d0e4a03 | 6793 | running_off, stride_step); |
7d75abc8 MM |
6794 | vect_finish_stmt_generation (stmt, incr, gsi); |
6795 | ||
6796 | running_off = newoff; | |
6797 | } | |
6798 | ||
7b5fc413 | 6799 | if (slp) |
b266b968 RB |
6800 | { |
6801 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
6802 | if (slp_perm) | |
6803 | dr_chain.quick_push (gimple_assign_lhs (new_stmt)); | |
6804 | } | |
7d75abc8 | 6805 | else |
225ce44b RB |
6806 | { |
6807 | if (j == 0) | |
6808 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6809 | else | |
6810 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6811 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6812 | } | |
7d75abc8 | 6813 | } |
b266b968 RB |
6814 | if (slp_perm) |
6815 | vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, | |
6816 | slp_node_instance, false); | |
7d75abc8 MM |
6817 | return true; |
6818 | } | |
aec7ae7d | 6819 | |
0d0293ac | 6820 | if (grouped_load) |
ebfd146a | 6821 | { |
e14c1050 | 6822 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
4f0a0218 | 6823 | /* For SLP vectorization we directly vectorize a subchain |
52eab378 RB |
6824 | without permutation. */ |
6825 | if (slp && ! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
4f0a0218 RB |
6826 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
6827 | /* For BB vectorization always use the first stmt to base | |
6828 | the data ref pointer on. */ | |
6829 | if (bb_vinfo) | |
6830 | first_stmt_for_drptr = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 6831 | |
ebfd146a | 6832 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
6833 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
6834 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
6835 | ??? But we can only do so if there is exactly one | |
6836 | as we have no way to get at the rest. Leave the CSE | |
6837 | opportunity alone. | |
6838 | ??? With the group load eventually participating | |
6839 | in multiple different permutations (having multiple | |
6840 | slp nodes which refer to the same group) the CSE | |
6841 | is even wrong code. See PR56270. */ | |
6842 | && !slp) | |
ebfd146a IR |
6843 | { |
6844 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
6845 | return true; | |
6846 | } | |
6847 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 6848 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
9b999e8c | 6849 | group_gap_adj = 0; |
ebfd146a IR |
6850 | |
6851 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
6852 | if (slp) | |
6853 | { | |
0d0293ac | 6854 | grouped_load = false; |
91ff1504 RB |
6855 | /* For SLP permutation support we need to load the whole group, |
6856 | not only the number of vector stmts the permutation result | |
6857 | fits in. */ | |
6858 | if (slp_perm) | |
6859 | vec_num = (group_size * vf + nunits - 1) / nunits; | |
6860 | else | |
6861 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9b999e8c | 6862 | group_gap_adj = vf * group_size - nunits * vec_num; |
a70d6342 | 6863 | } |
ebfd146a | 6864 | else |
9b999e8c | 6865 | vec_num = group_size; |
ebfd146a IR |
6866 | } |
6867 | else | |
6868 | { | |
6869 | first_stmt = stmt; | |
6870 | first_dr = dr; | |
6871 | group_size = vec_num = 1; | |
9b999e8c | 6872 | group_gap_adj = 0; |
ebfd146a IR |
6873 | } |
6874 | ||
720f5239 | 6875 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 6876 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
6877 | /* Targets with load-lane instructions must not require explicit |
6878 | realignment. */ | |
6879 | gcc_assert (!load_lanes_p | |
6880 | || alignment_support_scheme == dr_aligned | |
6881 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
6882 | |
6883 | /* In case the vectorization factor (VF) is bigger than the number | |
6884 | of elements that we can fit in a vectype (nunits), we have to generate | |
6885 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 6886 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 6887 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 6888 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 6889 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
6890 | stmts that use the defs of the current stmt. The example below |
6891 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
6892 | need to create 4 vectorized stmts): | |
ebfd146a IR |
6893 | |
6894 | before vectorization: | |
6895 | RELATED_STMT VEC_STMT | |
6896 | S1: x = memref - - | |
6897 | S2: z = x + 1 - - | |
6898 | ||
6899 | step 1: vectorize stmt S1: | |
6900 | We first create the vector stmt VS1_0, and, as usual, record a | |
6901 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
6902 | Next, we create the vector stmt VS1_1, and record a pointer to | |
6903 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 6904 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
6905 | stmts and pointers: |
6906 | RELATED_STMT VEC_STMT | |
6907 | VS1_0: vx0 = memref0 VS1_1 - | |
6908 | VS1_1: vx1 = memref1 VS1_2 - | |
6909 | VS1_2: vx2 = memref2 VS1_3 - | |
6910 | VS1_3: vx3 = memref3 - - | |
6911 | S1: x = load - VS1_0 | |
6912 | S2: z = x + 1 - - | |
6913 | ||
b8698a0f L |
6914 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
6915 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
6916 | stmt S2. */ |
6917 | ||
0d0293ac | 6918 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
6919 | |
6920 | S1: x2 = &base + 2 | |
6921 | S2: x0 = &base | |
6922 | S3: x1 = &base + 1 | |
6923 | S4: x3 = &base + 3 | |
6924 | ||
b8698a0f | 6925 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
6926 | starting from the access of the first stmt of the chain: |
6927 | ||
6928 | VS1: vx0 = &base | |
6929 | VS2: vx1 = &base + vec_size*1 | |
6930 | VS3: vx3 = &base + vec_size*2 | |
6931 | VS4: vx4 = &base + vec_size*3 | |
6932 | ||
6933 | Then permutation statements are generated: | |
6934 | ||
e2c83630 RH |
6935 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
6936 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
6937 | ... |
6938 | ||
6939 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
6940 | (the order of the data-refs in the output of vect_permute_load_chain | |
6941 | corresponds to the order of scalar stmts in the interleaving chain - see | |
6942 | the documentation of vect_permute_load_chain()). | |
6943 | The generation of permutation stmts and recording them in | |
0d0293ac | 6944 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 6945 | |
b8698a0f | 6946 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
6947 | permutation stmts above are created for every copy. The result vector |
6948 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
6949 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
6950 | |
6951 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
6952 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
6953 | we generate the following code: | |
6954 | p = initial_addr; | |
6955 | indx = 0; | |
6956 | loop { | |
6957 | p = p + indx * vectype_size; | |
6958 | vec_dest = *(p); | |
6959 | indx = indx + 1; | |
6960 | } | |
6961 | ||
6962 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 6963 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
6964 | then generate the following code, in which the data in each iteration is |
6965 | obtained by two vector loads, one from the previous iteration, and one | |
6966 | from the current iteration: | |
6967 | p1 = initial_addr; | |
6968 | msq_init = *(floor(p1)) | |
6969 | p2 = initial_addr + VS - 1; | |
6970 | realignment_token = call target_builtin; | |
6971 | indx = 0; | |
6972 | loop { | |
6973 | p2 = p2 + indx * vectype_size | |
6974 | lsq = *(floor(p2)) | |
6975 | vec_dest = realign_load (msq, lsq, realignment_token) | |
6976 | indx = indx + 1; | |
6977 | msq = lsq; | |
6978 | } */ | |
6979 | ||
6980 | /* If the misalignment remains the same throughout the execution of the | |
6981 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 6982 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
6983 | This can only occur when vectorizing memory accesses in the inner-loop |
6984 | nested within an outer-loop that is being vectorized. */ | |
6985 | ||
d1e4b493 | 6986 | if (nested_in_vect_loop |
211bea38 | 6987 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
ebfd146a IR |
6988 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) |
6989 | { | |
6990 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
6991 | compute_in_loop = true; | |
6992 | } | |
6993 | ||
6994 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
6995 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 6996 | && !compute_in_loop) |
ebfd146a IR |
6997 | { |
6998 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
6999 | alignment_support_scheme, NULL_TREE, | |
7000 | &at_loop); | |
7001 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7002 | { | |
538dd0b7 | 7003 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (msq)); |
356bbc4c JJ |
7004 | byte_offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
7005 | size_one_node); | |
ebfd146a IR |
7006 | } |
7007 | } | |
7008 | else | |
7009 | at_loop = loop; | |
7010 | ||
a1e53f3f L |
7011 | if (negative) |
7012 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
7013 | ||
272c6793 RS |
7014 | if (load_lanes_p) |
7015 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
7016 | else | |
7017 | aggr_type = vectype; | |
7018 | ||
ebfd146a IR |
7019 | prev_stmt_info = NULL; |
7020 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 7021 | { |
272c6793 | 7022 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 7023 | if (j == 0) |
74bf76ed JJ |
7024 | { |
7025 | bool simd_lane_access_p | |
7026 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
7027 | if (simd_lane_access_p | |
7028 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
7029 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
7030 | && integer_zerop (DR_OFFSET (first_dr)) | |
7031 | && integer_zerop (DR_INIT (first_dr)) | |
7032 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
7033 | get_alias_set (DR_REF (first_dr))) | |
7034 | && (alignment_support_scheme == dr_aligned | |
7035 | || alignment_support_scheme == dr_unaligned_supported)) | |
7036 | { | |
7037 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
7038 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
7039 | (DR_REF (first_dr)), 0); | |
8928eff3 | 7040 | inv_p = false; |
74bf76ed | 7041 | } |
4f0a0218 RB |
7042 | else if (first_stmt_for_drptr |
7043 | && first_stmt != first_stmt_for_drptr) | |
7044 | { | |
7045 | dataref_ptr | |
7046 | = vect_create_data_ref_ptr (first_stmt_for_drptr, aggr_type, | |
7047 | at_loop, offset, &dummy, gsi, | |
7048 | &ptr_incr, simd_lane_access_p, | |
7049 | &inv_p, byte_offset); | |
7050 | /* Adjust the pointer by the difference to first_stmt. */ | |
7051 | data_reference_p ptrdr | |
7052 | = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt_for_drptr)); | |
7053 | tree diff = fold_convert (sizetype, | |
7054 | size_binop (MINUS_EXPR, | |
7055 | DR_INIT (first_dr), | |
7056 | DR_INIT (ptrdr))); | |
7057 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
7058 | stmt, diff); | |
7059 | } | |
74bf76ed JJ |
7060 | else |
7061 | dataref_ptr | |
7062 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
7063 | offset, &dummy, gsi, &ptr_incr, | |
356bbc4c JJ |
7064 | simd_lane_access_p, &inv_p, |
7065 | byte_offset); | |
74bf76ed JJ |
7066 | } |
7067 | else if (dataref_offset) | |
7068 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
7069 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 7070 | else |
272c6793 RS |
7071 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
7072 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 7073 | |
0d0293ac | 7074 | if (grouped_load || slp_perm) |
9771b263 | 7075 | dr_chain.create (vec_num); |
5ce1ee7f | 7076 | |
272c6793 | 7077 | if (load_lanes_p) |
ebfd146a | 7078 | { |
272c6793 RS |
7079 | tree vec_array; |
7080 | ||
7081 | vec_array = create_vector_array (vectype, vec_num); | |
7082 | ||
7083 | /* Emit: | |
7084 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
7085 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
7086 | new_stmt = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
7087 | gimple_call_set_lhs (new_stmt, vec_array); | |
7088 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a | 7089 | |
272c6793 RS |
7090 | /* Extract each vector into an SSA_NAME. */ |
7091 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 7092 | { |
272c6793 RS |
7093 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
7094 | vec_array, i); | |
9771b263 | 7095 | dr_chain.quick_push (new_temp); |
272c6793 RS |
7096 | } |
7097 | ||
7098 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 7099 | vect_record_grouped_load_vectors (stmt, dr_chain); |
272c6793 RS |
7100 | } |
7101 | else | |
7102 | { | |
7103 | for (i = 0; i < vec_num; i++) | |
7104 | { | |
7105 | if (i > 0) | |
7106 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
7107 | stmt, NULL_TREE); | |
7108 | ||
7109 | /* 2. Create the vector-load in the loop. */ | |
7110 | switch (alignment_support_scheme) | |
7111 | { | |
7112 | case dr_aligned: | |
7113 | case dr_unaligned_supported: | |
be1ac4ec | 7114 | { |
644ffefd MJ |
7115 | unsigned int align, misalign; |
7116 | ||
272c6793 | 7117 | data_ref |
aed93b23 RB |
7118 | = fold_build2 (MEM_REF, vectype, dataref_ptr, |
7119 | dataref_offset | |
7120 | ? dataref_offset | |
7121 | : build_int_cst (reference_alias_ptr_type | |
7122 | (DR_REF (first_dr)), 0)); | |
644ffefd | 7123 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 RS |
7124 | if (alignment_support_scheme == dr_aligned) |
7125 | { | |
7126 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 7127 | misalign = 0; |
272c6793 RS |
7128 | } |
7129 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
7130 | { | |
52639a61 RB |
7131 | if (DR_VECT_AUX (first_dr)->base_element_aligned) |
7132 | align = TYPE_ALIGN_UNIT (elem_type); | |
7133 | else | |
7134 | align = (get_object_alignment (DR_REF (first_dr)) | |
7135 | / BITS_PER_UNIT); | |
7136 | misalign = 0; | |
272c6793 RS |
7137 | TREE_TYPE (data_ref) |
7138 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 7139 | align * BITS_PER_UNIT); |
272c6793 RS |
7140 | } |
7141 | else | |
7142 | { | |
7143 | TREE_TYPE (data_ref) | |
7144 | = build_aligned_type (TREE_TYPE (data_ref), | |
7145 | TYPE_ALIGN (elem_type)); | |
644ffefd | 7146 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 7147 | } |
aed93b23 RB |
7148 | if (dataref_offset == NULL_TREE |
7149 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
7150 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), |
7151 | align, misalign); | |
272c6793 | 7152 | break; |
be1ac4ec | 7153 | } |
272c6793 | 7154 | case dr_explicit_realign: |
267d3070 | 7155 | { |
272c6793 | 7156 | tree ptr, bump; |
272c6793 | 7157 | |
d88981fc | 7158 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
272c6793 RS |
7159 | |
7160 | if (compute_in_loop) | |
7161 | msq = vect_setup_realignment (first_stmt, gsi, | |
7162 | &realignment_token, | |
7163 | dr_explicit_realign, | |
7164 | dataref_ptr, NULL); | |
7165 | ||
aed93b23 RB |
7166 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
7167 | ptr = copy_ssa_name (dataref_ptr); | |
7168 | else | |
7169 | ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
0d0e4a03 JJ |
7170 | new_stmt = gimple_build_assign |
7171 | (ptr, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
7172 | build_int_cst |
7173 | (TREE_TYPE (dataref_ptr), | |
7174 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
7175 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7176 | data_ref | |
7177 | = build2 (MEM_REF, vectype, ptr, | |
7178 | build_int_cst (reference_alias_ptr_type | |
7179 | (DR_REF (first_dr)), 0)); | |
7180 | vec_dest = vect_create_destination_var (scalar_dest, | |
7181 | vectype); | |
7182 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
7183 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
7184 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7185 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
7186 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
7187 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7188 | msq = new_temp; | |
7189 | ||
d88981fc | 7190 | bump = size_binop (MULT_EXPR, vs, |
7b7b1813 | 7191 | TYPE_SIZE_UNIT (elem_type)); |
d88981fc | 7192 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
272c6793 | 7193 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
0d0e4a03 JJ |
7194 | new_stmt = gimple_build_assign |
7195 | (NULL_TREE, BIT_AND_EXPR, ptr, | |
272c6793 RS |
7196 | build_int_cst |
7197 | (TREE_TYPE (ptr), | |
7198 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
aed93b23 | 7199 | ptr = copy_ssa_name (ptr, new_stmt); |
272c6793 RS |
7200 | gimple_assign_set_lhs (new_stmt, ptr); |
7201 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7202 | data_ref | |
7203 | = build2 (MEM_REF, vectype, ptr, | |
7204 | build_int_cst (reference_alias_ptr_type | |
7205 | (DR_REF (first_dr)), 0)); | |
7206 | break; | |
267d3070 | 7207 | } |
272c6793 | 7208 | case dr_explicit_realign_optimized: |
aed93b23 RB |
7209 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
7210 | new_temp = copy_ssa_name (dataref_ptr); | |
7211 | else | |
7212 | new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
0d0e4a03 JJ |
7213 | new_stmt = gimple_build_assign |
7214 | (new_temp, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
7215 | build_int_cst |
7216 | (TREE_TYPE (dataref_ptr), | |
7217 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
7218 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7219 | data_ref | |
7220 | = build2 (MEM_REF, vectype, new_temp, | |
7221 | build_int_cst (reference_alias_ptr_type | |
7222 | (DR_REF (first_dr)), 0)); | |
7223 | break; | |
7224 | default: | |
7225 | gcc_unreachable (); | |
7226 | } | |
ebfd146a | 7227 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 7228 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
7229 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7230 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7231 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7232 | ||
272c6793 RS |
7233 | /* 3. Handle explicit realignment if necessary/supported. |
7234 | Create in loop: | |
7235 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
7236 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
7237 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 7238 | { |
272c6793 RS |
7239 | lsq = gimple_assign_lhs (new_stmt); |
7240 | if (!realignment_token) | |
7241 | realignment_token = dataref_ptr; | |
7242 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
0d0e4a03 JJ |
7243 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, |
7244 | msq, lsq, realignment_token); | |
272c6793 RS |
7245 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7246 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7247 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7248 | ||
7249 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7250 | { | |
7251 | gcc_assert (phi); | |
7252 | if (i == vec_num - 1 && j == ncopies - 1) | |
7253 | add_phi_arg (phi, lsq, | |
7254 | loop_latch_edge (containing_loop), | |
9e227d60 | 7255 | UNKNOWN_LOCATION); |
272c6793 RS |
7256 | msq = lsq; |
7257 | } | |
ebfd146a | 7258 | } |
ebfd146a | 7259 | |
59fd17e3 RB |
7260 | /* 4. Handle invariant-load. */ |
7261 | if (inv_p && !bb_vinfo) | |
7262 | { | |
59fd17e3 | 7263 | gcc_assert (!grouped_load); |
d1417442 JJ |
7264 | /* If we have versioned for aliasing or the loop doesn't |
7265 | have any data dependencies that would preclude this, | |
7266 | then we are sure this is a loop invariant load and | |
7267 | thus we can insert it on the preheader edge. */ | |
7268 | if (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) | |
7269 | && !nested_in_vect_loop | |
6b916b36 | 7270 | && hoist_defs_of_uses (stmt, loop)) |
a0e35eb0 RB |
7271 | { |
7272 | if (dump_enabled_p ()) | |
7273 | { | |
7274 | dump_printf_loc (MSG_NOTE, vect_location, | |
7275 | "hoisting out of the vectorized " | |
7276 | "loop: "); | |
7277 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
a0e35eb0 | 7278 | } |
b731b390 | 7279 | tree tem = copy_ssa_name (scalar_dest); |
a0e35eb0 RB |
7280 | gsi_insert_on_edge_immediate |
7281 | (loop_preheader_edge (loop), | |
7282 | gimple_build_assign (tem, | |
7283 | unshare_expr | |
7284 | (gimple_assign_rhs1 (stmt)))); | |
7285 | new_temp = vect_init_vector (stmt, tem, vectype, NULL); | |
34cd48e5 RB |
7286 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
7287 | set_vinfo_for_stmt (new_stmt, | |
7288 | new_stmt_vec_info (new_stmt, vinfo)); | |
a0e35eb0 RB |
7289 | } |
7290 | else | |
7291 | { | |
7292 | gimple_stmt_iterator gsi2 = *gsi; | |
7293 | gsi_next (&gsi2); | |
7294 | new_temp = vect_init_vector (stmt, scalar_dest, | |
7295 | vectype, &gsi2); | |
34cd48e5 | 7296 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
a0e35eb0 | 7297 | } |
59fd17e3 RB |
7298 | } |
7299 | ||
272c6793 RS |
7300 | if (negative) |
7301 | { | |
aec7ae7d JJ |
7302 | tree perm_mask = perm_mask_for_reverse (vectype); |
7303 | new_temp = permute_vec_elements (new_temp, new_temp, | |
7304 | perm_mask, stmt, gsi); | |
ebfd146a IR |
7305 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
7306 | } | |
267d3070 | 7307 | |
272c6793 | 7308 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
7309 | vect_transform_grouped_load (). */ |
7310 | if (grouped_load || slp_perm) | |
9771b263 | 7311 | dr_chain.quick_push (new_temp); |
267d3070 | 7312 | |
272c6793 RS |
7313 | /* Store vector loads in the corresponding SLP_NODE. */ |
7314 | if (slp && !slp_perm) | |
9771b263 | 7315 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
272c6793 | 7316 | } |
9b999e8c RB |
7317 | /* Bump the vector pointer to account for a gap or for excess |
7318 | elements loaded for a permuted SLP load. */ | |
7319 | if (group_gap_adj != 0) | |
a64b9c26 | 7320 | { |
9b999e8c RB |
7321 | bool ovf; |
7322 | tree bump | |
7323 | = wide_int_to_tree (sizetype, | |
7324 | wi::smul (TYPE_SIZE_UNIT (elem_type), | |
7325 | group_gap_adj, &ovf)); | |
a64b9c26 RB |
7326 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
7327 | stmt, bump); | |
7328 | } | |
ebfd146a IR |
7329 | } |
7330 | ||
7331 | if (slp && !slp_perm) | |
7332 | continue; | |
7333 | ||
7334 | if (slp_perm) | |
7335 | { | |
01d8bf07 | 7336 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
ebfd146a IR |
7337 | slp_node_instance, false)) |
7338 | { | |
9771b263 | 7339 | dr_chain.release (); |
ebfd146a IR |
7340 | return false; |
7341 | } | |
7342 | } | |
7343 | else | |
7344 | { | |
0d0293ac | 7345 | if (grouped_load) |
ebfd146a | 7346 | { |
272c6793 | 7347 | if (!load_lanes_p) |
0d0293ac | 7348 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 7349 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
7350 | } |
7351 | else | |
7352 | { | |
7353 | if (j == 0) | |
7354 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7355 | else | |
7356 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7357 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7358 | } | |
7359 | } | |
9771b263 | 7360 | dr_chain.release (); |
ebfd146a IR |
7361 | } |
7362 | ||
ebfd146a IR |
7363 | return true; |
7364 | } | |
7365 | ||
7366 | /* Function vect_is_simple_cond. | |
b8698a0f | 7367 | |
ebfd146a IR |
7368 | Input: |
7369 | LOOP - the loop that is being vectorized. | |
7370 | COND - Condition that is checked for simple use. | |
7371 | ||
e9e1d143 RG |
7372 | Output: |
7373 | *COMP_VECTYPE - the vector type for the comparison. | |
7374 | ||
ebfd146a IR |
7375 | Returns whether a COND can be vectorized. Checks whether |
7376 | condition operands are supportable using vec_is_simple_use. */ | |
7377 | ||
87aab9b2 | 7378 | static bool |
81c40241 | 7379 | vect_is_simple_cond (tree cond, vec_info *vinfo, tree *comp_vectype) |
ebfd146a IR |
7380 | { |
7381 | tree lhs, rhs; | |
ebfd146a | 7382 | enum vect_def_type dt; |
e9e1d143 | 7383 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a | 7384 | |
a414c77f IE |
7385 | /* Mask case. */ |
7386 | if (TREE_CODE (cond) == SSA_NAME | |
7387 | && TREE_CODE (TREE_TYPE (cond)) == BOOLEAN_TYPE) | |
7388 | { | |
7389 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (cond); | |
7390 | if (!vect_is_simple_use (cond, vinfo, &lhs_def_stmt, | |
7391 | &dt, comp_vectype) | |
7392 | || !*comp_vectype | |
7393 | || !VECTOR_BOOLEAN_TYPE_P (*comp_vectype)) | |
7394 | return false; | |
7395 | return true; | |
7396 | } | |
7397 | ||
ebfd146a IR |
7398 | if (!COMPARISON_CLASS_P (cond)) |
7399 | return false; | |
7400 | ||
7401 | lhs = TREE_OPERAND (cond, 0); | |
7402 | rhs = TREE_OPERAND (cond, 1); | |
7403 | ||
7404 | if (TREE_CODE (lhs) == SSA_NAME) | |
7405 | { | |
355fe088 | 7406 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); |
81c40241 | 7407 | if (!vect_is_simple_use (lhs, vinfo, &lhs_def_stmt, &dt, &vectype1)) |
ebfd146a IR |
7408 | return false; |
7409 | } | |
7410 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
7411 | && TREE_CODE (lhs) != FIXED_CST) | |
7412 | return false; | |
7413 | ||
7414 | if (TREE_CODE (rhs) == SSA_NAME) | |
7415 | { | |
355fe088 | 7416 | gimple *rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); |
81c40241 | 7417 | if (!vect_is_simple_use (rhs, vinfo, &rhs_def_stmt, &dt, &vectype2)) |
ebfd146a IR |
7418 | return false; |
7419 | } | |
f7e531cf | 7420 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST |
ebfd146a IR |
7421 | && TREE_CODE (rhs) != FIXED_CST) |
7422 | return false; | |
7423 | ||
e9e1d143 | 7424 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
ebfd146a IR |
7425 | return true; |
7426 | } | |
7427 | ||
7428 | /* vectorizable_condition. | |
7429 | ||
b8698a0f L |
7430 | Check if STMT is conditional modify expression that can be vectorized. |
7431 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
7432 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
7433 | at GSI. |
7434 | ||
7435 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
7436 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
0ad23163 | 7437 | else clause if it is 2). |
ebfd146a IR |
7438 | |
7439 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7440 | ||
4bbe8262 | 7441 | bool |
355fe088 TS |
7442 | vectorizable_condition (gimple *stmt, gimple_stmt_iterator *gsi, |
7443 | gimple **vec_stmt, tree reduc_def, int reduc_index, | |
f7e531cf | 7444 | slp_tree slp_node) |
ebfd146a IR |
7445 | { |
7446 | tree scalar_dest = NULL_TREE; | |
7447 | tree vec_dest = NULL_TREE; | |
ebfd146a IR |
7448 | tree cond_expr, then_clause, else_clause; |
7449 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
df11cc78 | 7450 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
7451 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
7452 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
ebfd146a IR |
7453 | tree vec_compare, vec_cond_expr; |
7454 | tree new_temp; | |
7455 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a855b1b1 | 7456 | enum vect_def_type dt, dts[4]; |
f7e531cf | 7457 | int ncopies; |
ebfd146a | 7458 | enum tree_code code; |
a855b1b1 | 7459 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
7460 | int i, j; |
7461 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
7462 | vec<tree> vec_oprnds0 = vNULL; |
7463 | vec<tree> vec_oprnds1 = vNULL; | |
7464 | vec<tree> vec_oprnds2 = vNULL; | |
7465 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 7466 | tree vec_cmp_type; |
a414c77f | 7467 | bool masked = false; |
b8698a0f | 7468 | |
f7e531cf IR |
7469 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
7470 | return false; | |
7471 | ||
af29617a AH |
7472 | if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) == TREE_CODE_REDUCTION) |
7473 | { | |
7474 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
7475 | return false; | |
ebfd146a | 7476 | |
af29617a AH |
7477 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7478 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
7479 | && reduc_def)) | |
7480 | return false; | |
ebfd146a | 7481 | |
af29617a AH |
7482 | /* FORNOW: not yet supported. */ |
7483 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
7484 | { | |
7485 | if (dump_enabled_p ()) | |
7486 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7487 | "value used after loop.\n"); | |
7488 | return false; | |
7489 | } | |
ebfd146a IR |
7490 | } |
7491 | ||
7492 | /* Is vectorizable conditional operation? */ | |
7493 | if (!is_gimple_assign (stmt)) | |
7494 | return false; | |
7495 | ||
7496 | code = gimple_assign_rhs_code (stmt); | |
7497 | ||
7498 | if (code != COND_EXPR) | |
7499 | return false; | |
7500 | ||
465c8c19 JJ |
7501 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
7502 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
7503 | ||
7504 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
7505 | ncopies = 1; | |
7506 | else | |
7507 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
7508 | ||
7509 | gcc_assert (ncopies >= 1); | |
7510 | if (reduc_index && ncopies > 1) | |
7511 | return false; /* FORNOW */ | |
7512 | ||
4e71066d RG |
7513 | cond_expr = gimple_assign_rhs1 (stmt); |
7514 | then_clause = gimple_assign_rhs2 (stmt); | |
7515 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 7516 | |
81c40241 | 7517 | if (!vect_is_simple_cond (cond_expr, stmt_info->vinfo, &comp_vectype) |
e9e1d143 | 7518 | || !comp_vectype) |
ebfd146a IR |
7519 | return false; |
7520 | ||
81c40241 RB |
7521 | gimple *def_stmt; |
7522 | if (!vect_is_simple_use (then_clause, stmt_info->vinfo, &def_stmt, &dt)) | |
ebfd146a | 7523 | return false; |
81c40241 | 7524 | if (!vect_is_simple_use (else_clause, stmt_info->vinfo, &def_stmt, &dt)) |
ebfd146a IR |
7525 | return false; |
7526 | ||
a414c77f IE |
7527 | if (VECTOR_BOOLEAN_TYPE_P (comp_vectype)) |
7528 | { | |
7529 | vec_cmp_type = comp_vectype; | |
7530 | masked = true; | |
7531 | } | |
7532 | else | |
7533 | vec_cmp_type = build_same_sized_truth_vector_type (comp_vectype); | |
74946978 MP |
7534 | if (vec_cmp_type == NULL_TREE) |
7535 | return false; | |
784fb9b3 | 7536 | |
b8698a0f | 7537 | if (!vec_stmt) |
ebfd146a IR |
7538 | { |
7539 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
e9e1d143 | 7540 | return expand_vec_cond_expr_p (vectype, comp_vectype); |
ebfd146a IR |
7541 | } |
7542 | ||
f7e531cf IR |
7543 | /* Transform. */ |
7544 | ||
7545 | if (!slp_node) | |
7546 | { | |
9771b263 DN |
7547 | vec_oprnds0.create (1); |
7548 | vec_oprnds1.create (1); | |
7549 | vec_oprnds2.create (1); | |
7550 | vec_oprnds3.create (1); | |
f7e531cf | 7551 | } |
ebfd146a IR |
7552 | |
7553 | /* Handle def. */ | |
7554 | scalar_dest = gimple_assign_lhs (stmt); | |
7555 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
7556 | ||
7557 | /* Handle cond expr. */ | |
a855b1b1 MM |
7558 | for (j = 0; j < ncopies; j++) |
7559 | { | |
538dd0b7 | 7560 | gassign *new_stmt = NULL; |
a855b1b1 MM |
7561 | if (j == 0) |
7562 | { | |
f7e531cf IR |
7563 | if (slp_node) |
7564 | { | |
00f96dc9 TS |
7565 | auto_vec<tree, 4> ops; |
7566 | auto_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 7567 | |
a414c77f IE |
7568 | if (masked) |
7569 | ops.safe_push (cond_expr); | |
7570 | else | |
7571 | { | |
7572 | ops.safe_push (TREE_OPERAND (cond_expr, 0)); | |
7573 | ops.safe_push (TREE_OPERAND (cond_expr, 1)); | |
7574 | } | |
9771b263 DN |
7575 | ops.safe_push (then_clause); |
7576 | ops.safe_push (else_clause); | |
f7e531cf | 7577 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); |
37b5ec8f JJ |
7578 | vec_oprnds3 = vec_defs.pop (); |
7579 | vec_oprnds2 = vec_defs.pop (); | |
a414c77f IE |
7580 | if (!masked) |
7581 | vec_oprnds1 = vec_defs.pop (); | |
37b5ec8f | 7582 | vec_oprnds0 = vec_defs.pop (); |
f7e531cf | 7583 | |
9771b263 DN |
7584 | ops.release (); |
7585 | vec_defs.release (); | |
f7e531cf IR |
7586 | } |
7587 | else | |
7588 | { | |
355fe088 | 7589 | gimple *gtemp; |
a414c77f IE |
7590 | if (masked) |
7591 | { | |
7592 | vec_cond_lhs | |
7593 | = vect_get_vec_def_for_operand (cond_expr, stmt, | |
7594 | comp_vectype); | |
7595 | vect_is_simple_use (cond_expr, stmt_info->vinfo, | |
7596 | >emp, &dts[0]); | |
7597 | } | |
7598 | else | |
7599 | { | |
7600 | vec_cond_lhs = | |
7601 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), | |
7602 | stmt, comp_vectype); | |
7603 | vect_is_simple_use (TREE_OPERAND (cond_expr, 0), | |
7604 | loop_vinfo, >emp, &dts[0]); | |
7605 | ||
7606 | vec_cond_rhs = | |
7607 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), | |
7608 | stmt, comp_vectype); | |
7609 | vect_is_simple_use (TREE_OPERAND (cond_expr, 1), | |
7610 | loop_vinfo, >emp, &dts[1]); | |
7611 | } | |
f7e531cf IR |
7612 | if (reduc_index == 1) |
7613 | vec_then_clause = reduc_def; | |
7614 | else | |
7615 | { | |
7616 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
81c40241 RB |
7617 | stmt); |
7618 | vect_is_simple_use (then_clause, loop_vinfo, | |
7619 | >emp, &dts[2]); | |
f7e531cf IR |
7620 | } |
7621 | if (reduc_index == 2) | |
7622 | vec_else_clause = reduc_def; | |
7623 | else | |
7624 | { | |
7625 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
81c40241 RB |
7626 | stmt); |
7627 | vect_is_simple_use (else_clause, loop_vinfo, >emp, &dts[3]); | |
f7e531cf | 7628 | } |
a855b1b1 MM |
7629 | } |
7630 | } | |
7631 | else | |
7632 | { | |
a414c77f IE |
7633 | vec_cond_lhs |
7634 | = vect_get_vec_def_for_stmt_copy (dts[0], | |
7635 | vec_oprnds0.pop ()); | |
7636 | if (!masked) | |
7637 | vec_cond_rhs | |
7638 | = vect_get_vec_def_for_stmt_copy (dts[1], | |
7639 | vec_oprnds1.pop ()); | |
7640 | ||
a855b1b1 | 7641 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 7642 | vec_oprnds2.pop ()); |
a855b1b1 | 7643 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 7644 | vec_oprnds3.pop ()); |
f7e531cf IR |
7645 | } |
7646 | ||
7647 | if (!slp_node) | |
7648 | { | |
9771b263 | 7649 | vec_oprnds0.quick_push (vec_cond_lhs); |
a414c77f IE |
7650 | if (!masked) |
7651 | vec_oprnds1.quick_push (vec_cond_rhs); | |
9771b263 DN |
7652 | vec_oprnds2.quick_push (vec_then_clause); |
7653 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
7654 | } |
7655 | ||
9dc3f7de | 7656 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 7657 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 7658 | { |
9771b263 DN |
7659 | vec_then_clause = vec_oprnds2[i]; |
7660 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 7661 | |
a414c77f IE |
7662 | if (masked) |
7663 | vec_compare = vec_cond_lhs; | |
7664 | else | |
7665 | { | |
7666 | vec_cond_rhs = vec_oprnds1[i]; | |
7667 | vec_compare = build2 (TREE_CODE (cond_expr), vec_cmp_type, | |
7668 | vec_cond_lhs, vec_cond_rhs); | |
7669 | } | |
f7e531cf IR |
7670 | vec_cond_expr = build3 (VEC_COND_EXPR, vectype, |
7671 | vec_compare, vec_then_clause, vec_else_clause); | |
a855b1b1 | 7672 | |
f7e531cf IR |
7673 | new_stmt = gimple_build_assign (vec_dest, vec_cond_expr); |
7674 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
7675 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7676 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7677 | if (slp_node) | |
9771b263 | 7678 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
7679 | } |
7680 | ||
7681 | if (slp_node) | |
7682 | continue; | |
7683 | ||
7684 | if (j == 0) | |
7685 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7686 | else | |
7687 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7688 | ||
7689 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 7690 | } |
b8698a0f | 7691 | |
9771b263 DN |
7692 | vec_oprnds0.release (); |
7693 | vec_oprnds1.release (); | |
7694 | vec_oprnds2.release (); | |
7695 | vec_oprnds3.release (); | |
f7e531cf | 7696 | |
ebfd146a IR |
7697 | return true; |
7698 | } | |
7699 | ||
42fd8198 IE |
7700 | /* vectorizable_comparison. |
7701 | ||
7702 | Check if STMT is comparison expression that can be vectorized. | |
7703 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
7704 | comparison, put it in VEC_STMT, and insert it at GSI. | |
7705 | ||
7706 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7707 | ||
7708 | bool | |
7709 | vectorizable_comparison (gimple *stmt, gimple_stmt_iterator *gsi, | |
7710 | gimple **vec_stmt, tree reduc_def, | |
7711 | slp_tree slp_node) | |
7712 | { | |
7713 | tree lhs, rhs1, rhs2; | |
7714 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
7715 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; | |
7716 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
7717 | tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE; | |
7718 | tree new_temp; | |
7719 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
7720 | enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
7721 | unsigned nunits; | |
7722 | int ncopies; | |
7723 | enum tree_code code; | |
7724 | stmt_vec_info prev_stmt_info = NULL; | |
7725 | int i, j; | |
7726 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
7727 | vec<tree> vec_oprnds0 = vNULL; | |
7728 | vec<tree> vec_oprnds1 = vNULL; | |
7729 | gimple *def_stmt; | |
7730 | tree mask_type; | |
7731 | tree mask; | |
7732 | ||
c245362b IE |
7733 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
7734 | return false; | |
7735 | ||
42fd8198 IE |
7736 | if (!VECTOR_BOOLEAN_TYPE_P (vectype)) |
7737 | return false; | |
7738 | ||
7739 | mask_type = vectype; | |
7740 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
7741 | ||
7742 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
7743 | ncopies = 1; | |
7744 | else | |
7745 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
7746 | ||
7747 | gcc_assert (ncopies >= 1); | |
42fd8198 IE |
7748 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7749 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
7750 | && reduc_def)) | |
7751 | return false; | |
7752 | ||
7753 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
7754 | { | |
7755 | if (dump_enabled_p ()) | |
7756 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7757 | "value used after loop.\n"); | |
7758 | return false; | |
7759 | } | |
7760 | ||
7761 | if (!is_gimple_assign (stmt)) | |
7762 | return false; | |
7763 | ||
7764 | code = gimple_assign_rhs_code (stmt); | |
7765 | ||
7766 | if (TREE_CODE_CLASS (code) != tcc_comparison) | |
7767 | return false; | |
7768 | ||
7769 | rhs1 = gimple_assign_rhs1 (stmt); | |
7770 | rhs2 = gimple_assign_rhs2 (stmt); | |
7771 | ||
7772 | if (!vect_is_simple_use (rhs1, stmt_info->vinfo, &def_stmt, | |
7773 | &dts[0], &vectype1)) | |
7774 | return false; | |
7775 | ||
7776 | if (!vect_is_simple_use (rhs2, stmt_info->vinfo, &def_stmt, | |
7777 | &dts[1], &vectype2)) | |
7778 | return false; | |
7779 | ||
7780 | if (vectype1 && vectype2 | |
7781 | && TYPE_VECTOR_SUBPARTS (vectype1) != TYPE_VECTOR_SUBPARTS (vectype2)) | |
7782 | return false; | |
7783 | ||
7784 | vectype = vectype1 ? vectype1 : vectype2; | |
7785 | ||
7786 | /* Invariant comparison. */ | |
7787 | if (!vectype) | |
7788 | { | |
7789 | vectype = build_vector_type (TREE_TYPE (rhs1), nunits); | |
7790 | if (tree_to_shwi (TYPE_SIZE_UNIT (vectype)) != current_vector_size) | |
7791 | return false; | |
7792 | } | |
7793 | else if (nunits != TYPE_VECTOR_SUBPARTS (vectype)) | |
7794 | return false; | |
7795 | ||
7796 | if (!vec_stmt) | |
7797 | { | |
7798 | STMT_VINFO_TYPE (stmt_info) = comparison_vec_info_type; | |
7799 | vect_model_simple_cost (stmt_info, ncopies, dts, NULL, NULL); | |
7800 | return expand_vec_cmp_expr_p (vectype, mask_type); | |
7801 | } | |
7802 | ||
7803 | /* Transform. */ | |
7804 | if (!slp_node) | |
7805 | { | |
7806 | vec_oprnds0.create (1); | |
7807 | vec_oprnds1.create (1); | |
7808 | } | |
7809 | ||
7810 | /* Handle def. */ | |
7811 | lhs = gimple_assign_lhs (stmt); | |
7812 | mask = vect_create_destination_var (lhs, mask_type); | |
7813 | ||
7814 | /* Handle cmp expr. */ | |
7815 | for (j = 0; j < ncopies; j++) | |
7816 | { | |
7817 | gassign *new_stmt = NULL; | |
7818 | if (j == 0) | |
7819 | { | |
7820 | if (slp_node) | |
7821 | { | |
7822 | auto_vec<tree, 2> ops; | |
7823 | auto_vec<vec<tree>, 2> vec_defs; | |
7824 | ||
7825 | ops.safe_push (rhs1); | |
7826 | ops.safe_push (rhs2); | |
7827 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); | |
7828 | vec_oprnds1 = vec_defs.pop (); | |
7829 | vec_oprnds0 = vec_defs.pop (); | |
7830 | } | |
7831 | else | |
7832 | { | |
e4af0bc4 IE |
7833 | vec_rhs1 = vect_get_vec_def_for_operand (rhs1, stmt, vectype); |
7834 | vec_rhs2 = vect_get_vec_def_for_operand (rhs2, stmt, vectype); | |
42fd8198 IE |
7835 | } |
7836 | } | |
7837 | else | |
7838 | { | |
7839 | vec_rhs1 = vect_get_vec_def_for_stmt_copy (dts[0], | |
7840 | vec_oprnds0.pop ()); | |
7841 | vec_rhs2 = vect_get_vec_def_for_stmt_copy (dts[1], | |
7842 | vec_oprnds1.pop ()); | |
7843 | } | |
7844 | ||
7845 | if (!slp_node) | |
7846 | { | |
7847 | vec_oprnds0.quick_push (vec_rhs1); | |
7848 | vec_oprnds1.quick_push (vec_rhs2); | |
7849 | } | |
7850 | ||
7851 | /* Arguments are ready. Create the new vector stmt. */ | |
7852 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1) | |
7853 | { | |
7854 | vec_rhs2 = vec_oprnds1[i]; | |
7855 | ||
7856 | new_temp = make_ssa_name (mask); | |
7857 | new_stmt = gimple_build_assign (new_temp, code, vec_rhs1, vec_rhs2); | |
7858 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7859 | if (slp_node) | |
7860 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
7861 | } | |
7862 | ||
7863 | if (slp_node) | |
7864 | continue; | |
7865 | ||
7866 | if (j == 0) | |
7867 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7868 | else | |
7869 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7870 | ||
7871 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7872 | } | |
7873 | ||
7874 | vec_oprnds0.release (); | |
7875 | vec_oprnds1.release (); | |
7876 | ||
7877 | return true; | |
7878 | } | |
ebfd146a | 7879 | |
8644a673 | 7880 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
7881 | |
7882 | bool | |
355fe088 | 7883 | vect_analyze_stmt (gimple *stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 7884 | { |
8644a673 | 7885 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 7886 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 7887 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 7888 | bool ok; |
a70d6342 | 7889 | tree scalar_type, vectype; |
355fe088 | 7890 | gimple *pattern_stmt; |
363477c0 | 7891 | gimple_seq pattern_def_seq; |
ebfd146a | 7892 | |
73fbfcad | 7893 | if (dump_enabled_p ()) |
ebfd146a | 7894 | { |
78c60e3d SS |
7895 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
7896 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
8644a673 | 7897 | } |
ebfd146a | 7898 | |
1825a1f3 | 7899 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 7900 | { |
73fbfcad | 7901 | if (dump_enabled_p ()) |
78c60e3d | 7902 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7903 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
7904 | |
7905 | return false; | |
7906 | } | |
b8698a0f L |
7907 | |
7908 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
7909 | to include: |
7910 | - the COND_EXPR which is the loop exit condition | |
7911 | - any LABEL_EXPRs in the loop | |
b8698a0f | 7912 | - computations that are used only for array indexing or loop control. |
8644a673 | 7913 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 7914 | instance, therefore, all the statements are relevant. |
ebfd146a | 7915 | |
d092494c | 7916 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 7917 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
7918 | statements. In basic blocks we are called from some SLP instance |
7919 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
7920 | already will be part of SLP instance. */ | |
83197f37 IR |
7921 | |
7922 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 7923 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 7924 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 7925 | { |
9d5e7640 | 7926 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 7927 | && pattern_stmt |
9d5e7640 IR |
7928 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
7929 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
7930 | { | |
83197f37 | 7931 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
7932 | stmt = pattern_stmt; |
7933 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 7934 | if (dump_enabled_p ()) |
9d5e7640 | 7935 | { |
78c60e3d SS |
7936 | dump_printf_loc (MSG_NOTE, vect_location, |
7937 | "==> examining pattern statement: "); | |
7938 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
9d5e7640 IR |
7939 | } |
7940 | } | |
7941 | else | |
7942 | { | |
73fbfcad | 7943 | if (dump_enabled_p ()) |
e645e942 | 7944 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 7945 | |
9d5e7640 IR |
7946 | return true; |
7947 | } | |
8644a673 | 7948 | } |
83197f37 | 7949 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 7950 | && node == NULL |
83197f37 IR |
7951 | && pattern_stmt |
7952 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
7953 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
7954 | { | |
7955 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 7956 | if (dump_enabled_p ()) |
83197f37 | 7957 | { |
78c60e3d SS |
7958 | dump_printf_loc (MSG_NOTE, vect_location, |
7959 | "==> examining pattern statement: "); | |
7960 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
83197f37 IR |
7961 | } |
7962 | ||
7963 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node)) | |
7964 | return false; | |
7965 | } | |
ebfd146a | 7966 | |
1107f3ae | 7967 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 7968 | && node == NULL |
363477c0 | 7969 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 7970 | { |
363477c0 | 7971 | gimple_stmt_iterator si; |
1107f3ae | 7972 | |
363477c0 JJ |
7973 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
7974 | { | |
355fe088 | 7975 | gimple *pattern_def_stmt = gsi_stmt (si); |
363477c0 JJ |
7976 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) |
7977 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
7978 | { | |
7979 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 7980 | if (dump_enabled_p ()) |
363477c0 | 7981 | { |
78c60e3d SS |
7982 | dump_printf_loc (MSG_NOTE, vect_location, |
7983 | "==> examining pattern def statement: "); | |
7984 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
363477c0 | 7985 | } |
1107f3ae | 7986 | |
363477c0 JJ |
7987 | if (!vect_analyze_stmt (pattern_def_stmt, |
7988 | need_to_vectorize, node)) | |
7989 | return false; | |
7990 | } | |
7991 | } | |
7992 | } | |
1107f3ae | 7993 | |
8644a673 IR |
7994 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
7995 | { | |
7996 | case vect_internal_def: | |
7997 | break; | |
ebfd146a | 7998 | |
8644a673 | 7999 | case vect_reduction_def: |
7c5222ff | 8000 | case vect_nested_cycle: |
14a61437 RB |
8001 | gcc_assert (!bb_vinfo |
8002 | && (relevance == vect_used_in_outer | |
8003 | || relevance == vect_used_in_outer_by_reduction | |
8004 | || relevance == vect_used_by_reduction | |
8005 | || relevance == vect_unused_in_scope)); | |
8644a673 IR |
8006 | break; |
8007 | ||
8008 | case vect_induction_def: | |
8009 | case vect_constant_def: | |
8010 | case vect_external_def: | |
8011 | case vect_unknown_def_type: | |
8012 | default: | |
8013 | gcc_unreachable (); | |
8014 | } | |
ebfd146a | 8015 | |
a70d6342 IR |
8016 | if (bb_vinfo) |
8017 | { | |
8018 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
8019 | ||
b690cc0f | 8020 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
73fbfcad | 8021 | if (dump_enabled_p ()) |
a70d6342 | 8022 | { |
78c60e3d SS |
8023 | dump_printf_loc (MSG_NOTE, vect_location, |
8024 | "get vectype for scalar type: "); | |
8025 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
e645e942 | 8026 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
8027 | } |
8028 | ||
8029 | vectype = get_vectype_for_scalar_type (scalar_type); | |
8030 | if (!vectype) | |
8031 | { | |
73fbfcad | 8032 | if (dump_enabled_p ()) |
a70d6342 | 8033 | { |
78c60e3d SS |
8034 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8035 | "not SLPed: unsupported data-type "); | |
8036 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
8037 | scalar_type); | |
e645e942 | 8038 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
a70d6342 IR |
8039 | } |
8040 | return false; | |
8041 | } | |
8042 | ||
73fbfcad | 8043 | if (dump_enabled_p ()) |
a70d6342 | 8044 | { |
78c60e3d SS |
8045 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); |
8046 | dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); | |
e645e942 | 8047 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
8048 | } |
8049 | ||
8050 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
8051 | } | |
8052 | ||
8644a673 | 8053 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 8054 | { |
8644a673 | 8055 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
0136f8f0 AH |
8056 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
8057 | || (is_gimple_call (stmt) | |
8058 | && gimple_call_lhs (stmt) == NULL_TREE)); | |
8644a673 | 8059 | *need_to_vectorize = true; |
ebfd146a IR |
8060 | } |
8061 | ||
b1af7da6 RB |
8062 | if (PURE_SLP_STMT (stmt_info) && !node) |
8063 | { | |
8064 | dump_printf_loc (MSG_NOTE, vect_location, | |
8065 | "handled only by SLP analysis\n"); | |
8066 | return true; | |
8067 | } | |
8068 | ||
8069 | ok = true; | |
8070 | if (!bb_vinfo | |
8071 | && (STMT_VINFO_RELEVANT_P (stmt_info) | |
8072 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
8073 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
8074 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
8075 | || vectorizable_shift (stmt, NULL, NULL, node) | |
8076 | || vectorizable_operation (stmt, NULL, NULL, node) | |
8077 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
8078 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
8079 | || vectorizable_call (stmt, NULL, NULL, node) | |
8080 | || vectorizable_store (stmt, NULL, NULL, node) | |
8081 | || vectorizable_reduction (stmt, NULL, NULL, node) | |
42fd8198 IE |
8082 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
8083 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 RB |
8084 | else |
8085 | { | |
8086 | if (bb_vinfo) | |
8087 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
8088 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
8089 | || vectorizable_shift (stmt, NULL, NULL, node) | |
8090 | || vectorizable_operation (stmt, NULL, NULL, node) | |
8091 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
8092 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
8093 | || vectorizable_call (stmt, NULL, NULL, node) | |
8094 | || vectorizable_store (stmt, NULL, NULL, node) | |
42fd8198 IE |
8095 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
8096 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 | 8097 | } |
8644a673 IR |
8098 | |
8099 | if (!ok) | |
ebfd146a | 8100 | { |
73fbfcad | 8101 | if (dump_enabled_p ()) |
8644a673 | 8102 | { |
78c60e3d SS |
8103 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8104 | "not vectorized: relevant stmt not "); | |
8105 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
8106 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 8107 | } |
b8698a0f | 8108 | |
ebfd146a IR |
8109 | return false; |
8110 | } | |
8111 | ||
a70d6342 IR |
8112 | if (bb_vinfo) |
8113 | return true; | |
8114 | ||
8644a673 IR |
8115 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
8116 | need extra handling, except for vectorizable reductions. */ | |
8117 | if (STMT_VINFO_LIVE_P (stmt_info) | |
8118 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
8119 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 8120 | |
8644a673 | 8121 | if (!ok) |
ebfd146a | 8122 | { |
73fbfcad | 8123 | if (dump_enabled_p ()) |
8644a673 | 8124 | { |
78c60e3d SS |
8125 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8126 | "not vectorized: live stmt not "); | |
8127 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
8128 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 8129 | } |
b8698a0f | 8130 | |
8644a673 | 8131 | return false; |
ebfd146a IR |
8132 | } |
8133 | ||
ebfd146a IR |
8134 | return true; |
8135 | } | |
8136 | ||
8137 | ||
8138 | /* Function vect_transform_stmt. | |
8139 | ||
8140 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
8141 | ||
8142 | bool | |
355fe088 | 8143 | vect_transform_stmt (gimple *stmt, gimple_stmt_iterator *gsi, |
0d0293ac | 8144 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
8145 | slp_instance slp_node_instance) |
8146 | { | |
8147 | bool is_store = false; | |
355fe088 | 8148 | gimple *vec_stmt = NULL; |
ebfd146a | 8149 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
ebfd146a | 8150 | bool done; |
ebfd146a | 8151 | |
355fe088 | 8152 | gimple *old_vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
225ce44b | 8153 | |
ebfd146a IR |
8154 | switch (STMT_VINFO_TYPE (stmt_info)) |
8155 | { | |
8156 | case type_demotion_vec_info_type: | |
ebfd146a | 8157 | case type_promotion_vec_info_type: |
ebfd146a IR |
8158 | case type_conversion_vec_info_type: |
8159 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
8160 | gcc_assert (done); | |
8161 | break; | |
8162 | ||
8163 | case induc_vec_info_type: | |
8164 | gcc_assert (!slp_node); | |
8165 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
8166 | gcc_assert (done); | |
8167 | break; | |
8168 | ||
9dc3f7de IR |
8169 | case shift_vec_info_type: |
8170 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
8171 | gcc_assert (done); | |
8172 | break; | |
8173 | ||
ebfd146a IR |
8174 | case op_vec_info_type: |
8175 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
8176 | gcc_assert (done); | |
8177 | break; | |
8178 | ||
8179 | case assignment_vec_info_type: | |
8180 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
8181 | gcc_assert (done); | |
8182 | break; | |
8183 | ||
8184 | case load_vec_info_type: | |
b8698a0f | 8185 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
8186 | slp_node_instance); |
8187 | gcc_assert (done); | |
8188 | break; | |
8189 | ||
8190 | case store_vec_info_type: | |
8191 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
8192 | gcc_assert (done); | |
0d0293ac | 8193 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
8194 | { |
8195 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 8196 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
8197 | one are skipped, and there vec_stmt_info shouldn't be freed |
8198 | meanwhile. */ | |
0d0293ac | 8199 | *grouped_store = true; |
ebfd146a IR |
8200 | if (STMT_VINFO_VEC_STMT (stmt_info)) |
8201 | is_store = true; | |
8202 | } | |
8203 | else | |
8204 | is_store = true; | |
8205 | break; | |
8206 | ||
8207 | case condition_vec_info_type: | |
f7e531cf | 8208 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
8209 | gcc_assert (done); |
8210 | break; | |
8211 | ||
42fd8198 IE |
8212 | case comparison_vec_info_type: |
8213 | done = vectorizable_comparison (stmt, gsi, &vec_stmt, NULL, slp_node); | |
8214 | gcc_assert (done); | |
8215 | break; | |
8216 | ||
ebfd146a | 8217 | case call_vec_info_type: |
190c2236 | 8218 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 8219 | stmt = gsi_stmt (*gsi); |
5ce9450f JJ |
8220 | if (is_gimple_call (stmt) |
8221 | && gimple_call_internal_p (stmt) | |
8222 | && gimple_call_internal_fn (stmt) == IFN_MASK_STORE) | |
8223 | is_store = true; | |
ebfd146a IR |
8224 | break; |
8225 | ||
0136f8f0 AH |
8226 | case call_simd_clone_vec_info_type: |
8227 | done = vectorizable_simd_clone_call (stmt, gsi, &vec_stmt, slp_node); | |
8228 | stmt = gsi_stmt (*gsi); | |
8229 | break; | |
8230 | ||
ebfd146a | 8231 | case reduc_vec_info_type: |
b5aeb3bb | 8232 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
8233 | gcc_assert (done); |
8234 | break; | |
8235 | ||
8236 | default: | |
8237 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
8238 | { | |
73fbfcad | 8239 | if (dump_enabled_p ()) |
78c60e3d | 8240 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8241 | "stmt not supported.\n"); |
ebfd146a IR |
8242 | gcc_unreachable (); |
8243 | } | |
8244 | } | |
8245 | ||
225ce44b RB |
8246 | /* Verify SLP vectorization doesn't mess with STMT_VINFO_VEC_STMT. |
8247 | This would break hybrid SLP vectorization. */ | |
8248 | if (slp_node) | |
d90f8440 RB |
8249 | gcc_assert (!vec_stmt |
8250 | && STMT_VINFO_VEC_STMT (stmt_info) == old_vec_stmt); | |
225ce44b | 8251 | |
ebfd146a IR |
8252 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that |
8253 | is being vectorized, but outside the immediately enclosing loop. */ | |
8254 | if (vec_stmt | |
a70d6342 IR |
8255 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
8256 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
8257 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
8258 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
8259 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 8260 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 8261 | vect_used_in_outer_by_reduction)) |
ebfd146a | 8262 | { |
a70d6342 IR |
8263 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
8264 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
8265 | imm_use_iterator imm_iter; |
8266 | use_operand_p use_p; | |
8267 | tree scalar_dest; | |
355fe088 | 8268 | gimple *exit_phi; |
ebfd146a | 8269 | |
73fbfcad | 8270 | if (dump_enabled_p ()) |
78c60e3d | 8271 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 8272 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
8273 | |
8274 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
8275 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
8276 | STMT). */ | |
8277 | if (gimple_code (stmt) == GIMPLE_PHI) | |
8278 | scalar_dest = PHI_RESULT (stmt); | |
8279 | else | |
8280 | scalar_dest = gimple_assign_lhs (stmt); | |
8281 | ||
8282 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
8283 | { | |
8284 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
8285 | { | |
8286 | exit_phi = USE_STMT (use_p); | |
8287 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
8288 | } | |
8289 | } | |
8290 | } | |
8291 | ||
8292 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
8293 | being vectorized. */ | |
8294 | if (STMT_VINFO_LIVE_P (stmt_info) | |
8295 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
8296 | { | |
8297 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
8298 | gcc_assert (done); | |
8299 | } | |
8300 | ||
8301 | if (vec_stmt) | |
83197f37 | 8302 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 8303 | |
b8698a0f | 8304 | return is_store; |
ebfd146a IR |
8305 | } |
8306 | ||
8307 | ||
b8698a0f | 8308 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
8309 | stmt_vec_info. */ |
8310 | ||
8311 | void | |
355fe088 | 8312 | vect_remove_stores (gimple *first_stmt) |
ebfd146a | 8313 | { |
355fe088 TS |
8314 | gimple *next = first_stmt; |
8315 | gimple *tmp; | |
ebfd146a IR |
8316 | gimple_stmt_iterator next_si; |
8317 | ||
8318 | while (next) | |
8319 | { | |
78048b1c JJ |
8320 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
8321 | ||
8322 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
8323 | if (is_pattern_stmt_p (stmt_info)) | |
8324 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
8325 | /* Free the attached stmt_vec_info and remove the stmt. */ |
8326 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 8327 | unlink_stmt_vdef (next); |
ebfd146a | 8328 | gsi_remove (&next_si, true); |
3d3f2249 | 8329 | release_defs (next); |
ebfd146a IR |
8330 | free_stmt_vec_info (next); |
8331 | next = tmp; | |
8332 | } | |
8333 | } | |
8334 | ||
8335 | ||
8336 | /* Function new_stmt_vec_info. | |
8337 | ||
8338 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
8339 | ||
8340 | stmt_vec_info | |
310213d4 | 8341 | new_stmt_vec_info (gimple *stmt, vec_info *vinfo) |
ebfd146a IR |
8342 | { |
8343 | stmt_vec_info res; | |
8344 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
8345 | ||
8346 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
8347 | STMT_VINFO_STMT (res) = stmt; | |
310213d4 | 8348 | res->vinfo = vinfo; |
8644a673 | 8349 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
8350 | STMT_VINFO_LIVE_P (res) = false; |
8351 | STMT_VINFO_VECTYPE (res) = NULL; | |
8352 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 8353 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
8354 | STMT_VINFO_IN_PATTERN_P (res) = false; |
8355 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 8356 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a | 8357 | STMT_VINFO_DATA_REF (res) = NULL; |
af29617a | 8358 | STMT_VINFO_VEC_REDUCTION_TYPE (res) = TREE_CODE_REDUCTION; |
ebfd146a IR |
8359 | |
8360 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
8361 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
8362 | STMT_VINFO_DR_INIT (res) = NULL; | |
8363 | STMT_VINFO_DR_STEP (res) = NULL; | |
8364 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
8365 | ||
8366 | if (gimple_code (stmt) == GIMPLE_PHI | |
8367 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
8368 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
8369 | else | |
8644a673 IR |
8370 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
8371 | ||
9771b263 | 8372 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 8373 | STMT_SLP_TYPE (res) = loop_vect; |
e14c1050 IR |
8374 | GROUP_FIRST_ELEMENT (res) = NULL; |
8375 | GROUP_NEXT_ELEMENT (res) = NULL; | |
8376 | GROUP_SIZE (res) = 0; | |
8377 | GROUP_STORE_COUNT (res) = 0; | |
8378 | GROUP_GAP (res) = 0; | |
8379 | GROUP_SAME_DR_STMT (res) = NULL; | |
ebfd146a IR |
8380 | |
8381 | return res; | |
8382 | } | |
8383 | ||
8384 | ||
8385 | /* Create a hash table for stmt_vec_info. */ | |
8386 | ||
8387 | void | |
8388 | init_stmt_vec_info_vec (void) | |
8389 | { | |
9771b263 DN |
8390 | gcc_assert (!stmt_vec_info_vec.exists ()); |
8391 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
8392 | } |
8393 | ||
8394 | ||
8395 | /* Free hash table for stmt_vec_info. */ | |
8396 | ||
8397 | void | |
8398 | free_stmt_vec_info_vec (void) | |
8399 | { | |
93675444 | 8400 | unsigned int i; |
3161455c | 8401 | stmt_vec_info info; |
93675444 JJ |
8402 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) |
8403 | if (info != NULL) | |
3161455c | 8404 | free_stmt_vec_info (STMT_VINFO_STMT (info)); |
9771b263 DN |
8405 | gcc_assert (stmt_vec_info_vec.exists ()); |
8406 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
8407 | } |
8408 | ||
8409 | ||
8410 | /* Free stmt vectorization related info. */ | |
8411 | ||
8412 | void | |
355fe088 | 8413 | free_stmt_vec_info (gimple *stmt) |
ebfd146a IR |
8414 | { |
8415 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8416 | ||
8417 | if (!stmt_info) | |
8418 | return; | |
8419 | ||
78048b1c JJ |
8420 | /* Check if this statement has a related "pattern stmt" |
8421 | (introduced by the vectorizer during the pattern recognition | |
8422 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
8423 | too. */ | |
8424 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
8425 | { | |
8426 | stmt_vec_info patt_info | |
8427 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
8428 | if (patt_info) | |
8429 | { | |
363477c0 | 8430 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
355fe088 | 8431 | gimple *patt_stmt = STMT_VINFO_STMT (patt_info); |
f0281fde RB |
8432 | gimple_set_bb (patt_stmt, NULL); |
8433 | tree lhs = gimple_get_lhs (patt_stmt); | |
e6f5c25d | 8434 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde | 8435 | release_ssa_name (lhs); |
363477c0 JJ |
8436 | if (seq) |
8437 | { | |
8438 | gimple_stmt_iterator si; | |
8439 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
f0281fde | 8440 | { |
355fe088 | 8441 | gimple *seq_stmt = gsi_stmt (si); |
f0281fde | 8442 | gimple_set_bb (seq_stmt, NULL); |
7532abf2 | 8443 | lhs = gimple_get_lhs (seq_stmt); |
e6f5c25d | 8444 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde RB |
8445 | release_ssa_name (lhs); |
8446 | free_stmt_vec_info (seq_stmt); | |
8447 | } | |
363477c0 | 8448 | } |
f0281fde | 8449 | free_stmt_vec_info (patt_stmt); |
78048b1c JJ |
8450 | } |
8451 | } | |
8452 | ||
9771b263 | 8453 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
6c9e85fb | 8454 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).release (); |
ebfd146a IR |
8455 | set_vinfo_for_stmt (stmt, NULL); |
8456 | free (stmt_info); | |
8457 | } | |
8458 | ||
8459 | ||
bb67d9c7 | 8460 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 8461 | |
bb67d9c7 | 8462 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
8463 | by the target. */ |
8464 | ||
bb67d9c7 RG |
8465 | static tree |
8466 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a | 8467 | { |
ef4bddc2 RS |
8468 | machine_mode inner_mode = TYPE_MODE (scalar_type); |
8469 | machine_mode simd_mode; | |
2f816591 | 8470 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
8471 | int nunits; |
8472 | tree vectype; | |
8473 | ||
cc4b5170 | 8474 | if (nbytes == 0) |
ebfd146a IR |
8475 | return NULL_TREE; |
8476 | ||
48f2e373 RB |
8477 | if (GET_MODE_CLASS (inner_mode) != MODE_INT |
8478 | && GET_MODE_CLASS (inner_mode) != MODE_FLOAT) | |
8479 | return NULL_TREE; | |
8480 | ||
7b7b1813 RG |
8481 | /* For vector types of elements whose mode precision doesn't |
8482 | match their types precision we use a element type of mode | |
8483 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
8484 | they support the proper result truncation/extension. |
8485 | We also make sure to build vector types with INTEGER_TYPE | |
8486 | component type only. */ | |
6d7971b8 | 8487 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
8488 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
8489 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
8490 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
8491 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 8492 | |
ccbf5bb4 RG |
8493 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
8494 | When the component mode passes the above test simply use a type | |
8495 | corresponding to that mode. The theory is that any use that | |
8496 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 8497 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 8498 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
8499 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
8500 | ||
8501 | /* We can't build a vector type of elements with alignment bigger than | |
8502 | their size. */ | |
dfc2e2ac | 8503 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
8504 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
8505 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 8506 | |
dfc2e2ac RB |
8507 | /* If we felt back to using the mode fail if there was |
8508 | no scalar type for it. */ | |
8509 | if (scalar_type == NULL_TREE) | |
8510 | return NULL_TREE; | |
8511 | ||
bb67d9c7 RG |
8512 | /* If no size was supplied use the mode the target prefers. Otherwise |
8513 | lookup a vector mode of the specified size. */ | |
8514 | if (size == 0) | |
8515 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
8516 | else | |
8517 | simd_mode = mode_for_vector (inner_mode, size / nbytes); | |
cc4b5170 RG |
8518 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
8519 | if (nunits <= 1) | |
8520 | return NULL_TREE; | |
ebfd146a IR |
8521 | |
8522 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
8523 | |
8524 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
8525 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 8526 | return NULL_TREE; |
ebfd146a IR |
8527 | |
8528 | return vectype; | |
8529 | } | |
8530 | ||
bb67d9c7 RG |
8531 | unsigned int current_vector_size; |
8532 | ||
8533 | /* Function get_vectype_for_scalar_type. | |
8534 | ||
8535 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
8536 | by the target. */ | |
8537 | ||
8538 | tree | |
8539 | get_vectype_for_scalar_type (tree scalar_type) | |
8540 | { | |
8541 | tree vectype; | |
8542 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
8543 | current_vector_size); | |
8544 | if (vectype | |
8545 | && current_vector_size == 0) | |
8546 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
8547 | return vectype; | |
8548 | } | |
8549 | ||
42fd8198 IE |
8550 | /* Function get_mask_type_for_scalar_type. |
8551 | ||
8552 | Returns the mask type corresponding to a result of comparison | |
8553 | of vectors of specified SCALAR_TYPE as supported by target. */ | |
8554 | ||
8555 | tree | |
8556 | get_mask_type_for_scalar_type (tree scalar_type) | |
8557 | { | |
8558 | tree vectype = get_vectype_for_scalar_type (scalar_type); | |
8559 | ||
8560 | if (!vectype) | |
8561 | return NULL; | |
8562 | ||
8563 | return build_truth_vector_type (TYPE_VECTOR_SUBPARTS (vectype), | |
8564 | current_vector_size); | |
8565 | } | |
8566 | ||
b690cc0f RG |
8567 | /* Function get_same_sized_vectype |
8568 | ||
8569 | Returns a vector type corresponding to SCALAR_TYPE of size | |
8570 | VECTOR_TYPE if supported by the target. */ | |
8571 | ||
8572 | tree | |
bb67d9c7 | 8573 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 8574 | { |
9f47c7e5 IE |
8575 | if (TREE_CODE (scalar_type) == BOOLEAN_TYPE) |
8576 | return build_same_sized_truth_vector_type (vector_type); | |
8577 | ||
bb67d9c7 RG |
8578 | return get_vectype_for_scalar_type_and_size |
8579 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
8580 | } |
8581 | ||
ebfd146a IR |
8582 | /* Function vect_is_simple_use. |
8583 | ||
8584 | Input: | |
81c40241 RB |
8585 | VINFO - the vect info of the loop or basic block that is being vectorized. |
8586 | OPERAND - operand in the loop or bb. | |
8587 | Output: | |
8588 | DEF_STMT - the defining stmt in case OPERAND is an SSA_NAME. | |
8589 | DT - the type of definition | |
ebfd146a IR |
8590 | |
8591 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 8592 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 8593 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 8594 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
8595 | is the case in reduction/induction computations). |
8596 | For basic blocks, supportable operands are constants and bb invariants. | |
8597 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
8598 | |
8599 | bool | |
81c40241 RB |
8600 | vect_is_simple_use (tree operand, vec_info *vinfo, |
8601 | gimple **def_stmt, enum vect_def_type *dt) | |
b8698a0f | 8602 | { |
ebfd146a | 8603 | *def_stmt = NULL; |
3fc356dc | 8604 | *dt = vect_unknown_def_type; |
b8698a0f | 8605 | |
73fbfcad | 8606 | if (dump_enabled_p ()) |
ebfd146a | 8607 | { |
78c60e3d SS |
8608 | dump_printf_loc (MSG_NOTE, vect_location, |
8609 | "vect_is_simple_use: operand "); | |
8610 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
e645e942 | 8611 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 8612 | } |
b8698a0f | 8613 | |
b758f602 | 8614 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
8615 | { |
8616 | *dt = vect_constant_def; | |
8617 | return true; | |
8618 | } | |
b8698a0f | 8619 | |
ebfd146a IR |
8620 | if (is_gimple_min_invariant (operand)) |
8621 | { | |
8644a673 | 8622 | *dt = vect_external_def; |
ebfd146a IR |
8623 | return true; |
8624 | } | |
8625 | ||
ebfd146a IR |
8626 | if (TREE_CODE (operand) != SSA_NAME) |
8627 | { | |
73fbfcad | 8628 | if (dump_enabled_p ()) |
af29617a AH |
8629 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8630 | "not ssa-name.\n"); | |
ebfd146a IR |
8631 | return false; |
8632 | } | |
b8698a0f | 8633 | |
3fc356dc | 8634 | if (SSA_NAME_IS_DEFAULT_DEF (operand)) |
ebfd146a | 8635 | { |
3fc356dc RB |
8636 | *dt = vect_external_def; |
8637 | return true; | |
ebfd146a IR |
8638 | } |
8639 | ||
3fc356dc | 8640 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
73fbfcad | 8641 | if (dump_enabled_p ()) |
ebfd146a | 8642 | { |
78c60e3d SS |
8643 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
8644 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
ebfd146a IR |
8645 | } |
8646 | ||
61d371eb | 8647 | if (! vect_stmt_in_region_p (vinfo, *def_stmt)) |
8644a673 | 8648 | *dt = vect_external_def; |
ebfd146a IR |
8649 | else |
8650 | { | |
3fc356dc | 8651 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (*def_stmt); |
603cca93 | 8652 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
ebfd146a IR |
8653 | } |
8654 | ||
2e8ab70c RB |
8655 | if (dump_enabled_p ()) |
8656 | { | |
8657 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: "); | |
8658 | switch (*dt) | |
8659 | { | |
8660 | case vect_uninitialized_def: | |
8661 | dump_printf (MSG_NOTE, "uninitialized\n"); | |
8662 | break; | |
8663 | case vect_constant_def: | |
8664 | dump_printf (MSG_NOTE, "constant\n"); | |
8665 | break; | |
8666 | case vect_external_def: | |
8667 | dump_printf (MSG_NOTE, "external\n"); | |
8668 | break; | |
8669 | case vect_internal_def: | |
8670 | dump_printf (MSG_NOTE, "internal\n"); | |
8671 | break; | |
8672 | case vect_induction_def: | |
8673 | dump_printf (MSG_NOTE, "induction\n"); | |
8674 | break; | |
8675 | case vect_reduction_def: | |
8676 | dump_printf (MSG_NOTE, "reduction\n"); | |
8677 | break; | |
8678 | case vect_double_reduction_def: | |
8679 | dump_printf (MSG_NOTE, "double reduction\n"); | |
8680 | break; | |
8681 | case vect_nested_cycle: | |
8682 | dump_printf (MSG_NOTE, "nested cycle\n"); | |
8683 | break; | |
8684 | case vect_unknown_def_type: | |
8685 | dump_printf (MSG_NOTE, "unknown\n"); | |
8686 | break; | |
8687 | } | |
8688 | } | |
8689 | ||
81c40241 | 8690 | if (*dt == vect_unknown_def_type) |
ebfd146a | 8691 | { |
73fbfcad | 8692 | if (dump_enabled_p ()) |
78c60e3d | 8693 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8694 | "Unsupported pattern.\n"); |
ebfd146a IR |
8695 | return false; |
8696 | } | |
8697 | ||
ebfd146a IR |
8698 | switch (gimple_code (*def_stmt)) |
8699 | { | |
8700 | case GIMPLE_PHI: | |
ebfd146a | 8701 | case GIMPLE_ASSIGN: |
ebfd146a | 8702 | case GIMPLE_CALL: |
81c40241 | 8703 | break; |
ebfd146a | 8704 | default: |
73fbfcad | 8705 | if (dump_enabled_p ()) |
78c60e3d | 8706 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8707 | "unsupported defining stmt:\n"); |
ebfd146a IR |
8708 | return false; |
8709 | } | |
8710 | ||
8711 | return true; | |
8712 | } | |
8713 | ||
81c40241 | 8714 | /* Function vect_is_simple_use. |
b690cc0f | 8715 | |
81c40241 | 8716 | Same as vect_is_simple_use but also determines the vector operand |
b690cc0f RG |
8717 | type of OPERAND and stores it to *VECTYPE. If the definition of |
8718 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
8719 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
8720 | is responsible to compute the best suited vector type for the | |
8721 | scalar operand. */ | |
8722 | ||
8723 | bool | |
81c40241 RB |
8724 | vect_is_simple_use (tree operand, vec_info *vinfo, |
8725 | gimple **def_stmt, enum vect_def_type *dt, tree *vectype) | |
b690cc0f | 8726 | { |
81c40241 | 8727 | if (!vect_is_simple_use (operand, vinfo, def_stmt, dt)) |
b690cc0f RG |
8728 | return false; |
8729 | ||
8730 | /* Now get a vector type if the def is internal, otherwise supply | |
8731 | NULL_TREE and leave it up to the caller to figure out a proper | |
8732 | type for the use stmt. */ | |
8733 | if (*dt == vect_internal_def | |
8734 | || *dt == vect_induction_def | |
8735 | || *dt == vect_reduction_def | |
8736 | || *dt == vect_double_reduction_def | |
8737 | || *dt == vect_nested_cycle) | |
8738 | { | |
8739 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
8740 | |
8741 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
8742 | && !STMT_VINFO_RELEVANT (stmt_info) | |
8743 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 8744 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 8745 | |
b690cc0f RG |
8746 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
8747 | gcc_assert (*vectype != NULL_TREE); | |
8748 | } | |
8749 | else if (*dt == vect_uninitialized_def | |
8750 | || *dt == vect_constant_def | |
8751 | || *dt == vect_external_def) | |
8752 | *vectype = NULL_TREE; | |
8753 | else | |
8754 | gcc_unreachable (); | |
8755 | ||
8756 | return true; | |
8757 | } | |
8758 | ||
ebfd146a IR |
8759 | |
8760 | /* Function supportable_widening_operation | |
8761 | ||
b8698a0f L |
8762 | Check whether an operation represented by the code CODE is a |
8763 | widening operation that is supported by the target platform in | |
b690cc0f RG |
8764 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
8765 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 8766 | |
ebfd146a IR |
8767 | Widening operations we currently support are NOP (CONVERT), FLOAT |
8768 | and WIDEN_MULT. This function checks if these operations are supported | |
8769 | by the target platform either directly (via vector tree-codes), or via | |
8770 | target builtins. | |
8771 | ||
8772 | Output: | |
b8698a0f L |
8773 | - CODE1 and CODE2 are codes of vector operations to be used when |
8774 | vectorizing the operation, if available. | |
ebfd146a IR |
8775 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
8776 | case of multi-step conversion (like char->short->int - in that case | |
8777 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
8778 | - INTERM_TYPES contains the intermediate type required to perform the |
8779 | widening operation (short in the above example). */ | |
ebfd146a IR |
8780 | |
8781 | bool | |
355fe088 | 8782 | supportable_widening_operation (enum tree_code code, gimple *stmt, |
b690cc0f | 8783 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
8784 | enum tree_code *code1, enum tree_code *code2, |
8785 | int *multi_step_cvt, | |
9771b263 | 8786 | vec<tree> *interm_types) |
ebfd146a IR |
8787 | { |
8788 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8789 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 8790 | struct loop *vect_loop = NULL; |
ef4bddc2 | 8791 | machine_mode vec_mode; |
81f40b79 | 8792 | enum insn_code icode1, icode2; |
ebfd146a | 8793 | optab optab1, optab2; |
b690cc0f RG |
8794 | tree vectype = vectype_in; |
8795 | tree wide_vectype = vectype_out; | |
ebfd146a | 8796 | enum tree_code c1, c2; |
4a00c761 JJ |
8797 | int i; |
8798 | tree prev_type, intermediate_type; | |
ef4bddc2 | 8799 | machine_mode intermediate_mode, prev_mode; |
4a00c761 | 8800 | optab optab3, optab4; |
ebfd146a | 8801 | |
4a00c761 | 8802 | *multi_step_cvt = 0; |
4ef69dfc IR |
8803 | if (loop_info) |
8804 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
8805 | ||
ebfd146a IR |
8806 | switch (code) |
8807 | { | |
8808 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
8809 | /* The result of a vectorized widening operation usually requires |
8810 | two vectors (because the widened results do not fit into one vector). | |
8811 | The generated vector results would normally be expected to be | |
8812 | generated in the same order as in the original scalar computation, | |
8813 | i.e. if 8 results are generated in each vector iteration, they are | |
8814 | to be organized as follows: | |
8815 | vect1: [res1,res2,res3,res4], | |
8816 | vect2: [res5,res6,res7,res8]. | |
8817 | ||
8818 | However, in the special case that the result of the widening | |
8819 | operation is used in a reduction computation only, the order doesn't | |
8820 | matter (because when vectorizing a reduction we change the order of | |
8821 | the computation). Some targets can take advantage of this and | |
8822 | generate more efficient code. For example, targets like Altivec, | |
8823 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
8824 | generate the following vectors: | |
8825 | vect1: [res1,res3,res5,res7], | |
8826 | vect2: [res2,res4,res6,res8]. | |
8827 | ||
8828 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
8829 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
8830 | iterations in parallel). We therefore don't allow to change the | |
8831 | order of the computation in the inner-loop during outer-loop | |
8832 | vectorization. */ | |
8833 | /* TODO: Another case in which order doesn't *really* matter is when we | |
8834 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
8835 | Normally, pack_trunc performs an even/odd permute, whereas the | |
8836 | repack from an even/odd expansion would be an interleave, which | |
8837 | would be significantly simpler for e.g. AVX2. */ | |
8838 | /* In any case, in order to avoid duplicating the code below, recurse | |
8839 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
8840 | are properly set up for the caller. If we fail, we'll continue with | |
8841 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
8842 | if (vect_loop | |
8843 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
8844 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
8845 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
8846 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
8847 | code1, code2, multi_step_cvt, |
8848 | interm_types)) | |
ebc047a2 CH |
8849 | { |
8850 | /* Elements in a vector with vect_used_by_reduction property cannot | |
8851 | be reordered if the use chain with this property does not have the | |
8852 | same operation. One such an example is s += a * b, where elements | |
8853 | in a and b cannot be reordered. Here we check if the vector defined | |
8854 | by STMT is only directly used in the reduction statement. */ | |
8855 | tree lhs = gimple_assign_lhs (stmt); | |
8856 | use_operand_p dummy; | |
355fe088 | 8857 | gimple *use_stmt; |
ebc047a2 CH |
8858 | stmt_vec_info use_stmt_info = NULL; |
8859 | if (single_imm_use (lhs, &dummy, &use_stmt) | |
8860 | && (use_stmt_info = vinfo_for_stmt (use_stmt)) | |
8861 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) | |
8862 | return true; | |
8863 | } | |
4a00c761 JJ |
8864 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
8865 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
8866 | break; |
8867 | ||
81c40241 RB |
8868 | case DOT_PROD_EXPR: |
8869 | c1 = DOT_PROD_EXPR; | |
8870 | c2 = DOT_PROD_EXPR; | |
8871 | break; | |
8872 | ||
8873 | case SAD_EXPR: | |
8874 | c1 = SAD_EXPR; | |
8875 | c2 = SAD_EXPR; | |
8876 | break; | |
8877 | ||
6ae6116f RH |
8878 | case VEC_WIDEN_MULT_EVEN_EXPR: |
8879 | /* Support the recursion induced just above. */ | |
8880 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
8881 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
8882 | break; | |
8883 | ||
36ba4aae | 8884 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
8885 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
8886 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
8887 | break; |
8888 | ||
ebfd146a | 8889 | CASE_CONVERT: |
4a00c761 JJ |
8890 | c1 = VEC_UNPACK_LO_EXPR; |
8891 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
8892 | break; |
8893 | ||
8894 | case FLOAT_EXPR: | |
4a00c761 JJ |
8895 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
8896 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
8897 | break; |
8898 | ||
8899 | case FIX_TRUNC_EXPR: | |
8900 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
8901 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
8902 | computing the operation. */ | |
8903 | return false; | |
8904 | ||
8905 | default: | |
8906 | gcc_unreachable (); | |
8907 | } | |
8908 | ||
6ae6116f | 8909 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
6b4db501 | 8910 | std::swap (c1, c2); |
4a00c761 | 8911 | |
ebfd146a IR |
8912 | if (code == FIX_TRUNC_EXPR) |
8913 | { | |
8914 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
8915 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
8916 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
8917 | } |
8918 | else | |
8919 | { | |
8920 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
8921 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
8922 | } | |
8923 | ||
8924 | if (!optab1 || !optab2) | |
8925 | return false; | |
8926 | ||
8927 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
8928 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
8929 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
8930 | return false; |
8931 | ||
4a00c761 JJ |
8932 | *code1 = c1; |
8933 | *code2 = c2; | |
8934 | ||
8935 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
8936 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
8937 | return true; | |
8938 | ||
b8698a0f | 8939 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 8940 | types. */ |
ebfd146a | 8941 | |
4a00c761 JJ |
8942 | prev_type = vectype; |
8943 | prev_mode = vec_mode; | |
b8698a0f | 8944 | |
4a00c761 JJ |
8945 | if (!CONVERT_EXPR_CODE_P (code)) |
8946 | return false; | |
b8698a0f | 8947 | |
4a00c761 JJ |
8948 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
8949 | intermediate steps in promotion sequence. We try | |
8950 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
8951 | not. */ | |
9771b263 | 8952 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
8953 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
8954 | { | |
8955 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
8956 | intermediate_type | |
8957 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
8958 | TYPE_UNSIGNED (prev_type)); | |
8959 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); | |
8960 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
8961 | ||
8962 | if (!optab3 || !optab4 | |
8963 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
8964 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
8965 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
8966 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
8967 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
8968 | == CODE_FOR_nothing) | |
8969 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
8970 | == CODE_FOR_nothing)) | |
8971 | break; | |
ebfd146a | 8972 | |
9771b263 | 8973 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
8974 | (*multi_step_cvt)++; |
8975 | ||
8976 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
8977 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
8978 | return true; | |
8979 | ||
8980 | prev_type = intermediate_type; | |
8981 | prev_mode = intermediate_mode; | |
ebfd146a IR |
8982 | } |
8983 | ||
9771b263 | 8984 | interm_types->release (); |
4a00c761 | 8985 | return false; |
ebfd146a IR |
8986 | } |
8987 | ||
8988 | ||
8989 | /* Function supportable_narrowing_operation | |
8990 | ||
b8698a0f L |
8991 | Check whether an operation represented by the code CODE is a |
8992 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
8993 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
8994 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 8995 | |
ebfd146a | 8996 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 8997 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
8998 | the target platform directly via vector tree-codes. |
8999 | ||
9000 | Output: | |
b8698a0f L |
9001 | - CODE1 is the code of a vector operation to be used when |
9002 | vectorizing the operation, if available. | |
ebfd146a IR |
9003 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
9004 | case of multi-step conversion (like int->short->char - in that case | |
9005 | MULTI_STEP_CVT will be 1). | |
9006 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 9007 | narrowing operation (short in the above example). */ |
ebfd146a IR |
9008 | |
9009 | bool | |
9010 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 9011 | tree vectype_out, tree vectype_in, |
ebfd146a | 9012 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 9013 | vec<tree> *interm_types) |
ebfd146a | 9014 | { |
ef4bddc2 | 9015 | machine_mode vec_mode; |
ebfd146a IR |
9016 | enum insn_code icode1; |
9017 | optab optab1, interm_optab; | |
b690cc0f RG |
9018 | tree vectype = vectype_in; |
9019 | tree narrow_vectype = vectype_out; | |
ebfd146a | 9020 | enum tree_code c1; |
4a00c761 | 9021 | tree intermediate_type; |
ef4bddc2 | 9022 | machine_mode intermediate_mode, prev_mode; |
ebfd146a | 9023 | int i; |
4a00c761 | 9024 | bool uns; |
ebfd146a | 9025 | |
4a00c761 | 9026 | *multi_step_cvt = 0; |
ebfd146a IR |
9027 | switch (code) |
9028 | { | |
9029 | CASE_CONVERT: | |
9030 | c1 = VEC_PACK_TRUNC_EXPR; | |
9031 | break; | |
9032 | ||
9033 | case FIX_TRUNC_EXPR: | |
9034 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
9035 | break; | |
9036 | ||
9037 | case FLOAT_EXPR: | |
9038 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
9039 | tree code and optabs used for computing the operation. */ | |
9040 | return false; | |
9041 | ||
9042 | default: | |
9043 | gcc_unreachable (); | |
9044 | } | |
9045 | ||
9046 | if (code == FIX_TRUNC_EXPR) | |
9047 | /* The signedness is determined from output operand. */ | |
b690cc0f | 9048 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
9049 | else |
9050 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
9051 | ||
9052 | if (!optab1) | |
9053 | return false; | |
9054 | ||
9055 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 9056 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
9057 | return false; |
9058 | ||
4a00c761 JJ |
9059 | *code1 = c1; |
9060 | ||
9061 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
9062 | return true; | |
9063 | ||
ebfd146a IR |
9064 | /* Check if it's a multi-step conversion that can be done using intermediate |
9065 | types. */ | |
4a00c761 JJ |
9066 | prev_mode = vec_mode; |
9067 | if (code == FIX_TRUNC_EXPR) | |
9068 | uns = TYPE_UNSIGNED (vectype_out); | |
9069 | else | |
9070 | uns = TYPE_UNSIGNED (vectype); | |
9071 | ||
9072 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
9073 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
9074 | costly than signed. */ | |
9075 | if (code == FIX_TRUNC_EXPR && uns) | |
9076 | { | |
9077 | enum insn_code icode2; | |
9078 | ||
9079 | intermediate_type | |
9080 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
9081 | interm_optab | |
9082 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 9083 | if (interm_optab != unknown_optab |
4a00c761 JJ |
9084 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
9085 | && insn_data[icode1].operand[0].mode | |
9086 | == insn_data[icode2].operand[0].mode) | |
9087 | { | |
9088 | uns = false; | |
9089 | optab1 = interm_optab; | |
9090 | icode1 = icode2; | |
9091 | } | |
9092 | } | |
ebfd146a | 9093 | |
4a00c761 JJ |
9094 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
9095 | intermediate steps in promotion sequence. We try | |
9096 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 9097 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
9098 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
9099 | { | |
9100 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
9101 | intermediate_type | |
9102 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
9103 | interm_optab | |
9104 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
9105 | optab_default); | |
9106 | if (!interm_optab | |
9107 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
9108 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
9109 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
9110 | == CODE_FOR_nothing)) | |
9111 | break; | |
9112 | ||
9771b263 | 9113 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
9114 | (*multi_step_cvt)++; |
9115 | ||
9116 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
9117 | return true; | |
9118 | ||
9119 | prev_mode = intermediate_mode; | |
9120 | optab1 = interm_optab; | |
ebfd146a IR |
9121 | } |
9122 | ||
9771b263 | 9123 | interm_types->release (); |
4a00c761 | 9124 | return false; |
ebfd146a | 9125 | } |