]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
5624e564 | 2 | Copyright (C) 2003-2015 Free Software Foundation, Inc. |
b8698a0f | 3 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
4 | and Ira Rosen <irar@il.ibm.com> |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 3, or (at your option) any later | |
11 | version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
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 | { |
5467ee52 | 1303 | if (CONSTANT_CLASS_P (val)) |
42fd8198 | 1304 | val = fold_convert (TREE_TYPE (type), val); |
418b7df3 RG |
1305 | else |
1306 | { | |
b731b390 | 1307 | new_temp = make_ssa_name (TREE_TYPE (type)); |
0d0e4a03 | 1308 | init_stmt = gimple_build_assign (new_temp, NOP_EXPR, val); |
418b7df3 | 1309 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1310 | val = new_temp; |
418b7df3 RG |
1311 | } |
1312 | } | |
5467ee52 | 1313 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1314 | } |
1315 | ||
0e22bb5a RB |
1316 | new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_"); |
1317 | init_stmt = gimple_build_assign (new_temp, val); | |
418b7df3 | 1318 | vect_init_vector_1 (stmt, init_stmt, gsi); |
0e22bb5a | 1319 | return new_temp; |
ebfd146a IR |
1320 | } |
1321 | ||
a70d6342 | 1322 | |
ebfd146a IR |
1323 | /* Function vect_get_vec_def_for_operand. |
1324 | ||
ff802fa1 | 1325 | OP is an operand in STMT. This function returns a (vector) def that will be |
ebfd146a IR |
1326 | used in the vectorized stmt for STMT. |
1327 | ||
1328 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1329 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1330 | ||
1331 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
42fd8198 IE |
1332 | needs to be introduced. VECTYPE may be used to specify a required type for |
1333 | vector invariant. */ | |
ebfd146a IR |
1334 | |
1335 | tree | |
42fd8198 | 1336 | vect_get_vec_def_for_operand (tree op, gimple *stmt, tree vectype) |
ebfd146a IR |
1337 | { |
1338 | tree vec_oprnd; | |
355fe088 TS |
1339 | gimple *vec_stmt; |
1340 | gimple *def_stmt; | |
ebfd146a IR |
1341 | stmt_vec_info def_stmt_info = NULL; |
1342 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
42fd8198 | 1343 | tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo); |
ebfd146a | 1344 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
ebfd146a IR |
1345 | enum vect_def_type dt; |
1346 | bool is_simple_use; | |
1347 | tree vector_type; | |
1348 | ||
73fbfcad | 1349 | if (dump_enabled_p ()) |
ebfd146a | 1350 | { |
78c60e3d SS |
1351 | dump_printf_loc (MSG_NOTE, vect_location, |
1352 | "vect_get_vec_def_for_operand: "); | |
1353 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
e645e942 | 1354 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1355 | } |
1356 | ||
81c40241 | 1357 | is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt); |
ebfd146a | 1358 | gcc_assert (is_simple_use); |
73fbfcad | 1359 | if (dump_enabled_p ()) |
ebfd146a | 1360 | { |
78c60e3d | 1361 | int loc_printed = 0; |
ebfd146a IR |
1362 | if (def_stmt) |
1363 | { | |
78c60e3d SS |
1364 | if (loc_printed) |
1365 | dump_printf (MSG_NOTE, " def_stmt = "); | |
1366 | else | |
1367 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1368 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
ebfd146a IR |
1369 | } |
1370 | } | |
1371 | ||
1372 | switch (dt) | |
1373 | { | |
81c40241 | 1374 | /* operand is a constant or a loop invariant. */ |
ebfd146a | 1375 | case vect_constant_def: |
81c40241 | 1376 | case vect_external_def: |
ebfd146a | 1377 | { |
42fd8198 IE |
1378 | if (vectype) |
1379 | vector_type = vectype; | |
1380 | else if (TREE_CODE (TREE_TYPE (op)) == BOOLEAN_TYPE | |
1381 | && VECTOR_BOOLEAN_TYPE_P (stmt_vectype)) | |
1382 | vector_type = build_same_sized_truth_vector_type (stmt_vectype); | |
1383 | else | |
1384 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
1385 | ||
7569a6cc | 1386 | gcc_assert (vector_type); |
418b7df3 | 1387 | return vect_init_vector (stmt, op, vector_type, NULL); |
ebfd146a IR |
1388 | } |
1389 | ||
81c40241 | 1390 | /* operand is defined inside the loop. */ |
8644a673 | 1391 | case vect_internal_def: |
ebfd146a | 1392 | { |
ebfd146a IR |
1393 | /* Get the def from the vectorized stmt. */ |
1394 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1395 | |
ebfd146a | 1396 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1397 | /* Get vectorized pattern statement. */ |
1398 | if (!vec_stmt | |
1399 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1400 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1401 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1402 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1403 | gcc_assert (vec_stmt); |
1404 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1405 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1406 | else if (is_gimple_call (vec_stmt)) | |
1407 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1408 | else | |
1409 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1410 | return vec_oprnd; | |
1411 | } | |
1412 | ||
81c40241 | 1413 | /* operand is defined by a loop header phi - reduction */ |
ebfd146a | 1414 | case vect_reduction_def: |
06066f92 | 1415 | case vect_double_reduction_def: |
7c5222ff | 1416 | case vect_nested_cycle: |
81c40241 RB |
1417 | /* Code should use get_initial_def_for_reduction. */ |
1418 | gcc_unreachable (); | |
ebfd146a | 1419 | |
81c40241 | 1420 | /* operand is defined by loop-header phi - induction. */ |
ebfd146a IR |
1421 | case vect_induction_def: |
1422 | { | |
1423 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1424 | ||
1425 | /* Get the def from the vectorized stmt. */ | |
1426 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1427 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1428 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1429 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1430 | else | |
1431 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1432 | return vec_oprnd; |
1433 | } | |
1434 | ||
1435 | default: | |
1436 | gcc_unreachable (); | |
1437 | } | |
1438 | } | |
1439 | ||
1440 | ||
1441 | /* Function vect_get_vec_def_for_stmt_copy | |
1442 | ||
ff802fa1 | 1443 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1444 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1445 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1446 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1447 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1448 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1449 | DT is the type of the vector def VEC_OPRND. |
1450 | ||
1451 | Context: | |
1452 | In case the vectorization factor (VF) is bigger than the number | |
1453 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1454 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1455 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1456 | smallest data-type determines the VF, and as a result, when vectorizing |
1457 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1458 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1459 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1460 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1461 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1462 | ||
1463 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1464 | |
ebfd146a IR |
1465 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1466 | VS1.1: vx.1 = memref1 VS1.2 | |
1467 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1468 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1469 | |
1470 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1471 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1472 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1473 | VSnew.3: vz3 = vx.3 + ... | |
1474 | ||
1475 | The vectorization of S1 is explained in vectorizable_load. | |
1476 | The vectorization of S2: | |
b8698a0f L |
1477 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1478 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1479 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1480 | returns the vector-def 'vx.0'. |
1481 | ||
b8698a0f L |
1482 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1483 | function is called to get the relevant vector-def for each operand. It is | |
1484 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1485 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1486 | ||
b8698a0f L |
1487 | For example, to obtain the vector-def 'vx.1' in order to create the |
1488 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1489 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1490 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1491 | and return its def ('vx.1'). | |
1492 | Overall, to create the above sequence this function will be called 3 times: | |
1493 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1494 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1495 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1496 | ||
1497 | tree | |
1498 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1499 | { | |
355fe088 | 1500 | gimple *vec_stmt_for_operand; |
ebfd146a IR |
1501 | stmt_vec_info def_stmt_info; |
1502 | ||
1503 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1504 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1505 | return vec_oprnd; |
1506 | ||
1507 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1508 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1509 | gcc_assert (def_stmt_info); | |
1510 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1511 | gcc_assert (vec_stmt_for_operand); | |
ebfd146a IR |
1512 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) |
1513 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1514 | else | |
1515 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1516 | return vec_oprnd; | |
1517 | } | |
1518 | ||
1519 | ||
1520 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1521 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a IR |
1522 | |
1523 | static void | |
b8698a0f | 1524 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1525 | vec<tree> *vec_oprnds0, |
1526 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1527 | { |
9771b263 | 1528 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1529 | |
1530 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1531 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1532 | |
9771b263 | 1533 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1534 | { |
9771b263 | 1535 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1536 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1537 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1538 | } |
1539 | } | |
1540 | ||
1541 | ||
d092494c IR |
1542 | /* Get vectorized definitions for OP0 and OP1. |
1543 | REDUC_INDEX is the index of reduction operand in case of reduction, | |
1544 | and -1 otherwise. */ | |
ebfd146a | 1545 | |
d092494c | 1546 | void |
355fe088 | 1547 | vect_get_vec_defs (tree op0, tree op1, gimple *stmt, |
9771b263 DN |
1548 | vec<tree> *vec_oprnds0, |
1549 | vec<tree> *vec_oprnds1, | |
d092494c | 1550 | slp_tree slp_node, int reduc_index) |
ebfd146a IR |
1551 | { |
1552 | if (slp_node) | |
d092494c IR |
1553 | { |
1554 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1555 | auto_vec<tree> ops (nops); |
1556 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1557 | |
9771b263 | 1558 | ops.quick_push (op0); |
d092494c | 1559 | if (op1) |
9771b263 | 1560 | ops.quick_push (op1); |
d092494c IR |
1561 | |
1562 | vect_get_slp_defs (ops, slp_node, &vec_defs, reduc_index); | |
1563 | ||
37b5ec8f | 1564 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1565 | if (op1) |
37b5ec8f | 1566 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1567 | } |
ebfd146a IR |
1568 | else |
1569 | { | |
1570 | tree vec_oprnd; | |
1571 | ||
9771b263 | 1572 | vec_oprnds0->create (1); |
81c40241 | 1573 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 1574 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1575 | |
1576 | if (op1) | |
1577 | { | |
9771b263 | 1578 | vec_oprnds1->create (1); |
81c40241 | 1579 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 1580 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1581 | } |
1582 | } | |
1583 | } | |
1584 | ||
1585 | ||
1586 | /* Function vect_finish_stmt_generation. | |
1587 | ||
1588 | Insert a new stmt. */ | |
1589 | ||
1590 | void | |
355fe088 | 1591 | vect_finish_stmt_generation (gimple *stmt, gimple *vec_stmt, |
ebfd146a IR |
1592 | gimple_stmt_iterator *gsi) |
1593 | { | |
1594 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
310213d4 | 1595 | vec_info *vinfo = stmt_info->vinfo; |
ebfd146a IR |
1596 | |
1597 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1598 | ||
54e8e2c3 RG |
1599 | if (!gsi_end_p (*gsi) |
1600 | && gimple_has_mem_ops (vec_stmt)) | |
1601 | { | |
355fe088 | 1602 | gimple *at_stmt = gsi_stmt (*gsi); |
54e8e2c3 RG |
1603 | tree vuse = gimple_vuse (at_stmt); |
1604 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1605 | { | |
1606 | tree vdef = gimple_vdef (at_stmt); | |
1607 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1608 | /* If we have an SSA vuse and insert a store, update virtual | |
1609 | SSA form to avoid triggering the renamer. Do so only | |
1610 | if we can easily see all uses - which is what almost always | |
1611 | happens with the way vectorized stmts are inserted. */ | |
1612 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1613 | && ((is_gimple_assign (vec_stmt) | |
1614 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1615 | || (is_gimple_call (vec_stmt) | |
1616 | && !(gimple_call_flags (vec_stmt) | |
1617 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1618 | { | |
1619 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1620 | gimple_set_vdef (vec_stmt, new_vdef); | |
1621 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1622 | } | |
1623 | } | |
1624 | } | |
ebfd146a IR |
1625 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1626 | ||
310213d4 | 1627 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, vinfo)); |
ebfd146a | 1628 | |
73fbfcad | 1629 | if (dump_enabled_p ()) |
ebfd146a | 1630 | { |
78c60e3d SS |
1631 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); |
1632 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
ebfd146a IR |
1633 | } |
1634 | ||
ad885386 | 1635 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
8e91d222 JJ |
1636 | |
1637 | /* While EH edges will generally prevent vectorization, stmt might | |
1638 | e.g. be in a must-not-throw region. Ensure newly created stmts | |
1639 | that could throw are part of the same region. */ | |
1640 | int lp_nr = lookup_stmt_eh_lp (stmt); | |
1641 | if (lp_nr != 0 && stmt_could_throw_p (vec_stmt)) | |
1642 | add_stmt_to_eh_lp (vec_stmt, lp_nr); | |
ebfd146a IR |
1643 | } |
1644 | ||
70439f0d RS |
1645 | /* We want to vectorize a call to combined function CFN with function |
1646 | decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN | |
1647 | as the types of all inputs. Check whether this is possible using | |
1648 | an internal function, returning its code if so or IFN_LAST if not. */ | |
ebfd146a | 1649 | |
70439f0d RS |
1650 | static internal_fn |
1651 | vectorizable_internal_function (combined_fn cfn, tree fndecl, | |
1652 | tree vectype_out, tree vectype_in) | |
ebfd146a | 1653 | { |
70439f0d RS |
1654 | internal_fn ifn; |
1655 | if (internal_fn_p (cfn)) | |
1656 | ifn = as_internal_fn (cfn); | |
1657 | else | |
1658 | ifn = associated_internal_fn (fndecl); | |
1659 | if (ifn != IFN_LAST && direct_internal_fn_p (ifn)) | |
1660 | { | |
1661 | const direct_internal_fn_info &info = direct_internal_fn (ifn); | |
1662 | if (info.vectorizable) | |
1663 | { | |
1664 | tree type0 = (info.type0 < 0 ? vectype_out : vectype_in); | |
1665 | tree type1 = (info.type1 < 0 ? vectype_out : vectype_in); | |
1666 | if (direct_internal_fn_supported_p (ifn, tree_pair (type0, type1))) | |
1667 | return ifn; | |
1668 | } | |
1669 | } | |
1670 | return IFN_LAST; | |
ebfd146a IR |
1671 | } |
1672 | ||
5ce9450f | 1673 | |
355fe088 | 1674 | static tree permute_vec_elements (tree, tree, tree, gimple *, |
5ce9450f JJ |
1675 | gimple_stmt_iterator *); |
1676 | ||
1677 | ||
1678 | /* Function vectorizable_mask_load_store. | |
1679 | ||
1680 | Check if STMT performs a conditional load or store that can be vectorized. | |
1681 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
1682 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. | |
1683 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1684 | ||
1685 | static bool | |
355fe088 TS |
1686 | vectorizable_mask_load_store (gimple *stmt, gimple_stmt_iterator *gsi, |
1687 | gimple **vec_stmt, slp_tree slp_node) | |
5ce9450f JJ |
1688 | { |
1689 | tree vec_dest = NULL; | |
1690 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1691 | stmt_vec_info prev_stmt_info; | |
1692 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1693 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1694 | bool nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
1695 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
1696 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
57e2f6ad | 1697 | tree rhs_vectype = NULL_TREE; |
045c1278 | 1698 | tree mask_vectype; |
5ce9450f | 1699 | tree elem_type; |
355fe088 | 1700 | gimple *new_stmt; |
5ce9450f JJ |
1701 | tree dummy; |
1702 | tree dataref_ptr = NULL_TREE; | |
355fe088 | 1703 | gimple *ptr_incr; |
5ce9450f JJ |
1704 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
1705 | int ncopies; | |
1706 | int i, j; | |
1707 | bool inv_p; | |
1708 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; | |
1709 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
1710 | int gather_scale = 1; | |
1711 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
1712 | bool is_store; | |
1713 | tree mask; | |
355fe088 | 1714 | gimple *def_stmt; |
5ce9450f JJ |
1715 | enum vect_def_type dt; |
1716 | ||
1717 | if (slp_node != NULL) | |
1718 | return false; | |
1719 | ||
1720 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
1721 | gcc_assert (ncopies >= 1); | |
1722 | ||
1723 | is_store = gimple_call_internal_fn (stmt) == IFN_MASK_STORE; | |
1724 | mask = gimple_call_arg (stmt, 2); | |
045c1278 IE |
1725 | |
1726 | if (TREE_CODE (TREE_TYPE (mask)) != BOOLEAN_TYPE) | |
5ce9450f JJ |
1727 | return false; |
1728 | ||
1729 | /* FORNOW. This restriction should be relaxed. */ | |
1730 | if (nested_in_vect_loop && ncopies > 1) | |
1731 | { | |
1732 | if (dump_enabled_p ()) | |
1733 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1734 | "multiple types in nested loop."); | |
1735 | return false; | |
1736 | } | |
1737 | ||
1738 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) | |
1739 | return false; | |
1740 | ||
1741 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
1742 | return false; | |
1743 | ||
1744 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
1745 | return false; | |
1746 | ||
1747 | elem_type = TREE_TYPE (vectype); | |
1748 | ||
1749 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1750 | return false; | |
1751 | ||
f2e2a985 | 1752 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
5ce9450f JJ |
1753 | return false; |
1754 | ||
045c1278 IE |
1755 | if (TREE_CODE (mask) != SSA_NAME) |
1756 | return false; | |
1757 | ||
1758 | if (!vect_is_simple_use (mask, loop_vinfo, &def_stmt, &dt, &mask_vectype)) | |
1759 | return false; | |
1760 | ||
1761 | if (!mask_vectype) | |
1762 | mask_vectype = get_mask_type_for_scalar_type (TREE_TYPE (vectype)); | |
1763 | ||
1764 | if (!mask_vectype) | |
1765 | return false; | |
1766 | ||
57e2f6ad IE |
1767 | if (is_store) |
1768 | { | |
1769 | tree rhs = gimple_call_arg (stmt, 3); | |
1770 | if (!vect_is_simple_use (rhs, loop_vinfo, &def_stmt, &dt, &rhs_vectype)) | |
1771 | return false; | |
1772 | } | |
1773 | ||
3bab6342 | 1774 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5ce9450f | 1775 | { |
355fe088 | 1776 | gimple *def_stmt; |
3bab6342 | 1777 | gather_decl = vect_check_gather_scatter (stmt, loop_vinfo, &gather_base, |
5ce9450f JJ |
1778 | &gather_off, &gather_scale); |
1779 | gcc_assert (gather_decl); | |
81c40241 RB |
1780 | if (!vect_is_simple_use (gather_off, loop_vinfo, &def_stmt, &gather_dt, |
1781 | &gather_off_vectype)) | |
5ce9450f JJ |
1782 | { |
1783 | if (dump_enabled_p ()) | |
1784 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1785 | "gather index use not simple."); | |
1786 | return false; | |
1787 | } | |
03b9e8e4 JJ |
1788 | |
1789 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
1790 | tree masktype | |
1791 | = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (TREE_CHAIN (arglist)))); | |
1792 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
1793 | { | |
1794 | if (dump_enabled_p ()) | |
1795 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1796 | "masked gather with integer mask not supported."); | |
1797 | return false; | |
1798 | } | |
5ce9450f JJ |
1799 | } |
1800 | else if (tree_int_cst_compare (nested_in_vect_loop | |
1801 | ? STMT_VINFO_DR_STEP (stmt_info) | |
1802 | : DR_STEP (dr), size_zero_node) <= 0) | |
1803 | return false; | |
1804 | else if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
045c1278 IE |
1805 | || !can_vec_mask_load_store_p (TYPE_MODE (vectype), |
1806 | TYPE_MODE (mask_vectype), | |
57e2f6ad IE |
1807 | !is_store) |
1808 | || (rhs_vectype | |
1809 | && !useless_type_conversion_p (vectype, rhs_vectype))) | |
5ce9450f JJ |
1810 | return false; |
1811 | ||
5ce9450f JJ |
1812 | if (!vec_stmt) /* transformation not required. */ |
1813 | { | |
1814 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
1815 | if (is_store) | |
1816 | vect_model_store_cost (stmt_info, ncopies, false, dt, | |
1817 | NULL, NULL, NULL); | |
1818 | else | |
1819 | vect_model_load_cost (stmt_info, ncopies, false, NULL, NULL, NULL); | |
1820 | return true; | |
1821 | } | |
1822 | ||
1823 | /** Transform. **/ | |
1824 | ||
3bab6342 | 1825 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5ce9450f JJ |
1826 | { |
1827 | tree vec_oprnd0 = NULL_TREE, op; | |
1828 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
1829 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
acdcd61b | 1830 | tree ptr, vec_mask = NULL_TREE, mask_op = NULL_TREE, var, scale; |
5ce9450f | 1831 | tree perm_mask = NULL_TREE, prev_res = NULL_TREE; |
acdcd61b | 1832 | tree mask_perm_mask = NULL_TREE; |
5ce9450f JJ |
1833 | edge pe = loop_preheader_edge (loop); |
1834 | gimple_seq seq; | |
1835 | basic_block new_bb; | |
1836 | enum { NARROW, NONE, WIDEN } modifier; | |
1837 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
1838 | ||
acdcd61b JJ |
1839 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); |
1840 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1841 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1842 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1843 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1844 | scaletype = TREE_VALUE (arglist); | |
1845 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
1846 | && types_compatible_p (srctype, masktype)); | |
1847 | ||
5ce9450f JJ |
1848 | if (nunits == gather_off_nunits) |
1849 | modifier = NONE; | |
1850 | else if (nunits == gather_off_nunits / 2) | |
1851 | { | |
1852 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
1853 | modifier = WIDEN; | |
1854 | ||
1855 | for (i = 0; i < gather_off_nunits; ++i) | |
1856 | sel[i] = i | nunits; | |
1857 | ||
557be5a8 | 1858 | perm_mask = vect_gen_perm_mask_checked (gather_off_vectype, sel); |
5ce9450f JJ |
1859 | } |
1860 | else if (nunits == gather_off_nunits * 2) | |
1861 | { | |
1862 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
1863 | modifier = NARROW; | |
1864 | ||
1865 | for (i = 0; i < nunits; ++i) | |
1866 | sel[i] = i < gather_off_nunits | |
1867 | ? i : i + nunits - gather_off_nunits; | |
1868 | ||
557be5a8 | 1869 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); |
5ce9450f | 1870 | ncopies *= 2; |
acdcd61b JJ |
1871 | for (i = 0; i < nunits; ++i) |
1872 | sel[i] = i | gather_off_nunits; | |
557be5a8 | 1873 | mask_perm_mask = vect_gen_perm_mask_checked (masktype, sel); |
5ce9450f JJ |
1874 | } |
1875 | else | |
1876 | gcc_unreachable (); | |
1877 | ||
5ce9450f JJ |
1878 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); |
1879 | ||
1880 | ptr = fold_convert (ptrtype, gather_base); | |
1881 | if (!is_gimple_min_invariant (ptr)) | |
1882 | { | |
1883 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
1884 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
1885 | gcc_assert (!new_bb); | |
1886 | } | |
1887 | ||
1888 | scale = build_int_cst (scaletype, gather_scale); | |
1889 | ||
1890 | prev_stmt_info = NULL; | |
1891 | for (j = 0; j < ncopies; ++j) | |
1892 | { | |
1893 | if (modifier == WIDEN && (j & 1)) | |
1894 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
1895 | perm_mask, stmt, gsi); | |
1896 | else if (j == 0) | |
1897 | op = vec_oprnd0 | |
81c40241 | 1898 | = vect_get_vec_def_for_operand (gather_off, stmt); |
5ce9450f JJ |
1899 | else |
1900 | op = vec_oprnd0 | |
1901 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
1902 | ||
1903 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
1904 | { | |
1905 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
1906 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 1907 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
5ce9450f JJ |
1908 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
1909 | new_stmt | |
0d0e4a03 | 1910 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
1911 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1912 | op = var; | |
1913 | } | |
1914 | ||
acdcd61b JJ |
1915 | if (mask_perm_mask && (j & 1)) |
1916 | mask_op = permute_vec_elements (mask_op, mask_op, | |
1917 | mask_perm_mask, stmt, gsi); | |
5ce9450f JJ |
1918 | else |
1919 | { | |
acdcd61b | 1920 | if (j == 0) |
81c40241 | 1921 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); |
acdcd61b JJ |
1922 | else |
1923 | { | |
81c40241 | 1924 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
acdcd61b JJ |
1925 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
1926 | } | |
5ce9450f | 1927 | |
acdcd61b JJ |
1928 | mask_op = vec_mask; |
1929 | if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask))) | |
1930 | { | |
1931 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)) | |
1932 | == TYPE_VECTOR_SUBPARTS (masktype)); | |
0e22bb5a | 1933 | var = vect_get_new_ssa_name (masktype, vect_simple_var); |
acdcd61b JJ |
1934 | mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op); |
1935 | new_stmt | |
0d0e4a03 | 1936 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, mask_op); |
acdcd61b JJ |
1937 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1938 | mask_op = var; | |
1939 | } | |
5ce9450f JJ |
1940 | } |
1941 | ||
1942 | new_stmt | |
1943 | = gimple_build_call (gather_decl, 5, mask_op, ptr, op, mask_op, | |
1944 | scale); | |
1945 | ||
1946 | if (!useless_type_conversion_p (vectype, rettype)) | |
1947 | { | |
1948 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
1949 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 1950 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
5ce9450f JJ |
1951 | gimple_call_set_lhs (new_stmt, op); |
1952 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 1953 | var = make_ssa_name (vec_dest); |
5ce9450f | 1954 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
0d0e4a03 | 1955 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
1956 | } |
1957 | else | |
1958 | { | |
1959 | var = make_ssa_name (vec_dest, new_stmt); | |
1960 | gimple_call_set_lhs (new_stmt, var); | |
1961 | } | |
1962 | ||
1963 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1964 | ||
1965 | if (modifier == NARROW) | |
1966 | { | |
1967 | if ((j & 1) == 0) | |
1968 | { | |
1969 | prev_res = var; | |
1970 | continue; | |
1971 | } | |
1972 | var = permute_vec_elements (prev_res, var, | |
1973 | perm_mask, stmt, gsi); | |
1974 | new_stmt = SSA_NAME_DEF_STMT (var); | |
1975 | } | |
1976 | ||
1977 | if (prev_stmt_info == NULL) | |
1978 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1979 | else | |
1980 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1981 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1982 | } | |
3efe2e2c JJ |
1983 | |
1984 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
1985 | from the IL. */ | |
e6f5c25d IE |
1986 | if (STMT_VINFO_RELATED_STMT (stmt_info)) |
1987 | { | |
1988 | stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
1989 | stmt_info = vinfo_for_stmt (stmt); | |
1990 | } | |
3efe2e2c JJ |
1991 | tree lhs = gimple_call_lhs (stmt); |
1992 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
1993 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
1994 | set_vinfo_for_stmt (stmt, NULL); | |
1995 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
1996 | gsi_replace (gsi, new_stmt, true); | |
5ce9450f JJ |
1997 | return true; |
1998 | } | |
1999 | else if (is_store) | |
2000 | { | |
2001 | tree vec_rhs = NULL_TREE, vec_mask = NULL_TREE; | |
2002 | prev_stmt_info = NULL; | |
2003 | for (i = 0; i < ncopies; i++) | |
2004 | { | |
2005 | unsigned align, misalign; | |
2006 | ||
2007 | if (i == 0) | |
2008 | { | |
2009 | tree rhs = gimple_call_arg (stmt, 3); | |
81c40241 RB |
2010 | vec_rhs = vect_get_vec_def_for_operand (rhs, stmt); |
2011 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); | |
5ce9450f JJ |
2012 | /* We should have catched mismatched types earlier. */ |
2013 | gcc_assert (useless_type_conversion_p (vectype, | |
2014 | TREE_TYPE (vec_rhs))); | |
2015 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, | |
2016 | NULL_TREE, &dummy, gsi, | |
2017 | &ptr_incr, false, &inv_p); | |
2018 | gcc_assert (!inv_p); | |
2019 | } | |
2020 | else | |
2021 | { | |
81c40241 | 2022 | vect_is_simple_use (vec_rhs, loop_vinfo, &def_stmt, &dt); |
5ce9450f | 2023 | vec_rhs = vect_get_vec_def_for_stmt_copy (dt, vec_rhs); |
81c40241 | 2024 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2025 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2026 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2027 | TYPE_SIZE_UNIT (vectype)); | |
2028 | } | |
2029 | ||
2030 | align = TYPE_ALIGN_UNIT (vectype); | |
2031 | if (aligned_access_p (dr)) | |
2032 | misalign = 0; | |
2033 | else if (DR_MISALIGNMENT (dr) == -1) | |
2034 | { | |
2035 | align = TYPE_ALIGN_UNIT (elem_type); | |
2036 | misalign = 0; | |
2037 | } | |
2038 | else | |
2039 | misalign = DR_MISALIGNMENT (dr); | |
2040 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2041 | misalign); | |
2042 | new_stmt | |
2043 | = gimple_build_call_internal (IFN_MASK_STORE, 4, dataref_ptr, | |
2044 | gimple_call_arg (stmt, 1), | |
2045 | vec_mask, vec_rhs); | |
2046 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2047 | if (i == 0) | |
2048 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2049 | else | |
2050 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2051 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2052 | } | |
2053 | } | |
2054 | else | |
2055 | { | |
2056 | tree vec_mask = NULL_TREE; | |
2057 | prev_stmt_info = NULL; | |
2058 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); | |
2059 | for (i = 0; i < ncopies; i++) | |
2060 | { | |
2061 | unsigned align, misalign; | |
2062 | ||
2063 | if (i == 0) | |
2064 | { | |
81c40241 | 2065 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); |
5ce9450f JJ |
2066 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, |
2067 | NULL_TREE, &dummy, gsi, | |
2068 | &ptr_incr, false, &inv_p); | |
2069 | gcc_assert (!inv_p); | |
2070 | } | |
2071 | else | |
2072 | { | |
81c40241 | 2073 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2074 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2075 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2076 | TYPE_SIZE_UNIT (vectype)); | |
2077 | } | |
2078 | ||
2079 | align = TYPE_ALIGN_UNIT (vectype); | |
2080 | if (aligned_access_p (dr)) | |
2081 | misalign = 0; | |
2082 | else if (DR_MISALIGNMENT (dr) == -1) | |
2083 | { | |
2084 | align = TYPE_ALIGN_UNIT (elem_type); | |
2085 | misalign = 0; | |
2086 | } | |
2087 | else | |
2088 | misalign = DR_MISALIGNMENT (dr); | |
2089 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2090 | misalign); | |
2091 | new_stmt | |
2092 | = gimple_build_call_internal (IFN_MASK_LOAD, 3, dataref_ptr, | |
2093 | gimple_call_arg (stmt, 1), | |
2094 | vec_mask); | |
b731b390 | 2095 | gimple_call_set_lhs (new_stmt, make_ssa_name (vec_dest)); |
5ce9450f JJ |
2096 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2097 | if (i == 0) | |
2098 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2099 | else | |
2100 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2101 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2102 | } | |
2103 | } | |
2104 | ||
3efe2e2c JJ |
2105 | if (!is_store) |
2106 | { | |
2107 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
2108 | from the IL. */ | |
e6f5c25d IE |
2109 | if (STMT_VINFO_RELATED_STMT (stmt_info)) |
2110 | { | |
2111 | stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
2112 | stmt_info = vinfo_for_stmt (stmt); | |
2113 | } | |
3efe2e2c JJ |
2114 | tree lhs = gimple_call_lhs (stmt); |
2115 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
2116 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
2117 | set_vinfo_for_stmt (stmt, NULL); | |
2118 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2119 | gsi_replace (gsi, new_stmt, true); | |
2120 | } | |
2121 | ||
5ce9450f JJ |
2122 | return true; |
2123 | } | |
2124 | ||
2125 | ||
ebfd146a IR |
2126 | /* Function vectorizable_call. |
2127 | ||
538dd0b7 | 2128 | Check if GS performs a function call that can be vectorized. |
b8698a0f | 2129 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
2130 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2131 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2132 | ||
2133 | static bool | |
355fe088 | 2134 | vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
190c2236 | 2135 | slp_tree slp_node) |
ebfd146a | 2136 | { |
538dd0b7 | 2137 | gcall *stmt; |
ebfd146a IR |
2138 | tree vec_dest; |
2139 | tree scalar_dest; | |
2140 | tree op, type; | |
2141 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
538dd0b7 | 2142 | stmt_vec_info stmt_info = vinfo_for_stmt (gs), prev_stmt_info; |
ebfd146a IR |
2143 | tree vectype_out, vectype_in; |
2144 | int nunits_in; | |
2145 | int nunits_out; | |
2146 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
190c2236 | 2147 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 2148 | vec_info *vinfo = stmt_info->vinfo; |
81c40241 | 2149 | tree fndecl, new_temp, rhs_type; |
355fe088 | 2150 | gimple *def_stmt; |
0502fb85 UB |
2151 | enum vect_def_type dt[3] |
2152 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
355fe088 | 2153 | gimple *new_stmt = NULL; |
ebfd146a | 2154 | int ncopies, j; |
6e1aa848 | 2155 | vec<tree> vargs = vNULL; |
ebfd146a IR |
2156 | enum { NARROW, NONE, WIDEN } modifier; |
2157 | size_t i, nargs; | |
9d5e7640 | 2158 | tree lhs; |
ebfd146a | 2159 | |
190c2236 | 2160 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2161 | return false; |
2162 | ||
8644a673 | 2163 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2164 | return false; |
2165 | ||
538dd0b7 DM |
2166 | /* Is GS a vectorizable call? */ |
2167 | stmt = dyn_cast <gcall *> (gs); | |
2168 | if (!stmt) | |
ebfd146a IR |
2169 | return false; |
2170 | ||
5ce9450f JJ |
2171 | if (gimple_call_internal_p (stmt) |
2172 | && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD | |
2173 | || gimple_call_internal_fn (stmt) == IFN_MASK_STORE)) | |
2174 | return vectorizable_mask_load_store (stmt, gsi, vec_stmt, | |
2175 | slp_node); | |
2176 | ||
0136f8f0 AH |
2177 | if (gimple_call_lhs (stmt) == NULL_TREE |
2178 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
ebfd146a IR |
2179 | return false; |
2180 | ||
0136f8f0 | 2181 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); |
5a2c1986 | 2182 | |
b690cc0f RG |
2183 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
2184 | ||
ebfd146a IR |
2185 | /* Process function arguments. */ |
2186 | rhs_type = NULL_TREE; | |
b690cc0f | 2187 | vectype_in = NULL_TREE; |
ebfd146a IR |
2188 | nargs = gimple_call_num_args (stmt); |
2189 | ||
1b1562a5 MM |
2190 | /* Bail out if the function has more than three arguments, we do not have |
2191 | interesting builtin functions to vectorize with more than two arguments | |
2192 | except for fma. No arguments is also not good. */ | |
2193 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
2194 | return false; |
2195 | ||
74bf76ed JJ |
2196 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
2197 | if (gimple_call_internal_p (stmt) | |
2198 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2199 | { | |
2200 | nargs = 0; | |
2201 | rhs_type = unsigned_type_node; | |
2202 | } | |
2203 | ||
ebfd146a IR |
2204 | for (i = 0; i < nargs; i++) |
2205 | { | |
b690cc0f RG |
2206 | tree opvectype; |
2207 | ||
ebfd146a IR |
2208 | op = gimple_call_arg (stmt, i); |
2209 | ||
2210 | /* We can only handle calls with arguments of the same type. */ | |
2211 | if (rhs_type | |
8533c9d8 | 2212 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 2213 | { |
73fbfcad | 2214 | if (dump_enabled_p ()) |
78c60e3d | 2215 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2216 | "argument types differ.\n"); |
ebfd146a IR |
2217 | return false; |
2218 | } | |
b690cc0f RG |
2219 | if (!rhs_type) |
2220 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 2221 | |
81c40241 | 2222 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[i], &opvectype)) |
ebfd146a | 2223 | { |
73fbfcad | 2224 | if (dump_enabled_p ()) |
78c60e3d | 2225 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2226 | "use not simple.\n"); |
ebfd146a IR |
2227 | return false; |
2228 | } | |
ebfd146a | 2229 | |
b690cc0f RG |
2230 | if (!vectype_in) |
2231 | vectype_in = opvectype; | |
2232 | else if (opvectype | |
2233 | && opvectype != vectype_in) | |
2234 | { | |
73fbfcad | 2235 | if (dump_enabled_p ()) |
78c60e3d | 2236 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2237 | "argument vector types differ.\n"); |
b690cc0f RG |
2238 | return false; |
2239 | } | |
2240 | } | |
2241 | /* If all arguments are external or constant defs use a vector type with | |
2242 | the same size as the output vector type. */ | |
ebfd146a | 2243 | if (!vectype_in) |
b690cc0f | 2244 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
2245 | if (vec_stmt) |
2246 | gcc_assert (vectype_in); | |
2247 | if (!vectype_in) | |
2248 | { | |
73fbfcad | 2249 | if (dump_enabled_p ()) |
7d8930a0 | 2250 | { |
78c60e3d SS |
2251 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2252 | "no vectype for scalar type "); | |
2253 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 2254 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
2255 | } |
2256 | ||
2257 | return false; | |
2258 | } | |
ebfd146a IR |
2259 | |
2260 | /* FORNOW */ | |
b690cc0f RG |
2261 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
2262 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
2263 | if (nunits_in == nunits_out / 2) |
2264 | modifier = NARROW; | |
2265 | else if (nunits_out == nunits_in) | |
2266 | modifier = NONE; | |
2267 | else if (nunits_out == nunits_in / 2) | |
2268 | modifier = WIDEN; | |
2269 | else | |
2270 | return false; | |
2271 | ||
70439f0d RS |
2272 | /* We only handle functions that do not read or clobber memory. */ |
2273 | if (gimple_vuse (stmt)) | |
2274 | { | |
2275 | if (dump_enabled_p ()) | |
2276 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2277 | "function reads from or writes to memory.\n"); | |
2278 | return false; | |
2279 | } | |
2280 | ||
ebfd146a IR |
2281 | /* For now, we only vectorize functions if a target specific builtin |
2282 | is available. TODO -- in some cases, it might be profitable to | |
2283 | insert the calls for pieces of the vector, in order to be able | |
2284 | to vectorize other operations in the loop. */ | |
70439f0d RS |
2285 | fndecl = NULL_TREE; |
2286 | internal_fn ifn = IFN_LAST; | |
2287 | combined_fn cfn = gimple_call_combined_fn (stmt); | |
2288 | tree callee = gimple_call_fndecl (stmt); | |
2289 | ||
2290 | /* First try using an internal function. */ | |
2291 | if (cfn != CFN_LAST) | |
2292 | ifn = vectorizable_internal_function (cfn, callee, vectype_out, | |
2293 | vectype_in); | |
2294 | ||
2295 | /* If that fails, try asking for a target-specific built-in function. */ | |
2296 | if (ifn == IFN_LAST) | |
2297 | { | |
2298 | if (cfn != CFN_LAST) | |
2299 | fndecl = targetm.vectorize.builtin_vectorized_function | |
2300 | (cfn, vectype_out, vectype_in); | |
2301 | else | |
2302 | fndecl = targetm.vectorize.builtin_md_vectorized_function | |
2303 | (callee, vectype_out, vectype_in); | |
2304 | } | |
2305 | ||
2306 | if (ifn == IFN_LAST && !fndecl) | |
ebfd146a | 2307 | { |
70439f0d | 2308 | if (cfn == CFN_GOMP_SIMD_LANE |
74bf76ed JJ |
2309 | && !slp_node |
2310 | && loop_vinfo | |
2311 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2312 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
2313 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2314 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
2315 | { | |
2316 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
2317 | { 0, 1, 2, ... vf - 1 } vector. */ | |
2318 | gcc_assert (nargs == 0); | |
2319 | } | |
2320 | else | |
2321 | { | |
2322 | if (dump_enabled_p ()) | |
2323 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 2324 | "function is not vectorizable.\n"); |
74bf76ed JJ |
2325 | return false; |
2326 | } | |
ebfd146a IR |
2327 | } |
2328 | ||
190c2236 JJ |
2329 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
2330 | ncopies = 1; | |
2331 | else if (modifier == NARROW) | |
ebfd146a IR |
2332 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; |
2333 | else | |
2334 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2335 | ||
2336 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
2337 | needs to be generated. */ | |
2338 | gcc_assert (ncopies >= 1); | |
2339 | ||
2340 | if (!vec_stmt) /* transformation not required. */ | |
2341 | { | |
2342 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 2343 | if (dump_enabled_p ()) |
e645e942 TJ |
2344 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ===" |
2345 | "\n"); | |
c3e7ee41 | 2346 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
2347 | return true; |
2348 | } | |
2349 | ||
2350 | /** Transform. **/ | |
2351 | ||
73fbfcad | 2352 | if (dump_enabled_p ()) |
e645e942 | 2353 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
2354 | |
2355 | /* Handle def. */ | |
2356 | scalar_dest = gimple_call_lhs (stmt); | |
2357 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
2358 | ||
2359 | prev_stmt_info = NULL; | |
2360 | switch (modifier) | |
2361 | { | |
2362 | case NONE: | |
2363 | for (j = 0; j < ncopies; ++j) | |
2364 | { | |
2365 | /* Build argument list for the vectorized call. */ | |
2366 | if (j == 0) | |
9771b263 | 2367 | vargs.create (nargs); |
ebfd146a | 2368 | else |
9771b263 | 2369 | vargs.truncate (0); |
ebfd146a | 2370 | |
190c2236 JJ |
2371 | if (slp_node) |
2372 | { | |
ef062b13 | 2373 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2374 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2375 | |
2376 | for (i = 0; i < nargs; i++) | |
9771b263 | 2377 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2378 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2379 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2380 | |
2381 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2382 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
2383 | { |
2384 | size_t k; | |
2385 | for (k = 0; k < nargs; k++) | |
2386 | { | |
37b5ec8f | 2387 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 2388 | vargs[k] = vec_oprndsk[i]; |
190c2236 | 2389 | } |
70439f0d RS |
2390 | if (ifn != IFN_LAST) |
2391 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2392 | else | |
2393 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
190c2236 JJ |
2394 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2395 | gimple_call_set_lhs (new_stmt, new_temp); | |
2396 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 2397 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2398 | } |
2399 | ||
2400 | for (i = 0; i < nargs; i++) | |
2401 | { | |
37b5ec8f | 2402 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2403 | vec_oprndsi.release (); |
190c2236 | 2404 | } |
190c2236 JJ |
2405 | continue; |
2406 | } | |
2407 | ||
ebfd146a IR |
2408 | for (i = 0; i < nargs; i++) |
2409 | { | |
2410 | op = gimple_call_arg (stmt, i); | |
2411 | if (j == 0) | |
2412 | vec_oprnd0 | |
81c40241 | 2413 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 2414 | else |
63827fb8 IR |
2415 | { |
2416 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
2417 | vec_oprnd0 | |
2418 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
2419 | } | |
ebfd146a | 2420 | |
9771b263 | 2421 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
2422 | } |
2423 | ||
74bf76ed JJ |
2424 | if (gimple_call_internal_p (stmt) |
2425 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2426 | { | |
2427 | tree *v = XALLOCAVEC (tree, nunits_out); | |
2428 | int k; | |
2429 | for (k = 0; k < nunits_out; ++k) | |
2430 | v[k] = build_int_cst (unsigned_type_node, j * nunits_out + k); | |
2431 | tree cst = build_vector (vectype_out, v); | |
2432 | tree new_var | |
0e22bb5a | 2433 | = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_"); |
355fe088 | 2434 | gimple *init_stmt = gimple_build_assign (new_var, cst); |
74bf76ed | 2435 | vect_init_vector_1 (stmt, init_stmt, NULL); |
b731b390 | 2436 | new_temp = make_ssa_name (vec_dest); |
0e22bb5a | 2437 | new_stmt = gimple_build_assign (new_temp, new_var); |
74bf76ed JJ |
2438 | } |
2439 | else | |
2440 | { | |
70439f0d RS |
2441 | if (ifn != IFN_LAST) |
2442 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2443 | else | |
2444 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
74bf76ed JJ |
2445 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2446 | gimple_call_set_lhs (new_stmt, new_temp); | |
2447 | } | |
ebfd146a IR |
2448 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2449 | ||
2450 | if (j == 0) | |
2451 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2452 | else | |
2453 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2454 | ||
2455 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2456 | } | |
2457 | ||
2458 | break; | |
2459 | ||
2460 | case NARROW: | |
2461 | for (j = 0; j < ncopies; ++j) | |
2462 | { | |
2463 | /* Build argument list for the vectorized call. */ | |
2464 | if (j == 0) | |
9771b263 | 2465 | vargs.create (nargs * 2); |
ebfd146a | 2466 | else |
9771b263 | 2467 | vargs.truncate (0); |
ebfd146a | 2468 | |
190c2236 JJ |
2469 | if (slp_node) |
2470 | { | |
ef062b13 | 2471 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2472 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2473 | |
2474 | for (i = 0; i < nargs; i++) | |
9771b263 | 2475 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2476 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2477 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2478 | |
2479 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2480 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
2481 | { |
2482 | size_t k; | |
9771b263 | 2483 | vargs.truncate (0); |
190c2236 JJ |
2484 | for (k = 0; k < nargs; k++) |
2485 | { | |
37b5ec8f | 2486 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
2487 | vargs.quick_push (vec_oprndsk[i]); |
2488 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 | 2489 | } |
70439f0d RS |
2490 | if (ifn != IFN_LAST) |
2491 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2492 | else | |
2493 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
190c2236 JJ |
2494 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2495 | gimple_call_set_lhs (new_stmt, new_temp); | |
2496 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 2497 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2498 | } |
2499 | ||
2500 | for (i = 0; i < nargs; i++) | |
2501 | { | |
37b5ec8f | 2502 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2503 | vec_oprndsi.release (); |
190c2236 | 2504 | } |
190c2236 JJ |
2505 | continue; |
2506 | } | |
2507 | ||
ebfd146a IR |
2508 | for (i = 0; i < nargs; i++) |
2509 | { | |
2510 | op = gimple_call_arg (stmt, i); | |
2511 | if (j == 0) | |
2512 | { | |
2513 | vec_oprnd0 | |
81c40241 | 2514 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 2515 | vec_oprnd1 |
63827fb8 | 2516 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2517 | } |
2518 | else | |
2519 | { | |
336ecb65 | 2520 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 2521 | vec_oprnd0 |
63827fb8 | 2522 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 2523 | vec_oprnd1 |
63827fb8 | 2524 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2525 | } |
2526 | ||
9771b263 DN |
2527 | vargs.quick_push (vec_oprnd0); |
2528 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
2529 | } |
2530 | ||
70439f0d RS |
2531 | if (ifn != IFN_LAST) |
2532 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2533 | else | |
2534 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
ebfd146a IR |
2535 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2536 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
2537 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2538 | ||
2539 | if (j == 0) | |
2540 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2541 | else | |
2542 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2543 | ||
2544 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2545 | } | |
2546 | ||
2547 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
2548 | ||
2549 | break; | |
2550 | ||
2551 | case WIDEN: | |
2552 | /* No current target implements this case. */ | |
2553 | return false; | |
2554 | } | |
2555 | ||
9771b263 | 2556 | vargs.release (); |
ebfd146a | 2557 | |
ebfd146a IR |
2558 | /* The call in STMT might prevent it from being removed in dce. |
2559 | We however cannot remove it here, due to the way the ssa name | |
2560 | it defines is mapped to the new definition. So just replace | |
2561 | rhs of the statement with something harmless. */ | |
2562 | ||
dd34c087 JJ |
2563 | if (slp_node) |
2564 | return true; | |
2565 | ||
ebfd146a | 2566 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
2567 | if (is_pattern_stmt_p (stmt_info)) |
2568 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
2569 | else | |
2570 | lhs = gimple_call_lhs (stmt); | |
3cc2fa2a JJ |
2571 | |
2572 | if (gimple_call_internal_p (stmt) | |
2573 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2574 | { | |
2575 | /* Replace uses of the lhs of GOMP_SIMD_LANE call outside the loop | |
2576 | with vf - 1 rather than 0, that is the last iteration of the | |
2577 | vectorized loop. */ | |
2578 | imm_use_iterator iter; | |
2579 | use_operand_p use_p; | |
355fe088 | 2580 | gimple *use_stmt; |
3cc2fa2a JJ |
2581 | FOR_EACH_IMM_USE_STMT (use_stmt, iter, lhs) |
2582 | { | |
2583 | basic_block use_bb = gimple_bb (use_stmt); | |
2584 | if (use_bb | |
2585 | && !flow_bb_inside_loop_p (LOOP_VINFO_LOOP (loop_vinfo), use_bb)) | |
2586 | { | |
2587 | FOR_EACH_IMM_USE_ON_STMT (use_p, iter) | |
2588 | SET_USE (use_p, build_int_cst (TREE_TYPE (lhs), | |
2589 | ncopies * nunits_out - 1)); | |
2590 | update_stmt (use_stmt); | |
2591 | } | |
2592 | } | |
2593 | } | |
2594 | ||
9d5e7640 | 2595 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); |
ebfd146a | 2596 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 2597 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
2598 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
2599 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
2600 | |
2601 | return true; | |
2602 | } | |
2603 | ||
2604 | ||
0136f8f0 AH |
2605 | struct simd_call_arg_info |
2606 | { | |
2607 | tree vectype; | |
2608 | tree op; | |
2609 | enum vect_def_type dt; | |
2610 | HOST_WIDE_INT linear_step; | |
2611 | unsigned int align; | |
17b658af | 2612 | bool simd_lane_linear; |
0136f8f0 AH |
2613 | }; |
2614 | ||
17b658af JJ |
2615 | /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME, |
2616 | is linear within simd lane (but not within whole loop), note it in | |
2617 | *ARGINFO. */ | |
2618 | ||
2619 | static void | |
2620 | vect_simd_lane_linear (tree op, struct loop *loop, | |
2621 | struct simd_call_arg_info *arginfo) | |
2622 | { | |
355fe088 | 2623 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
17b658af JJ |
2624 | |
2625 | if (!is_gimple_assign (def_stmt) | |
2626 | || gimple_assign_rhs_code (def_stmt) != POINTER_PLUS_EXPR | |
2627 | || !is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))) | |
2628 | return; | |
2629 | ||
2630 | tree base = gimple_assign_rhs1 (def_stmt); | |
2631 | HOST_WIDE_INT linear_step = 0; | |
2632 | tree v = gimple_assign_rhs2 (def_stmt); | |
2633 | while (TREE_CODE (v) == SSA_NAME) | |
2634 | { | |
2635 | tree t; | |
2636 | def_stmt = SSA_NAME_DEF_STMT (v); | |
2637 | if (is_gimple_assign (def_stmt)) | |
2638 | switch (gimple_assign_rhs_code (def_stmt)) | |
2639 | { | |
2640 | case PLUS_EXPR: | |
2641 | t = gimple_assign_rhs2 (def_stmt); | |
2642 | if (linear_step || TREE_CODE (t) != INTEGER_CST) | |
2643 | return; | |
2644 | base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t); | |
2645 | v = gimple_assign_rhs1 (def_stmt); | |
2646 | continue; | |
2647 | case MULT_EXPR: | |
2648 | t = gimple_assign_rhs2 (def_stmt); | |
2649 | if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t)) | |
2650 | return; | |
2651 | linear_step = tree_to_shwi (t); | |
2652 | v = gimple_assign_rhs1 (def_stmt); | |
2653 | continue; | |
2654 | CASE_CONVERT: | |
2655 | t = gimple_assign_rhs1 (def_stmt); | |
2656 | if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE | |
2657 | || (TYPE_PRECISION (TREE_TYPE (v)) | |
2658 | < TYPE_PRECISION (TREE_TYPE (t)))) | |
2659 | return; | |
2660 | if (!linear_step) | |
2661 | linear_step = 1; | |
2662 | v = t; | |
2663 | continue; | |
2664 | default: | |
2665 | return; | |
2666 | } | |
2667 | else if (is_gimple_call (def_stmt) | |
2668 | && gimple_call_internal_p (def_stmt) | |
2669 | && gimple_call_internal_fn (def_stmt) == IFN_GOMP_SIMD_LANE | |
2670 | && loop->simduid | |
2671 | && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME | |
2672 | && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0)) | |
2673 | == loop->simduid)) | |
2674 | { | |
2675 | if (!linear_step) | |
2676 | linear_step = 1; | |
2677 | arginfo->linear_step = linear_step; | |
2678 | arginfo->op = base; | |
2679 | arginfo->simd_lane_linear = true; | |
2680 | return; | |
2681 | } | |
2682 | } | |
2683 | } | |
2684 | ||
0136f8f0 AH |
2685 | /* Function vectorizable_simd_clone_call. |
2686 | ||
2687 | Check if STMT performs a function call that can be vectorized | |
2688 | by calling a simd clone of the function. | |
2689 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2690 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2691 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2692 | ||
2693 | static bool | |
355fe088 TS |
2694 | vectorizable_simd_clone_call (gimple *stmt, gimple_stmt_iterator *gsi, |
2695 | gimple **vec_stmt, slp_tree slp_node) | |
0136f8f0 AH |
2696 | { |
2697 | tree vec_dest; | |
2698 | tree scalar_dest; | |
2699 | tree op, type; | |
2700 | tree vec_oprnd0 = NULL_TREE; | |
2701 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
2702 | tree vectype; | |
2703 | unsigned int nunits; | |
2704 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2705 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
310213d4 | 2706 | vec_info *vinfo = stmt_info->vinfo; |
0136f8f0 | 2707 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
81c40241 | 2708 | tree fndecl, new_temp; |
355fe088 TS |
2709 | gimple *def_stmt; |
2710 | gimple *new_stmt = NULL; | |
0136f8f0 AH |
2711 | int ncopies, j; |
2712 | vec<simd_call_arg_info> arginfo = vNULL; | |
2713 | vec<tree> vargs = vNULL; | |
2714 | size_t i, nargs; | |
2715 | tree lhs, rtype, ratype; | |
2716 | vec<constructor_elt, va_gc> *ret_ctor_elts; | |
2717 | ||
2718 | /* Is STMT a vectorizable call? */ | |
2719 | if (!is_gimple_call (stmt)) | |
2720 | return false; | |
2721 | ||
2722 | fndecl = gimple_call_fndecl (stmt); | |
2723 | if (fndecl == NULL_TREE) | |
2724 | return false; | |
2725 | ||
d52f5295 | 2726 | struct cgraph_node *node = cgraph_node::get (fndecl); |
0136f8f0 AH |
2727 | if (node == NULL || node->simd_clones == NULL) |
2728 | return false; | |
2729 | ||
2730 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
2731 | return false; | |
2732 | ||
2733 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
2734 | return false; | |
2735 | ||
2736 | if (gimple_call_lhs (stmt) | |
2737 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
2738 | return false; | |
2739 | ||
2740 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); | |
2741 | ||
2742 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2743 | ||
2744 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt)) | |
2745 | return false; | |
2746 | ||
2747 | /* FORNOW */ | |
2748 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
2749 | return false; | |
2750 | ||
2751 | /* Process function arguments. */ | |
2752 | nargs = gimple_call_num_args (stmt); | |
2753 | ||
2754 | /* Bail out if the function has zero arguments. */ | |
2755 | if (nargs == 0) | |
2756 | return false; | |
2757 | ||
2758 | arginfo.create (nargs); | |
2759 | ||
2760 | for (i = 0; i < nargs; i++) | |
2761 | { | |
2762 | simd_call_arg_info thisarginfo; | |
2763 | affine_iv iv; | |
2764 | ||
2765 | thisarginfo.linear_step = 0; | |
2766 | thisarginfo.align = 0; | |
2767 | thisarginfo.op = NULL_TREE; | |
17b658af | 2768 | thisarginfo.simd_lane_linear = false; |
0136f8f0 AH |
2769 | |
2770 | op = gimple_call_arg (stmt, i); | |
81c40241 RB |
2771 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &thisarginfo.dt, |
2772 | &thisarginfo.vectype) | |
0136f8f0 AH |
2773 | || thisarginfo.dt == vect_uninitialized_def) |
2774 | { | |
2775 | if (dump_enabled_p ()) | |
2776 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2777 | "use not simple.\n"); | |
2778 | arginfo.release (); | |
2779 | return false; | |
2780 | } | |
2781 | ||
2782 | if (thisarginfo.dt == vect_constant_def | |
2783 | || thisarginfo.dt == vect_external_def) | |
2784 | gcc_assert (thisarginfo.vectype == NULL_TREE); | |
2785 | else | |
2786 | gcc_assert (thisarginfo.vectype != NULL_TREE); | |
2787 | ||
6c9e85fb JJ |
2788 | /* For linear arguments, the analyze phase should have saved |
2789 | the base and step in STMT_VINFO_SIMD_CLONE_INFO. */ | |
17b658af JJ |
2790 | if (i * 3 + 4 <= STMT_VINFO_SIMD_CLONE_INFO (stmt_info).length () |
2791 | && STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]) | |
6c9e85fb JJ |
2792 | { |
2793 | gcc_assert (vec_stmt); | |
2794 | thisarginfo.linear_step | |
17b658af | 2795 | = tree_to_shwi (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]); |
6c9e85fb | 2796 | thisarginfo.op |
17b658af JJ |
2797 | = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 1]; |
2798 | thisarginfo.simd_lane_linear | |
2799 | = (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 3] | |
2800 | == boolean_true_node); | |
6c9e85fb JJ |
2801 | /* If loop has been peeled for alignment, we need to adjust it. */ |
2802 | tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo); | |
2803 | tree n2 = LOOP_VINFO_NITERS (loop_vinfo); | |
17b658af | 2804 | if (n1 != n2 && !thisarginfo.simd_lane_linear) |
6c9e85fb JJ |
2805 | { |
2806 | tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2); | |
17b658af | 2807 | tree step = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]; |
6c9e85fb JJ |
2808 | tree opt = TREE_TYPE (thisarginfo.op); |
2809 | bias = fold_convert (TREE_TYPE (step), bias); | |
2810 | bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step); | |
2811 | thisarginfo.op | |
2812 | = fold_build2 (POINTER_TYPE_P (opt) | |
2813 | ? POINTER_PLUS_EXPR : PLUS_EXPR, opt, | |
2814 | thisarginfo.op, bias); | |
2815 | } | |
2816 | } | |
2817 | else if (!vec_stmt | |
2818 | && thisarginfo.dt != vect_constant_def | |
2819 | && thisarginfo.dt != vect_external_def | |
2820 | && loop_vinfo | |
2821 | && TREE_CODE (op) == SSA_NAME | |
2822 | && simple_iv (loop, loop_containing_stmt (stmt), op, | |
2823 | &iv, false) | |
2824 | && tree_fits_shwi_p (iv.step)) | |
0136f8f0 AH |
2825 | { |
2826 | thisarginfo.linear_step = tree_to_shwi (iv.step); | |
2827 | thisarginfo.op = iv.base; | |
2828 | } | |
2829 | else if ((thisarginfo.dt == vect_constant_def | |
2830 | || thisarginfo.dt == vect_external_def) | |
2831 | && POINTER_TYPE_P (TREE_TYPE (op))) | |
2832 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; | |
17b658af JJ |
2833 | /* Addresses of array elements indexed by GOMP_SIMD_LANE are |
2834 | linear too. */ | |
2835 | if (POINTER_TYPE_P (TREE_TYPE (op)) | |
2836 | && !thisarginfo.linear_step | |
2837 | && !vec_stmt | |
2838 | && thisarginfo.dt != vect_constant_def | |
2839 | && thisarginfo.dt != vect_external_def | |
2840 | && loop_vinfo | |
2841 | && !slp_node | |
2842 | && TREE_CODE (op) == SSA_NAME) | |
2843 | vect_simd_lane_linear (op, loop, &thisarginfo); | |
0136f8f0 AH |
2844 | |
2845 | arginfo.quick_push (thisarginfo); | |
2846 | } | |
2847 | ||
2848 | unsigned int badness = 0; | |
2849 | struct cgraph_node *bestn = NULL; | |
6c9e85fb JJ |
2850 | if (STMT_VINFO_SIMD_CLONE_INFO (stmt_info).exists ()) |
2851 | bestn = cgraph_node::get (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[0]); | |
0136f8f0 AH |
2852 | else |
2853 | for (struct cgraph_node *n = node->simd_clones; n != NULL; | |
2854 | n = n->simdclone->next_clone) | |
2855 | { | |
2856 | unsigned int this_badness = 0; | |
2857 | if (n->simdclone->simdlen | |
2858 | > (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
2859 | || n->simdclone->nargs != nargs) | |
2860 | continue; | |
2861 | if (n->simdclone->simdlen | |
2862 | < (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
2863 | this_badness += (exact_log2 (LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
2864 | - exact_log2 (n->simdclone->simdlen)) * 1024; | |
2865 | if (n->simdclone->inbranch) | |
2866 | this_badness += 2048; | |
2867 | int target_badness = targetm.simd_clone.usable (n); | |
2868 | if (target_badness < 0) | |
2869 | continue; | |
2870 | this_badness += target_badness * 512; | |
2871 | /* FORNOW: Have to add code to add the mask argument. */ | |
2872 | if (n->simdclone->inbranch) | |
2873 | continue; | |
2874 | for (i = 0; i < nargs; i++) | |
2875 | { | |
2876 | switch (n->simdclone->args[i].arg_type) | |
2877 | { | |
2878 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
2879 | if (!useless_type_conversion_p | |
2880 | (n->simdclone->args[i].orig_type, | |
2881 | TREE_TYPE (gimple_call_arg (stmt, i)))) | |
2882 | i = -1; | |
2883 | else if (arginfo[i].dt == vect_constant_def | |
2884 | || arginfo[i].dt == vect_external_def | |
2885 | || arginfo[i].linear_step) | |
2886 | this_badness += 64; | |
2887 | break; | |
2888 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
2889 | if (arginfo[i].dt != vect_constant_def | |
2890 | && arginfo[i].dt != vect_external_def) | |
2891 | i = -1; | |
2892 | break; | |
2893 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
d9a6bd32 | 2894 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
2895 | if (arginfo[i].dt == vect_constant_def |
2896 | || arginfo[i].dt == vect_external_def | |
2897 | || (arginfo[i].linear_step | |
2898 | != n->simdclone->args[i].linear_step)) | |
2899 | i = -1; | |
2900 | break; | |
2901 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
d9a6bd32 JJ |
2902 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
2903 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
e01d41e5 JJ |
2904 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
2905 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
2906 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
2907 | /* FORNOW */ |
2908 | i = -1; | |
2909 | break; | |
2910 | case SIMD_CLONE_ARG_TYPE_MASK: | |
2911 | gcc_unreachable (); | |
2912 | } | |
2913 | if (i == (size_t) -1) | |
2914 | break; | |
2915 | if (n->simdclone->args[i].alignment > arginfo[i].align) | |
2916 | { | |
2917 | i = -1; | |
2918 | break; | |
2919 | } | |
2920 | if (arginfo[i].align) | |
2921 | this_badness += (exact_log2 (arginfo[i].align) | |
2922 | - exact_log2 (n->simdclone->args[i].alignment)); | |
2923 | } | |
2924 | if (i == (size_t) -1) | |
2925 | continue; | |
2926 | if (bestn == NULL || this_badness < badness) | |
2927 | { | |
2928 | bestn = n; | |
2929 | badness = this_badness; | |
2930 | } | |
2931 | } | |
2932 | ||
2933 | if (bestn == NULL) | |
2934 | { | |
2935 | arginfo.release (); | |
2936 | return false; | |
2937 | } | |
2938 | ||
2939 | for (i = 0; i < nargs; i++) | |
2940 | if ((arginfo[i].dt == vect_constant_def | |
2941 | || arginfo[i].dt == vect_external_def) | |
2942 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
2943 | { | |
2944 | arginfo[i].vectype | |
2945 | = get_vectype_for_scalar_type (TREE_TYPE (gimple_call_arg (stmt, | |
2946 | i))); | |
2947 | if (arginfo[i].vectype == NULL | |
2948 | || (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
2949 | > bestn->simdclone->simdlen)) | |
2950 | { | |
2951 | arginfo.release (); | |
2952 | return false; | |
2953 | } | |
2954 | } | |
2955 | ||
2956 | fndecl = bestn->decl; | |
2957 | nunits = bestn->simdclone->simdlen; | |
2958 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
2959 | ||
2960 | /* If the function isn't const, only allow it in simd loops where user | |
2961 | has asserted that at least nunits consecutive iterations can be | |
2962 | performed using SIMD instructions. */ | |
2963 | if ((loop == NULL || (unsigned) loop->safelen < nunits) | |
2964 | && gimple_vuse (stmt)) | |
2965 | { | |
2966 | arginfo.release (); | |
2967 | return false; | |
2968 | } | |
2969 | ||
2970 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
2971 | needs to be generated. */ | |
2972 | gcc_assert (ncopies >= 1); | |
2973 | ||
2974 | if (!vec_stmt) /* transformation not required. */ | |
2975 | { | |
6c9e85fb JJ |
2976 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (bestn->decl); |
2977 | for (i = 0; i < nargs; i++) | |
2978 | if (bestn->simdclone->args[i].arg_type | |
2979 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
2980 | { | |
17b658af | 2981 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_grow_cleared (i * 3 |
6c9e85fb JJ |
2982 | + 1); |
2983 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (arginfo[i].op); | |
2984 | tree lst = POINTER_TYPE_P (TREE_TYPE (arginfo[i].op)) | |
2985 | ? size_type_node : TREE_TYPE (arginfo[i].op); | |
2986 | tree ls = build_int_cst (lst, arginfo[i].linear_step); | |
2987 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (ls); | |
17b658af JJ |
2988 | tree sll = arginfo[i].simd_lane_linear |
2989 | ? boolean_true_node : boolean_false_node; | |
2990 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (sll); | |
6c9e85fb | 2991 | } |
0136f8f0 AH |
2992 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; |
2993 | if (dump_enabled_p ()) | |
2994 | dump_printf_loc (MSG_NOTE, vect_location, | |
2995 | "=== vectorizable_simd_clone_call ===\n"); | |
2996 | /* vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); */ | |
2997 | arginfo.release (); | |
2998 | return true; | |
2999 | } | |
3000 | ||
3001 | /** Transform. **/ | |
3002 | ||
3003 | if (dump_enabled_p ()) | |
3004 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); | |
3005 | ||
3006 | /* Handle def. */ | |
3007 | scalar_dest = gimple_call_lhs (stmt); | |
3008 | vec_dest = NULL_TREE; | |
3009 | rtype = NULL_TREE; | |
3010 | ratype = NULL_TREE; | |
3011 | if (scalar_dest) | |
3012 | { | |
3013 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3014 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); | |
3015 | if (TREE_CODE (rtype) == ARRAY_TYPE) | |
3016 | { | |
3017 | ratype = rtype; | |
3018 | rtype = TREE_TYPE (ratype); | |
3019 | } | |
3020 | } | |
3021 | ||
3022 | prev_stmt_info = NULL; | |
3023 | for (j = 0; j < ncopies; ++j) | |
3024 | { | |
3025 | /* Build argument list for the vectorized call. */ | |
3026 | if (j == 0) | |
3027 | vargs.create (nargs); | |
3028 | else | |
3029 | vargs.truncate (0); | |
3030 | ||
3031 | for (i = 0; i < nargs; i++) | |
3032 | { | |
3033 | unsigned int k, l, m, o; | |
3034 | tree atype; | |
3035 | op = gimple_call_arg (stmt, i); | |
3036 | switch (bestn->simdclone->args[i].arg_type) | |
3037 | { | |
3038 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3039 | atype = bestn->simdclone->args[i].vector_type; | |
3040 | o = nunits / TYPE_VECTOR_SUBPARTS (atype); | |
3041 | for (m = j * o; m < (j + 1) * o; m++) | |
3042 | { | |
3043 | if (TYPE_VECTOR_SUBPARTS (atype) | |
3044 | < TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)) | |
3045 | { | |
3046 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); | |
3047 | k = (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
3048 | / TYPE_VECTOR_SUBPARTS (atype)); | |
3049 | gcc_assert ((k & (k - 1)) == 0); | |
3050 | if (m == 0) | |
3051 | vec_oprnd0 | |
81c40241 | 3052 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3053 | else |
3054 | { | |
3055 | vec_oprnd0 = arginfo[i].op; | |
3056 | if ((m & (k - 1)) == 0) | |
3057 | vec_oprnd0 | |
3058 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3059 | vec_oprnd0); | |
3060 | } | |
3061 | arginfo[i].op = vec_oprnd0; | |
3062 | vec_oprnd0 | |
3063 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, | |
3064 | size_int (prec), | |
3065 | bitsize_int ((m & (k - 1)) * prec)); | |
3066 | new_stmt | |
b731b390 | 3067 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3068 | vec_oprnd0); |
3069 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3070 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3071 | } | |
3072 | else | |
3073 | { | |
3074 | k = (TYPE_VECTOR_SUBPARTS (atype) | |
3075 | / TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)); | |
3076 | gcc_assert ((k & (k - 1)) == 0); | |
3077 | vec<constructor_elt, va_gc> *ctor_elts; | |
3078 | if (k != 1) | |
3079 | vec_alloc (ctor_elts, k); | |
3080 | else | |
3081 | ctor_elts = NULL; | |
3082 | for (l = 0; l < k; l++) | |
3083 | { | |
3084 | if (m == 0 && l == 0) | |
3085 | vec_oprnd0 | |
81c40241 | 3086 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3087 | else |
3088 | vec_oprnd0 | |
3089 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3090 | arginfo[i].op); | |
3091 | arginfo[i].op = vec_oprnd0; | |
3092 | if (k == 1) | |
3093 | break; | |
3094 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, | |
3095 | vec_oprnd0); | |
3096 | } | |
3097 | if (k == 1) | |
3098 | vargs.safe_push (vec_oprnd0); | |
3099 | else | |
3100 | { | |
3101 | vec_oprnd0 = build_constructor (atype, ctor_elts); | |
3102 | new_stmt | |
b731b390 | 3103 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3104 | vec_oprnd0); |
3105 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3106 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3107 | } | |
3108 | } | |
3109 | } | |
3110 | break; | |
3111 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3112 | vargs.safe_push (op); | |
3113 | break; | |
3114 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
3115 | if (j == 0) | |
3116 | { | |
3117 | gimple_seq stmts; | |
3118 | arginfo[i].op | |
3119 | = force_gimple_operand (arginfo[i].op, &stmts, true, | |
3120 | NULL_TREE); | |
3121 | if (stmts != NULL) | |
3122 | { | |
3123 | basic_block new_bb; | |
3124 | edge pe = loop_preheader_edge (loop); | |
3125 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
3126 | gcc_assert (!new_bb); | |
3127 | } | |
17b658af JJ |
3128 | if (arginfo[i].simd_lane_linear) |
3129 | { | |
3130 | vargs.safe_push (arginfo[i].op); | |
3131 | break; | |
3132 | } | |
b731b390 | 3133 | tree phi_res = copy_ssa_name (op); |
538dd0b7 | 3134 | gphi *new_phi = create_phi_node (phi_res, loop->header); |
0136f8f0 | 3135 | set_vinfo_for_stmt (new_phi, |
310213d4 | 3136 | new_stmt_vec_info (new_phi, loop_vinfo)); |
0136f8f0 AH |
3137 | add_phi_arg (new_phi, arginfo[i].op, |
3138 | loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
3139 | enum tree_code code | |
3140 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3141 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3142 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3143 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3144 | widest_int cst |
3145 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3146 | ncopies * nunits); | |
3147 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3148 | tree phi_arg = copy_ssa_name (op); |
0d0e4a03 JJ |
3149 | new_stmt |
3150 | = gimple_build_assign (phi_arg, code, phi_res, tcst); | |
0136f8f0 AH |
3151 | gimple_stmt_iterator si = gsi_after_labels (loop->header); |
3152 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); | |
3153 | set_vinfo_for_stmt (new_stmt, | |
310213d4 | 3154 | new_stmt_vec_info (new_stmt, loop_vinfo)); |
0136f8f0 AH |
3155 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), |
3156 | UNKNOWN_LOCATION); | |
3157 | arginfo[i].op = phi_res; | |
3158 | vargs.safe_push (phi_res); | |
3159 | } | |
3160 | else | |
3161 | { | |
3162 | enum tree_code code | |
3163 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3164 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3165 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3166 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3167 | widest_int cst |
3168 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3169 | j * nunits); | |
3170 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3171 | new_temp = make_ssa_name (TREE_TYPE (op)); |
0d0e4a03 JJ |
3172 | new_stmt = gimple_build_assign (new_temp, code, |
3173 | arginfo[i].op, tcst); | |
0136f8f0 AH |
3174 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3175 | vargs.safe_push (new_temp); | |
3176 | } | |
3177 | break; | |
3178 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
e01d41e5 JJ |
3179 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3180 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3181 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3182 | default: |
3183 | gcc_unreachable (); | |
3184 | } | |
3185 | } | |
3186 | ||
3187 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
3188 | if (vec_dest) | |
3189 | { | |
3190 | gcc_assert (ratype || TYPE_VECTOR_SUBPARTS (rtype) == nunits); | |
3191 | if (ratype) | |
b731b390 | 3192 | new_temp = create_tmp_var (ratype); |
0136f8f0 AH |
3193 | else if (TYPE_VECTOR_SUBPARTS (vectype) |
3194 | == TYPE_VECTOR_SUBPARTS (rtype)) | |
3195 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3196 | else | |
3197 | new_temp = make_ssa_name (rtype, new_stmt); | |
3198 | gimple_call_set_lhs (new_stmt, new_temp); | |
3199 | } | |
3200 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3201 | ||
3202 | if (vec_dest) | |
3203 | { | |
3204 | if (TYPE_VECTOR_SUBPARTS (vectype) < nunits) | |
3205 | { | |
3206 | unsigned int k, l; | |
3207 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); | |
3208 | k = nunits / TYPE_VECTOR_SUBPARTS (vectype); | |
3209 | gcc_assert ((k & (k - 1)) == 0); | |
3210 | for (l = 0; l < k; l++) | |
3211 | { | |
3212 | tree t; | |
3213 | if (ratype) | |
3214 | { | |
3215 | t = build_fold_addr_expr (new_temp); | |
3216 | t = build2 (MEM_REF, vectype, t, | |
3217 | build_int_cst (TREE_TYPE (t), | |
3218 | l * prec / BITS_PER_UNIT)); | |
3219 | } | |
3220 | else | |
3221 | t = build3 (BIT_FIELD_REF, vectype, new_temp, | |
3222 | size_int (prec), bitsize_int (l * prec)); | |
3223 | new_stmt | |
b731b390 | 3224 | = gimple_build_assign (make_ssa_name (vectype), t); |
0136f8f0 AH |
3225 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3226 | if (j == 0 && l == 0) | |
3227 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3228 | else | |
3229 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3230 | ||
3231 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3232 | } | |
3233 | ||
3234 | if (ratype) | |
3235 | { | |
3236 | tree clobber = build_constructor (ratype, NULL); | |
3237 | TREE_THIS_VOLATILE (clobber) = 1; | |
3238 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3239 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3240 | } | |
3241 | continue; | |
3242 | } | |
3243 | else if (TYPE_VECTOR_SUBPARTS (vectype) > nunits) | |
3244 | { | |
3245 | unsigned int k = (TYPE_VECTOR_SUBPARTS (vectype) | |
3246 | / TYPE_VECTOR_SUBPARTS (rtype)); | |
3247 | gcc_assert ((k & (k - 1)) == 0); | |
3248 | if ((j & (k - 1)) == 0) | |
3249 | vec_alloc (ret_ctor_elts, k); | |
3250 | if (ratype) | |
3251 | { | |
3252 | unsigned int m, o = nunits / TYPE_VECTOR_SUBPARTS (rtype); | |
3253 | for (m = 0; m < o; m++) | |
3254 | { | |
3255 | tree tem = build4 (ARRAY_REF, rtype, new_temp, | |
3256 | size_int (m), NULL_TREE, NULL_TREE); | |
3257 | new_stmt | |
b731b390 | 3258 | = gimple_build_assign (make_ssa_name (rtype), tem); |
0136f8f0 AH |
3259 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3260 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, | |
3261 | gimple_assign_lhs (new_stmt)); | |
3262 | } | |
3263 | tree clobber = build_constructor (ratype, NULL); | |
3264 | TREE_THIS_VOLATILE (clobber) = 1; | |
3265 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3266 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3267 | } | |
3268 | else | |
3269 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); | |
3270 | if ((j & (k - 1)) != k - 1) | |
3271 | continue; | |
3272 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); | |
3273 | new_stmt | |
b731b390 | 3274 | = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0); |
0136f8f0 AH |
3275 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3276 | ||
3277 | if ((unsigned) j == k - 1) | |
3278 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3279 | else | |
3280 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3281 | ||
3282 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3283 | continue; | |
3284 | } | |
3285 | else if (ratype) | |
3286 | { | |
3287 | tree t = build_fold_addr_expr (new_temp); | |
3288 | t = build2 (MEM_REF, vectype, t, | |
3289 | build_int_cst (TREE_TYPE (t), 0)); | |
3290 | new_stmt | |
b731b390 | 3291 | = gimple_build_assign (make_ssa_name (vec_dest), t); |
0136f8f0 AH |
3292 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3293 | tree clobber = build_constructor (ratype, NULL); | |
3294 | TREE_THIS_VOLATILE (clobber) = 1; | |
3295 | vect_finish_stmt_generation (stmt, | |
3296 | gimple_build_assign (new_temp, | |
3297 | clobber), gsi); | |
3298 | } | |
3299 | } | |
3300 | ||
3301 | if (j == 0) | |
3302 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3303 | else | |
3304 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3305 | ||
3306 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3307 | } | |
3308 | ||
3309 | vargs.release (); | |
3310 | ||
3311 | /* The call in STMT might prevent it from being removed in dce. | |
3312 | We however cannot remove it here, due to the way the ssa name | |
3313 | it defines is mapped to the new definition. So just replace | |
3314 | rhs of the statement with something harmless. */ | |
3315 | ||
3316 | if (slp_node) | |
3317 | return true; | |
3318 | ||
3319 | if (scalar_dest) | |
3320 | { | |
3321 | type = TREE_TYPE (scalar_dest); | |
3322 | if (is_pattern_stmt_p (stmt_info)) | |
3323 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
3324 | else | |
3325 | lhs = gimple_call_lhs (stmt); | |
3326 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
3327 | } | |
3328 | else | |
3329 | new_stmt = gimple_build_nop (); | |
3330 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
3331 | set_vinfo_for_stmt (stmt, NULL); | |
3332 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2865f32a | 3333 | gsi_replace (gsi, new_stmt, true); |
0136f8f0 AH |
3334 | unlink_stmt_vdef (stmt); |
3335 | ||
3336 | return true; | |
3337 | } | |
3338 | ||
3339 | ||
ebfd146a IR |
3340 | /* Function vect_gen_widened_results_half |
3341 | ||
3342 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 3343 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 3344 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
3345 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
3346 | needs to be created (DECL is a function-decl of a target-builtin). | |
3347 | STMT is the original scalar stmt that we are vectorizing. */ | |
3348 | ||
355fe088 | 3349 | static gimple * |
ebfd146a IR |
3350 | vect_gen_widened_results_half (enum tree_code code, |
3351 | tree decl, | |
3352 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
3353 | tree vec_dest, gimple_stmt_iterator *gsi, | |
355fe088 | 3354 | gimple *stmt) |
b8698a0f | 3355 | { |
355fe088 | 3356 | gimple *new_stmt; |
b8698a0f L |
3357 | tree new_temp; |
3358 | ||
3359 | /* Generate half of the widened result: */ | |
3360 | if (code == CALL_EXPR) | |
3361 | { | |
3362 | /* Target specific support */ | |
ebfd146a IR |
3363 | if (op_type == binary_op) |
3364 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
3365 | else | |
3366 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
3367 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3368 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
3369 | } |
3370 | else | |
ebfd146a | 3371 | { |
b8698a0f L |
3372 | /* Generic support */ |
3373 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
3374 | if (op_type != binary_op) |
3375 | vec_oprnd1 = NULL; | |
0d0e4a03 | 3376 | new_stmt = gimple_build_assign (vec_dest, code, vec_oprnd0, vec_oprnd1); |
ebfd146a IR |
3377 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3378 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 3379 | } |
ebfd146a IR |
3380 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3381 | ||
ebfd146a IR |
3382 | return new_stmt; |
3383 | } | |
3384 | ||
4a00c761 JJ |
3385 | |
3386 | /* Get vectorized definitions for loop-based vectorization. For the first | |
3387 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
3388 | scalar operand), and for the rest we get a copy with | |
3389 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
3390 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
3391 | The vectors are collected into VEC_OPRNDS. */ | |
3392 | ||
3393 | static void | |
355fe088 | 3394 | vect_get_loop_based_defs (tree *oprnd, gimple *stmt, enum vect_def_type dt, |
9771b263 | 3395 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
3396 | { |
3397 | tree vec_oprnd; | |
3398 | ||
3399 | /* Get first vector operand. */ | |
3400 | /* All the vector operands except the very first one (that is scalar oprnd) | |
3401 | are stmt copies. */ | |
3402 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
81c40241 | 3403 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt); |
4a00c761 JJ |
3404 | else |
3405 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
3406 | ||
9771b263 | 3407 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3408 | |
3409 | /* Get second vector operand. */ | |
3410 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 3411 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3412 | |
3413 | *oprnd = vec_oprnd; | |
3414 | ||
3415 | /* For conversion in multiple steps, continue to get operands | |
3416 | recursively. */ | |
3417 | if (multi_step_cvt) | |
3418 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
3419 | } | |
3420 | ||
3421 | ||
3422 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
3423 | For multi-step conversions store the resulting vectors and call the function | |
3424 | recursively. */ | |
3425 | ||
3426 | static void | |
9771b263 | 3427 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
355fe088 | 3428 | int multi_step_cvt, gimple *stmt, |
9771b263 | 3429 | vec<tree> vec_dsts, |
4a00c761 JJ |
3430 | gimple_stmt_iterator *gsi, |
3431 | slp_tree slp_node, enum tree_code code, | |
3432 | stmt_vec_info *prev_stmt_info) | |
3433 | { | |
3434 | unsigned int i; | |
3435 | tree vop0, vop1, new_tmp, vec_dest; | |
355fe088 | 3436 | gimple *new_stmt; |
4a00c761 JJ |
3437 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
3438 | ||
9771b263 | 3439 | vec_dest = vec_dsts.pop (); |
4a00c761 | 3440 | |
9771b263 | 3441 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
3442 | { |
3443 | /* Create demotion operation. */ | |
9771b263 DN |
3444 | vop0 = (*vec_oprnds)[i]; |
3445 | vop1 = (*vec_oprnds)[i + 1]; | |
0d0e4a03 | 3446 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
4a00c761 JJ |
3447 | new_tmp = make_ssa_name (vec_dest, new_stmt); |
3448 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
3449 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3450 | ||
3451 | if (multi_step_cvt) | |
3452 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 3453 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
3454 | else |
3455 | { | |
3456 | /* This is the last step of the conversion sequence. Store the | |
3457 | vectors in SLP_NODE or in vector info of the scalar statement | |
3458 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
3459 | if (slp_node) | |
9771b263 | 3460 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 3461 | else |
c689ce1e RB |
3462 | { |
3463 | if (!*prev_stmt_info) | |
3464 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3465 | else | |
3466 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
4a00c761 | 3467 | |
c689ce1e RB |
3468 | *prev_stmt_info = vinfo_for_stmt (new_stmt); |
3469 | } | |
4a00c761 JJ |
3470 | } |
3471 | } | |
3472 | ||
3473 | /* For multi-step demotion operations we first generate demotion operations | |
3474 | from the source type to the intermediate types, and then combine the | |
3475 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
3476 | type. */ | |
3477 | if (multi_step_cvt) | |
3478 | { | |
3479 | /* At each level of recursion we have half of the operands we had at the | |
3480 | previous level. */ | |
9771b263 | 3481 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
3482 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
3483 | stmt, vec_dsts, gsi, slp_node, | |
3484 | VEC_PACK_TRUNC_EXPR, | |
3485 | prev_stmt_info); | |
3486 | } | |
3487 | ||
9771b263 | 3488 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3489 | } |
3490 | ||
3491 | ||
3492 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
3493 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
3494 | the resulting vectors and call the function recursively. */ | |
3495 | ||
3496 | static void | |
9771b263 DN |
3497 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
3498 | vec<tree> *vec_oprnds1, | |
355fe088 | 3499 | gimple *stmt, tree vec_dest, |
4a00c761 JJ |
3500 | gimple_stmt_iterator *gsi, |
3501 | enum tree_code code1, | |
3502 | enum tree_code code2, tree decl1, | |
3503 | tree decl2, int op_type) | |
3504 | { | |
3505 | int i; | |
3506 | tree vop0, vop1, new_tmp1, new_tmp2; | |
355fe088 | 3507 | gimple *new_stmt1, *new_stmt2; |
6e1aa848 | 3508 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 3509 | |
9771b263 DN |
3510 | vec_tmp.create (vec_oprnds0->length () * 2); |
3511 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
3512 | { |
3513 | if (op_type == binary_op) | |
9771b263 | 3514 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
3515 | else |
3516 | vop1 = NULL_TREE; | |
3517 | ||
3518 | /* Generate the two halves of promotion operation. */ | |
3519 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
3520 | op_type, vec_dest, gsi, stmt); | |
3521 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
3522 | op_type, vec_dest, gsi, stmt); | |
3523 | if (is_gimple_call (new_stmt1)) | |
3524 | { | |
3525 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
3526 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
3527 | } | |
3528 | else | |
3529 | { | |
3530 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
3531 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
3532 | } | |
3533 | ||
3534 | /* Store the results for the next step. */ | |
9771b263 DN |
3535 | vec_tmp.quick_push (new_tmp1); |
3536 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
3537 | } |
3538 | ||
689eaba3 | 3539 | vec_oprnds0->release (); |
4a00c761 JJ |
3540 | *vec_oprnds0 = vec_tmp; |
3541 | } | |
3542 | ||
3543 | ||
b8698a0f L |
3544 | /* Check if STMT performs a conversion operation, that can be vectorized. |
3545 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 3546 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
3547 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
3548 | ||
3549 | static bool | |
355fe088 TS |
3550 | vectorizable_conversion (gimple *stmt, gimple_stmt_iterator *gsi, |
3551 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
3552 | { |
3553 | tree vec_dest; | |
3554 | tree scalar_dest; | |
4a00c761 | 3555 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
3556 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
3557 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3558 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3559 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 3560 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
3561 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
3562 | tree new_temp; | |
355fe088 | 3563 | gimple *def_stmt; |
ebfd146a | 3564 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
355fe088 | 3565 | gimple *new_stmt = NULL; |
ebfd146a IR |
3566 | stmt_vec_info prev_stmt_info; |
3567 | int nunits_in; | |
3568 | int nunits_out; | |
3569 | tree vectype_out, vectype_in; | |
4a00c761 JJ |
3570 | int ncopies, i, j; |
3571 | tree lhs_type, rhs_type; | |
ebfd146a | 3572 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
3573 | vec<tree> vec_oprnds0 = vNULL; |
3574 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 3575 | tree vop0; |
4a00c761 | 3576 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 3577 | vec_info *vinfo = stmt_info->vinfo; |
4a00c761 | 3578 | int multi_step_cvt = 0; |
6e1aa848 DN |
3579 | vec<tree> vec_dsts = vNULL; |
3580 | vec<tree> interm_types = vNULL; | |
4a00c761 JJ |
3581 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
3582 | int op_type; | |
ef4bddc2 | 3583 | machine_mode rhs_mode; |
4a00c761 | 3584 | unsigned short fltsz; |
ebfd146a IR |
3585 | |
3586 | /* Is STMT a vectorizable conversion? */ | |
3587 | ||
4a00c761 | 3588 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3589 | return false; |
3590 | ||
8644a673 | 3591 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3592 | return false; |
3593 | ||
3594 | if (!is_gimple_assign (stmt)) | |
3595 | return false; | |
3596 | ||
3597 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3598 | return false; | |
3599 | ||
3600 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
3601 | if (!CONVERT_EXPR_CODE_P (code) |
3602 | && code != FIX_TRUNC_EXPR | |
3603 | && code != FLOAT_EXPR | |
3604 | && code != WIDEN_MULT_EXPR | |
3605 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
3606 | return false; |
3607 | ||
4a00c761 JJ |
3608 | op_type = TREE_CODE_LENGTH (code); |
3609 | ||
ebfd146a | 3610 | /* Check types of lhs and rhs. */ |
b690cc0f | 3611 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 3612 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
3613 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
3614 | ||
ebfd146a IR |
3615 | op0 = gimple_assign_rhs1 (stmt); |
3616 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
3617 | |
3618 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
3619 | && !((INTEGRAL_TYPE_P (lhs_type) | |
3620 | && INTEGRAL_TYPE_P (rhs_type)) | |
3621 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
3622 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
3623 | return false; | |
3624 | ||
e6f5c25d IE |
3625 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
3626 | && ((INTEGRAL_TYPE_P (lhs_type) | |
3627 | && (TYPE_PRECISION (lhs_type) | |
3628 | != GET_MODE_PRECISION (TYPE_MODE (lhs_type)))) | |
3629 | || (INTEGRAL_TYPE_P (rhs_type) | |
3630 | && (TYPE_PRECISION (rhs_type) | |
3631 | != GET_MODE_PRECISION (TYPE_MODE (rhs_type)))))) | |
4a00c761 | 3632 | { |
73fbfcad | 3633 | if (dump_enabled_p ()) |
78c60e3d | 3634 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
3635 | "type conversion to/from bit-precision unsupported." |
3636 | "\n"); | |
4a00c761 JJ |
3637 | return false; |
3638 | } | |
3639 | ||
b690cc0f | 3640 | /* Check the operands of the operation. */ |
81c40241 | 3641 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype_in)) |
b690cc0f | 3642 | { |
73fbfcad | 3643 | if (dump_enabled_p ()) |
78c60e3d | 3644 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3645 | "use not simple.\n"); |
b690cc0f RG |
3646 | return false; |
3647 | } | |
4a00c761 JJ |
3648 | if (op_type == binary_op) |
3649 | { | |
3650 | bool ok; | |
3651 | ||
3652 | op1 = gimple_assign_rhs2 (stmt); | |
3653 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
3654 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
3655 | OP1. */ | |
3656 | if (CONSTANT_CLASS_P (op0)) | |
81c40241 | 3657 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &vectype_in); |
4a00c761 | 3658 | else |
81c40241 | 3659 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1]); |
4a00c761 JJ |
3660 | |
3661 | if (!ok) | |
3662 | { | |
73fbfcad | 3663 | if (dump_enabled_p ()) |
78c60e3d | 3664 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3665 | "use not simple.\n"); |
4a00c761 JJ |
3666 | return false; |
3667 | } | |
3668 | } | |
3669 | ||
b690cc0f RG |
3670 | /* If op0 is an external or constant defs use a vector type of |
3671 | the same size as the output vector type. */ | |
ebfd146a | 3672 | if (!vectype_in) |
b690cc0f | 3673 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
3674 | if (vec_stmt) |
3675 | gcc_assert (vectype_in); | |
3676 | if (!vectype_in) | |
3677 | { | |
73fbfcad | 3678 | if (dump_enabled_p ()) |
4a00c761 | 3679 | { |
78c60e3d SS |
3680 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3681 | "no vectype for scalar type "); | |
3682 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 3683 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 3684 | } |
7d8930a0 IR |
3685 | |
3686 | return false; | |
3687 | } | |
ebfd146a | 3688 | |
e6f5c25d IE |
3689 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
3690 | && !VECTOR_BOOLEAN_TYPE_P (vectype_in)) | |
3691 | { | |
3692 | if (dump_enabled_p ()) | |
3693 | { | |
3694 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3695 | "can't convert between boolean and non " | |
3696 | "boolean vectors"); | |
3697 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
3698 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
3699 | } | |
3700 | ||
3701 | return false; | |
3702 | } | |
3703 | ||
b690cc0f RG |
3704 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
3705 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4a00c761 | 3706 | if (nunits_in < nunits_out) |
ebfd146a IR |
3707 | modifier = NARROW; |
3708 | else if (nunits_out == nunits_in) | |
3709 | modifier = NONE; | |
ebfd146a | 3710 | else |
4a00c761 | 3711 | modifier = WIDEN; |
ebfd146a | 3712 | |
ff802fa1 IR |
3713 | /* Multiple types in SLP are handled by creating the appropriate number of |
3714 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3715 | case of SLP. */ | |
437f4a00 | 3716 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a | 3717 | ncopies = 1; |
4a00c761 JJ |
3718 | else if (modifier == NARROW) |
3719 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
3720 | else | |
3721 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
b8698a0f | 3722 | |
ebfd146a IR |
3723 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
3724 | needs to be generated. */ | |
3725 | gcc_assert (ncopies >= 1); | |
3726 | ||
ebfd146a | 3727 | /* Supportable by target? */ |
4a00c761 | 3728 | switch (modifier) |
ebfd146a | 3729 | { |
4a00c761 JJ |
3730 | case NONE: |
3731 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
3732 | return false; | |
3733 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
3734 | &decl1, &code1)) | |
3735 | break; | |
3736 | /* FALLTHRU */ | |
3737 | unsupported: | |
73fbfcad | 3738 | if (dump_enabled_p ()) |
78c60e3d | 3739 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3740 | "conversion not supported by target.\n"); |
ebfd146a | 3741 | return false; |
ebfd146a | 3742 | |
4a00c761 JJ |
3743 | case WIDEN: |
3744 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
3745 | &code1, &code2, &multi_step_cvt, |
3746 | &interm_types)) | |
4a00c761 JJ |
3747 | { |
3748 | /* Binary widening operation can only be supported directly by the | |
3749 | architecture. */ | |
3750 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
3751 | break; | |
3752 | } | |
3753 | ||
3754 | if (code != FLOAT_EXPR | |
3755 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
3756 | <= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
3757 | goto unsupported; | |
3758 | ||
3759 | rhs_mode = TYPE_MODE (rhs_type); | |
3760 | fltsz = GET_MODE_SIZE (TYPE_MODE (lhs_type)); | |
3761 | for (rhs_mode = GET_MODE_2XWIDER_MODE (TYPE_MODE (rhs_type)); | |
3762 | rhs_mode != VOIDmode && GET_MODE_SIZE (rhs_mode) <= fltsz; | |
3763 | rhs_mode = GET_MODE_2XWIDER_MODE (rhs_mode)) | |
3764 | { | |
3765 | cvt_type | |
3766 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
3767 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
3768 | if (cvt_type == NULL_TREE) | |
3769 | goto unsupported; | |
3770 | ||
3771 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
3772 | { | |
3773 | if (!supportable_convert_operation (code, vectype_out, | |
3774 | cvt_type, &decl1, &codecvt1)) | |
3775 | goto unsupported; | |
3776 | } | |
3777 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
3778 | cvt_type, &codecvt1, |
3779 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
3780 | &interm_types)) |
3781 | continue; | |
3782 | else | |
3783 | gcc_assert (multi_step_cvt == 0); | |
3784 | ||
3785 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
3786 | vectype_in, &code1, &code2, |
3787 | &multi_step_cvt, &interm_types)) | |
4a00c761 JJ |
3788 | break; |
3789 | } | |
3790 | ||
3791 | if (rhs_mode == VOIDmode || GET_MODE_SIZE (rhs_mode) > fltsz) | |
3792 | goto unsupported; | |
3793 | ||
3794 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
3795 | codecvt2 = ERROR_MARK; | |
3796 | else | |
3797 | { | |
3798 | multi_step_cvt++; | |
9771b263 | 3799 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
3800 | cvt_type = NULL_TREE; |
3801 | } | |
3802 | break; | |
3803 | ||
3804 | case NARROW: | |
3805 | gcc_assert (op_type == unary_op); | |
3806 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
3807 | &code1, &multi_step_cvt, | |
3808 | &interm_types)) | |
3809 | break; | |
3810 | ||
3811 | if (code != FIX_TRUNC_EXPR | |
3812 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
3813 | >= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
3814 | goto unsupported; | |
3815 | ||
3816 | rhs_mode = TYPE_MODE (rhs_type); | |
3817 | cvt_type | |
3818 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
3819 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
3820 | if (cvt_type == NULL_TREE) | |
3821 | goto unsupported; | |
3822 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
3823 | &decl1, &codecvt1)) | |
3824 | goto unsupported; | |
3825 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
3826 | &code1, &multi_step_cvt, | |
3827 | &interm_types)) | |
3828 | break; | |
3829 | goto unsupported; | |
3830 | ||
3831 | default: | |
3832 | gcc_unreachable (); | |
ebfd146a IR |
3833 | } |
3834 | ||
3835 | if (!vec_stmt) /* transformation not required. */ | |
3836 | { | |
73fbfcad | 3837 | if (dump_enabled_p ()) |
78c60e3d | 3838 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3839 | "=== vectorizable_conversion ===\n"); |
4a00c761 | 3840 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
3841 | { |
3842 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
c3e7ee41 | 3843 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
8bd37302 | 3844 | } |
4a00c761 JJ |
3845 | else if (modifier == NARROW) |
3846 | { | |
3847 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
8bd37302 | 3848 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 JJ |
3849 | } |
3850 | else | |
3851 | { | |
3852 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
8bd37302 | 3853 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 | 3854 | } |
9771b263 | 3855 | interm_types.release (); |
ebfd146a IR |
3856 | return true; |
3857 | } | |
3858 | ||
3859 | /** Transform. **/ | |
73fbfcad | 3860 | if (dump_enabled_p ()) |
78c60e3d | 3861 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3862 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 3863 | |
4a00c761 JJ |
3864 | if (op_type == binary_op) |
3865 | { | |
3866 | if (CONSTANT_CLASS_P (op0)) | |
3867 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
3868 | else if (CONSTANT_CLASS_P (op1)) | |
3869 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
3870 | } | |
3871 | ||
3872 | /* In case of multi-step conversion, we first generate conversion operations | |
3873 | to the intermediate types, and then from that types to the final one. | |
3874 | We create vector destinations for the intermediate type (TYPES) received | |
3875 | from supportable_*_operation, and store them in the correct order | |
3876 | for future use in vect_create_vectorized_*_stmts (). */ | |
9771b263 | 3877 | vec_dsts.create (multi_step_cvt + 1); |
82294ec1 JJ |
3878 | vec_dest = vect_create_destination_var (scalar_dest, |
3879 | (cvt_type && modifier == WIDEN) | |
3880 | ? cvt_type : vectype_out); | |
9771b263 | 3881 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3882 | |
3883 | if (multi_step_cvt) | |
3884 | { | |
9771b263 DN |
3885 | for (i = interm_types.length () - 1; |
3886 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
3887 | { |
3888 | vec_dest = vect_create_destination_var (scalar_dest, | |
3889 | intermediate_type); | |
9771b263 | 3890 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3891 | } |
3892 | } | |
ebfd146a | 3893 | |
4a00c761 | 3894 | if (cvt_type) |
82294ec1 JJ |
3895 | vec_dest = vect_create_destination_var (scalar_dest, |
3896 | modifier == WIDEN | |
3897 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
3898 | |
3899 | if (!slp_node) | |
3900 | { | |
30862efc | 3901 | if (modifier == WIDEN) |
4a00c761 | 3902 | { |
c3284718 | 3903 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 3904 | if (op_type == binary_op) |
9771b263 | 3905 | vec_oprnds1.create (1); |
4a00c761 | 3906 | } |
30862efc | 3907 | else if (modifier == NARROW) |
9771b263 DN |
3908 | vec_oprnds0.create ( |
3909 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
3910 | } |
3911 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 3912 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 3913 | |
4a00c761 | 3914 | last_oprnd = op0; |
ebfd146a IR |
3915 | prev_stmt_info = NULL; |
3916 | switch (modifier) | |
3917 | { | |
3918 | case NONE: | |
3919 | for (j = 0; j < ncopies; j++) | |
3920 | { | |
ebfd146a | 3921 | if (j == 0) |
d092494c IR |
3922 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node, |
3923 | -1); | |
ebfd146a IR |
3924 | else |
3925 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
3926 | ||
9771b263 | 3927 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
3928 | { |
3929 | /* Arguments are ready, create the new vector stmt. */ | |
3930 | if (code1 == CALL_EXPR) | |
3931 | { | |
3932 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
3933 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3934 | gimple_call_set_lhs (new_stmt, new_temp); | |
3935 | } | |
3936 | else | |
3937 | { | |
3938 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
0d0e4a03 | 3939 | new_stmt = gimple_build_assign (vec_dest, code1, vop0); |
4a00c761 JJ |
3940 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3941 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3942 | } | |
3943 | ||
3944 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3945 | if (slp_node) | |
9771b263 | 3946 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
225ce44b RB |
3947 | else |
3948 | { | |
3949 | if (!prev_stmt_info) | |
3950 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3951 | else | |
3952 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3953 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3954 | } | |
4a00c761 | 3955 | } |
ebfd146a IR |
3956 | } |
3957 | break; | |
3958 | ||
3959 | case WIDEN: | |
3960 | /* In case the vectorization factor (VF) is bigger than the number | |
3961 | of elements that we can fit in a vectype (nunits), we have to | |
3962 | generate more than one vector stmt - i.e - we need to "unroll" | |
3963 | the vector stmt by a factor VF/nunits. */ | |
3964 | for (j = 0; j < ncopies; j++) | |
3965 | { | |
4a00c761 | 3966 | /* Handle uses. */ |
ebfd146a | 3967 | if (j == 0) |
4a00c761 JJ |
3968 | { |
3969 | if (slp_node) | |
3970 | { | |
3971 | if (code == WIDEN_LSHIFT_EXPR) | |
3972 | { | |
3973 | unsigned int k; | |
ebfd146a | 3974 | |
4a00c761 JJ |
3975 | vec_oprnd1 = op1; |
3976 | /* Store vec_oprnd1 for every vector stmt to be created | |
3977 | for SLP_NODE. We check during the analysis that all | |
3978 | the shift arguments are the same. */ | |
3979 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 3980 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
3981 | |
3982 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
3983 | slp_node, -1); | |
3984 | } | |
3985 | else | |
3986 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
3987 | &vec_oprnds1, slp_node, -1); | |
3988 | } | |
3989 | else | |
3990 | { | |
81c40241 | 3991 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 3992 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
3993 | if (op_type == binary_op) |
3994 | { | |
3995 | if (code == WIDEN_LSHIFT_EXPR) | |
3996 | vec_oprnd1 = op1; | |
3997 | else | |
81c40241 | 3998 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 3999 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4000 | } |
4001 | } | |
4002 | } | |
ebfd146a | 4003 | else |
4a00c761 JJ |
4004 | { |
4005 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
4006 | vec_oprnds0.truncate (0); |
4007 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
4008 | if (op_type == binary_op) |
4009 | { | |
4010 | if (code == WIDEN_LSHIFT_EXPR) | |
4011 | vec_oprnd1 = op1; | |
4012 | else | |
4013 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
4014 | vec_oprnd1); | |
9771b263 DN |
4015 | vec_oprnds1.truncate (0); |
4016 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
4017 | } |
4018 | } | |
ebfd146a | 4019 | |
4a00c761 JJ |
4020 | /* Arguments are ready. Create the new vector stmts. */ |
4021 | for (i = multi_step_cvt; i >= 0; i--) | |
4022 | { | |
9771b263 | 4023 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
4024 | enum tree_code c1 = code1, c2 = code2; |
4025 | if (i == 0 && codecvt2 != ERROR_MARK) | |
4026 | { | |
4027 | c1 = codecvt1; | |
4028 | c2 = codecvt2; | |
4029 | } | |
4030 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
4031 | &vec_oprnds1, | |
4032 | stmt, this_dest, gsi, | |
4033 | c1, c2, decl1, decl2, | |
4034 | op_type); | |
4035 | } | |
4036 | ||
9771b263 | 4037 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4038 | { |
4039 | if (cvt_type) | |
4040 | { | |
4041 | if (codecvt1 == CALL_EXPR) | |
4042 | { | |
4043 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4044 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4045 | gimple_call_set_lhs (new_stmt, new_temp); | |
4046 | } | |
4047 | else | |
4048 | { | |
4049 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4050 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4051 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4052 | vop0); | |
4a00c761 JJ |
4053 | } |
4054 | ||
4055 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4056 | } | |
4057 | else | |
4058 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
4059 | ||
4060 | if (slp_node) | |
9771b263 | 4061 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 4062 | else |
c689ce1e RB |
4063 | { |
4064 | if (!prev_stmt_info) | |
4065 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
4066 | else | |
4067 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4068 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4069 | } | |
4a00c761 | 4070 | } |
ebfd146a | 4071 | } |
4a00c761 JJ |
4072 | |
4073 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
4074 | break; |
4075 | ||
4076 | case NARROW: | |
4077 | /* In case the vectorization factor (VF) is bigger than the number | |
4078 | of elements that we can fit in a vectype (nunits), we have to | |
4079 | generate more than one vector stmt - i.e - we need to "unroll" | |
4080 | the vector stmt by a factor VF/nunits. */ | |
4081 | for (j = 0; j < ncopies; j++) | |
4082 | { | |
4083 | /* Handle uses. */ | |
4a00c761 JJ |
4084 | if (slp_node) |
4085 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
4086 | slp_node, -1); | |
ebfd146a IR |
4087 | else |
4088 | { | |
9771b263 | 4089 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
4090 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
4091 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
4092 | } |
4093 | ||
4a00c761 JJ |
4094 | /* Arguments are ready. Create the new vector stmts. */ |
4095 | if (cvt_type) | |
9771b263 | 4096 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4097 | { |
4098 | if (codecvt1 == CALL_EXPR) | |
4099 | { | |
4100 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4101 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4102 | gimple_call_set_lhs (new_stmt, new_temp); | |
4103 | } | |
4104 | else | |
4105 | { | |
4106 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4107 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4108 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4109 | vop0); | |
4a00c761 | 4110 | } |
ebfd146a | 4111 | |
4a00c761 | 4112 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 4113 | vec_oprnds0[i] = new_temp; |
4a00c761 | 4114 | } |
ebfd146a | 4115 | |
4a00c761 JJ |
4116 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
4117 | stmt, vec_dsts, gsi, | |
4118 | slp_node, code1, | |
4119 | &prev_stmt_info); | |
ebfd146a IR |
4120 | } |
4121 | ||
4122 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 4123 | break; |
ebfd146a IR |
4124 | } |
4125 | ||
9771b263 DN |
4126 | vec_oprnds0.release (); |
4127 | vec_oprnds1.release (); | |
4128 | vec_dsts.release (); | |
4129 | interm_types.release (); | |
ebfd146a IR |
4130 | |
4131 | return true; | |
4132 | } | |
ff802fa1 IR |
4133 | |
4134 | ||
ebfd146a IR |
4135 | /* Function vectorizable_assignment. |
4136 | ||
b8698a0f L |
4137 | Check if STMT performs an assignment (copy) that can be vectorized. |
4138 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4139 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4140 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4141 | ||
4142 | static bool | |
355fe088 TS |
4143 | vectorizable_assignment (gimple *stmt, gimple_stmt_iterator *gsi, |
4144 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
4145 | { |
4146 | tree vec_dest; | |
4147 | tree scalar_dest; | |
4148 | tree op; | |
4149 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a IR |
4150 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4151 | tree new_temp; | |
355fe088 | 4152 | gimple *def_stmt; |
ebfd146a | 4153 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
ebfd146a | 4154 | int ncopies; |
f18b55bd | 4155 | int i, j; |
6e1aa848 | 4156 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 4157 | tree vop; |
a70d6342 | 4158 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4159 | vec_info *vinfo = stmt_info->vinfo; |
355fe088 | 4160 | gimple *new_stmt = NULL; |
f18b55bd | 4161 | stmt_vec_info prev_stmt_info = NULL; |
fde9c428 RG |
4162 | enum tree_code code; |
4163 | tree vectype_in; | |
ebfd146a | 4164 | |
a70d6342 | 4165 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4166 | return false; |
4167 | ||
8644a673 | 4168 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
4169 | return false; |
4170 | ||
4171 | /* Is vectorizable assignment? */ | |
4172 | if (!is_gimple_assign (stmt)) | |
4173 | return false; | |
4174 | ||
4175 | scalar_dest = gimple_assign_lhs (stmt); | |
4176 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4177 | return false; | |
4178 | ||
fde9c428 | 4179 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 4180 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
4181 | || code == PAREN_EXPR |
4182 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
4183 | op = gimple_assign_rhs1 (stmt); |
4184 | else | |
4185 | return false; | |
4186 | ||
7b7ec6c5 RG |
4187 | if (code == VIEW_CONVERT_EXPR) |
4188 | op = TREE_OPERAND (op, 0); | |
4189 | ||
465c8c19 JJ |
4190 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
4191 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4192 | ||
4193 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4194 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4195 | case of SLP. */ | |
4196 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
4197 | ncopies = 1; | |
4198 | else | |
4199 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4200 | ||
4201 | gcc_assert (ncopies >= 1); | |
4202 | ||
81c40241 | 4203 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[0], &vectype_in)) |
ebfd146a | 4204 | { |
73fbfcad | 4205 | if (dump_enabled_p ()) |
78c60e3d | 4206 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4207 | "use not simple.\n"); |
ebfd146a IR |
4208 | return false; |
4209 | } | |
4210 | ||
fde9c428 RG |
4211 | /* We can handle NOP_EXPR conversions that do not change the number |
4212 | of elements or the vector size. */ | |
7b7ec6c5 RG |
4213 | if ((CONVERT_EXPR_CODE_P (code) |
4214 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 RG |
4215 | && (!vectype_in |
4216 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
4217 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
4218 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
4219 | return false; | |
4220 | ||
7b7b1813 RG |
4221 | /* We do not handle bit-precision changes. */ |
4222 | if ((CONVERT_EXPR_CODE_P (code) | |
4223 | || code == VIEW_CONVERT_EXPR) | |
4224 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
4225 | && ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4226 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4227 | || ((TYPE_PRECISION (TREE_TYPE (op)) | |
4228 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))) | |
4229 | /* But a conversion that does not change the bit-pattern is ok. */ | |
4230 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4231 | > TYPE_PRECISION (TREE_TYPE (op))) | |
4232 | && TYPE_UNSIGNED (TREE_TYPE (op)))) | |
4233 | { | |
73fbfcad | 4234 | if (dump_enabled_p ()) |
78c60e3d SS |
4235 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4236 | "type conversion to/from bit-precision " | |
e645e942 | 4237 | "unsupported.\n"); |
7b7b1813 RG |
4238 | return false; |
4239 | } | |
4240 | ||
ebfd146a IR |
4241 | if (!vec_stmt) /* transformation not required. */ |
4242 | { | |
4243 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 4244 | if (dump_enabled_p ()) |
78c60e3d | 4245 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4246 | "=== vectorizable_assignment ===\n"); |
c3e7ee41 | 4247 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
4248 | return true; |
4249 | } | |
4250 | ||
4251 | /** Transform. **/ | |
73fbfcad | 4252 | if (dump_enabled_p ()) |
e645e942 | 4253 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
4254 | |
4255 | /* Handle def. */ | |
4256 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4257 | ||
4258 | /* Handle use. */ | |
f18b55bd | 4259 | for (j = 0; j < ncopies; j++) |
ebfd146a | 4260 | { |
f18b55bd IR |
4261 | /* Handle uses. */ |
4262 | if (j == 0) | |
d092494c | 4263 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node, -1); |
f18b55bd IR |
4264 | else |
4265 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
4266 | ||
4267 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 4268 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 4269 | { |
7b7ec6c5 RG |
4270 | if (CONVERT_EXPR_CODE_P (code) |
4271 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 4272 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
4273 | new_stmt = gimple_build_assign (vec_dest, vop); |
4274 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4275 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4276 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4277 | if (slp_node) | |
9771b263 | 4278 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 4279 | } |
ebfd146a IR |
4280 | |
4281 | if (slp_node) | |
f18b55bd IR |
4282 | continue; |
4283 | ||
4284 | if (j == 0) | |
4285 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4286 | else | |
4287 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4288 | ||
4289 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4290 | } | |
b8698a0f | 4291 | |
9771b263 | 4292 | vec_oprnds.release (); |
ebfd146a IR |
4293 | return true; |
4294 | } | |
4295 | ||
9dc3f7de | 4296 | |
1107f3ae IR |
4297 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
4298 | either as shift by a scalar or by a vector. */ | |
4299 | ||
4300 | bool | |
4301 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
4302 | { | |
4303 | ||
ef4bddc2 | 4304 | machine_mode vec_mode; |
1107f3ae IR |
4305 | optab optab; |
4306 | int icode; | |
4307 | tree vectype; | |
4308 | ||
4309 | vectype = get_vectype_for_scalar_type (scalar_type); | |
4310 | if (!vectype) | |
4311 | return false; | |
4312 | ||
4313 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4314 | if (!optab | |
4315 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
4316 | { | |
4317 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4318 | if (!optab | |
4319 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
4320 | == CODE_FOR_nothing)) | |
4321 | return false; | |
4322 | } | |
4323 | ||
4324 | vec_mode = TYPE_MODE (vectype); | |
4325 | icode = (int) optab_handler (optab, vec_mode); | |
4326 | if (icode == CODE_FOR_nothing) | |
4327 | return false; | |
4328 | ||
4329 | return true; | |
4330 | } | |
4331 | ||
4332 | ||
9dc3f7de IR |
4333 | /* Function vectorizable_shift. |
4334 | ||
4335 | Check if STMT performs a shift operation that can be vectorized. | |
4336 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4337 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
4338 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4339 | ||
4340 | static bool | |
355fe088 TS |
4341 | vectorizable_shift (gimple *stmt, gimple_stmt_iterator *gsi, |
4342 | gimple **vec_stmt, slp_tree slp_node) | |
9dc3f7de IR |
4343 | { |
4344 | tree vec_dest; | |
4345 | tree scalar_dest; | |
4346 | tree op0, op1 = NULL; | |
4347 | tree vec_oprnd1 = NULL_TREE; | |
4348 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4349 | tree vectype; | |
4350 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4351 | enum tree_code code; | |
ef4bddc2 | 4352 | machine_mode vec_mode; |
9dc3f7de IR |
4353 | tree new_temp; |
4354 | optab optab; | |
4355 | int icode; | |
ef4bddc2 | 4356 | machine_mode optab_op2_mode; |
355fe088 | 4357 | gimple *def_stmt; |
9dc3f7de | 4358 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
355fe088 | 4359 | gimple *new_stmt = NULL; |
9dc3f7de IR |
4360 | stmt_vec_info prev_stmt_info; |
4361 | int nunits_in; | |
4362 | int nunits_out; | |
4363 | tree vectype_out; | |
cede2577 | 4364 | tree op1_vectype; |
9dc3f7de IR |
4365 | int ncopies; |
4366 | int j, i; | |
6e1aa848 DN |
4367 | vec<tree> vec_oprnds0 = vNULL; |
4368 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
4369 | tree vop0, vop1; |
4370 | unsigned int k; | |
49eab32e | 4371 | bool scalar_shift_arg = true; |
9dc3f7de | 4372 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4373 | vec_info *vinfo = stmt_info->vinfo; |
9dc3f7de IR |
4374 | int vf; |
4375 | ||
4376 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
4377 | return false; | |
4378 | ||
4379 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
4380 | return false; | |
4381 | ||
4382 | /* Is STMT a vectorizable binary/unary operation? */ | |
4383 | if (!is_gimple_assign (stmt)) | |
4384 | return false; | |
4385 | ||
4386 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4387 | return false; | |
4388 | ||
4389 | code = gimple_assign_rhs_code (stmt); | |
4390 | ||
4391 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
4392 | || code == RROTATE_EXPR)) | |
4393 | return false; | |
4394 | ||
4395 | scalar_dest = gimple_assign_lhs (stmt); | |
4396 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
7b7b1813 RG |
4397 | if (TYPE_PRECISION (TREE_TYPE (scalar_dest)) |
4398 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4399 | { | |
73fbfcad | 4400 | if (dump_enabled_p ()) |
78c60e3d | 4401 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4402 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
4403 | return false; |
4404 | } | |
9dc3f7de IR |
4405 | |
4406 | op0 = gimple_assign_rhs1 (stmt); | |
81c40241 | 4407 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
9dc3f7de | 4408 | { |
73fbfcad | 4409 | if (dump_enabled_p ()) |
78c60e3d | 4410 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4411 | "use not simple.\n"); |
9dc3f7de IR |
4412 | return false; |
4413 | } | |
4414 | /* If op0 is an external or constant def use a vector type with | |
4415 | the same size as the output vector type. */ | |
4416 | if (!vectype) | |
4417 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
4418 | if (vec_stmt) | |
4419 | gcc_assert (vectype); | |
4420 | if (!vectype) | |
4421 | { | |
73fbfcad | 4422 | if (dump_enabled_p ()) |
78c60e3d | 4423 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4424 | "no vectype for scalar type\n"); |
9dc3f7de IR |
4425 | return false; |
4426 | } | |
4427 | ||
4428 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4429 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4430 | if (nunits_out != nunits_in) | |
4431 | return false; | |
4432 | ||
4433 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 4434 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &op1_vectype)) |
9dc3f7de | 4435 | { |
73fbfcad | 4436 | if (dump_enabled_p ()) |
78c60e3d | 4437 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4438 | "use not simple.\n"); |
9dc3f7de IR |
4439 | return false; |
4440 | } | |
4441 | ||
4442 | if (loop_vinfo) | |
4443 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4444 | else | |
4445 | vf = 1; | |
4446 | ||
4447 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4448 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4449 | case of SLP. */ | |
437f4a00 | 4450 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
9dc3f7de IR |
4451 | ncopies = 1; |
4452 | else | |
4453 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
4454 | ||
4455 | gcc_assert (ncopies >= 1); | |
4456 | ||
4457 | /* Determine whether the shift amount is a vector, or scalar. If the | |
4458 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
4459 | ||
dbfa87aa YR |
4460 | if ((dt[1] == vect_internal_def |
4461 | || dt[1] == vect_induction_def) | |
4462 | && !slp_node) | |
49eab32e JJ |
4463 | scalar_shift_arg = false; |
4464 | else if (dt[1] == vect_constant_def | |
4465 | || dt[1] == vect_external_def | |
4466 | || dt[1] == vect_internal_def) | |
4467 | { | |
4468 | /* In SLP, need to check whether the shift count is the same, | |
4469 | in loops if it is a constant or invariant, it is always | |
4470 | a scalar shift. */ | |
4471 | if (slp_node) | |
4472 | { | |
355fe088 TS |
4473 | vec<gimple *> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
4474 | gimple *slpstmt; | |
49eab32e | 4475 | |
9771b263 | 4476 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
4477 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
4478 | scalar_shift_arg = false; | |
4479 | } | |
4480 | } | |
4481 | else | |
4482 | { | |
73fbfcad | 4483 | if (dump_enabled_p ()) |
78c60e3d | 4484 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4485 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
4486 | return false; |
4487 | } | |
4488 | ||
9dc3f7de | 4489 | /* Vector shifted by vector. */ |
49eab32e | 4490 | if (!scalar_shift_arg) |
9dc3f7de IR |
4491 | { |
4492 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 4493 | if (dump_enabled_p ()) |
78c60e3d | 4494 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4495 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 4496 | |
aa948027 JJ |
4497 | if (!op1_vectype) |
4498 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
4499 | if (op1_vectype == NULL_TREE | |
4500 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 4501 | { |
73fbfcad | 4502 | if (dump_enabled_p ()) |
78c60e3d SS |
4503 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4504 | "unusable type for last operand in" | |
e645e942 | 4505 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
4506 | return false; |
4507 | } | |
9dc3f7de IR |
4508 | } |
4509 | /* See if the machine has a vector shifted by scalar insn and if not | |
4510 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 4511 | else |
9dc3f7de IR |
4512 | { |
4513 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4514 | if (optab | |
4515 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
4516 | { | |
73fbfcad | 4517 | if (dump_enabled_p ()) |
78c60e3d | 4518 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4519 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
4520 | } |
4521 | else | |
4522 | { | |
4523 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4524 | if (optab | |
4525 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
4526 | != CODE_FOR_nothing)) | |
4527 | { | |
49eab32e JJ |
4528 | scalar_shift_arg = false; |
4529 | ||
73fbfcad | 4530 | if (dump_enabled_p ()) |
78c60e3d | 4531 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4532 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
4533 | |
4534 | /* Unlike the other binary operators, shifts/rotates have | |
4535 | the rhs being int, instead of the same type as the lhs, | |
4536 | so make sure the scalar is the right type if we are | |
aa948027 | 4537 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
4538 | if (dt[1] == vect_constant_def) |
4539 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
4540 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
4541 | TREE_TYPE (op1))) | |
4542 | { | |
4543 | if (slp_node | |
4544 | && TYPE_MODE (TREE_TYPE (vectype)) | |
4545 | != TYPE_MODE (TREE_TYPE (op1))) | |
4546 | { | |
73fbfcad | 4547 | if (dump_enabled_p ()) |
78c60e3d SS |
4548 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4549 | "unusable type for last operand in" | |
e645e942 | 4550 | " vector/vector shift/rotate.\n"); |
21c0a521 | 4551 | return false; |
aa948027 JJ |
4552 | } |
4553 | if (vec_stmt && !slp_node) | |
4554 | { | |
4555 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
4556 | op1 = vect_init_vector (stmt, op1, | |
4557 | TREE_TYPE (vectype), NULL); | |
4558 | } | |
4559 | } | |
9dc3f7de IR |
4560 | } |
4561 | } | |
4562 | } | |
9dc3f7de IR |
4563 | |
4564 | /* Supportable by target? */ | |
4565 | if (!optab) | |
4566 | { | |
73fbfcad | 4567 | if (dump_enabled_p ()) |
78c60e3d | 4568 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4569 | "no optab.\n"); |
9dc3f7de IR |
4570 | return false; |
4571 | } | |
4572 | vec_mode = TYPE_MODE (vectype); | |
4573 | icode = (int) optab_handler (optab, vec_mode); | |
4574 | if (icode == CODE_FOR_nothing) | |
4575 | { | |
73fbfcad | 4576 | if (dump_enabled_p ()) |
78c60e3d | 4577 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4578 | "op not supported by target.\n"); |
9dc3f7de IR |
4579 | /* Check only during analysis. */ |
4580 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
4581 | || (vf < vect_min_worthwhile_factor (code) | |
4582 | && !vec_stmt)) | |
4583 | return false; | |
73fbfcad | 4584 | if (dump_enabled_p ()) |
e645e942 TJ |
4585 | dump_printf_loc (MSG_NOTE, vect_location, |
4586 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
4587 | } |
4588 | ||
4589 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
4590 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
4591 | && vf < vect_min_worthwhile_factor (code) | |
4592 | && !vec_stmt) | |
4593 | { | |
73fbfcad | 4594 | if (dump_enabled_p ()) |
78c60e3d | 4595 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4596 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
4597 | return false; |
4598 | } | |
4599 | ||
4600 | if (!vec_stmt) /* transformation not required. */ | |
4601 | { | |
4602 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 4603 | if (dump_enabled_p ()) |
e645e942 TJ |
4604 | dump_printf_loc (MSG_NOTE, vect_location, |
4605 | "=== vectorizable_shift ===\n"); | |
c3e7ee41 | 4606 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
9dc3f7de IR |
4607 | return true; |
4608 | } | |
4609 | ||
4610 | /** Transform. **/ | |
4611 | ||
73fbfcad | 4612 | if (dump_enabled_p ()) |
78c60e3d | 4613 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4614 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
4615 | |
4616 | /* Handle def. */ | |
4617 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4618 | ||
9dc3f7de IR |
4619 | prev_stmt_info = NULL; |
4620 | for (j = 0; j < ncopies; j++) | |
4621 | { | |
4622 | /* Handle uses. */ | |
4623 | if (j == 0) | |
4624 | { | |
4625 | if (scalar_shift_arg) | |
4626 | { | |
4627 | /* Vector shl and shr insn patterns can be defined with scalar | |
4628 | operand 2 (shift operand). In this case, use constant or loop | |
4629 | invariant op1 directly, without extending it to vector mode | |
4630 | first. */ | |
4631 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
4632 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
4633 | { | |
73fbfcad | 4634 | if (dump_enabled_p ()) |
78c60e3d | 4635 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4636 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 4637 | vec_oprnd1 = op1; |
8930f723 | 4638 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 4639 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
4640 | if (slp_node) |
4641 | { | |
4642 | /* Store vec_oprnd1 for every vector stmt to be created | |
4643 | for SLP_NODE. We check during the analysis that all | |
4644 | the shift arguments are the same. | |
4645 | TODO: Allow different constants for different vector | |
4646 | stmts generated for an SLP instance. */ | |
4647 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4648 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
4649 | } |
4650 | } | |
4651 | } | |
4652 | ||
4653 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
4654 | (a special case for certain kind of vector shifts); otherwise, | |
4655 | operand 1 should be of a vector type (the usual case). */ | |
4656 | if (vec_oprnd1) | |
4657 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
d092494c | 4658 | slp_node, -1); |
9dc3f7de IR |
4659 | else |
4660 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
d092494c | 4661 | slp_node, -1); |
9dc3f7de IR |
4662 | } |
4663 | else | |
4664 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
4665 | ||
4666 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 4667 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 4668 | { |
9771b263 | 4669 | vop1 = vec_oprnds1[i]; |
0d0e4a03 | 4670 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
9dc3f7de IR |
4671 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4672 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4673 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4674 | if (slp_node) | |
9771b263 | 4675 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
4676 | } |
4677 | ||
4678 | if (slp_node) | |
4679 | continue; | |
4680 | ||
4681 | if (j == 0) | |
4682 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4683 | else | |
4684 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4685 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4686 | } | |
4687 | ||
9771b263 DN |
4688 | vec_oprnds0.release (); |
4689 | vec_oprnds1.release (); | |
9dc3f7de IR |
4690 | |
4691 | return true; | |
4692 | } | |
4693 | ||
4694 | ||
ebfd146a IR |
4695 | /* Function vectorizable_operation. |
4696 | ||
16949072 RG |
4697 | Check if STMT performs a binary, unary or ternary operation that can |
4698 | be vectorized. | |
b8698a0f | 4699 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
4700 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4701 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4702 | ||
4703 | static bool | |
355fe088 TS |
4704 | vectorizable_operation (gimple *stmt, gimple_stmt_iterator *gsi, |
4705 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a | 4706 | { |
00f07b86 | 4707 | tree vec_dest; |
ebfd146a | 4708 | tree scalar_dest; |
16949072 | 4709 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 4710 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 4711 | tree vectype; |
ebfd146a IR |
4712 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4713 | enum tree_code code; | |
ef4bddc2 | 4714 | machine_mode vec_mode; |
ebfd146a IR |
4715 | tree new_temp; |
4716 | int op_type; | |
00f07b86 | 4717 | optab optab; |
523ba738 | 4718 | bool target_support_p; |
355fe088 | 4719 | gimple *def_stmt; |
16949072 RG |
4720 | enum vect_def_type dt[3] |
4721 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
355fe088 | 4722 | gimple *new_stmt = NULL; |
ebfd146a | 4723 | stmt_vec_info prev_stmt_info; |
b690cc0f | 4724 | int nunits_in; |
ebfd146a IR |
4725 | int nunits_out; |
4726 | tree vectype_out; | |
4727 | int ncopies; | |
4728 | int j, i; | |
6e1aa848 DN |
4729 | vec<tree> vec_oprnds0 = vNULL; |
4730 | vec<tree> vec_oprnds1 = vNULL; | |
4731 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 4732 | tree vop0, vop1, vop2; |
a70d6342 | 4733 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4734 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 IR |
4735 | int vf; |
4736 | ||
a70d6342 | 4737 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4738 | return false; |
4739 | ||
8644a673 | 4740 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
4741 | return false; |
4742 | ||
4743 | /* Is STMT a vectorizable binary/unary operation? */ | |
4744 | if (!is_gimple_assign (stmt)) | |
4745 | return false; | |
4746 | ||
4747 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4748 | return false; | |
4749 | ||
ebfd146a IR |
4750 | code = gimple_assign_rhs_code (stmt); |
4751 | ||
4752 | /* For pointer addition, we should use the normal plus for | |
4753 | the vector addition. */ | |
4754 | if (code == POINTER_PLUS_EXPR) | |
4755 | code = PLUS_EXPR; | |
4756 | ||
4757 | /* Support only unary or binary operations. */ | |
4758 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 4759 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 4760 | { |
73fbfcad | 4761 | if (dump_enabled_p ()) |
78c60e3d | 4762 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4763 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 4764 | op_type); |
ebfd146a IR |
4765 | return false; |
4766 | } | |
4767 | ||
b690cc0f RG |
4768 | scalar_dest = gimple_assign_lhs (stmt); |
4769 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
4770 | ||
7b7b1813 RG |
4771 | /* Most operations cannot handle bit-precision types without extra |
4772 | truncations. */ | |
045c1278 IE |
4773 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4774 | && (TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4775 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
7b7b1813 RG |
4776 | /* Exception are bitwise binary operations. */ |
4777 | && code != BIT_IOR_EXPR | |
4778 | && code != BIT_XOR_EXPR | |
4779 | && code != BIT_AND_EXPR) | |
4780 | { | |
73fbfcad | 4781 | if (dump_enabled_p ()) |
78c60e3d | 4782 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4783 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
4784 | return false; |
4785 | } | |
4786 | ||
ebfd146a | 4787 | op0 = gimple_assign_rhs1 (stmt); |
81c40241 | 4788 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
ebfd146a | 4789 | { |
73fbfcad | 4790 | if (dump_enabled_p ()) |
78c60e3d | 4791 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4792 | "use not simple.\n"); |
ebfd146a IR |
4793 | return false; |
4794 | } | |
b690cc0f RG |
4795 | /* If op0 is an external or constant def use a vector type with |
4796 | the same size as the output vector type. */ | |
4797 | if (!vectype) | |
b036c6c5 IE |
4798 | { |
4799 | /* For boolean type we cannot determine vectype by | |
4800 | invariant value (don't know whether it is a vector | |
4801 | of booleans or vector of integers). We use output | |
4802 | vectype because operations on boolean don't change | |
4803 | type. */ | |
4804 | if (TREE_CODE (TREE_TYPE (op0)) == BOOLEAN_TYPE) | |
4805 | { | |
4806 | if (TREE_CODE (TREE_TYPE (scalar_dest)) != BOOLEAN_TYPE) | |
4807 | { | |
4808 | if (dump_enabled_p ()) | |
4809 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
4810 | "not supported operation on bool value.\n"); | |
4811 | return false; | |
4812 | } | |
4813 | vectype = vectype_out; | |
4814 | } | |
4815 | else | |
4816 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
4817 | } | |
7d8930a0 IR |
4818 | if (vec_stmt) |
4819 | gcc_assert (vectype); | |
4820 | if (!vectype) | |
4821 | { | |
73fbfcad | 4822 | if (dump_enabled_p ()) |
7d8930a0 | 4823 | { |
78c60e3d SS |
4824 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4825 | "no vectype for scalar type "); | |
4826 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
4827 | TREE_TYPE (op0)); | |
e645e942 | 4828 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
4829 | } |
4830 | ||
4831 | return false; | |
4832 | } | |
b690cc0f RG |
4833 | |
4834 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4835 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4836 | if (nunits_out != nunits_in) | |
4837 | return false; | |
ebfd146a | 4838 | |
16949072 | 4839 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
4840 | { |
4841 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 4842 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1])) |
ebfd146a | 4843 | { |
73fbfcad | 4844 | if (dump_enabled_p ()) |
78c60e3d | 4845 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4846 | "use not simple.\n"); |
ebfd146a IR |
4847 | return false; |
4848 | } | |
4849 | } | |
16949072 RG |
4850 | if (op_type == ternary_op) |
4851 | { | |
4852 | op2 = gimple_assign_rhs3 (stmt); | |
81c40241 | 4853 | if (!vect_is_simple_use (op2, vinfo, &def_stmt, &dt[2])) |
16949072 | 4854 | { |
73fbfcad | 4855 | if (dump_enabled_p ()) |
78c60e3d | 4856 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4857 | "use not simple.\n"); |
16949072 RG |
4858 | return false; |
4859 | } | |
4860 | } | |
ebfd146a | 4861 | |
b690cc0f RG |
4862 | if (loop_vinfo) |
4863 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4864 | else | |
4865 | vf = 1; | |
4866 | ||
4867 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 4868 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 4869 | case of SLP. */ |
437f4a00 | 4870 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
b690cc0f RG |
4871 | ncopies = 1; |
4872 | else | |
4873 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
4874 | ||
4875 | gcc_assert (ncopies >= 1); | |
4876 | ||
9dc3f7de | 4877 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
4878 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
4879 | || code == RROTATE_EXPR) | |
9dc3f7de | 4880 | return false; |
ebfd146a | 4881 | |
ebfd146a | 4882 | /* Supportable by target? */ |
00f07b86 RH |
4883 | |
4884 | vec_mode = TYPE_MODE (vectype); | |
4885 | if (code == MULT_HIGHPART_EXPR) | |
523ba738 | 4886 | target_support_p = can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype)); |
00f07b86 RH |
4887 | else |
4888 | { | |
4889 | optab = optab_for_tree_code (code, vectype, optab_default); | |
4890 | if (!optab) | |
5deb57cb | 4891 | { |
73fbfcad | 4892 | if (dump_enabled_p ()) |
78c60e3d | 4893 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4894 | "no optab.\n"); |
00f07b86 | 4895 | return false; |
5deb57cb | 4896 | } |
523ba738 RS |
4897 | target_support_p = (optab_handler (optab, vec_mode) |
4898 | != CODE_FOR_nothing); | |
5deb57cb JJ |
4899 | } |
4900 | ||
523ba738 | 4901 | if (!target_support_p) |
ebfd146a | 4902 | { |
73fbfcad | 4903 | if (dump_enabled_p ()) |
78c60e3d | 4904 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4905 | "op not supported by target.\n"); |
ebfd146a IR |
4906 | /* Check only during analysis. */ |
4907 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
5deb57cb | 4908 | || (!vec_stmt && vf < vect_min_worthwhile_factor (code))) |
ebfd146a | 4909 | return false; |
73fbfcad | 4910 | if (dump_enabled_p ()) |
e645e942 TJ |
4911 | dump_printf_loc (MSG_NOTE, vect_location, |
4912 | "proceeding using word mode.\n"); | |
383d9c83 IR |
4913 | } |
4914 | ||
4a00c761 | 4915 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
4916 | if (!VECTOR_MODE_P (vec_mode) |
4917 | && !vec_stmt | |
4918 | && vf < vect_min_worthwhile_factor (code)) | |
7d8930a0 | 4919 | { |
73fbfcad | 4920 | if (dump_enabled_p ()) |
78c60e3d | 4921 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4922 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 4923 | return false; |
7d8930a0 | 4924 | } |
ebfd146a | 4925 | |
ebfd146a IR |
4926 | if (!vec_stmt) /* transformation not required. */ |
4927 | { | |
4a00c761 | 4928 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 4929 | if (dump_enabled_p ()) |
78c60e3d | 4930 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4931 | "=== vectorizable_operation ===\n"); |
c3e7ee41 | 4932 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
4933 | return true; |
4934 | } | |
4935 | ||
4936 | /** Transform. **/ | |
4937 | ||
73fbfcad | 4938 | if (dump_enabled_p ()) |
78c60e3d | 4939 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4940 | "transform binary/unary operation.\n"); |
383d9c83 | 4941 | |
ebfd146a | 4942 | /* Handle def. */ |
00f07b86 | 4943 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 4944 | |
ebfd146a IR |
4945 | /* In case the vectorization factor (VF) is bigger than the number |
4946 | of elements that we can fit in a vectype (nunits), we have to generate | |
4947 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
4948 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
4949 | from one copy of the vector stmt to the next, in the field | |
4950 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
4951 | stages to find the correct vector defs to be used when vectorizing | |
4952 | stmts that use the defs of the current stmt. The example below | |
4953 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
4954 | we need to create 4 vectorized stmts): | |
4955 | ||
4956 | before vectorization: | |
4957 | RELATED_STMT VEC_STMT | |
4958 | S1: x = memref - - | |
4959 | S2: z = x + 1 - - | |
4960 | ||
4961 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
4962 | there): | |
4963 | RELATED_STMT VEC_STMT | |
4964 | VS1_0: vx0 = memref0 VS1_1 - | |
4965 | VS1_1: vx1 = memref1 VS1_2 - | |
4966 | VS1_2: vx2 = memref2 VS1_3 - | |
4967 | VS1_3: vx3 = memref3 - - | |
4968 | S1: x = load - VS1_0 | |
4969 | S2: z = x + 1 - - | |
4970 | ||
4971 | step2: vectorize stmt S2 (done here): | |
4972 | To vectorize stmt S2 we first need to find the relevant vector | |
4973 | def for the first operand 'x'. This is, as usual, obtained from | |
4974 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
4975 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
4976 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
4977 | the vector stmt VS2_0, and as usual, record it in the | |
4978 | STMT_VINFO_VEC_STMT of stmt S2. | |
4979 | When creating the second copy (VS2_1), we obtain the relevant vector | |
4980 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
4981 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
4982 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
4983 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
4984 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
4985 | chain of stmts and pointers: | |
4986 | RELATED_STMT VEC_STMT | |
4987 | VS1_0: vx0 = memref0 VS1_1 - | |
4988 | VS1_1: vx1 = memref1 VS1_2 - | |
4989 | VS1_2: vx2 = memref2 VS1_3 - | |
4990 | VS1_3: vx3 = memref3 - - | |
4991 | S1: x = load - VS1_0 | |
4992 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
4993 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
4994 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
4995 | VS2_3: vz3 = vx3 + v1 - - | |
4996 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
4997 | |
4998 | prev_stmt_info = NULL; | |
4999 | for (j = 0; j < ncopies; j++) | |
5000 | { | |
5001 | /* Handle uses. */ | |
5002 | if (j == 0) | |
4a00c761 JJ |
5003 | { |
5004 | if (op_type == binary_op || op_type == ternary_op) | |
5005 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
5006 | slp_node, -1); | |
5007 | else | |
5008 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
5009 | slp_node, -1); | |
5010 | if (op_type == ternary_op) | |
36ba4aae | 5011 | { |
9771b263 DN |
5012 | vec_oprnds2.create (1); |
5013 | vec_oprnds2.quick_push (vect_get_vec_def_for_operand (op2, | |
81c40241 | 5014 | stmt)); |
36ba4aae | 5015 | } |
4a00c761 | 5016 | } |
ebfd146a | 5017 | else |
4a00c761 JJ |
5018 | { |
5019 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5020 | if (op_type == ternary_op) | |
5021 | { | |
9771b263 DN |
5022 | tree vec_oprnd = vec_oprnds2.pop (); |
5023 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
5024 | vec_oprnd)); | |
4a00c761 JJ |
5025 | } |
5026 | } | |
5027 | ||
5028 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 5029 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 5030 | { |
4a00c761 | 5031 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 5032 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 5033 | vop2 = ((op_type == ternary_op) |
9771b263 | 5034 | ? vec_oprnds2[i] : NULL_TREE); |
0d0e4a03 | 5035 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1, vop2); |
4a00c761 JJ |
5036 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5037 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5038 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5039 | if (slp_node) | |
9771b263 | 5040 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
5041 | } |
5042 | ||
4a00c761 JJ |
5043 | if (slp_node) |
5044 | continue; | |
5045 | ||
5046 | if (j == 0) | |
5047 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5048 | else | |
5049 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5050 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
5051 | } |
5052 | ||
9771b263 DN |
5053 | vec_oprnds0.release (); |
5054 | vec_oprnds1.release (); | |
5055 | vec_oprnds2.release (); | |
ebfd146a | 5056 | |
ebfd146a IR |
5057 | return true; |
5058 | } | |
5059 | ||
c716e67f XDL |
5060 | /* A helper function to ensure data reference DR's base alignment |
5061 | for STMT_INFO. */ | |
5062 | ||
5063 | static void | |
5064 | ensure_base_align (stmt_vec_info stmt_info, struct data_reference *dr) | |
5065 | { | |
5066 | if (!dr->aux) | |
5067 | return; | |
5068 | ||
52639a61 | 5069 | if (DR_VECT_AUX (dr)->base_misaligned) |
c716e67f XDL |
5070 | { |
5071 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
52639a61 | 5072 | tree base_decl = DR_VECT_AUX (dr)->base_decl; |
c716e67f | 5073 | |
428f0c67 JH |
5074 | if (decl_in_symtab_p (base_decl)) |
5075 | symtab_node::get (base_decl)->increase_alignment (TYPE_ALIGN (vectype)); | |
5076 | else | |
5077 | { | |
5078 | DECL_ALIGN (base_decl) = TYPE_ALIGN (vectype); | |
5079 | DECL_USER_ALIGN (base_decl) = 1; | |
5080 | } | |
52639a61 | 5081 | DR_VECT_AUX (dr)->base_misaligned = false; |
c716e67f XDL |
5082 | } |
5083 | } | |
5084 | ||
ebfd146a | 5085 | |
09dfa495 BM |
5086 | /* Given a vector type VECTYPE returns the VECTOR_CST mask that implements |
5087 | reversal of the vector elements. If that is impossible to do, | |
5088 | returns NULL. */ | |
5089 | ||
5090 | static tree | |
5091 | perm_mask_for_reverse (tree vectype) | |
5092 | { | |
5093 | int i, nunits; | |
5094 | unsigned char *sel; | |
5095 | ||
5096 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
5097 | sel = XALLOCAVEC (unsigned char, nunits); | |
5098 | ||
5099 | for (i = 0; i < nunits; ++i) | |
5100 | sel[i] = nunits - 1 - i; | |
5101 | ||
557be5a8 AL |
5102 | if (!can_vec_perm_p (TYPE_MODE (vectype), false, sel)) |
5103 | return NULL_TREE; | |
5104 | return vect_gen_perm_mask_checked (vectype, sel); | |
09dfa495 BM |
5105 | } |
5106 | ||
ebfd146a IR |
5107 | /* Function vectorizable_store. |
5108 | ||
b8698a0f L |
5109 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
5110 | can be vectorized. | |
5111 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
5112 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5113 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5114 | ||
5115 | static bool | |
355fe088 | 5116 | vectorizable_store (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 5117 | slp_tree slp_node) |
ebfd146a IR |
5118 | { |
5119 | tree scalar_dest; | |
5120 | tree data_ref; | |
5121 | tree op; | |
5122 | tree vec_oprnd = NULL_TREE; | |
5123 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5124 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
272c6793 | 5125 | tree elem_type; |
ebfd146a | 5126 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 5127 | struct loop *loop = NULL; |
ef4bddc2 | 5128 | machine_mode vec_mode; |
ebfd146a IR |
5129 | tree dummy; |
5130 | enum dr_alignment_support alignment_support_scheme; | |
355fe088 | 5131 | gimple *def_stmt; |
ebfd146a IR |
5132 | enum vect_def_type dt; |
5133 | stmt_vec_info prev_stmt_info = NULL; | |
5134 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 5135 | tree dataref_offset = NULL_TREE; |
355fe088 | 5136 | gimple *ptr_incr = NULL; |
ebfd146a IR |
5137 | int ncopies; |
5138 | int j; | |
355fe088 | 5139 | gimple *next_stmt, *first_stmt = NULL; |
0d0293ac | 5140 | bool grouped_store = false; |
272c6793 | 5141 | bool store_lanes_p = false; |
ebfd146a | 5142 | unsigned int group_size, i; |
6e1aa848 DN |
5143 | vec<tree> dr_chain = vNULL; |
5144 | vec<tree> oprnds = vNULL; | |
5145 | vec<tree> result_chain = vNULL; | |
ebfd146a | 5146 | bool inv_p; |
09dfa495 BM |
5147 | bool negative = false; |
5148 | tree offset = NULL_TREE; | |
6e1aa848 | 5149 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 5150 | bool slp = (slp_node != NULL); |
ebfd146a | 5151 | unsigned int vec_num; |
a70d6342 | 5152 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5153 | vec_info *vinfo = stmt_info->vinfo; |
272c6793 | 5154 | tree aggr_type; |
3bab6342 AT |
5155 | tree scatter_base = NULL_TREE, scatter_off = NULL_TREE; |
5156 | tree scatter_off_vectype = NULL_TREE, scatter_decl = NULL_TREE; | |
5157 | int scatter_scale = 1; | |
5158 | enum vect_def_type scatter_idx_dt = vect_unknown_def_type; | |
5159 | enum vect_def_type scatter_src_dt = vect_unknown_def_type; | |
355fe088 | 5160 | gimple *new_stmt; |
a70d6342 | 5161 | |
a70d6342 | 5162 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5163 | return false; |
5164 | ||
8644a673 | 5165 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
5166 | return false; |
5167 | ||
5168 | /* Is vectorizable store? */ | |
5169 | ||
5170 | if (!is_gimple_assign (stmt)) | |
5171 | return false; | |
5172 | ||
5173 | scalar_dest = gimple_assign_lhs (stmt); | |
ab0ef706 JJ |
5174 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
5175 | && is_pattern_stmt_p (stmt_info)) | |
5176 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
ebfd146a | 5177 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
38000232 | 5178 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF |
ebfd146a | 5179 | && TREE_CODE (scalar_dest) != INDIRECT_REF |
e9dbe7bb IR |
5180 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
5181 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
5182 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
5183 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
5184 | return false; |
5185 | ||
5186 | gcc_assert (gimple_assign_single_p (stmt)); | |
465c8c19 JJ |
5187 | |
5188 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
5189 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
5190 | ||
5191 | if (loop_vinfo) | |
5192 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
5193 | ||
5194 | /* Multiple types in SLP are handled by creating the appropriate number of | |
5195 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5196 | case of SLP. */ | |
5197 | if (slp || PURE_SLP_STMT (stmt_info)) | |
5198 | ncopies = 1; | |
5199 | else | |
5200 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
5201 | ||
5202 | gcc_assert (ncopies >= 1); | |
5203 | ||
5204 | /* FORNOW. This restriction should be relaxed. */ | |
5205 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) | |
5206 | { | |
5207 | if (dump_enabled_p ()) | |
5208 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5209 | "multiple types in nested loop.\n"); | |
5210 | return false; | |
5211 | } | |
5212 | ||
ebfd146a | 5213 | op = gimple_assign_rhs1 (stmt); |
81c40241 | 5214 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt)) |
ebfd146a | 5215 | { |
73fbfcad | 5216 | if (dump_enabled_p ()) |
78c60e3d | 5217 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5218 | "use not simple.\n"); |
ebfd146a IR |
5219 | return false; |
5220 | } | |
5221 | ||
272c6793 | 5222 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 5223 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 5224 | |
ebfd146a IR |
5225 | /* FORNOW. In some cases can vectorize even if data-type not supported |
5226 | (e.g. - array initialization with 0). */ | |
947131ba | 5227 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
5228 | return false; |
5229 | ||
5230 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
5231 | return false; | |
5232 | ||
f2e2a985 | 5233 | if (!STMT_VINFO_STRIDED_P (stmt_info)) |
09dfa495 | 5234 | { |
f2e2a985 MM |
5235 | negative = |
5236 | tree_int_cst_compare (loop && nested_in_vect_loop_p (loop, stmt) | |
5237 | ? STMT_VINFO_DR_STEP (stmt_info) : DR_STEP (dr), | |
5238 | size_zero_node) < 0; | |
5239 | if (negative && ncopies > 1) | |
09dfa495 BM |
5240 | { |
5241 | if (dump_enabled_p ()) | |
5242 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
f2e2a985 | 5243 | "multiple types with negative step.\n"); |
09dfa495 BM |
5244 | return false; |
5245 | } | |
f2e2a985 | 5246 | if (negative) |
09dfa495 | 5247 | { |
f2e2a985 MM |
5248 | gcc_assert (!grouped_store); |
5249 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
5250 | if (alignment_support_scheme != dr_aligned | |
5251 | && alignment_support_scheme != dr_unaligned_supported) | |
5252 | { | |
5253 | if (dump_enabled_p ()) | |
5254 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5255 | "negative step but alignment required.\n"); | |
5256 | return false; | |
5257 | } | |
5258 | if (dt != vect_constant_def | |
5259 | && dt != vect_external_def | |
5260 | && !perm_mask_for_reverse (vectype)) | |
5261 | { | |
5262 | if (dump_enabled_p ()) | |
5263 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5264 | "negative step and reversing not supported.\n"); | |
5265 | return false; | |
5266 | } | |
09dfa495 BM |
5267 | } |
5268 | } | |
5269 | ||
0d0293ac | 5270 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 5271 | { |
0d0293ac | 5272 | grouped_store = true; |
e14c1050 | 5273 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
cee62fee MM |
5274 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
5275 | if (!slp | |
5276 | && !PURE_SLP_STMT (stmt_info) | |
5277 | && !STMT_VINFO_STRIDED_P (stmt_info)) | |
b602d918 | 5278 | { |
272c6793 RS |
5279 | if (vect_store_lanes_supported (vectype, group_size)) |
5280 | store_lanes_p = true; | |
0d0293ac | 5281 | else if (!vect_grouped_store_supported (vectype, group_size)) |
b602d918 RS |
5282 | return false; |
5283 | } | |
b8698a0f | 5284 | |
cee62fee MM |
5285 | if (STMT_VINFO_STRIDED_P (stmt_info) |
5286 | && (slp || PURE_SLP_STMT (stmt_info)) | |
5287 | && (group_size > nunits | |
5288 | || nunits % group_size != 0)) | |
5289 | { | |
5290 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5291 | "unhandled strided group store\n"); | |
5292 | return false; | |
5293 | } | |
5294 | ||
ebfd146a IR |
5295 | if (first_stmt == stmt) |
5296 | { | |
5297 | /* STMT is the leader of the group. Check the operands of all the | |
5298 | stmts of the group. */ | |
e14c1050 | 5299 | next_stmt = GROUP_NEXT_ELEMENT (stmt_info); |
ebfd146a IR |
5300 | while (next_stmt) |
5301 | { | |
5302 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
5303 | op = gimple_assign_rhs1 (next_stmt); | |
81c40241 | 5304 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt)) |
ebfd146a | 5305 | { |
73fbfcad | 5306 | if (dump_enabled_p ()) |
78c60e3d | 5307 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5308 | "use not simple.\n"); |
ebfd146a IR |
5309 | return false; |
5310 | } | |
e14c1050 | 5311 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
5312 | } |
5313 | } | |
5314 | } | |
5315 | ||
3bab6342 AT |
5316 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5317 | { | |
355fe088 | 5318 | gimple *def_stmt; |
3bab6342 AT |
5319 | scatter_decl = vect_check_gather_scatter (stmt, loop_vinfo, &scatter_base, |
5320 | &scatter_off, &scatter_scale); | |
5321 | gcc_assert (scatter_decl); | |
81c40241 RB |
5322 | if (!vect_is_simple_use (scatter_off, vinfo, &def_stmt, &scatter_idx_dt, |
5323 | &scatter_off_vectype)) | |
3bab6342 AT |
5324 | { |
5325 | if (dump_enabled_p ()) | |
5326 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5327 | "scatter index use not simple."); | |
5328 | return false; | |
5329 | } | |
5330 | } | |
5331 | ||
ebfd146a IR |
5332 | if (!vec_stmt) /* transformation not required. */ |
5333 | { | |
5334 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
2e8ab70c RB |
5335 | /* The SLP costs are calculated during SLP analysis. */ |
5336 | if (!PURE_SLP_STMT (stmt_info)) | |
5337 | vect_model_store_cost (stmt_info, ncopies, store_lanes_p, dt, | |
5338 | NULL, NULL, NULL); | |
ebfd146a IR |
5339 | return true; |
5340 | } | |
5341 | ||
5342 | /** Transform. **/ | |
5343 | ||
c716e67f XDL |
5344 | ensure_base_align (stmt_info, dr); |
5345 | ||
3bab6342 AT |
5346 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5347 | { | |
5348 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE, op, src; | |
5349 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (scatter_decl)); | |
5350 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
5351 | tree ptr, mask, var, scale, perm_mask = NULL_TREE; | |
5352 | edge pe = loop_preheader_edge (loop); | |
5353 | gimple_seq seq; | |
5354 | basic_block new_bb; | |
5355 | enum { NARROW, NONE, WIDEN } modifier; | |
5356 | int scatter_off_nunits = TYPE_VECTOR_SUBPARTS (scatter_off_vectype); | |
5357 | ||
5358 | if (nunits == (unsigned int) scatter_off_nunits) | |
5359 | modifier = NONE; | |
5360 | else if (nunits == (unsigned int) scatter_off_nunits / 2) | |
5361 | { | |
5362 | unsigned char *sel = XALLOCAVEC (unsigned char, scatter_off_nunits); | |
5363 | modifier = WIDEN; | |
5364 | ||
5365 | for (i = 0; i < (unsigned int) scatter_off_nunits; ++i) | |
5366 | sel[i] = i | nunits; | |
5367 | ||
5368 | perm_mask = vect_gen_perm_mask_checked (scatter_off_vectype, sel); | |
5369 | gcc_assert (perm_mask != NULL_TREE); | |
5370 | } | |
5371 | else if (nunits == (unsigned int) scatter_off_nunits * 2) | |
5372 | { | |
5373 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
5374 | modifier = NARROW; | |
5375 | ||
5376 | for (i = 0; i < (unsigned int) nunits; ++i) | |
5377 | sel[i] = i | scatter_off_nunits; | |
5378 | ||
5379 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); | |
5380 | gcc_assert (perm_mask != NULL_TREE); | |
5381 | ncopies *= 2; | |
5382 | } | |
5383 | else | |
5384 | gcc_unreachable (); | |
5385 | ||
5386 | rettype = TREE_TYPE (TREE_TYPE (scatter_decl)); | |
5387 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5388 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5389 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5390 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5391 | scaletype = TREE_VALUE (arglist); | |
5392 | ||
5393 | gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE | |
5394 | && TREE_CODE (rettype) == VOID_TYPE); | |
5395 | ||
5396 | ptr = fold_convert (ptrtype, scatter_base); | |
5397 | if (!is_gimple_min_invariant (ptr)) | |
5398 | { | |
5399 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
5400 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
5401 | gcc_assert (!new_bb); | |
5402 | } | |
5403 | ||
5404 | /* Currently we support only unconditional scatter stores, | |
5405 | so mask should be all ones. */ | |
5406 | mask = build_int_cst (masktype, -1); | |
5407 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
5408 | ||
5409 | scale = build_int_cst (scaletype, scatter_scale); | |
5410 | ||
5411 | prev_stmt_info = NULL; | |
5412 | for (j = 0; j < ncopies; ++j) | |
5413 | { | |
5414 | if (j == 0) | |
5415 | { | |
5416 | src = vec_oprnd1 | |
81c40241 | 5417 | = vect_get_vec_def_for_operand (gimple_assign_rhs1 (stmt), stmt); |
3bab6342 | 5418 | op = vec_oprnd0 |
81c40241 | 5419 | = vect_get_vec_def_for_operand (scatter_off, stmt); |
3bab6342 AT |
5420 | } |
5421 | else if (modifier != NONE && (j & 1)) | |
5422 | { | |
5423 | if (modifier == WIDEN) | |
5424 | { | |
5425 | src = vec_oprnd1 | |
5426 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5427 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, perm_mask, | |
5428 | stmt, gsi); | |
5429 | } | |
5430 | else if (modifier == NARROW) | |
5431 | { | |
5432 | src = permute_vec_elements (vec_oprnd1, vec_oprnd1, perm_mask, | |
5433 | stmt, gsi); | |
5434 | op = vec_oprnd0 | |
5435 | = vect_get_vec_def_for_stmt_copy (scatter_idx_dt, vec_oprnd0); | |
5436 | } | |
5437 | else | |
5438 | gcc_unreachable (); | |
5439 | } | |
5440 | else | |
5441 | { | |
5442 | src = vec_oprnd1 | |
5443 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5444 | op = vec_oprnd0 | |
5445 | = vect_get_vec_def_for_stmt_copy (scatter_idx_dt, vec_oprnd0); | |
5446 | } | |
5447 | ||
5448 | if (!useless_type_conversion_p (srctype, TREE_TYPE (src))) | |
5449 | { | |
5450 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)) | |
5451 | == TYPE_VECTOR_SUBPARTS (srctype)); | |
0e22bb5a | 5452 | var = vect_get_new_ssa_name (srctype, vect_simple_var); |
3bab6342 AT |
5453 | src = build1 (VIEW_CONVERT_EXPR, srctype, src); |
5454 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, src); | |
5455 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5456 | src = var; | |
5457 | } | |
5458 | ||
5459 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
5460 | { | |
5461 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
5462 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 5463 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
3bab6342 AT |
5464 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
5465 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
5466 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5467 | op = var; | |
5468 | } | |
5469 | ||
5470 | new_stmt | |
5471 | = gimple_build_call (scatter_decl, 5, ptr, mask, op, src, scale); | |
5472 | ||
5473 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5474 | ||
5475 | if (prev_stmt_info == NULL) | |
5476 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5477 | else | |
5478 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5479 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5480 | } | |
5481 | return true; | |
5482 | } | |
5483 | ||
0d0293ac | 5484 | if (grouped_store) |
ebfd146a IR |
5485 | { |
5486 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 5487 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 5488 | |
e14c1050 | 5489 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
5490 | |
5491 | /* FORNOW */ | |
a70d6342 | 5492 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
5493 | |
5494 | /* We vectorize all the stmts of the interleaving group when we | |
5495 | reach the last stmt in the group. */ | |
e14c1050 IR |
5496 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
5497 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
5498 | && !slp) |
5499 | { | |
5500 | *vec_stmt = NULL; | |
5501 | return true; | |
5502 | } | |
5503 | ||
5504 | if (slp) | |
4b5caab7 | 5505 | { |
0d0293ac | 5506 | grouped_store = false; |
4b5caab7 IR |
5507 | /* VEC_NUM is the number of vect stmts to be created for this |
5508 | group. */ | |
5509 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 5510 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
52eab378 | 5511 | gcc_assert (GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_stmt)) == first_stmt); |
4b5caab7 | 5512 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
d092494c | 5513 | op = gimple_assign_rhs1 (first_stmt); |
4b5caab7 | 5514 | } |
ebfd146a | 5515 | else |
4b5caab7 IR |
5516 | /* VEC_NUM is the number of vect stmts to be created for this |
5517 | group. */ | |
ebfd146a IR |
5518 | vec_num = group_size; |
5519 | } | |
b8698a0f | 5520 | else |
ebfd146a IR |
5521 | { |
5522 | first_stmt = stmt; | |
5523 | first_dr = dr; | |
5524 | group_size = vec_num = 1; | |
ebfd146a | 5525 | } |
b8698a0f | 5526 | |
73fbfcad | 5527 | if (dump_enabled_p ()) |
78c60e3d | 5528 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5529 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 5530 | |
f2e2a985 MM |
5531 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
5532 | { | |
5533 | gimple_stmt_iterator incr_gsi; | |
5534 | bool insert_after; | |
355fe088 | 5535 | gimple *incr; |
f2e2a985 MM |
5536 | tree offvar; |
5537 | tree ivstep; | |
5538 | tree running_off; | |
5539 | gimple_seq stmts = NULL; | |
5540 | tree stride_base, stride_step, alias_off; | |
5541 | tree vec_oprnd; | |
f502d50e | 5542 | unsigned int g; |
f2e2a985 MM |
5543 | |
5544 | gcc_assert (!nested_in_vect_loop_p (loop, stmt)); | |
5545 | ||
5546 | stride_base | |
5547 | = fold_build_pointer_plus | |
f502d50e | 5548 | (unshare_expr (DR_BASE_ADDRESS (first_dr)), |
f2e2a985 | 5549 | size_binop (PLUS_EXPR, |
f502d50e MM |
5550 | convert_to_ptrofftype (unshare_expr (DR_OFFSET (first_dr))), |
5551 | convert_to_ptrofftype (DR_INIT(first_dr)))); | |
5552 | stride_step = fold_convert (sizetype, unshare_expr (DR_STEP (first_dr))); | |
f2e2a985 MM |
5553 | |
5554 | /* For a store with loop-invariant (but other than power-of-2) | |
5555 | stride (i.e. not a grouped access) like so: | |
5556 | ||
5557 | for (i = 0; i < n; i += stride) | |
5558 | array[i] = ...; | |
5559 | ||
5560 | we generate a new induction variable and new stores from | |
5561 | the components of the (vectorized) rhs: | |
5562 | ||
5563 | for (j = 0; ; j += VF*stride) | |
5564 | vectemp = ...; | |
5565 | tmp1 = vectemp[0]; | |
5566 | array[j] = tmp1; | |
5567 | tmp2 = vectemp[1]; | |
5568 | array[j + stride] = tmp2; | |
5569 | ... | |
5570 | */ | |
5571 | ||
cee62fee MM |
5572 | unsigned nstores = nunits; |
5573 | tree ltype = elem_type; | |
5574 | if (slp) | |
5575 | { | |
5576 | nstores = nunits / group_size; | |
5577 | if (group_size < nunits) | |
5578 | ltype = build_vector_type (elem_type, group_size); | |
5579 | else | |
5580 | ltype = vectype; | |
5581 | ltype = build_aligned_type (ltype, TYPE_ALIGN (elem_type)); | |
5582 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
f502d50e | 5583 | group_size = 1; |
cee62fee MM |
5584 | } |
5585 | ||
f2e2a985 MM |
5586 | ivstep = stride_step; |
5587 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
5588 | build_int_cst (TREE_TYPE (ivstep), | |
cee62fee | 5589 | ncopies * nstores)); |
f2e2a985 MM |
5590 | |
5591 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
5592 | ||
5593 | create_iv (stride_base, ivstep, NULL, | |
5594 | loop, &incr_gsi, insert_after, | |
5595 | &offvar, NULL); | |
5596 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 5597 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
f2e2a985 MM |
5598 | |
5599 | stride_step = force_gimple_operand (stride_step, &stmts, true, NULL_TREE); | |
5600 | if (stmts) | |
5601 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
5602 | ||
5603 | prev_stmt_info = NULL; | |
f502d50e MM |
5604 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (first_dr)), 0); |
5605 | next_stmt = first_stmt; | |
5606 | for (g = 0; g < group_size; g++) | |
f2e2a985 | 5607 | { |
f502d50e MM |
5608 | running_off = offvar; |
5609 | if (g) | |
f2e2a985 | 5610 | { |
f502d50e MM |
5611 | tree size = TYPE_SIZE_UNIT (ltype); |
5612 | tree pos = fold_build2 (MULT_EXPR, sizetype, size_int (g), | |
f2e2a985 | 5613 | size); |
f502d50e | 5614 | tree newoff = copy_ssa_name (running_off, NULL); |
f2e2a985 | 5615 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
f502d50e | 5616 | running_off, pos); |
f2e2a985 | 5617 | vect_finish_stmt_generation (stmt, incr, gsi); |
f2e2a985 | 5618 | running_off = newoff; |
f502d50e MM |
5619 | } |
5620 | for (j = 0; j < ncopies; j++) | |
5621 | { | |
5622 | /* We've set op and dt above, from gimple_assign_rhs1(stmt), | |
5623 | and first_stmt == stmt. */ | |
5624 | if (j == 0) | |
5625 | { | |
5626 | if (slp) | |
5627 | { | |
5628 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, NULL, | |
5629 | slp_node, -1); | |
5630 | vec_oprnd = vec_oprnds[0]; | |
5631 | } | |
5632 | else | |
5633 | { | |
5634 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
5635 | op = gimple_assign_rhs1 (next_stmt); | |
81c40241 | 5636 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
f502d50e MM |
5637 | } |
5638 | } | |
f2e2a985 | 5639 | else |
f502d50e MM |
5640 | { |
5641 | if (slp) | |
5642 | vec_oprnd = vec_oprnds[j]; | |
5643 | else | |
c079cbac | 5644 | { |
81c40241 | 5645 | vect_is_simple_use (vec_oprnd, vinfo, &def_stmt, &dt); |
c079cbac RB |
5646 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); |
5647 | } | |
f502d50e MM |
5648 | } |
5649 | ||
5650 | for (i = 0; i < nstores; i++) | |
5651 | { | |
5652 | tree newref, newoff; | |
355fe088 | 5653 | gimple *incr, *assign; |
f502d50e MM |
5654 | tree size = TYPE_SIZE (ltype); |
5655 | /* Extract the i'th component. */ | |
5656 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, | |
5657 | bitsize_int (i), size); | |
5658 | tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd, | |
5659 | size, pos); | |
5660 | ||
5661 | elem = force_gimple_operand_gsi (gsi, elem, true, | |
5662 | NULL_TREE, true, | |
5663 | GSI_SAME_STMT); | |
5664 | ||
5665 | newref = build2 (MEM_REF, ltype, | |
5666 | running_off, alias_off); | |
5667 | ||
5668 | /* And store it to *running_off. */ | |
5669 | assign = gimple_build_assign (newref, elem); | |
5670 | vect_finish_stmt_generation (stmt, assign, gsi); | |
5671 | ||
5672 | newoff = copy_ssa_name (running_off, NULL); | |
5673 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
5674 | running_off, stride_step); | |
5675 | vect_finish_stmt_generation (stmt, incr, gsi); | |
5676 | ||
5677 | running_off = newoff; | |
225ce44b RB |
5678 | if (g == group_size - 1 |
5679 | && !slp) | |
f502d50e MM |
5680 | { |
5681 | if (j == 0 && i == 0) | |
225ce44b RB |
5682 | STMT_VINFO_VEC_STMT (stmt_info) |
5683 | = *vec_stmt = assign; | |
f502d50e MM |
5684 | else |
5685 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = assign; | |
5686 | prev_stmt_info = vinfo_for_stmt (assign); | |
5687 | } | |
5688 | } | |
f2e2a985 | 5689 | } |
f502d50e | 5690 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
f2e2a985 MM |
5691 | } |
5692 | return true; | |
5693 | } | |
5694 | ||
9771b263 DN |
5695 | dr_chain.create (group_size); |
5696 | oprnds.create (group_size); | |
ebfd146a | 5697 | |
720f5239 | 5698 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 5699 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
5700 | /* Targets with store-lane instructions must not require explicit |
5701 | realignment. */ | |
5702 | gcc_assert (!store_lanes_p | |
5703 | || alignment_support_scheme == dr_aligned | |
5704 | || alignment_support_scheme == dr_unaligned_supported); | |
5705 | ||
09dfa495 BM |
5706 | if (negative) |
5707 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
5708 | ||
272c6793 RS |
5709 | if (store_lanes_p) |
5710 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
5711 | else | |
5712 | aggr_type = vectype; | |
ebfd146a IR |
5713 | |
5714 | /* In case the vectorization factor (VF) is bigger than the number | |
5715 | of elements that we can fit in a vectype (nunits), we have to generate | |
5716 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 5717 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
5718 | vect_get_vec_def_for_copy_stmt. */ |
5719 | ||
0d0293ac | 5720 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
5721 | |
5722 | S1: &base + 2 = x2 | |
5723 | S2: &base = x0 | |
5724 | S3: &base + 1 = x1 | |
5725 | S4: &base + 3 = x3 | |
5726 | ||
5727 | We create vectorized stores starting from base address (the access of the | |
5728 | first stmt in the chain (S2 in the above example), when the last store stmt | |
5729 | of the chain (S4) is reached: | |
5730 | ||
5731 | VS1: &base = vx2 | |
5732 | VS2: &base + vec_size*1 = vx0 | |
5733 | VS3: &base + vec_size*2 = vx1 | |
5734 | VS4: &base + vec_size*3 = vx3 | |
5735 | ||
5736 | Then permutation statements are generated: | |
5737 | ||
3fcc1b55 JJ |
5738 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
5739 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 5740 | ... |
b8698a0f | 5741 | |
ebfd146a IR |
5742 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
5743 | (the order of the data-refs in the output of vect_permute_store_chain | |
5744 | corresponds to the order of scalar stmts in the interleaving chain - see | |
5745 | the documentation of vect_permute_store_chain()). | |
5746 | ||
5747 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 5748 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 5749 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 5750 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
5751 | */ |
5752 | ||
5753 | prev_stmt_info = NULL; | |
5754 | for (j = 0; j < ncopies; j++) | |
5755 | { | |
ebfd146a IR |
5756 | |
5757 | if (j == 0) | |
5758 | { | |
5759 | if (slp) | |
5760 | { | |
5761 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c IR |
5762 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
5763 | NULL, slp_node, -1); | |
ebfd146a | 5764 | |
9771b263 | 5765 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
5766 | } |
5767 | else | |
5768 | { | |
b8698a0f L |
5769 | /* For interleaved stores we collect vectorized defs for all the |
5770 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
5771 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
5772 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
5773 | ||
0d0293ac | 5774 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 5775 | OPRNDS are of size 1. */ |
b8698a0f | 5776 | next_stmt = first_stmt; |
ebfd146a IR |
5777 | for (i = 0; i < group_size; i++) |
5778 | { | |
b8698a0f L |
5779 | /* Since gaps are not supported for interleaved stores, |
5780 | GROUP_SIZE is the exact number of stmts in the chain. | |
5781 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
5782 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
5783 | iteration of the loop will be executed. */ |
5784 | gcc_assert (next_stmt | |
5785 | && gimple_assign_single_p (next_stmt)); | |
5786 | op = gimple_assign_rhs1 (next_stmt); | |
5787 | ||
81c40241 | 5788 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
9771b263 DN |
5789 | dr_chain.quick_push (vec_oprnd); |
5790 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 5791 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
5792 | } |
5793 | } | |
5794 | ||
5795 | /* We should have catched mismatched types earlier. */ | |
5796 | gcc_assert (useless_type_conversion_p (vectype, | |
5797 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
5798 | bool simd_lane_access_p |
5799 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
5800 | if (simd_lane_access_p | |
5801 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
5802 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
5803 | && integer_zerop (DR_OFFSET (first_dr)) | |
5804 | && integer_zerop (DR_INIT (first_dr)) | |
5805 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
5806 | get_alias_set (DR_REF (first_dr)))) | |
5807 | { | |
5808 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
5809 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
5810 | (DR_REF (first_dr)), 0); | |
8928eff3 | 5811 | inv_p = false; |
74bf76ed JJ |
5812 | } |
5813 | else | |
5814 | dataref_ptr | |
5815 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
5816 | simd_lane_access_p ? loop : NULL, | |
09dfa495 | 5817 | offset, &dummy, gsi, &ptr_incr, |
74bf76ed | 5818 | simd_lane_access_p, &inv_p); |
a70d6342 | 5819 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 5820 | } |
b8698a0f | 5821 | else |
ebfd146a | 5822 | { |
b8698a0f L |
5823 | /* For interleaved stores we created vectorized defs for all the |
5824 | defs stored in OPRNDS in the previous iteration (previous copy). | |
5825 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
5826 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
5827 | next copy. | |
0d0293ac | 5828 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
5829 | OPRNDS are of size 1. */ |
5830 | for (i = 0; i < group_size; i++) | |
5831 | { | |
9771b263 | 5832 | op = oprnds[i]; |
81c40241 | 5833 | vect_is_simple_use (op, vinfo, &def_stmt, &dt); |
b8698a0f | 5834 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
9771b263 DN |
5835 | dr_chain[i] = vec_oprnd; |
5836 | oprnds[i] = vec_oprnd; | |
ebfd146a | 5837 | } |
74bf76ed JJ |
5838 | if (dataref_offset) |
5839 | dataref_offset | |
5840 | = int_const_binop (PLUS_EXPR, dataref_offset, | |
5841 | TYPE_SIZE_UNIT (aggr_type)); | |
5842 | else | |
5843 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
5844 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
5845 | } |
5846 | ||
272c6793 | 5847 | if (store_lanes_p) |
ebfd146a | 5848 | { |
272c6793 | 5849 | tree vec_array; |
267d3070 | 5850 | |
272c6793 RS |
5851 | /* Combine all the vectors into an array. */ |
5852 | vec_array = create_vector_array (vectype, vec_num); | |
5853 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 5854 | { |
9771b263 | 5855 | vec_oprnd = dr_chain[i]; |
272c6793 | 5856 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 5857 | } |
b8698a0f | 5858 | |
272c6793 RS |
5859 | /* Emit: |
5860 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
5861 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
5862 | new_stmt = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array); | |
5863 | gimple_call_set_lhs (new_stmt, data_ref); | |
267d3070 | 5864 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
5865 | } |
5866 | else | |
5867 | { | |
5868 | new_stmt = NULL; | |
0d0293ac | 5869 | if (grouped_store) |
272c6793 | 5870 | { |
b6b9227d JJ |
5871 | if (j == 0) |
5872 | result_chain.create (group_size); | |
272c6793 RS |
5873 | /* Permute. */ |
5874 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
5875 | &result_chain); | |
5876 | } | |
c2d7ab2a | 5877 | |
272c6793 RS |
5878 | next_stmt = first_stmt; |
5879 | for (i = 0; i < vec_num; i++) | |
5880 | { | |
644ffefd | 5881 | unsigned align, misalign; |
272c6793 RS |
5882 | |
5883 | if (i > 0) | |
5884 | /* Bump the vector pointer. */ | |
5885 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
5886 | stmt, NULL_TREE); | |
5887 | ||
5888 | if (slp) | |
9771b263 | 5889 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
5890 | else if (grouped_store) |
5891 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 5892 | vect_permute_store_chain(). */ |
9771b263 | 5893 | vec_oprnd = result_chain[i]; |
272c6793 | 5894 | |
aed93b23 RB |
5895 | data_ref = fold_build2 (MEM_REF, TREE_TYPE (vec_oprnd), |
5896 | dataref_ptr, | |
5897 | dataref_offset | |
5898 | ? dataref_offset | |
5899 | : build_int_cst (reference_alias_ptr_type | |
5900 | (DR_REF (first_dr)), 0)); | |
644ffefd | 5901 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 | 5902 | if (aligned_access_p (first_dr)) |
644ffefd | 5903 | misalign = 0; |
272c6793 RS |
5904 | else if (DR_MISALIGNMENT (first_dr) == -1) |
5905 | { | |
52639a61 RB |
5906 | if (DR_VECT_AUX (first_dr)->base_element_aligned) |
5907 | align = TYPE_ALIGN_UNIT (elem_type); | |
5908 | else | |
5909 | align = get_object_alignment (DR_REF (first_dr)) | |
5910 | / BITS_PER_UNIT; | |
5911 | misalign = 0; | |
272c6793 RS |
5912 | TREE_TYPE (data_ref) |
5913 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 5914 | align * BITS_PER_UNIT); |
272c6793 RS |
5915 | } |
5916 | else | |
5917 | { | |
5918 | TREE_TYPE (data_ref) | |
5919 | = build_aligned_type (TREE_TYPE (data_ref), | |
5920 | TYPE_ALIGN (elem_type)); | |
644ffefd | 5921 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 5922 | } |
aed93b23 RB |
5923 | if (dataref_offset == NULL_TREE |
5924 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
5925 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
5926 | misalign); | |
c2d7ab2a | 5927 | |
f234d260 BM |
5928 | if (negative |
5929 | && dt != vect_constant_def | |
5930 | && dt != vect_external_def) | |
09dfa495 BM |
5931 | { |
5932 | tree perm_mask = perm_mask_for_reverse (vectype); | |
5933 | tree perm_dest | |
5934 | = vect_create_destination_var (gimple_assign_rhs1 (stmt), | |
5935 | vectype); | |
b731b390 | 5936 | tree new_temp = make_ssa_name (perm_dest); |
09dfa495 BM |
5937 | |
5938 | /* Generate the permute statement. */ | |
355fe088 | 5939 | gimple *perm_stmt |
0d0e4a03 JJ |
5940 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
5941 | vec_oprnd, perm_mask); | |
09dfa495 BM |
5942 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
5943 | ||
5944 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); | |
5945 | vec_oprnd = new_temp; | |
5946 | } | |
5947 | ||
272c6793 RS |
5948 | /* Arguments are ready. Create the new vector stmt. */ |
5949 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
5950 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
5951 | |
5952 | if (slp) | |
5953 | continue; | |
5954 | ||
e14c1050 | 5955 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
5956 | if (!next_stmt) |
5957 | break; | |
5958 | } | |
ebfd146a | 5959 | } |
1da0876c RS |
5960 | if (!slp) |
5961 | { | |
5962 | if (j == 0) | |
5963 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5964 | else | |
5965 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5966 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5967 | } | |
ebfd146a IR |
5968 | } |
5969 | ||
9771b263 DN |
5970 | dr_chain.release (); |
5971 | oprnds.release (); | |
5972 | result_chain.release (); | |
5973 | vec_oprnds.release (); | |
ebfd146a IR |
5974 | |
5975 | return true; | |
5976 | } | |
5977 | ||
557be5a8 AL |
5978 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
5979 | VECTOR_CST mask. No checks are made that the target platform supports the | |
5980 | mask, so callers may wish to test can_vec_perm_p separately, or use | |
5981 | vect_gen_perm_mask_checked. */ | |
a1e53f3f | 5982 | |
3fcc1b55 | 5983 | tree |
557be5a8 | 5984 | vect_gen_perm_mask_any (tree vectype, const unsigned char *sel) |
a1e53f3f | 5985 | { |
d2a12ae7 | 5986 | tree mask_elt_type, mask_type, mask_vec, *mask_elts; |
2635892a | 5987 | int i, nunits; |
a1e53f3f | 5988 | |
22e4dee7 | 5989 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
22e4dee7 | 5990 | |
96f9265a RG |
5991 | mask_elt_type = lang_hooks.types.type_for_mode |
5992 | (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype))), 1); | |
22e4dee7 | 5993 | mask_type = get_vectype_for_scalar_type (mask_elt_type); |
a1e53f3f | 5994 | |
d2a12ae7 | 5995 | mask_elts = XALLOCAVEC (tree, nunits); |
aec7ae7d | 5996 | for (i = nunits - 1; i >= 0; i--) |
d2a12ae7 RG |
5997 | mask_elts[i] = build_int_cst (mask_elt_type, sel[i]); |
5998 | mask_vec = build_vector (mask_type, mask_elts); | |
a1e53f3f | 5999 | |
2635892a | 6000 | return mask_vec; |
a1e53f3f L |
6001 | } |
6002 | ||
cf7aa6a3 AL |
6003 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_p, |
6004 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ | |
557be5a8 AL |
6005 | |
6006 | tree | |
6007 | vect_gen_perm_mask_checked (tree vectype, const unsigned char *sel) | |
6008 | { | |
6009 | gcc_assert (can_vec_perm_p (TYPE_MODE (vectype), false, sel)); | |
6010 | return vect_gen_perm_mask_any (vectype, sel); | |
6011 | } | |
6012 | ||
aec7ae7d JJ |
6013 | /* Given a vector variable X and Y, that was generated for the scalar |
6014 | STMT, generate instructions to permute the vector elements of X and Y | |
6015 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
6016 | permuted vector variable. */ | |
a1e53f3f L |
6017 | |
6018 | static tree | |
355fe088 | 6019 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple *stmt, |
aec7ae7d | 6020 | gimple_stmt_iterator *gsi) |
a1e53f3f L |
6021 | { |
6022 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 6023 | tree perm_dest, data_ref; |
355fe088 | 6024 | gimple *perm_stmt; |
a1e53f3f | 6025 | |
acdcd61b | 6026 | perm_dest = vect_create_destination_var (gimple_get_lhs (stmt), vectype); |
b731b390 | 6027 | data_ref = make_ssa_name (perm_dest); |
a1e53f3f L |
6028 | |
6029 | /* Generate the permute statement. */ | |
0d0e4a03 | 6030 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
a1e53f3f L |
6031 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
6032 | ||
6033 | return data_ref; | |
6034 | } | |
6035 | ||
6b916b36 RB |
6036 | /* Hoist the definitions of all SSA uses on STMT out of the loop LOOP, |
6037 | inserting them on the loops preheader edge. Returns true if we | |
6038 | were successful in doing so (and thus STMT can be moved then), | |
6039 | otherwise returns false. */ | |
6040 | ||
6041 | static bool | |
355fe088 | 6042 | hoist_defs_of_uses (gimple *stmt, struct loop *loop) |
6b916b36 RB |
6043 | { |
6044 | ssa_op_iter i; | |
6045 | tree op; | |
6046 | bool any = false; | |
6047 | ||
6048 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6049 | { | |
355fe088 | 6050 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6051 | if (!gimple_nop_p (def_stmt) |
6052 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6053 | { | |
6054 | /* Make sure we don't need to recurse. While we could do | |
6055 | so in simple cases when there are more complex use webs | |
6056 | we don't have an easy way to preserve stmt order to fulfil | |
6057 | dependencies within them. */ | |
6058 | tree op2; | |
6059 | ssa_op_iter i2; | |
d1417442 JJ |
6060 | if (gimple_code (def_stmt) == GIMPLE_PHI) |
6061 | return false; | |
6b916b36 RB |
6062 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
6063 | { | |
355fe088 | 6064 | gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2); |
6b916b36 RB |
6065 | if (!gimple_nop_p (def_stmt2) |
6066 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2))) | |
6067 | return false; | |
6068 | } | |
6069 | any = true; | |
6070 | } | |
6071 | } | |
6072 | ||
6073 | if (!any) | |
6074 | return true; | |
6075 | ||
6076 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6077 | { | |
355fe088 | 6078 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6079 | if (!gimple_nop_p (def_stmt) |
6080 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6081 | { | |
6082 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); | |
6083 | gsi_remove (&gsi, false); | |
6084 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); | |
6085 | } | |
6086 | } | |
6087 | ||
6088 | return true; | |
6089 | } | |
6090 | ||
ebfd146a IR |
6091 | /* vectorizable_load. |
6092 | ||
b8698a0f L |
6093 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
6094 | can be vectorized. | |
6095 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
6096 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
6097 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
6098 | ||
6099 | static bool | |
355fe088 | 6100 | vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 6101 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a IR |
6102 | { |
6103 | tree scalar_dest; | |
6104 | tree vec_dest = NULL; | |
6105 | tree data_ref = NULL; | |
6106 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 6107 | stmt_vec_info prev_stmt_info; |
ebfd146a | 6108 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 6109 | struct loop *loop = NULL; |
ebfd146a | 6110 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 6111 | bool nested_in_vect_loop = false; |
c716e67f | 6112 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
272c6793 | 6113 | tree elem_type; |
ebfd146a | 6114 | tree new_temp; |
ef4bddc2 | 6115 | machine_mode mode; |
355fe088 | 6116 | gimple *new_stmt = NULL; |
ebfd146a IR |
6117 | tree dummy; |
6118 | enum dr_alignment_support alignment_support_scheme; | |
6119 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 6120 | tree dataref_offset = NULL_TREE; |
355fe088 | 6121 | gimple *ptr_incr = NULL; |
ebfd146a | 6122 | int ncopies; |
9b999e8c | 6123 | int i, j, group_size = -1, group_gap_adj; |
ebfd146a IR |
6124 | tree msq = NULL_TREE, lsq; |
6125 | tree offset = NULL_TREE; | |
356bbc4c | 6126 | tree byte_offset = NULL_TREE; |
ebfd146a | 6127 | tree realignment_token = NULL_TREE; |
538dd0b7 | 6128 | gphi *phi = NULL; |
6e1aa848 | 6129 | vec<tree> dr_chain = vNULL; |
0d0293ac | 6130 | bool grouped_load = false; |
272c6793 | 6131 | bool load_lanes_p = false; |
355fe088 | 6132 | gimple *first_stmt; |
ebfd146a | 6133 | bool inv_p; |
319e6439 | 6134 | bool negative = false; |
ebfd146a IR |
6135 | bool compute_in_loop = false; |
6136 | struct loop *at_loop; | |
6137 | int vec_num; | |
6138 | bool slp = (slp_node != NULL); | |
6139 | bool slp_perm = false; | |
6140 | enum tree_code code; | |
a70d6342 IR |
6141 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
6142 | int vf; | |
272c6793 | 6143 | tree aggr_type; |
aec7ae7d JJ |
6144 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; |
6145 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
6146 | int gather_scale = 1; | |
6147 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
310213d4 | 6148 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 6149 | |
465c8c19 JJ |
6150 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
6151 | return false; | |
6152 | ||
6153 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
6154 | return false; | |
6155 | ||
6156 | /* Is vectorizable load? */ | |
6157 | if (!is_gimple_assign (stmt)) | |
6158 | return false; | |
6159 | ||
6160 | scalar_dest = gimple_assign_lhs (stmt); | |
6161 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
6162 | return false; | |
6163 | ||
6164 | code = gimple_assign_rhs_code (stmt); | |
6165 | if (code != ARRAY_REF | |
6166 | && code != BIT_FIELD_REF | |
6167 | && code != INDIRECT_REF | |
6168 | && code != COMPONENT_REF | |
6169 | && code != IMAGPART_EXPR | |
6170 | && code != REALPART_EXPR | |
6171 | && code != MEM_REF | |
6172 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
6173 | return false; | |
6174 | ||
6175 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
6176 | return false; | |
6177 | ||
6178 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
6179 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
6180 | ||
a70d6342 IR |
6181 | if (loop_vinfo) |
6182 | { | |
6183 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
6184 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
6185 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
6186 | } | |
6187 | else | |
3533e503 | 6188 | vf = 1; |
ebfd146a IR |
6189 | |
6190 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 6191 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 6192 | case of SLP. */ |
437f4a00 | 6193 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
6194 | ncopies = 1; |
6195 | else | |
6196 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
6197 | ||
6198 | gcc_assert (ncopies >= 1); | |
6199 | ||
6200 | /* FORNOW. This restriction should be relaxed. */ | |
6201 | if (nested_in_vect_loop && ncopies > 1) | |
6202 | { | |
73fbfcad | 6203 | if (dump_enabled_p ()) |
78c60e3d | 6204 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6205 | "multiple types in nested loop.\n"); |
ebfd146a IR |
6206 | return false; |
6207 | } | |
6208 | ||
f2556b68 RB |
6209 | /* Invalidate assumptions made by dependence analysis when vectorization |
6210 | on the unrolled body effectively re-orders stmts. */ | |
6211 | if (ncopies > 1 | |
6212 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
6213 | && ((unsigned)LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
6214 | > STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
6215 | { | |
6216 | if (dump_enabled_p ()) | |
6217 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6218 | "cannot perform implicit CSE when unrolling " | |
6219 | "with negative dependence distance\n"); | |
6220 | return false; | |
6221 | } | |
6222 | ||
7b7b1813 | 6223 | elem_type = TREE_TYPE (vectype); |
947131ba | 6224 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
6225 | |
6226 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
6227 | (e.g. - data copies). */ | |
947131ba | 6228 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 6229 | { |
73fbfcad | 6230 | if (dump_enabled_p ()) |
78c60e3d | 6231 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6232 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
6233 | return false; |
6234 | } | |
6235 | ||
ebfd146a | 6236 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 6237 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 6238 | { |
0d0293ac | 6239 | grouped_load = true; |
ebfd146a | 6240 | /* FORNOW */ |
3bab6342 | 6241 | gcc_assert (!nested_in_vect_loop && !STMT_VINFO_GATHER_SCATTER_P (stmt_info)); |
ebfd146a | 6242 | |
e14c1050 | 6243 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
d5f035ea RB |
6244 | |
6245 | /* If this is single-element interleaving with an element distance | |
6246 | that leaves unused vector loads around punt - we at least create | |
6247 | very sub-optimal code in that case (and blow up memory, | |
6248 | see PR65518). */ | |
6249 | if (first_stmt == stmt | |
6250 | && !GROUP_NEXT_ELEMENT (stmt_info) | |
6251 | && GROUP_SIZE (stmt_info) > TYPE_VECTOR_SUBPARTS (vectype)) | |
6252 | { | |
6253 | if (dump_enabled_p ()) | |
6254 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6255 | "single-element interleaving not supported " | |
6256 | "for not adjacent vector loads\n"); | |
6257 | return false; | |
6258 | } | |
6259 | ||
b1af7da6 RB |
6260 | if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
6261 | slp_perm = true; | |
6262 | ||
7b5fc413 RB |
6263 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
6264 | if (!slp | |
6265 | && !PURE_SLP_STMT (stmt_info) | |
f2e2a985 | 6266 | && !STMT_VINFO_STRIDED_P (stmt_info)) |
b602d918 | 6267 | { |
272c6793 RS |
6268 | if (vect_load_lanes_supported (vectype, group_size)) |
6269 | load_lanes_p = true; | |
0d0293ac | 6270 | else if (!vect_grouped_load_supported (vectype, group_size)) |
b602d918 RS |
6271 | return false; |
6272 | } | |
f2556b68 RB |
6273 | |
6274 | /* Invalidate assumptions made by dependence analysis when vectorization | |
6275 | on the unrolled body effectively re-orders stmts. */ | |
6276 | if (!PURE_SLP_STMT (stmt_info) | |
6277 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
6278 | && ((unsigned)LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
6279 | > STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
6280 | { | |
6281 | if (dump_enabled_p ()) | |
6282 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6283 | "cannot perform implicit CSE when performing " | |
6284 | "group loads with negative dependence distance\n"); | |
6285 | return false; | |
6286 | } | |
96bb56b2 RB |
6287 | |
6288 | /* Similarly when the stmt is a load that is both part of a SLP | |
6289 | instance and a loop vectorized stmt via the same-dr mechanism | |
6290 | we have to give up. */ | |
6291 | if (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info) | |
6292 | && (STMT_SLP_TYPE (stmt_info) | |
6293 | != STMT_SLP_TYPE (vinfo_for_stmt | |
6294 | (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info))))) | |
6295 | { | |
6296 | if (dump_enabled_p ()) | |
6297 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6298 | "conflicting SLP types for CSEd load\n"); | |
6299 | return false; | |
6300 | } | |
ebfd146a IR |
6301 | } |
6302 | ||
a1e53f3f | 6303 | |
3bab6342 | 6304 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
aec7ae7d | 6305 | { |
355fe088 | 6306 | gimple *def_stmt; |
3bab6342 AT |
6307 | gather_decl = vect_check_gather_scatter (stmt, loop_vinfo, &gather_base, |
6308 | &gather_off, &gather_scale); | |
aec7ae7d | 6309 | gcc_assert (gather_decl); |
81c40241 RB |
6310 | if (!vect_is_simple_use (gather_off, vinfo, &def_stmt, &gather_dt, |
6311 | &gather_off_vectype)) | |
aec7ae7d | 6312 | { |
73fbfcad | 6313 | if (dump_enabled_p ()) |
78c60e3d | 6314 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6315 | "gather index use not simple.\n"); |
aec7ae7d JJ |
6316 | return false; |
6317 | } | |
6318 | } | |
f2e2a985 | 6319 | else if (STMT_VINFO_STRIDED_P (stmt_info)) |
7b5fc413 RB |
6320 | { |
6321 | if ((grouped_load | |
6322 | && (slp || PURE_SLP_STMT (stmt_info))) | |
6323 | && (group_size > nunits | |
b266b968 | 6324 | || nunits % group_size != 0)) |
7b5fc413 RB |
6325 | { |
6326 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6327 | "unhandled strided group load\n"); | |
6328 | return false; | |
6329 | } | |
6330 | } | |
319e6439 RG |
6331 | else |
6332 | { | |
6333 | negative = tree_int_cst_compare (nested_in_vect_loop | |
6334 | ? STMT_VINFO_DR_STEP (stmt_info) | |
6335 | : DR_STEP (dr), | |
6336 | size_zero_node) < 0; | |
6337 | if (negative && ncopies > 1) | |
6338 | { | |
73fbfcad | 6339 | if (dump_enabled_p ()) |
78c60e3d | 6340 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6341 | "multiple types with negative step.\n"); |
319e6439 RG |
6342 | return false; |
6343 | } | |
6344 | ||
6345 | if (negative) | |
6346 | { | |
08940f33 RB |
6347 | if (grouped_load) |
6348 | { | |
6349 | if (dump_enabled_p ()) | |
6350 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 TJ |
6351 | "negative step for group load not supported" |
6352 | "\n"); | |
08940f33 RB |
6353 | return false; |
6354 | } | |
319e6439 RG |
6355 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
6356 | if (alignment_support_scheme != dr_aligned | |
6357 | && alignment_support_scheme != dr_unaligned_supported) | |
6358 | { | |
73fbfcad | 6359 | if (dump_enabled_p ()) |
78c60e3d | 6360 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6361 | "negative step but alignment required.\n"); |
319e6439 RG |
6362 | return false; |
6363 | } | |
6364 | if (!perm_mask_for_reverse (vectype)) | |
6365 | { | |
73fbfcad | 6366 | if (dump_enabled_p ()) |
78c60e3d | 6367 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
6368 | "negative step and reversing not supported." |
6369 | "\n"); | |
319e6439 RG |
6370 | return false; |
6371 | } | |
6372 | } | |
7d75abc8 | 6373 | } |
aec7ae7d | 6374 | |
ebfd146a IR |
6375 | if (!vec_stmt) /* transformation not required. */ |
6376 | { | |
6377 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
2e8ab70c RB |
6378 | /* The SLP costs are calculated during SLP analysis. */ |
6379 | if (!PURE_SLP_STMT (stmt_info)) | |
6380 | vect_model_load_cost (stmt_info, ncopies, load_lanes_p, | |
6381 | NULL, NULL, NULL); | |
ebfd146a IR |
6382 | return true; |
6383 | } | |
6384 | ||
73fbfcad | 6385 | if (dump_enabled_p ()) |
78c60e3d | 6386 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6387 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a IR |
6388 | |
6389 | /** Transform. **/ | |
6390 | ||
c716e67f XDL |
6391 | ensure_base_align (stmt_info, dr); |
6392 | ||
3bab6342 | 6393 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
aec7ae7d JJ |
6394 | { |
6395 | tree vec_oprnd0 = NULL_TREE, op; | |
6396 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
6397 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
d3c2fee0 | 6398 | tree ptr, mask, var, scale, merge, perm_mask = NULL_TREE, prev_res = NULL_TREE; |
aec7ae7d JJ |
6399 | edge pe = loop_preheader_edge (loop); |
6400 | gimple_seq seq; | |
6401 | basic_block new_bb; | |
6402 | enum { NARROW, NONE, WIDEN } modifier; | |
6403 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
6404 | ||
6405 | if (nunits == gather_off_nunits) | |
6406 | modifier = NONE; | |
6407 | else if (nunits == gather_off_nunits / 2) | |
6408 | { | |
6409 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
6410 | modifier = WIDEN; | |
6411 | ||
6412 | for (i = 0; i < gather_off_nunits; ++i) | |
6413 | sel[i] = i | nunits; | |
6414 | ||
557be5a8 | 6415 | perm_mask = vect_gen_perm_mask_checked (gather_off_vectype, sel); |
aec7ae7d JJ |
6416 | } |
6417 | else if (nunits == gather_off_nunits * 2) | |
6418 | { | |
6419 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
6420 | modifier = NARROW; | |
6421 | ||
6422 | for (i = 0; i < nunits; ++i) | |
6423 | sel[i] = i < gather_off_nunits | |
6424 | ? i : i + nunits - gather_off_nunits; | |
6425 | ||
557be5a8 | 6426 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); |
aec7ae7d JJ |
6427 | ncopies *= 2; |
6428 | } | |
6429 | else | |
6430 | gcc_unreachable (); | |
6431 | ||
6432 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); | |
6433 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6434 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6435 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6436 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6437 | scaletype = TREE_VALUE (arglist); | |
d3c2fee0 | 6438 | gcc_checking_assert (types_compatible_p (srctype, rettype)); |
aec7ae7d JJ |
6439 | |
6440 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
6441 | ||
6442 | ptr = fold_convert (ptrtype, gather_base); | |
6443 | if (!is_gimple_min_invariant (ptr)) | |
6444 | { | |
6445 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
6446 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
6447 | gcc_assert (!new_bb); | |
6448 | } | |
6449 | ||
6450 | /* Currently we support only unconditional gather loads, | |
6451 | so mask should be all ones. */ | |
d3c2fee0 AI |
6452 | if (TREE_CODE (masktype) == INTEGER_TYPE) |
6453 | mask = build_int_cst (masktype, -1); | |
6454 | else if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
6455 | { | |
6456 | mask = build_int_cst (TREE_TYPE (masktype), -1); | |
6457 | mask = build_vector_from_val (masktype, mask); | |
03b9e8e4 | 6458 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
d3c2fee0 | 6459 | } |
aec7ae7d JJ |
6460 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) |
6461 | { | |
6462 | REAL_VALUE_TYPE r; | |
6463 | long tmp[6]; | |
6464 | for (j = 0; j < 6; ++j) | |
6465 | tmp[j] = -1; | |
6466 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
6467 | mask = build_real (TREE_TYPE (masktype), r); | |
d3c2fee0 | 6468 | mask = build_vector_from_val (masktype, mask); |
03b9e8e4 | 6469 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
aec7ae7d JJ |
6470 | } |
6471 | else | |
6472 | gcc_unreachable (); | |
aec7ae7d JJ |
6473 | |
6474 | scale = build_int_cst (scaletype, gather_scale); | |
6475 | ||
d3c2fee0 AI |
6476 | if (TREE_CODE (TREE_TYPE (rettype)) == INTEGER_TYPE) |
6477 | merge = build_int_cst (TREE_TYPE (rettype), 0); | |
6478 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (rettype))) | |
6479 | { | |
6480 | REAL_VALUE_TYPE r; | |
6481 | long tmp[6]; | |
6482 | for (j = 0; j < 6; ++j) | |
6483 | tmp[j] = 0; | |
6484 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (rettype))); | |
6485 | merge = build_real (TREE_TYPE (rettype), r); | |
6486 | } | |
6487 | else | |
6488 | gcc_unreachable (); | |
6489 | merge = build_vector_from_val (rettype, merge); | |
6490 | merge = vect_init_vector (stmt, merge, rettype, NULL); | |
6491 | ||
aec7ae7d JJ |
6492 | prev_stmt_info = NULL; |
6493 | for (j = 0; j < ncopies; ++j) | |
6494 | { | |
6495 | if (modifier == WIDEN && (j & 1)) | |
6496 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
6497 | perm_mask, stmt, gsi); | |
6498 | else if (j == 0) | |
6499 | op = vec_oprnd0 | |
81c40241 | 6500 | = vect_get_vec_def_for_operand (gather_off, stmt); |
aec7ae7d JJ |
6501 | else |
6502 | op = vec_oprnd0 | |
6503 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
6504 | ||
6505 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
6506 | { | |
6507 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
6508 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 6509 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
aec7ae7d JJ |
6510 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
6511 | new_stmt | |
0d0e4a03 | 6512 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
6513 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6514 | op = var; | |
6515 | } | |
6516 | ||
6517 | new_stmt | |
d3c2fee0 | 6518 | = gimple_build_call (gather_decl, 5, merge, ptr, op, mask, scale); |
aec7ae7d JJ |
6519 | |
6520 | if (!useless_type_conversion_p (vectype, rettype)) | |
6521 | { | |
6522 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
6523 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 6524 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
aec7ae7d JJ |
6525 | gimple_call_set_lhs (new_stmt, op); |
6526 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 6527 | var = make_ssa_name (vec_dest); |
aec7ae7d JJ |
6528 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
6529 | new_stmt | |
0d0e4a03 | 6530 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
6531 | } |
6532 | else | |
6533 | { | |
6534 | var = make_ssa_name (vec_dest, new_stmt); | |
6535 | gimple_call_set_lhs (new_stmt, var); | |
6536 | } | |
6537 | ||
6538 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6539 | ||
6540 | if (modifier == NARROW) | |
6541 | { | |
6542 | if ((j & 1) == 0) | |
6543 | { | |
6544 | prev_res = var; | |
6545 | continue; | |
6546 | } | |
6547 | var = permute_vec_elements (prev_res, var, | |
6548 | perm_mask, stmt, gsi); | |
6549 | new_stmt = SSA_NAME_DEF_STMT (var); | |
6550 | } | |
6551 | ||
6552 | if (prev_stmt_info == NULL) | |
6553 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6554 | else | |
6555 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6556 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6557 | } | |
6558 | return true; | |
6559 | } | |
f2e2a985 | 6560 | else if (STMT_VINFO_STRIDED_P (stmt_info)) |
7d75abc8 MM |
6561 | { |
6562 | gimple_stmt_iterator incr_gsi; | |
6563 | bool insert_after; | |
355fe088 | 6564 | gimple *incr; |
7d75abc8 | 6565 | tree offvar; |
7d75abc8 MM |
6566 | tree ivstep; |
6567 | tree running_off; | |
9771b263 | 6568 | vec<constructor_elt, va_gc> *v = NULL; |
7d75abc8 | 6569 | gimple_seq stmts = NULL; |
14ac6aa2 RB |
6570 | tree stride_base, stride_step, alias_off; |
6571 | ||
6572 | gcc_assert (!nested_in_vect_loop); | |
7d75abc8 | 6573 | |
f502d50e | 6574 | if (slp && grouped_load) |
ab313a8c RB |
6575 | first_dr = STMT_VINFO_DATA_REF |
6576 | (vinfo_for_stmt (GROUP_FIRST_ELEMENT (stmt_info))); | |
6577 | else | |
6578 | first_dr = dr; | |
6579 | ||
14ac6aa2 RB |
6580 | stride_base |
6581 | = fold_build_pointer_plus | |
ab313a8c | 6582 | (DR_BASE_ADDRESS (first_dr), |
14ac6aa2 | 6583 | size_binop (PLUS_EXPR, |
ab313a8c RB |
6584 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
6585 | convert_to_ptrofftype (DR_INIT (first_dr)))); | |
6586 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); | |
7d75abc8 MM |
6587 | |
6588 | /* For a load with loop-invariant (but other than power-of-2) | |
6589 | stride (i.e. not a grouped access) like so: | |
6590 | ||
6591 | for (i = 0; i < n; i += stride) | |
6592 | ... = array[i]; | |
6593 | ||
6594 | we generate a new induction variable and new accesses to | |
6595 | form a new vector (or vectors, depending on ncopies): | |
6596 | ||
6597 | for (j = 0; ; j += VF*stride) | |
6598 | tmp1 = array[j]; | |
6599 | tmp2 = array[j + stride]; | |
6600 | ... | |
6601 | vectemp = {tmp1, tmp2, ...} | |
6602 | */ | |
6603 | ||
ab313a8c RB |
6604 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step, |
6605 | build_int_cst (TREE_TYPE (stride_step), vf)); | |
7d75abc8 MM |
6606 | |
6607 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
6608 | ||
ab313a8c | 6609 | create_iv (unshare_expr (stride_base), unshare_expr (ivstep), NULL, |
7d75abc8 MM |
6610 | loop, &incr_gsi, insert_after, |
6611 | &offvar, NULL); | |
6612 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 6613 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
7d75abc8 | 6614 | |
ab313a8c RB |
6615 | stride_step = force_gimple_operand (unshare_expr (stride_step), |
6616 | &stmts, true, NULL_TREE); | |
7d75abc8 MM |
6617 | if (stmts) |
6618 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
6619 | ||
6620 | prev_stmt_info = NULL; | |
6621 | running_off = offvar; | |
ab313a8c | 6622 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (first_dr)), 0); |
7b5fc413 RB |
6623 | int nloads = nunits; |
6624 | tree ltype = TREE_TYPE (vectype); | |
b266b968 | 6625 | auto_vec<tree> dr_chain; |
7b5fc413 RB |
6626 | if (slp) |
6627 | { | |
6628 | nloads = nunits / group_size; | |
6629 | if (group_size < nunits) | |
6630 | ltype = build_vector_type (TREE_TYPE (vectype), group_size); | |
6631 | else | |
6632 | ltype = vectype; | |
6633 | ltype = build_aligned_type (ltype, TYPE_ALIGN (TREE_TYPE (vectype))); | |
6634 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
b266b968 RB |
6635 | if (slp_perm) |
6636 | dr_chain.create (ncopies); | |
7b5fc413 | 6637 | } |
7d75abc8 MM |
6638 | for (j = 0; j < ncopies; j++) |
6639 | { | |
6640 | tree vec_inv; | |
6641 | ||
7b5fc413 RB |
6642 | if (nloads > 1) |
6643 | { | |
6644 | vec_alloc (v, nloads); | |
6645 | for (i = 0; i < nloads; i++) | |
6646 | { | |
6647 | tree newref, newoff; | |
355fe088 | 6648 | gimple *incr; |
7b5fc413 RB |
6649 | newref = build2 (MEM_REF, ltype, running_off, alias_off); |
6650 | ||
6651 | newref = force_gimple_operand_gsi (gsi, newref, true, | |
6652 | NULL_TREE, true, | |
6653 | GSI_SAME_STMT); | |
6654 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, newref); | |
6655 | newoff = copy_ssa_name (running_off); | |
6656 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
6657 | running_off, stride_step); | |
6658 | vect_finish_stmt_generation (stmt, incr, gsi); | |
6659 | ||
6660 | running_off = newoff; | |
6661 | } | |
6662 | ||
6663 | vec_inv = build_constructor (vectype, v); | |
6664 | new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi); | |
6665 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
6666 | } | |
6667 | else | |
7d75abc8 | 6668 | { |
7b5fc413 RB |
6669 | new_stmt = gimple_build_assign (make_ssa_name (ltype), |
6670 | build2 (MEM_REF, ltype, | |
6671 | running_off, alias_off)); | |
6672 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6673 | ||
6674 | tree newoff = copy_ssa_name (running_off); | |
355fe088 | 6675 | gimple *incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
0d0e4a03 | 6676 | running_off, stride_step); |
7d75abc8 MM |
6677 | vect_finish_stmt_generation (stmt, incr, gsi); |
6678 | ||
6679 | running_off = newoff; | |
6680 | } | |
6681 | ||
7b5fc413 | 6682 | if (slp) |
b266b968 RB |
6683 | { |
6684 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
6685 | if (slp_perm) | |
6686 | dr_chain.quick_push (gimple_assign_lhs (new_stmt)); | |
6687 | } | |
7d75abc8 | 6688 | else |
225ce44b RB |
6689 | { |
6690 | if (j == 0) | |
6691 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6692 | else | |
6693 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6694 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6695 | } | |
7d75abc8 | 6696 | } |
b266b968 RB |
6697 | if (slp_perm) |
6698 | vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, | |
6699 | slp_node_instance, false); | |
7d75abc8 MM |
6700 | return true; |
6701 | } | |
aec7ae7d | 6702 | |
0d0293ac | 6703 | if (grouped_load) |
ebfd146a | 6704 | { |
e14c1050 | 6705 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
52eab378 RB |
6706 | /* For BB vectorization we directly vectorize a subchain |
6707 | without permutation. */ | |
6708 | if (slp && ! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
9771b263 | 6709 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
6aa904c4 | 6710 | |
ebfd146a | 6711 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
6712 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
6713 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
6714 | ??? But we can only do so if there is exactly one | |
6715 | as we have no way to get at the rest. Leave the CSE | |
6716 | opportunity alone. | |
6717 | ??? With the group load eventually participating | |
6718 | in multiple different permutations (having multiple | |
6719 | slp nodes which refer to the same group) the CSE | |
6720 | is even wrong code. See PR56270. */ | |
6721 | && !slp) | |
ebfd146a IR |
6722 | { |
6723 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
6724 | return true; | |
6725 | } | |
6726 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 6727 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
9b999e8c | 6728 | group_gap_adj = 0; |
ebfd146a IR |
6729 | |
6730 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
6731 | if (slp) | |
6732 | { | |
0d0293ac | 6733 | grouped_load = false; |
91ff1504 RB |
6734 | /* For SLP permutation support we need to load the whole group, |
6735 | not only the number of vector stmts the permutation result | |
6736 | fits in. */ | |
6737 | if (slp_perm) | |
6738 | vec_num = (group_size * vf + nunits - 1) / nunits; | |
6739 | else | |
6740 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9b999e8c | 6741 | group_gap_adj = vf * group_size - nunits * vec_num; |
a70d6342 | 6742 | } |
ebfd146a | 6743 | else |
9b999e8c | 6744 | vec_num = group_size; |
ebfd146a IR |
6745 | } |
6746 | else | |
6747 | { | |
6748 | first_stmt = stmt; | |
6749 | first_dr = dr; | |
6750 | group_size = vec_num = 1; | |
9b999e8c | 6751 | group_gap_adj = 0; |
ebfd146a IR |
6752 | } |
6753 | ||
720f5239 | 6754 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 6755 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
6756 | /* Targets with load-lane instructions must not require explicit |
6757 | realignment. */ | |
6758 | gcc_assert (!load_lanes_p | |
6759 | || alignment_support_scheme == dr_aligned | |
6760 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
6761 | |
6762 | /* In case the vectorization factor (VF) is bigger than the number | |
6763 | of elements that we can fit in a vectype (nunits), we have to generate | |
6764 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 6765 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 6766 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 6767 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 6768 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
6769 | stmts that use the defs of the current stmt. The example below |
6770 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
6771 | need to create 4 vectorized stmts): | |
ebfd146a IR |
6772 | |
6773 | before vectorization: | |
6774 | RELATED_STMT VEC_STMT | |
6775 | S1: x = memref - - | |
6776 | S2: z = x + 1 - - | |
6777 | ||
6778 | step 1: vectorize stmt S1: | |
6779 | We first create the vector stmt VS1_0, and, as usual, record a | |
6780 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
6781 | Next, we create the vector stmt VS1_1, and record a pointer to | |
6782 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 6783 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
6784 | stmts and pointers: |
6785 | RELATED_STMT VEC_STMT | |
6786 | VS1_0: vx0 = memref0 VS1_1 - | |
6787 | VS1_1: vx1 = memref1 VS1_2 - | |
6788 | VS1_2: vx2 = memref2 VS1_3 - | |
6789 | VS1_3: vx3 = memref3 - - | |
6790 | S1: x = load - VS1_0 | |
6791 | S2: z = x + 1 - - | |
6792 | ||
b8698a0f L |
6793 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
6794 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
6795 | stmt S2. */ |
6796 | ||
0d0293ac | 6797 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
6798 | |
6799 | S1: x2 = &base + 2 | |
6800 | S2: x0 = &base | |
6801 | S3: x1 = &base + 1 | |
6802 | S4: x3 = &base + 3 | |
6803 | ||
b8698a0f | 6804 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
6805 | starting from the access of the first stmt of the chain: |
6806 | ||
6807 | VS1: vx0 = &base | |
6808 | VS2: vx1 = &base + vec_size*1 | |
6809 | VS3: vx3 = &base + vec_size*2 | |
6810 | VS4: vx4 = &base + vec_size*3 | |
6811 | ||
6812 | Then permutation statements are generated: | |
6813 | ||
e2c83630 RH |
6814 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
6815 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
6816 | ... |
6817 | ||
6818 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
6819 | (the order of the data-refs in the output of vect_permute_load_chain | |
6820 | corresponds to the order of scalar stmts in the interleaving chain - see | |
6821 | the documentation of vect_permute_load_chain()). | |
6822 | The generation of permutation stmts and recording them in | |
0d0293ac | 6823 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 6824 | |
b8698a0f | 6825 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
6826 | permutation stmts above are created for every copy. The result vector |
6827 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
6828 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
6829 | |
6830 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
6831 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
6832 | we generate the following code: | |
6833 | p = initial_addr; | |
6834 | indx = 0; | |
6835 | loop { | |
6836 | p = p + indx * vectype_size; | |
6837 | vec_dest = *(p); | |
6838 | indx = indx + 1; | |
6839 | } | |
6840 | ||
6841 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 6842 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
6843 | then generate the following code, in which the data in each iteration is |
6844 | obtained by two vector loads, one from the previous iteration, and one | |
6845 | from the current iteration: | |
6846 | p1 = initial_addr; | |
6847 | msq_init = *(floor(p1)) | |
6848 | p2 = initial_addr + VS - 1; | |
6849 | realignment_token = call target_builtin; | |
6850 | indx = 0; | |
6851 | loop { | |
6852 | p2 = p2 + indx * vectype_size | |
6853 | lsq = *(floor(p2)) | |
6854 | vec_dest = realign_load (msq, lsq, realignment_token) | |
6855 | indx = indx + 1; | |
6856 | msq = lsq; | |
6857 | } */ | |
6858 | ||
6859 | /* If the misalignment remains the same throughout the execution of the | |
6860 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 6861 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
6862 | This can only occur when vectorizing memory accesses in the inner-loop |
6863 | nested within an outer-loop that is being vectorized. */ | |
6864 | ||
d1e4b493 | 6865 | if (nested_in_vect_loop |
211bea38 | 6866 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
ebfd146a IR |
6867 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) |
6868 | { | |
6869 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
6870 | compute_in_loop = true; | |
6871 | } | |
6872 | ||
6873 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
6874 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 6875 | && !compute_in_loop) |
ebfd146a IR |
6876 | { |
6877 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
6878 | alignment_support_scheme, NULL_TREE, | |
6879 | &at_loop); | |
6880 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
6881 | { | |
538dd0b7 | 6882 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (msq)); |
356bbc4c JJ |
6883 | byte_offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
6884 | size_one_node); | |
ebfd146a IR |
6885 | } |
6886 | } | |
6887 | else | |
6888 | at_loop = loop; | |
6889 | ||
a1e53f3f L |
6890 | if (negative) |
6891 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
6892 | ||
272c6793 RS |
6893 | if (load_lanes_p) |
6894 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
6895 | else | |
6896 | aggr_type = vectype; | |
6897 | ||
ebfd146a IR |
6898 | prev_stmt_info = NULL; |
6899 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 6900 | { |
272c6793 | 6901 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 6902 | if (j == 0) |
74bf76ed JJ |
6903 | { |
6904 | bool simd_lane_access_p | |
6905 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
6906 | if (simd_lane_access_p | |
6907 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
6908 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
6909 | && integer_zerop (DR_OFFSET (first_dr)) | |
6910 | && integer_zerop (DR_INIT (first_dr)) | |
6911 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
6912 | get_alias_set (DR_REF (first_dr))) | |
6913 | && (alignment_support_scheme == dr_aligned | |
6914 | || alignment_support_scheme == dr_unaligned_supported)) | |
6915 | { | |
6916 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
6917 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
6918 | (DR_REF (first_dr)), 0); | |
8928eff3 | 6919 | inv_p = false; |
74bf76ed JJ |
6920 | } |
6921 | else | |
6922 | dataref_ptr | |
6923 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
6924 | offset, &dummy, gsi, &ptr_incr, | |
356bbc4c JJ |
6925 | simd_lane_access_p, &inv_p, |
6926 | byte_offset); | |
74bf76ed JJ |
6927 | } |
6928 | else if (dataref_offset) | |
6929 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
6930 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 6931 | else |
272c6793 RS |
6932 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
6933 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 6934 | |
0d0293ac | 6935 | if (grouped_load || slp_perm) |
9771b263 | 6936 | dr_chain.create (vec_num); |
5ce1ee7f | 6937 | |
272c6793 | 6938 | if (load_lanes_p) |
ebfd146a | 6939 | { |
272c6793 RS |
6940 | tree vec_array; |
6941 | ||
6942 | vec_array = create_vector_array (vectype, vec_num); | |
6943 | ||
6944 | /* Emit: | |
6945 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
6946 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
6947 | new_stmt = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
6948 | gimple_call_set_lhs (new_stmt, vec_array); | |
6949 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a | 6950 | |
272c6793 RS |
6951 | /* Extract each vector into an SSA_NAME. */ |
6952 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 6953 | { |
272c6793 RS |
6954 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
6955 | vec_array, i); | |
9771b263 | 6956 | dr_chain.quick_push (new_temp); |
272c6793 RS |
6957 | } |
6958 | ||
6959 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 6960 | vect_record_grouped_load_vectors (stmt, dr_chain); |
272c6793 RS |
6961 | } |
6962 | else | |
6963 | { | |
6964 | for (i = 0; i < vec_num; i++) | |
6965 | { | |
6966 | if (i > 0) | |
6967 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
6968 | stmt, NULL_TREE); | |
6969 | ||
6970 | /* 2. Create the vector-load in the loop. */ | |
6971 | switch (alignment_support_scheme) | |
6972 | { | |
6973 | case dr_aligned: | |
6974 | case dr_unaligned_supported: | |
be1ac4ec | 6975 | { |
644ffefd MJ |
6976 | unsigned int align, misalign; |
6977 | ||
272c6793 | 6978 | data_ref |
aed93b23 RB |
6979 | = fold_build2 (MEM_REF, vectype, dataref_ptr, |
6980 | dataref_offset | |
6981 | ? dataref_offset | |
6982 | : build_int_cst (reference_alias_ptr_type | |
6983 | (DR_REF (first_dr)), 0)); | |
644ffefd | 6984 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 RS |
6985 | if (alignment_support_scheme == dr_aligned) |
6986 | { | |
6987 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 6988 | misalign = 0; |
272c6793 RS |
6989 | } |
6990 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
6991 | { | |
52639a61 RB |
6992 | if (DR_VECT_AUX (first_dr)->base_element_aligned) |
6993 | align = TYPE_ALIGN_UNIT (elem_type); | |
6994 | else | |
6995 | align = (get_object_alignment (DR_REF (first_dr)) | |
6996 | / BITS_PER_UNIT); | |
6997 | misalign = 0; | |
272c6793 RS |
6998 | TREE_TYPE (data_ref) |
6999 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 7000 | align * BITS_PER_UNIT); |
272c6793 RS |
7001 | } |
7002 | else | |
7003 | { | |
7004 | TREE_TYPE (data_ref) | |
7005 | = build_aligned_type (TREE_TYPE (data_ref), | |
7006 | TYPE_ALIGN (elem_type)); | |
644ffefd | 7007 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 7008 | } |
aed93b23 RB |
7009 | if (dataref_offset == NULL_TREE |
7010 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
7011 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), |
7012 | align, misalign); | |
272c6793 | 7013 | break; |
be1ac4ec | 7014 | } |
272c6793 | 7015 | case dr_explicit_realign: |
267d3070 | 7016 | { |
272c6793 | 7017 | tree ptr, bump; |
272c6793 | 7018 | |
d88981fc | 7019 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
272c6793 RS |
7020 | |
7021 | if (compute_in_loop) | |
7022 | msq = vect_setup_realignment (first_stmt, gsi, | |
7023 | &realignment_token, | |
7024 | dr_explicit_realign, | |
7025 | dataref_ptr, NULL); | |
7026 | ||
aed93b23 RB |
7027 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
7028 | ptr = copy_ssa_name (dataref_ptr); | |
7029 | else | |
7030 | ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
0d0e4a03 JJ |
7031 | new_stmt = gimple_build_assign |
7032 | (ptr, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
7033 | build_int_cst |
7034 | (TREE_TYPE (dataref_ptr), | |
7035 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
7036 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7037 | data_ref | |
7038 | = build2 (MEM_REF, vectype, ptr, | |
7039 | build_int_cst (reference_alias_ptr_type | |
7040 | (DR_REF (first_dr)), 0)); | |
7041 | vec_dest = vect_create_destination_var (scalar_dest, | |
7042 | vectype); | |
7043 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
7044 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
7045 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7046 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
7047 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
7048 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7049 | msq = new_temp; | |
7050 | ||
d88981fc | 7051 | bump = size_binop (MULT_EXPR, vs, |
7b7b1813 | 7052 | TYPE_SIZE_UNIT (elem_type)); |
d88981fc | 7053 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
272c6793 | 7054 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
0d0e4a03 JJ |
7055 | new_stmt = gimple_build_assign |
7056 | (NULL_TREE, BIT_AND_EXPR, ptr, | |
272c6793 RS |
7057 | build_int_cst |
7058 | (TREE_TYPE (ptr), | |
7059 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
aed93b23 | 7060 | ptr = copy_ssa_name (ptr, new_stmt); |
272c6793 RS |
7061 | gimple_assign_set_lhs (new_stmt, ptr); |
7062 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7063 | data_ref | |
7064 | = build2 (MEM_REF, vectype, ptr, | |
7065 | build_int_cst (reference_alias_ptr_type | |
7066 | (DR_REF (first_dr)), 0)); | |
7067 | break; | |
267d3070 | 7068 | } |
272c6793 | 7069 | case dr_explicit_realign_optimized: |
aed93b23 RB |
7070 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
7071 | new_temp = copy_ssa_name (dataref_ptr); | |
7072 | else | |
7073 | new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
0d0e4a03 JJ |
7074 | new_stmt = gimple_build_assign |
7075 | (new_temp, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
7076 | build_int_cst |
7077 | (TREE_TYPE (dataref_ptr), | |
7078 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
7079 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7080 | data_ref | |
7081 | = build2 (MEM_REF, vectype, new_temp, | |
7082 | build_int_cst (reference_alias_ptr_type | |
7083 | (DR_REF (first_dr)), 0)); | |
7084 | break; | |
7085 | default: | |
7086 | gcc_unreachable (); | |
7087 | } | |
ebfd146a | 7088 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 7089 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
7090 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7091 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7092 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7093 | ||
272c6793 RS |
7094 | /* 3. Handle explicit realignment if necessary/supported. |
7095 | Create in loop: | |
7096 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
7097 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
7098 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 7099 | { |
272c6793 RS |
7100 | lsq = gimple_assign_lhs (new_stmt); |
7101 | if (!realignment_token) | |
7102 | realignment_token = dataref_ptr; | |
7103 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
0d0e4a03 JJ |
7104 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, |
7105 | msq, lsq, realignment_token); | |
272c6793 RS |
7106 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7107 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7108 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7109 | ||
7110 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7111 | { | |
7112 | gcc_assert (phi); | |
7113 | if (i == vec_num - 1 && j == ncopies - 1) | |
7114 | add_phi_arg (phi, lsq, | |
7115 | loop_latch_edge (containing_loop), | |
9e227d60 | 7116 | UNKNOWN_LOCATION); |
272c6793 RS |
7117 | msq = lsq; |
7118 | } | |
ebfd146a | 7119 | } |
ebfd146a | 7120 | |
59fd17e3 RB |
7121 | /* 4. Handle invariant-load. */ |
7122 | if (inv_p && !bb_vinfo) | |
7123 | { | |
59fd17e3 | 7124 | gcc_assert (!grouped_load); |
d1417442 JJ |
7125 | /* If we have versioned for aliasing or the loop doesn't |
7126 | have any data dependencies that would preclude this, | |
7127 | then we are sure this is a loop invariant load and | |
7128 | thus we can insert it on the preheader edge. */ | |
7129 | if (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) | |
7130 | && !nested_in_vect_loop | |
6b916b36 | 7131 | && hoist_defs_of_uses (stmt, loop)) |
a0e35eb0 RB |
7132 | { |
7133 | if (dump_enabled_p ()) | |
7134 | { | |
7135 | dump_printf_loc (MSG_NOTE, vect_location, | |
7136 | "hoisting out of the vectorized " | |
7137 | "loop: "); | |
7138 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
a0e35eb0 | 7139 | } |
b731b390 | 7140 | tree tem = copy_ssa_name (scalar_dest); |
a0e35eb0 RB |
7141 | gsi_insert_on_edge_immediate |
7142 | (loop_preheader_edge (loop), | |
7143 | gimple_build_assign (tem, | |
7144 | unshare_expr | |
7145 | (gimple_assign_rhs1 (stmt)))); | |
7146 | new_temp = vect_init_vector (stmt, tem, vectype, NULL); | |
7147 | } | |
7148 | else | |
7149 | { | |
7150 | gimple_stmt_iterator gsi2 = *gsi; | |
7151 | gsi_next (&gsi2); | |
7152 | new_temp = vect_init_vector (stmt, scalar_dest, | |
7153 | vectype, &gsi2); | |
7154 | } | |
59fd17e3 | 7155 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
a0e35eb0 | 7156 | set_vinfo_for_stmt (new_stmt, |
310213d4 | 7157 | new_stmt_vec_info (new_stmt, vinfo)); |
59fd17e3 RB |
7158 | } |
7159 | ||
272c6793 RS |
7160 | if (negative) |
7161 | { | |
aec7ae7d JJ |
7162 | tree perm_mask = perm_mask_for_reverse (vectype); |
7163 | new_temp = permute_vec_elements (new_temp, new_temp, | |
7164 | perm_mask, stmt, gsi); | |
ebfd146a IR |
7165 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
7166 | } | |
267d3070 | 7167 | |
272c6793 | 7168 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
7169 | vect_transform_grouped_load (). */ |
7170 | if (grouped_load || slp_perm) | |
9771b263 | 7171 | dr_chain.quick_push (new_temp); |
267d3070 | 7172 | |
272c6793 RS |
7173 | /* Store vector loads in the corresponding SLP_NODE. */ |
7174 | if (slp && !slp_perm) | |
9771b263 | 7175 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
272c6793 | 7176 | } |
9b999e8c RB |
7177 | /* Bump the vector pointer to account for a gap or for excess |
7178 | elements loaded for a permuted SLP load. */ | |
7179 | if (group_gap_adj != 0) | |
a64b9c26 | 7180 | { |
9b999e8c RB |
7181 | bool ovf; |
7182 | tree bump | |
7183 | = wide_int_to_tree (sizetype, | |
7184 | wi::smul (TYPE_SIZE_UNIT (elem_type), | |
7185 | group_gap_adj, &ovf)); | |
a64b9c26 RB |
7186 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
7187 | stmt, bump); | |
7188 | } | |
ebfd146a IR |
7189 | } |
7190 | ||
7191 | if (slp && !slp_perm) | |
7192 | continue; | |
7193 | ||
7194 | if (slp_perm) | |
7195 | { | |
01d8bf07 | 7196 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
ebfd146a IR |
7197 | slp_node_instance, false)) |
7198 | { | |
9771b263 | 7199 | dr_chain.release (); |
ebfd146a IR |
7200 | return false; |
7201 | } | |
7202 | } | |
7203 | else | |
7204 | { | |
0d0293ac | 7205 | if (grouped_load) |
ebfd146a | 7206 | { |
272c6793 | 7207 | if (!load_lanes_p) |
0d0293ac | 7208 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 7209 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
7210 | } |
7211 | else | |
7212 | { | |
7213 | if (j == 0) | |
7214 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7215 | else | |
7216 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7217 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7218 | } | |
7219 | } | |
9771b263 | 7220 | dr_chain.release (); |
ebfd146a IR |
7221 | } |
7222 | ||
ebfd146a IR |
7223 | return true; |
7224 | } | |
7225 | ||
7226 | /* Function vect_is_simple_cond. | |
b8698a0f | 7227 | |
ebfd146a IR |
7228 | Input: |
7229 | LOOP - the loop that is being vectorized. | |
7230 | COND - Condition that is checked for simple use. | |
7231 | ||
e9e1d143 RG |
7232 | Output: |
7233 | *COMP_VECTYPE - the vector type for the comparison. | |
7234 | ||
ebfd146a IR |
7235 | Returns whether a COND can be vectorized. Checks whether |
7236 | condition operands are supportable using vec_is_simple_use. */ | |
7237 | ||
87aab9b2 | 7238 | static bool |
81c40241 | 7239 | vect_is_simple_cond (tree cond, vec_info *vinfo, tree *comp_vectype) |
ebfd146a IR |
7240 | { |
7241 | tree lhs, rhs; | |
ebfd146a | 7242 | enum vect_def_type dt; |
e9e1d143 | 7243 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a | 7244 | |
a414c77f IE |
7245 | /* Mask case. */ |
7246 | if (TREE_CODE (cond) == SSA_NAME | |
7247 | && TREE_CODE (TREE_TYPE (cond)) == BOOLEAN_TYPE) | |
7248 | { | |
7249 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (cond); | |
7250 | if (!vect_is_simple_use (cond, vinfo, &lhs_def_stmt, | |
7251 | &dt, comp_vectype) | |
7252 | || !*comp_vectype | |
7253 | || !VECTOR_BOOLEAN_TYPE_P (*comp_vectype)) | |
7254 | return false; | |
7255 | return true; | |
7256 | } | |
7257 | ||
ebfd146a IR |
7258 | if (!COMPARISON_CLASS_P (cond)) |
7259 | return false; | |
7260 | ||
7261 | lhs = TREE_OPERAND (cond, 0); | |
7262 | rhs = TREE_OPERAND (cond, 1); | |
7263 | ||
7264 | if (TREE_CODE (lhs) == SSA_NAME) | |
7265 | { | |
355fe088 | 7266 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); |
81c40241 | 7267 | if (!vect_is_simple_use (lhs, vinfo, &lhs_def_stmt, &dt, &vectype1)) |
ebfd146a IR |
7268 | return false; |
7269 | } | |
7270 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
7271 | && TREE_CODE (lhs) != FIXED_CST) | |
7272 | return false; | |
7273 | ||
7274 | if (TREE_CODE (rhs) == SSA_NAME) | |
7275 | { | |
355fe088 | 7276 | gimple *rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); |
81c40241 | 7277 | if (!vect_is_simple_use (rhs, vinfo, &rhs_def_stmt, &dt, &vectype2)) |
ebfd146a IR |
7278 | return false; |
7279 | } | |
f7e531cf | 7280 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST |
ebfd146a IR |
7281 | && TREE_CODE (rhs) != FIXED_CST) |
7282 | return false; | |
7283 | ||
e9e1d143 | 7284 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
ebfd146a IR |
7285 | return true; |
7286 | } | |
7287 | ||
7288 | /* vectorizable_condition. | |
7289 | ||
b8698a0f L |
7290 | Check if STMT is conditional modify expression that can be vectorized. |
7291 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
7292 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
7293 | at GSI. |
7294 | ||
7295 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
7296 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
0ad23163 | 7297 | else clause if it is 2). |
ebfd146a IR |
7298 | |
7299 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7300 | ||
4bbe8262 | 7301 | bool |
355fe088 TS |
7302 | vectorizable_condition (gimple *stmt, gimple_stmt_iterator *gsi, |
7303 | gimple **vec_stmt, tree reduc_def, int reduc_index, | |
f7e531cf | 7304 | slp_tree slp_node) |
ebfd146a IR |
7305 | { |
7306 | tree scalar_dest = NULL_TREE; | |
7307 | tree vec_dest = NULL_TREE; | |
ebfd146a IR |
7308 | tree cond_expr, then_clause, else_clause; |
7309 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
df11cc78 | 7310 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
7311 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
7312 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
ebfd146a IR |
7313 | tree vec_compare, vec_cond_expr; |
7314 | tree new_temp; | |
7315 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a855b1b1 | 7316 | enum vect_def_type dt, dts[4]; |
f7e531cf | 7317 | int ncopies; |
ebfd146a | 7318 | enum tree_code code; |
a855b1b1 | 7319 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
7320 | int i, j; |
7321 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
7322 | vec<tree> vec_oprnds0 = vNULL; |
7323 | vec<tree> vec_oprnds1 = vNULL; | |
7324 | vec<tree> vec_oprnds2 = vNULL; | |
7325 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 7326 | tree vec_cmp_type; |
a414c77f | 7327 | bool masked = false; |
b8698a0f | 7328 | |
f7e531cf IR |
7329 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
7330 | return false; | |
7331 | ||
af29617a AH |
7332 | if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) == TREE_CODE_REDUCTION) |
7333 | { | |
7334 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
7335 | return false; | |
ebfd146a | 7336 | |
af29617a AH |
7337 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7338 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
7339 | && reduc_def)) | |
7340 | return false; | |
ebfd146a | 7341 | |
af29617a AH |
7342 | /* FORNOW: not yet supported. */ |
7343 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
7344 | { | |
7345 | if (dump_enabled_p ()) | |
7346 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7347 | "value used after loop.\n"); | |
7348 | return false; | |
7349 | } | |
ebfd146a IR |
7350 | } |
7351 | ||
7352 | /* Is vectorizable conditional operation? */ | |
7353 | if (!is_gimple_assign (stmt)) | |
7354 | return false; | |
7355 | ||
7356 | code = gimple_assign_rhs_code (stmt); | |
7357 | ||
7358 | if (code != COND_EXPR) | |
7359 | return false; | |
7360 | ||
465c8c19 JJ |
7361 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
7362 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
7363 | ||
7364 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
7365 | ncopies = 1; | |
7366 | else | |
7367 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
7368 | ||
7369 | gcc_assert (ncopies >= 1); | |
7370 | if (reduc_index && ncopies > 1) | |
7371 | return false; /* FORNOW */ | |
7372 | ||
4e71066d RG |
7373 | cond_expr = gimple_assign_rhs1 (stmt); |
7374 | then_clause = gimple_assign_rhs2 (stmt); | |
7375 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 7376 | |
81c40241 | 7377 | if (!vect_is_simple_cond (cond_expr, stmt_info->vinfo, &comp_vectype) |
e9e1d143 | 7378 | || !comp_vectype) |
ebfd146a IR |
7379 | return false; |
7380 | ||
81c40241 RB |
7381 | gimple *def_stmt; |
7382 | if (!vect_is_simple_use (then_clause, stmt_info->vinfo, &def_stmt, &dt)) | |
ebfd146a | 7383 | return false; |
81c40241 | 7384 | if (!vect_is_simple_use (else_clause, stmt_info->vinfo, &def_stmt, &dt)) |
ebfd146a IR |
7385 | return false; |
7386 | ||
a414c77f IE |
7387 | if (VECTOR_BOOLEAN_TYPE_P (comp_vectype)) |
7388 | { | |
7389 | vec_cmp_type = comp_vectype; | |
7390 | masked = true; | |
7391 | } | |
7392 | else | |
7393 | vec_cmp_type = build_same_sized_truth_vector_type (comp_vectype); | |
74946978 MP |
7394 | if (vec_cmp_type == NULL_TREE) |
7395 | return false; | |
784fb9b3 | 7396 | |
b8698a0f | 7397 | if (!vec_stmt) |
ebfd146a IR |
7398 | { |
7399 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
e9e1d143 | 7400 | return expand_vec_cond_expr_p (vectype, comp_vectype); |
ebfd146a IR |
7401 | } |
7402 | ||
f7e531cf IR |
7403 | /* Transform. */ |
7404 | ||
7405 | if (!slp_node) | |
7406 | { | |
9771b263 DN |
7407 | vec_oprnds0.create (1); |
7408 | vec_oprnds1.create (1); | |
7409 | vec_oprnds2.create (1); | |
7410 | vec_oprnds3.create (1); | |
f7e531cf | 7411 | } |
ebfd146a IR |
7412 | |
7413 | /* Handle def. */ | |
7414 | scalar_dest = gimple_assign_lhs (stmt); | |
7415 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
7416 | ||
7417 | /* Handle cond expr. */ | |
a855b1b1 MM |
7418 | for (j = 0; j < ncopies; j++) |
7419 | { | |
538dd0b7 | 7420 | gassign *new_stmt = NULL; |
a855b1b1 MM |
7421 | if (j == 0) |
7422 | { | |
f7e531cf IR |
7423 | if (slp_node) |
7424 | { | |
00f96dc9 TS |
7425 | auto_vec<tree, 4> ops; |
7426 | auto_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 7427 | |
a414c77f IE |
7428 | if (masked) |
7429 | ops.safe_push (cond_expr); | |
7430 | else | |
7431 | { | |
7432 | ops.safe_push (TREE_OPERAND (cond_expr, 0)); | |
7433 | ops.safe_push (TREE_OPERAND (cond_expr, 1)); | |
7434 | } | |
9771b263 DN |
7435 | ops.safe_push (then_clause); |
7436 | ops.safe_push (else_clause); | |
f7e531cf | 7437 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); |
37b5ec8f JJ |
7438 | vec_oprnds3 = vec_defs.pop (); |
7439 | vec_oprnds2 = vec_defs.pop (); | |
a414c77f IE |
7440 | if (!masked) |
7441 | vec_oprnds1 = vec_defs.pop (); | |
37b5ec8f | 7442 | vec_oprnds0 = vec_defs.pop (); |
f7e531cf | 7443 | |
9771b263 DN |
7444 | ops.release (); |
7445 | vec_defs.release (); | |
f7e531cf IR |
7446 | } |
7447 | else | |
7448 | { | |
355fe088 | 7449 | gimple *gtemp; |
a414c77f IE |
7450 | if (masked) |
7451 | { | |
7452 | vec_cond_lhs | |
7453 | = vect_get_vec_def_for_operand (cond_expr, stmt, | |
7454 | comp_vectype); | |
7455 | vect_is_simple_use (cond_expr, stmt_info->vinfo, | |
7456 | >emp, &dts[0]); | |
7457 | } | |
7458 | else | |
7459 | { | |
7460 | vec_cond_lhs = | |
7461 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), | |
7462 | stmt, comp_vectype); | |
7463 | vect_is_simple_use (TREE_OPERAND (cond_expr, 0), | |
7464 | loop_vinfo, >emp, &dts[0]); | |
7465 | ||
7466 | vec_cond_rhs = | |
7467 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), | |
7468 | stmt, comp_vectype); | |
7469 | vect_is_simple_use (TREE_OPERAND (cond_expr, 1), | |
7470 | loop_vinfo, >emp, &dts[1]); | |
7471 | } | |
f7e531cf IR |
7472 | if (reduc_index == 1) |
7473 | vec_then_clause = reduc_def; | |
7474 | else | |
7475 | { | |
7476 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
81c40241 RB |
7477 | stmt); |
7478 | vect_is_simple_use (then_clause, loop_vinfo, | |
7479 | >emp, &dts[2]); | |
f7e531cf IR |
7480 | } |
7481 | if (reduc_index == 2) | |
7482 | vec_else_clause = reduc_def; | |
7483 | else | |
7484 | { | |
7485 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
81c40241 RB |
7486 | stmt); |
7487 | vect_is_simple_use (else_clause, loop_vinfo, >emp, &dts[3]); | |
f7e531cf | 7488 | } |
a855b1b1 MM |
7489 | } |
7490 | } | |
7491 | else | |
7492 | { | |
a414c77f IE |
7493 | vec_cond_lhs |
7494 | = vect_get_vec_def_for_stmt_copy (dts[0], | |
7495 | vec_oprnds0.pop ()); | |
7496 | if (!masked) | |
7497 | vec_cond_rhs | |
7498 | = vect_get_vec_def_for_stmt_copy (dts[1], | |
7499 | vec_oprnds1.pop ()); | |
7500 | ||
a855b1b1 | 7501 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 7502 | vec_oprnds2.pop ()); |
a855b1b1 | 7503 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 7504 | vec_oprnds3.pop ()); |
f7e531cf IR |
7505 | } |
7506 | ||
7507 | if (!slp_node) | |
7508 | { | |
9771b263 | 7509 | vec_oprnds0.quick_push (vec_cond_lhs); |
a414c77f IE |
7510 | if (!masked) |
7511 | vec_oprnds1.quick_push (vec_cond_rhs); | |
9771b263 DN |
7512 | vec_oprnds2.quick_push (vec_then_clause); |
7513 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
7514 | } |
7515 | ||
9dc3f7de | 7516 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 7517 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 7518 | { |
9771b263 DN |
7519 | vec_then_clause = vec_oprnds2[i]; |
7520 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 7521 | |
a414c77f IE |
7522 | if (masked) |
7523 | vec_compare = vec_cond_lhs; | |
7524 | else | |
7525 | { | |
7526 | vec_cond_rhs = vec_oprnds1[i]; | |
7527 | vec_compare = build2 (TREE_CODE (cond_expr), vec_cmp_type, | |
7528 | vec_cond_lhs, vec_cond_rhs); | |
7529 | } | |
f7e531cf IR |
7530 | vec_cond_expr = build3 (VEC_COND_EXPR, vectype, |
7531 | vec_compare, vec_then_clause, vec_else_clause); | |
a855b1b1 | 7532 | |
f7e531cf IR |
7533 | new_stmt = gimple_build_assign (vec_dest, vec_cond_expr); |
7534 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
7535 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7536 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7537 | if (slp_node) | |
9771b263 | 7538 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
7539 | } |
7540 | ||
7541 | if (slp_node) | |
7542 | continue; | |
7543 | ||
7544 | if (j == 0) | |
7545 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7546 | else | |
7547 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7548 | ||
7549 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 7550 | } |
b8698a0f | 7551 | |
9771b263 DN |
7552 | vec_oprnds0.release (); |
7553 | vec_oprnds1.release (); | |
7554 | vec_oprnds2.release (); | |
7555 | vec_oprnds3.release (); | |
f7e531cf | 7556 | |
ebfd146a IR |
7557 | return true; |
7558 | } | |
7559 | ||
42fd8198 IE |
7560 | /* vectorizable_comparison. |
7561 | ||
7562 | Check if STMT is comparison expression that can be vectorized. | |
7563 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
7564 | comparison, put it in VEC_STMT, and insert it at GSI. | |
7565 | ||
7566 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7567 | ||
7568 | bool | |
7569 | vectorizable_comparison (gimple *stmt, gimple_stmt_iterator *gsi, | |
7570 | gimple **vec_stmt, tree reduc_def, | |
7571 | slp_tree slp_node) | |
7572 | { | |
7573 | tree lhs, rhs1, rhs2; | |
7574 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
7575 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; | |
7576 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
7577 | tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE; | |
7578 | tree new_temp; | |
7579 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
7580 | enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
7581 | unsigned nunits; | |
7582 | int ncopies; | |
7583 | enum tree_code code; | |
7584 | stmt_vec_info prev_stmt_info = NULL; | |
7585 | int i, j; | |
7586 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
7587 | vec<tree> vec_oprnds0 = vNULL; | |
7588 | vec<tree> vec_oprnds1 = vNULL; | |
7589 | gimple *def_stmt; | |
7590 | tree mask_type; | |
7591 | tree mask; | |
7592 | ||
7593 | if (!VECTOR_BOOLEAN_TYPE_P (vectype)) | |
7594 | return false; | |
7595 | ||
7596 | mask_type = vectype; | |
7597 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
7598 | ||
7599 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
7600 | ncopies = 1; | |
7601 | else | |
7602 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
7603 | ||
7604 | gcc_assert (ncopies >= 1); | |
7605 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
7606 | return false; | |
7607 | ||
7608 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def | |
7609 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
7610 | && reduc_def)) | |
7611 | return false; | |
7612 | ||
7613 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
7614 | { | |
7615 | if (dump_enabled_p ()) | |
7616 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7617 | "value used after loop.\n"); | |
7618 | return false; | |
7619 | } | |
7620 | ||
7621 | if (!is_gimple_assign (stmt)) | |
7622 | return false; | |
7623 | ||
7624 | code = gimple_assign_rhs_code (stmt); | |
7625 | ||
7626 | if (TREE_CODE_CLASS (code) != tcc_comparison) | |
7627 | return false; | |
7628 | ||
7629 | rhs1 = gimple_assign_rhs1 (stmt); | |
7630 | rhs2 = gimple_assign_rhs2 (stmt); | |
7631 | ||
7632 | if (!vect_is_simple_use (rhs1, stmt_info->vinfo, &def_stmt, | |
7633 | &dts[0], &vectype1)) | |
7634 | return false; | |
7635 | ||
7636 | if (!vect_is_simple_use (rhs2, stmt_info->vinfo, &def_stmt, | |
7637 | &dts[1], &vectype2)) | |
7638 | return false; | |
7639 | ||
7640 | if (vectype1 && vectype2 | |
7641 | && TYPE_VECTOR_SUBPARTS (vectype1) != TYPE_VECTOR_SUBPARTS (vectype2)) | |
7642 | return false; | |
7643 | ||
7644 | vectype = vectype1 ? vectype1 : vectype2; | |
7645 | ||
7646 | /* Invariant comparison. */ | |
7647 | if (!vectype) | |
7648 | { | |
7649 | vectype = build_vector_type (TREE_TYPE (rhs1), nunits); | |
7650 | if (tree_to_shwi (TYPE_SIZE_UNIT (vectype)) != current_vector_size) | |
7651 | return false; | |
7652 | } | |
7653 | else if (nunits != TYPE_VECTOR_SUBPARTS (vectype)) | |
7654 | return false; | |
7655 | ||
7656 | if (!vec_stmt) | |
7657 | { | |
7658 | STMT_VINFO_TYPE (stmt_info) = comparison_vec_info_type; | |
7659 | vect_model_simple_cost (stmt_info, ncopies, dts, NULL, NULL); | |
7660 | return expand_vec_cmp_expr_p (vectype, mask_type); | |
7661 | } | |
7662 | ||
7663 | /* Transform. */ | |
7664 | if (!slp_node) | |
7665 | { | |
7666 | vec_oprnds0.create (1); | |
7667 | vec_oprnds1.create (1); | |
7668 | } | |
7669 | ||
7670 | /* Handle def. */ | |
7671 | lhs = gimple_assign_lhs (stmt); | |
7672 | mask = vect_create_destination_var (lhs, mask_type); | |
7673 | ||
7674 | /* Handle cmp expr. */ | |
7675 | for (j = 0; j < ncopies; j++) | |
7676 | { | |
7677 | gassign *new_stmt = NULL; | |
7678 | if (j == 0) | |
7679 | { | |
7680 | if (slp_node) | |
7681 | { | |
7682 | auto_vec<tree, 2> ops; | |
7683 | auto_vec<vec<tree>, 2> vec_defs; | |
7684 | ||
7685 | ops.safe_push (rhs1); | |
7686 | ops.safe_push (rhs2); | |
7687 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); | |
7688 | vec_oprnds1 = vec_defs.pop (); | |
7689 | vec_oprnds0 = vec_defs.pop (); | |
7690 | } | |
7691 | else | |
7692 | { | |
e4af0bc4 IE |
7693 | vec_rhs1 = vect_get_vec_def_for_operand (rhs1, stmt, vectype); |
7694 | vec_rhs2 = vect_get_vec_def_for_operand (rhs2, stmt, vectype); | |
42fd8198 IE |
7695 | } |
7696 | } | |
7697 | else | |
7698 | { | |
7699 | vec_rhs1 = vect_get_vec_def_for_stmt_copy (dts[0], | |
7700 | vec_oprnds0.pop ()); | |
7701 | vec_rhs2 = vect_get_vec_def_for_stmt_copy (dts[1], | |
7702 | vec_oprnds1.pop ()); | |
7703 | } | |
7704 | ||
7705 | if (!slp_node) | |
7706 | { | |
7707 | vec_oprnds0.quick_push (vec_rhs1); | |
7708 | vec_oprnds1.quick_push (vec_rhs2); | |
7709 | } | |
7710 | ||
7711 | /* Arguments are ready. Create the new vector stmt. */ | |
7712 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1) | |
7713 | { | |
7714 | vec_rhs2 = vec_oprnds1[i]; | |
7715 | ||
7716 | new_temp = make_ssa_name (mask); | |
7717 | new_stmt = gimple_build_assign (new_temp, code, vec_rhs1, vec_rhs2); | |
7718 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7719 | if (slp_node) | |
7720 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
7721 | } | |
7722 | ||
7723 | if (slp_node) | |
7724 | continue; | |
7725 | ||
7726 | if (j == 0) | |
7727 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7728 | else | |
7729 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7730 | ||
7731 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7732 | } | |
7733 | ||
7734 | vec_oprnds0.release (); | |
7735 | vec_oprnds1.release (); | |
7736 | ||
7737 | return true; | |
7738 | } | |
ebfd146a | 7739 | |
8644a673 | 7740 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
7741 | |
7742 | bool | |
355fe088 | 7743 | vect_analyze_stmt (gimple *stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 7744 | { |
8644a673 | 7745 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 7746 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 7747 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 7748 | bool ok; |
a70d6342 | 7749 | tree scalar_type, vectype; |
355fe088 | 7750 | gimple *pattern_stmt; |
363477c0 | 7751 | gimple_seq pattern_def_seq; |
ebfd146a | 7752 | |
73fbfcad | 7753 | if (dump_enabled_p ()) |
ebfd146a | 7754 | { |
78c60e3d SS |
7755 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
7756 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
8644a673 | 7757 | } |
ebfd146a | 7758 | |
1825a1f3 | 7759 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 7760 | { |
73fbfcad | 7761 | if (dump_enabled_p ()) |
78c60e3d | 7762 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7763 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
7764 | |
7765 | return false; | |
7766 | } | |
b8698a0f L |
7767 | |
7768 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
7769 | to include: |
7770 | - the COND_EXPR which is the loop exit condition | |
7771 | - any LABEL_EXPRs in the loop | |
b8698a0f | 7772 | - computations that are used only for array indexing or loop control. |
8644a673 | 7773 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 7774 | instance, therefore, all the statements are relevant. |
ebfd146a | 7775 | |
d092494c | 7776 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 7777 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
7778 | statements. In basic blocks we are called from some SLP instance |
7779 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
7780 | already will be part of SLP instance. */ | |
83197f37 IR |
7781 | |
7782 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 7783 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 7784 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 7785 | { |
9d5e7640 | 7786 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 7787 | && pattern_stmt |
9d5e7640 IR |
7788 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
7789 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
7790 | { | |
83197f37 | 7791 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
7792 | stmt = pattern_stmt; |
7793 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 7794 | if (dump_enabled_p ()) |
9d5e7640 | 7795 | { |
78c60e3d SS |
7796 | dump_printf_loc (MSG_NOTE, vect_location, |
7797 | "==> examining pattern statement: "); | |
7798 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
9d5e7640 IR |
7799 | } |
7800 | } | |
7801 | else | |
7802 | { | |
73fbfcad | 7803 | if (dump_enabled_p ()) |
e645e942 | 7804 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 7805 | |
9d5e7640 IR |
7806 | return true; |
7807 | } | |
8644a673 | 7808 | } |
83197f37 | 7809 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 7810 | && node == NULL |
83197f37 IR |
7811 | && pattern_stmt |
7812 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
7813 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
7814 | { | |
7815 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 7816 | if (dump_enabled_p ()) |
83197f37 | 7817 | { |
78c60e3d SS |
7818 | dump_printf_loc (MSG_NOTE, vect_location, |
7819 | "==> examining pattern statement: "); | |
7820 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
83197f37 IR |
7821 | } |
7822 | ||
7823 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node)) | |
7824 | return false; | |
7825 | } | |
ebfd146a | 7826 | |
1107f3ae | 7827 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 7828 | && node == NULL |
363477c0 | 7829 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 7830 | { |
363477c0 | 7831 | gimple_stmt_iterator si; |
1107f3ae | 7832 | |
363477c0 JJ |
7833 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
7834 | { | |
355fe088 | 7835 | gimple *pattern_def_stmt = gsi_stmt (si); |
363477c0 JJ |
7836 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) |
7837 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
7838 | { | |
7839 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 7840 | if (dump_enabled_p ()) |
363477c0 | 7841 | { |
78c60e3d SS |
7842 | dump_printf_loc (MSG_NOTE, vect_location, |
7843 | "==> examining pattern def statement: "); | |
7844 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
363477c0 | 7845 | } |
1107f3ae | 7846 | |
363477c0 JJ |
7847 | if (!vect_analyze_stmt (pattern_def_stmt, |
7848 | need_to_vectorize, node)) | |
7849 | return false; | |
7850 | } | |
7851 | } | |
7852 | } | |
1107f3ae | 7853 | |
8644a673 IR |
7854 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
7855 | { | |
7856 | case vect_internal_def: | |
7857 | break; | |
ebfd146a | 7858 | |
8644a673 | 7859 | case vect_reduction_def: |
7c5222ff | 7860 | case vect_nested_cycle: |
14a61437 RB |
7861 | gcc_assert (!bb_vinfo |
7862 | && (relevance == vect_used_in_outer | |
7863 | || relevance == vect_used_in_outer_by_reduction | |
7864 | || relevance == vect_used_by_reduction | |
7865 | || relevance == vect_unused_in_scope)); | |
8644a673 IR |
7866 | break; |
7867 | ||
7868 | case vect_induction_def: | |
7869 | case vect_constant_def: | |
7870 | case vect_external_def: | |
7871 | case vect_unknown_def_type: | |
7872 | default: | |
7873 | gcc_unreachable (); | |
7874 | } | |
ebfd146a | 7875 | |
a70d6342 IR |
7876 | if (bb_vinfo) |
7877 | { | |
7878 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
7879 | ||
b690cc0f | 7880 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
73fbfcad | 7881 | if (dump_enabled_p ()) |
a70d6342 | 7882 | { |
78c60e3d SS |
7883 | dump_printf_loc (MSG_NOTE, vect_location, |
7884 | "get vectype for scalar type: "); | |
7885 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
e645e942 | 7886 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
7887 | } |
7888 | ||
7889 | vectype = get_vectype_for_scalar_type (scalar_type); | |
7890 | if (!vectype) | |
7891 | { | |
73fbfcad | 7892 | if (dump_enabled_p ()) |
a70d6342 | 7893 | { |
78c60e3d SS |
7894 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
7895 | "not SLPed: unsupported data-type "); | |
7896 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
7897 | scalar_type); | |
e645e942 | 7898 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
a70d6342 IR |
7899 | } |
7900 | return false; | |
7901 | } | |
7902 | ||
73fbfcad | 7903 | if (dump_enabled_p ()) |
a70d6342 | 7904 | { |
78c60e3d SS |
7905 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); |
7906 | dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); | |
e645e942 | 7907 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
7908 | } |
7909 | ||
7910 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
7911 | } | |
7912 | ||
8644a673 | 7913 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 7914 | { |
8644a673 | 7915 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
0136f8f0 AH |
7916 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
7917 | || (is_gimple_call (stmt) | |
7918 | && gimple_call_lhs (stmt) == NULL_TREE)); | |
8644a673 | 7919 | *need_to_vectorize = true; |
ebfd146a IR |
7920 | } |
7921 | ||
b1af7da6 RB |
7922 | if (PURE_SLP_STMT (stmt_info) && !node) |
7923 | { | |
7924 | dump_printf_loc (MSG_NOTE, vect_location, | |
7925 | "handled only by SLP analysis\n"); | |
7926 | return true; | |
7927 | } | |
7928 | ||
7929 | ok = true; | |
7930 | if (!bb_vinfo | |
7931 | && (STMT_VINFO_RELEVANT_P (stmt_info) | |
7932 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
7933 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
7934 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
7935 | || vectorizable_shift (stmt, NULL, NULL, node) | |
7936 | || vectorizable_operation (stmt, NULL, NULL, node) | |
7937 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
7938 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
7939 | || vectorizable_call (stmt, NULL, NULL, node) | |
7940 | || vectorizable_store (stmt, NULL, NULL, node) | |
7941 | || vectorizable_reduction (stmt, NULL, NULL, node) | |
42fd8198 IE |
7942 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
7943 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 RB |
7944 | else |
7945 | { | |
7946 | if (bb_vinfo) | |
7947 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
7948 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
7949 | || vectorizable_shift (stmt, NULL, NULL, node) | |
7950 | || vectorizable_operation (stmt, NULL, NULL, node) | |
7951 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
7952 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
7953 | || vectorizable_call (stmt, NULL, NULL, node) | |
7954 | || vectorizable_store (stmt, NULL, NULL, node) | |
42fd8198 IE |
7955 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
7956 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 | 7957 | } |
8644a673 IR |
7958 | |
7959 | if (!ok) | |
ebfd146a | 7960 | { |
73fbfcad | 7961 | if (dump_enabled_p ()) |
8644a673 | 7962 | { |
78c60e3d SS |
7963 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
7964 | "not vectorized: relevant stmt not "); | |
7965 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
7966 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 7967 | } |
b8698a0f | 7968 | |
ebfd146a IR |
7969 | return false; |
7970 | } | |
7971 | ||
a70d6342 IR |
7972 | if (bb_vinfo) |
7973 | return true; | |
7974 | ||
8644a673 IR |
7975 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
7976 | need extra handling, except for vectorizable reductions. */ | |
7977 | if (STMT_VINFO_LIVE_P (stmt_info) | |
7978 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
7979 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 7980 | |
8644a673 | 7981 | if (!ok) |
ebfd146a | 7982 | { |
73fbfcad | 7983 | if (dump_enabled_p ()) |
8644a673 | 7984 | { |
78c60e3d SS |
7985 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
7986 | "not vectorized: live stmt not "); | |
7987 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
7988 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 7989 | } |
b8698a0f | 7990 | |
8644a673 | 7991 | return false; |
ebfd146a IR |
7992 | } |
7993 | ||
ebfd146a IR |
7994 | return true; |
7995 | } | |
7996 | ||
7997 | ||
7998 | /* Function vect_transform_stmt. | |
7999 | ||
8000 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
8001 | ||
8002 | bool | |
355fe088 | 8003 | vect_transform_stmt (gimple *stmt, gimple_stmt_iterator *gsi, |
0d0293ac | 8004 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
8005 | slp_instance slp_node_instance) |
8006 | { | |
8007 | bool is_store = false; | |
355fe088 | 8008 | gimple *vec_stmt = NULL; |
ebfd146a | 8009 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
ebfd146a | 8010 | bool done; |
ebfd146a | 8011 | |
355fe088 | 8012 | gimple *old_vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
225ce44b | 8013 | |
ebfd146a IR |
8014 | switch (STMT_VINFO_TYPE (stmt_info)) |
8015 | { | |
8016 | case type_demotion_vec_info_type: | |
ebfd146a | 8017 | case type_promotion_vec_info_type: |
ebfd146a IR |
8018 | case type_conversion_vec_info_type: |
8019 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
8020 | gcc_assert (done); | |
8021 | break; | |
8022 | ||
8023 | case induc_vec_info_type: | |
8024 | gcc_assert (!slp_node); | |
8025 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
8026 | gcc_assert (done); | |
8027 | break; | |
8028 | ||
9dc3f7de IR |
8029 | case shift_vec_info_type: |
8030 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
8031 | gcc_assert (done); | |
8032 | break; | |
8033 | ||
ebfd146a IR |
8034 | case op_vec_info_type: |
8035 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
8036 | gcc_assert (done); | |
8037 | break; | |
8038 | ||
8039 | case assignment_vec_info_type: | |
8040 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
8041 | gcc_assert (done); | |
8042 | break; | |
8043 | ||
8044 | case load_vec_info_type: | |
b8698a0f | 8045 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
8046 | slp_node_instance); |
8047 | gcc_assert (done); | |
8048 | break; | |
8049 | ||
8050 | case store_vec_info_type: | |
8051 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
8052 | gcc_assert (done); | |
0d0293ac | 8053 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
8054 | { |
8055 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 8056 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
8057 | one are skipped, and there vec_stmt_info shouldn't be freed |
8058 | meanwhile. */ | |
0d0293ac | 8059 | *grouped_store = true; |
ebfd146a IR |
8060 | if (STMT_VINFO_VEC_STMT (stmt_info)) |
8061 | is_store = true; | |
8062 | } | |
8063 | else | |
8064 | is_store = true; | |
8065 | break; | |
8066 | ||
8067 | case condition_vec_info_type: | |
f7e531cf | 8068 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
8069 | gcc_assert (done); |
8070 | break; | |
8071 | ||
42fd8198 IE |
8072 | case comparison_vec_info_type: |
8073 | done = vectorizable_comparison (stmt, gsi, &vec_stmt, NULL, slp_node); | |
8074 | gcc_assert (done); | |
8075 | break; | |
8076 | ||
ebfd146a | 8077 | case call_vec_info_type: |
190c2236 | 8078 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 8079 | stmt = gsi_stmt (*gsi); |
5ce9450f JJ |
8080 | if (is_gimple_call (stmt) |
8081 | && gimple_call_internal_p (stmt) | |
8082 | && gimple_call_internal_fn (stmt) == IFN_MASK_STORE) | |
8083 | is_store = true; | |
ebfd146a IR |
8084 | break; |
8085 | ||
0136f8f0 AH |
8086 | case call_simd_clone_vec_info_type: |
8087 | done = vectorizable_simd_clone_call (stmt, gsi, &vec_stmt, slp_node); | |
8088 | stmt = gsi_stmt (*gsi); | |
8089 | break; | |
8090 | ||
ebfd146a | 8091 | case reduc_vec_info_type: |
b5aeb3bb | 8092 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
8093 | gcc_assert (done); |
8094 | break; | |
8095 | ||
8096 | default: | |
8097 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
8098 | { | |
73fbfcad | 8099 | if (dump_enabled_p ()) |
78c60e3d | 8100 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8101 | "stmt not supported.\n"); |
ebfd146a IR |
8102 | gcc_unreachable (); |
8103 | } | |
8104 | } | |
8105 | ||
225ce44b RB |
8106 | /* Verify SLP vectorization doesn't mess with STMT_VINFO_VEC_STMT. |
8107 | This would break hybrid SLP vectorization. */ | |
8108 | if (slp_node) | |
d90f8440 RB |
8109 | gcc_assert (!vec_stmt |
8110 | && STMT_VINFO_VEC_STMT (stmt_info) == old_vec_stmt); | |
225ce44b | 8111 | |
ebfd146a IR |
8112 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that |
8113 | is being vectorized, but outside the immediately enclosing loop. */ | |
8114 | if (vec_stmt | |
a70d6342 IR |
8115 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
8116 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
8117 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
8118 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
8119 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 8120 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 8121 | vect_used_in_outer_by_reduction)) |
ebfd146a | 8122 | { |
a70d6342 IR |
8123 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
8124 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
8125 | imm_use_iterator imm_iter; |
8126 | use_operand_p use_p; | |
8127 | tree scalar_dest; | |
355fe088 | 8128 | gimple *exit_phi; |
ebfd146a | 8129 | |
73fbfcad | 8130 | if (dump_enabled_p ()) |
78c60e3d | 8131 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 8132 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
8133 | |
8134 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
8135 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
8136 | STMT). */ | |
8137 | if (gimple_code (stmt) == GIMPLE_PHI) | |
8138 | scalar_dest = PHI_RESULT (stmt); | |
8139 | else | |
8140 | scalar_dest = gimple_assign_lhs (stmt); | |
8141 | ||
8142 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
8143 | { | |
8144 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
8145 | { | |
8146 | exit_phi = USE_STMT (use_p); | |
8147 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
8148 | } | |
8149 | } | |
8150 | } | |
8151 | ||
8152 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
8153 | being vectorized. */ | |
8154 | if (STMT_VINFO_LIVE_P (stmt_info) | |
8155 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
8156 | { | |
8157 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
8158 | gcc_assert (done); | |
8159 | } | |
8160 | ||
8161 | if (vec_stmt) | |
83197f37 | 8162 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 8163 | |
b8698a0f | 8164 | return is_store; |
ebfd146a IR |
8165 | } |
8166 | ||
8167 | ||
b8698a0f | 8168 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
8169 | stmt_vec_info. */ |
8170 | ||
8171 | void | |
355fe088 | 8172 | vect_remove_stores (gimple *first_stmt) |
ebfd146a | 8173 | { |
355fe088 TS |
8174 | gimple *next = first_stmt; |
8175 | gimple *tmp; | |
ebfd146a IR |
8176 | gimple_stmt_iterator next_si; |
8177 | ||
8178 | while (next) | |
8179 | { | |
78048b1c JJ |
8180 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
8181 | ||
8182 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
8183 | if (is_pattern_stmt_p (stmt_info)) | |
8184 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
8185 | /* Free the attached stmt_vec_info and remove the stmt. */ |
8186 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 8187 | unlink_stmt_vdef (next); |
ebfd146a | 8188 | gsi_remove (&next_si, true); |
3d3f2249 | 8189 | release_defs (next); |
ebfd146a IR |
8190 | free_stmt_vec_info (next); |
8191 | next = tmp; | |
8192 | } | |
8193 | } | |
8194 | ||
8195 | ||
8196 | /* Function new_stmt_vec_info. | |
8197 | ||
8198 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
8199 | ||
8200 | stmt_vec_info | |
310213d4 | 8201 | new_stmt_vec_info (gimple *stmt, vec_info *vinfo) |
ebfd146a IR |
8202 | { |
8203 | stmt_vec_info res; | |
8204 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
8205 | ||
8206 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
8207 | STMT_VINFO_STMT (res) = stmt; | |
310213d4 | 8208 | res->vinfo = vinfo; |
8644a673 | 8209 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
8210 | STMT_VINFO_LIVE_P (res) = false; |
8211 | STMT_VINFO_VECTYPE (res) = NULL; | |
8212 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 8213 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
8214 | STMT_VINFO_IN_PATTERN_P (res) = false; |
8215 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 8216 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a | 8217 | STMT_VINFO_DATA_REF (res) = NULL; |
af29617a | 8218 | STMT_VINFO_VEC_REDUCTION_TYPE (res) = TREE_CODE_REDUCTION; |
ebfd146a IR |
8219 | |
8220 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
8221 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
8222 | STMT_VINFO_DR_INIT (res) = NULL; | |
8223 | STMT_VINFO_DR_STEP (res) = NULL; | |
8224 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
8225 | ||
8226 | if (gimple_code (stmt) == GIMPLE_PHI | |
8227 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
8228 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
8229 | else | |
8644a673 IR |
8230 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
8231 | ||
9771b263 | 8232 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 8233 | STMT_SLP_TYPE (res) = loop_vect; |
e14c1050 IR |
8234 | GROUP_FIRST_ELEMENT (res) = NULL; |
8235 | GROUP_NEXT_ELEMENT (res) = NULL; | |
8236 | GROUP_SIZE (res) = 0; | |
8237 | GROUP_STORE_COUNT (res) = 0; | |
8238 | GROUP_GAP (res) = 0; | |
8239 | GROUP_SAME_DR_STMT (res) = NULL; | |
ebfd146a IR |
8240 | |
8241 | return res; | |
8242 | } | |
8243 | ||
8244 | ||
8245 | /* Create a hash table for stmt_vec_info. */ | |
8246 | ||
8247 | void | |
8248 | init_stmt_vec_info_vec (void) | |
8249 | { | |
9771b263 DN |
8250 | gcc_assert (!stmt_vec_info_vec.exists ()); |
8251 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
8252 | } |
8253 | ||
8254 | ||
8255 | /* Free hash table for stmt_vec_info. */ | |
8256 | ||
8257 | void | |
8258 | free_stmt_vec_info_vec (void) | |
8259 | { | |
93675444 | 8260 | unsigned int i; |
3161455c | 8261 | stmt_vec_info info; |
93675444 JJ |
8262 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) |
8263 | if (info != NULL) | |
3161455c | 8264 | free_stmt_vec_info (STMT_VINFO_STMT (info)); |
9771b263 DN |
8265 | gcc_assert (stmt_vec_info_vec.exists ()); |
8266 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
8267 | } |
8268 | ||
8269 | ||
8270 | /* Free stmt vectorization related info. */ | |
8271 | ||
8272 | void | |
355fe088 | 8273 | free_stmt_vec_info (gimple *stmt) |
ebfd146a IR |
8274 | { |
8275 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8276 | ||
8277 | if (!stmt_info) | |
8278 | return; | |
8279 | ||
78048b1c JJ |
8280 | /* Check if this statement has a related "pattern stmt" |
8281 | (introduced by the vectorizer during the pattern recognition | |
8282 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
8283 | too. */ | |
8284 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
8285 | { | |
8286 | stmt_vec_info patt_info | |
8287 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
8288 | if (patt_info) | |
8289 | { | |
363477c0 | 8290 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
355fe088 | 8291 | gimple *patt_stmt = STMT_VINFO_STMT (patt_info); |
f0281fde RB |
8292 | gimple_set_bb (patt_stmt, NULL); |
8293 | tree lhs = gimple_get_lhs (patt_stmt); | |
e6f5c25d | 8294 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde | 8295 | release_ssa_name (lhs); |
363477c0 JJ |
8296 | if (seq) |
8297 | { | |
8298 | gimple_stmt_iterator si; | |
8299 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
f0281fde | 8300 | { |
355fe088 | 8301 | gimple *seq_stmt = gsi_stmt (si); |
f0281fde | 8302 | gimple_set_bb (seq_stmt, NULL); |
7532abf2 | 8303 | lhs = gimple_get_lhs (seq_stmt); |
e6f5c25d | 8304 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde RB |
8305 | release_ssa_name (lhs); |
8306 | free_stmt_vec_info (seq_stmt); | |
8307 | } | |
363477c0 | 8308 | } |
f0281fde | 8309 | free_stmt_vec_info (patt_stmt); |
78048b1c JJ |
8310 | } |
8311 | } | |
8312 | ||
9771b263 | 8313 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
6c9e85fb | 8314 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).release (); |
ebfd146a IR |
8315 | set_vinfo_for_stmt (stmt, NULL); |
8316 | free (stmt_info); | |
8317 | } | |
8318 | ||
8319 | ||
bb67d9c7 | 8320 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 8321 | |
bb67d9c7 | 8322 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
8323 | by the target. */ |
8324 | ||
bb67d9c7 RG |
8325 | static tree |
8326 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a | 8327 | { |
ef4bddc2 RS |
8328 | machine_mode inner_mode = TYPE_MODE (scalar_type); |
8329 | machine_mode simd_mode; | |
2f816591 | 8330 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
8331 | int nunits; |
8332 | tree vectype; | |
8333 | ||
cc4b5170 | 8334 | if (nbytes == 0) |
ebfd146a IR |
8335 | return NULL_TREE; |
8336 | ||
48f2e373 RB |
8337 | if (GET_MODE_CLASS (inner_mode) != MODE_INT |
8338 | && GET_MODE_CLASS (inner_mode) != MODE_FLOAT) | |
8339 | return NULL_TREE; | |
8340 | ||
7b7b1813 RG |
8341 | /* For vector types of elements whose mode precision doesn't |
8342 | match their types precision we use a element type of mode | |
8343 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
8344 | they support the proper result truncation/extension. |
8345 | We also make sure to build vector types with INTEGER_TYPE | |
8346 | component type only. */ | |
6d7971b8 | 8347 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
8348 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
8349 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
8350 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
8351 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 8352 | |
ccbf5bb4 RG |
8353 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
8354 | When the component mode passes the above test simply use a type | |
8355 | corresponding to that mode. The theory is that any use that | |
8356 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 8357 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 8358 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
8359 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
8360 | ||
8361 | /* We can't build a vector type of elements with alignment bigger than | |
8362 | their size. */ | |
dfc2e2ac | 8363 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
8364 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
8365 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 8366 | |
dfc2e2ac RB |
8367 | /* If we felt back to using the mode fail if there was |
8368 | no scalar type for it. */ | |
8369 | if (scalar_type == NULL_TREE) | |
8370 | return NULL_TREE; | |
8371 | ||
bb67d9c7 RG |
8372 | /* If no size was supplied use the mode the target prefers. Otherwise |
8373 | lookup a vector mode of the specified size. */ | |
8374 | if (size == 0) | |
8375 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
8376 | else | |
8377 | simd_mode = mode_for_vector (inner_mode, size / nbytes); | |
cc4b5170 RG |
8378 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
8379 | if (nunits <= 1) | |
8380 | return NULL_TREE; | |
ebfd146a IR |
8381 | |
8382 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
8383 | |
8384 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
8385 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 8386 | return NULL_TREE; |
ebfd146a IR |
8387 | |
8388 | return vectype; | |
8389 | } | |
8390 | ||
bb67d9c7 RG |
8391 | unsigned int current_vector_size; |
8392 | ||
8393 | /* Function get_vectype_for_scalar_type. | |
8394 | ||
8395 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
8396 | by the target. */ | |
8397 | ||
8398 | tree | |
8399 | get_vectype_for_scalar_type (tree scalar_type) | |
8400 | { | |
8401 | tree vectype; | |
8402 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
8403 | current_vector_size); | |
8404 | if (vectype | |
8405 | && current_vector_size == 0) | |
8406 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
8407 | return vectype; | |
8408 | } | |
8409 | ||
42fd8198 IE |
8410 | /* Function get_mask_type_for_scalar_type. |
8411 | ||
8412 | Returns the mask type corresponding to a result of comparison | |
8413 | of vectors of specified SCALAR_TYPE as supported by target. */ | |
8414 | ||
8415 | tree | |
8416 | get_mask_type_for_scalar_type (tree scalar_type) | |
8417 | { | |
8418 | tree vectype = get_vectype_for_scalar_type (scalar_type); | |
8419 | ||
8420 | if (!vectype) | |
8421 | return NULL; | |
8422 | ||
8423 | return build_truth_vector_type (TYPE_VECTOR_SUBPARTS (vectype), | |
8424 | current_vector_size); | |
8425 | } | |
8426 | ||
b690cc0f RG |
8427 | /* Function get_same_sized_vectype |
8428 | ||
8429 | Returns a vector type corresponding to SCALAR_TYPE of size | |
8430 | VECTOR_TYPE if supported by the target. */ | |
8431 | ||
8432 | tree | |
bb67d9c7 | 8433 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 8434 | { |
9f47c7e5 IE |
8435 | if (TREE_CODE (scalar_type) == BOOLEAN_TYPE) |
8436 | return build_same_sized_truth_vector_type (vector_type); | |
8437 | ||
bb67d9c7 RG |
8438 | return get_vectype_for_scalar_type_and_size |
8439 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
8440 | } |
8441 | ||
ebfd146a IR |
8442 | /* Function vect_is_simple_use. |
8443 | ||
8444 | Input: | |
81c40241 RB |
8445 | VINFO - the vect info of the loop or basic block that is being vectorized. |
8446 | OPERAND - operand in the loop or bb. | |
8447 | Output: | |
8448 | DEF_STMT - the defining stmt in case OPERAND is an SSA_NAME. | |
8449 | DT - the type of definition | |
ebfd146a IR |
8450 | |
8451 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 8452 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 8453 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 8454 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
8455 | is the case in reduction/induction computations). |
8456 | For basic blocks, supportable operands are constants and bb invariants. | |
8457 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
8458 | |
8459 | bool | |
81c40241 RB |
8460 | vect_is_simple_use (tree operand, vec_info *vinfo, |
8461 | gimple **def_stmt, enum vect_def_type *dt) | |
b8698a0f | 8462 | { |
ebfd146a | 8463 | *def_stmt = NULL; |
3fc356dc | 8464 | *dt = vect_unknown_def_type; |
b8698a0f | 8465 | |
73fbfcad | 8466 | if (dump_enabled_p ()) |
ebfd146a | 8467 | { |
78c60e3d SS |
8468 | dump_printf_loc (MSG_NOTE, vect_location, |
8469 | "vect_is_simple_use: operand "); | |
8470 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
e645e942 | 8471 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 8472 | } |
b8698a0f | 8473 | |
b758f602 | 8474 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
8475 | { |
8476 | *dt = vect_constant_def; | |
8477 | return true; | |
8478 | } | |
b8698a0f | 8479 | |
ebfd146a IR |
8480 | if (is_gimple_min_invariant (operand)) |
8481 | { | |
8644a673 | 8482 | *dt = vect_external_def; |
ebfd146a IR |
8483 | return true; |
8484 | } | |
8485 | ||
ebfd146a IR |
8486 | if (TREE_CODE (operand) != SSA_NAME) |
8487 | { | |
73fbfcad | 8488 | if (dump_enabled_p ()) |
af29617a AH |
8489 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8490 | "not ssa-name.\n"); | |
ebfd146a IR |
8491 | return false; |
8492 | } | |
b8698a0f | 8493 | |
3fc356dc | 8494 | if (SSA_NAME_IS_DEFAULT_DEF (operand)) |
ebfd146a | 8495 | { |
3fc356dc RB |
8496 | *dt = vect_external_def; |
8497 | return true; | |
ebfd146a IR |
8498 | } |
8499 | ||
3fc356dc | 8500 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
73fbfcad | 8501 | if (dump_enabled_p ()) |
ebfd146a | 8502 | { |
78c60e3d SS |
8503 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
8504 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
ebfd146a IR |
8505 | } |
8506 | ||
61d371eb | 8507 | if (! vect_stmt_in_region_p (vinfo, *def_stmt)) |
8644a673 | 8508 | *dt = vect_external_def; |
ebfd146a IR |
8509 | else |
8510 | { | |
3fc356dc | 8511 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (*def_stmt); |
310213d4 | 8512 | if (is_a <bb_vec_info> (vinfo) && !STMT_VINFO_VECTORIZABLE (stmt_vinfo)) |
90dd6e3d RB |
8513 | *dt = vect_external_def; |
8514 | else | |
8515 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
ebfd146a IR |
8516 | } |
8517 | ||
2e8ab70c RB |
8518 | if (dump_enabled_p ()) |
8519 | { | |
8520 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: "); | |
8521 | switch (*dt) | |
8522 | { | |
8523 | case vect_uninitialized_def: | |
8524 | dump_printf (MSG_NOTE, "uninitialized\n"); | |
8525 | break; | |
8526 | case vect_constant_def: | |
8527 | dump_printf (MSG_NOTE, "constant\n"); | |
8528 | break; | |
8529 | case vect_external_def: | |
8530 | dump_printf (MSG_NOTE, "external\n"); | |
8531 | break; | |
8532 | case vect_internal_def: | |
8533 | dump_printf (MSG_NOTE, "internal\n"); | |
8534 | break; | |
8535 | case vect_induction_def: | |
8536 | dump_printf (MSG_NOTE, "induction\n"); | |
8537 | break; | |
8538 | case vect_reduction_def: | |
8539 | dump_printf (MSG_NOTE, "reduction\n"); | |
8540 | break; | |
8541 | case vect_double_reduction_def: | |
8542 | dump_printf (MSG_NOTE, "double reduction\n"); | |
8543 | break; | |
8544 | case vect_nested_cycle: | |
8545 | dump_printf (MSG_NOTE, "nested cycle\n"); | |
8546 | break; | |
8547 | case vect_unknown_def_type: | |
8548 | dump_printf (MSG_NOTE, "unknown\n"); | |
8549 | break; | |
8550 | } | |
8551 | } | |
8552 | ||
81c40241 | 8553 | if (*dt == vect_unknown_def_type) |
ebfd146a | 8554 | { |
73fbfcad | 8555 | if (dump_enabled_p ()) |
78c60e3d | 8556 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8557 | "Unsupported pattern.\n"); |
ebfd146a IR |
8558 | return false; |
8559 | } | |
8560 | ||
ebfd146a IR |
8561 | switch (gimple_code (*def_stmt)) |
8562 | { | |
8563 | case GIMPLE_PHI: | |
ebfd146a | 8564 | case GIMPLE_ASSIGN: |
ebfd146a | 8565 | case GIMPLE_CALL: |
81c40241 | 8566 | break; |
ebfd146a | 8567 | default: |
73fbfcad | 8568 | if (dump_enabled_p ()) |
78c60e3d | 8569 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8570 | "unsupported defining stmt:\n"); |
ebfd146a IR |
8571 | return false; |
8572 | } | |
8573 | ||
8574 | return true; | |
8575 | } | |
8576 | ||
81c40241 | 8577 | /* Function vect_is_simple_use. |
b690cc0f | 8578 | |
81c40241 | 8579 | Same as vect_is_simple_use but also determines the vector operand |
b690cc0f RG |
8580 | type of OPERAND and stores it to *VECTYPE. If the definition of |
8581 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
8582 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
8583 | is responsible to compute the best suited vector type for the | |
8584 | scalar operand. */ | |
8585 | ||
8586 | bool | |
81c40241 RB |
8587 | vect_is_simple_use (tree operand, vec_info *vinfo, |
8588 | gimple **def_stmt, enum vect_def_type *dt, tree *vectype) | |
b690cc0f | 8589 | { |
81c40241 | 8590 | if (!vect_is_simple_use (operand, vinfo, def_stmt, dt)) |
b690cc0f RG |
8591 | return false; |
8592 | ||
8593 | /* Now get a vector type if the def is internal, otherwise supply | |
8594 | NULL_TREE and leave it up to the caller to figure out a proper | |
8595 | type for the use stmt. */ | |
8596 | if (*dt == vect_internal_def | |
8597 | || *dt == vect_induction_def | |
8598 | || *dt == vect_reduction_def | |
8599 | || *dt == vect_double_reduction_def | |
8600 | || *dt == vect_nested_cycle) | |
8601 | { | |
8602 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
8603 | |
8604 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
8605 | && !STMT_VINFO_RELEVANT (stmt_info) | |
8606 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 8607 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 8608 | |
b690cc0f RG |
8609 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
8610 | gcc_assert (*vectype != NULL_TREE); | |
8611 | } | |
8612 | else if (*dt == vect_uninitialized_def | |
8613 | || *dt == vect_constant_def | |
8614 | || *dt == vect_external_def) | |
8615 | *vectype = NULL_TREE; | |
8616 | else | |
8617 | gcc_unreachable (); | |
8618 | ||
8619 | return true; | |
8620 | } | |
8621 | ||
ebfd146a IR |
8622 | |
8623 | /* Function supportable_widening_operation | |
8624 | ||
b8698a0f L |
8625 | Check whether an operation represented by the code CODE is a |
8626 | widening operation that is supported by the target platform in | |
b690cc0f RG |
8627 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
8628 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 8629 | |
ebfd146a IR |
8630 | Widening operations we currently support are NOP (CONVERT), FLOAT |
8631 | and WIDEN_MULT. This function checks if these operations are supported | |
8632 | by the target platform either directly (via vector tree-codes), or via | |
8633 | target builtins. | |
8634 | ||
8635 | Output: | |
b8698a0f L |
8636 | - CODE1 and CODE2 are codes of vector operations to be used when |
8637 | vectorizing the operation, if available. | |
ebfd146a IR |
8638 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
8639 | case of multi-step conversion (like char->short->int - in that case | |
8640 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
8641 | - INTERM_TYPES contains the intermediate type required to perform the |
8642 | widening operation (short in the above example). */ | |
ebfd146a IR |
8643 | |
8644 | bool | |
355fe088 | 8645 | supportable_widening_operation (enum tree_code code, gimple *stmt, |
b690cc0f | 8646 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
8647 | enum tree_code *code1, enum tree_code *code2, |
8648 | int *multi_step_cvt, | |
9771b263 | 8649 | vec<tree> *interm_types) |
ebfd146a IR |
8650 | { |
8651 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8652 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 8653 | struct loop *vect_loop = NULL; |
ef4bddc2 | 8654 | machine_mode vec_mode; |
81f40b79 | 8655 | enum insn_code icode1, icode2; |
ebfd146a | 8656 | optab optab1, optab2; |
b690cc0f RG |
8657 | tree vectype = vectype_in; |
8658 | tree wide_vectype = vectype_out; | |
ebfd146a | 8659 | enum tree_code c1, c2; |
4a00c761 JJ |
8660 | int i; |
8661 | tree prev_type, intermediate_type; | |
ef4bddc2 | 8662 | machine_mode intermediate_mode, prev_mode; |
4a00c761 | 8663 | optab optab3, optab4; |
ebfd146a | 8664 | |
4a00c761 | 8665 | *multi_step_cvt = 0; |
4ef69dfc IR |
8666 | if (loop_info) |
8667 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
8668 | ||
ebfd146a IR |
8669 | switch (code) |
8670 | { | |
8671 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
8672 | /* The result of a vectorized widening operation usually requires |
8673 | two vectors (because the widened results do not fit into one vector). | |
8674 | The generated vector results would normally be expected to be | |
8675 | generated in the same order as in the original scalar computation, | |
8676 | i.e. if 8 results are generated in each vector iteration, they are | |
8677 | to be organized as follows: | |
8678 | vect1: [res1,res2,res3,res4], | |
8679 | vect2: [res5,res6,res7,res8]. | |
8680 | ||
8681 | However, in the special case that the result of the widening | |
8682 | operation is used in a reduction computation only, the order doesn't | |
8683 | matter (because when vectorizing a reduction we change the order of | |
8684 | the computation). Some targets can take advantage of this and | |
8685 | generate more efficient code. For example, targets like Altivec, | |
8686 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
8687 | generate the following vectors: | |
8688 | vect1: [res1,res3,res5,res7], | |
8689 | vect2: [res2,res4,res6,res8]. | |
8690 | ||
8691 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
8692 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
8693 | iterations in parallel). We therefore don't allow to change the | |
8694 | order of the computation in the inner-loop during outer-loop | |
8695 | vectorization. */ | |
8696 | /* TODO: Another case in which order doesn't *really* matter is when we | |
8697 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
8698 | Normally, pack_trunc performs an even/odd permute, whereas the | |
8699 | repack from an even/odd expansion would be an interleave, which | |
8700 | would be significantly simpler for e.g. AVX2. */ | |
8701 | /* In any case, in order to avoid duplicating the code below, recurse | |
8702 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
8703 | are properly set up for the caller. If we fail, we'll continue with | |
8704 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
8705 | if (vect_loop | |
8706 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
8707 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
8708 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
8709 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
8710 | code1, code2, multi_step_cvt, |
8711 | interm_types)) | |
ebc047a2 CH |
8712 | { |
8713 | /* Elements in a vector with vect_used_by_reduction property cannot | |
8714 | be reordered if the use chain with this property does not have the | |
8715 | same operation. One such an example is s += a * b, where elements | |
8716 | in a and b cannot be reordered. Here we check if the vector defined | |
8717 | by STMT is only directly used in the reduction statement. */ | |
8718 | tree lhs = gimple_assign_lhs (stmt); | |
8719 | use_operand_p dummy; | |
355fe088 | 8720 | gimple *use_stmt; |
ebc047a2 CH |
8721 | stmt_vec_info use_stmt_info = NULL; |
8722 | if (single_imm_use (lhs, &dummy, &use_stmt) | |
8723 | && (use_stmt_info = vinfo_for_stmt (use_stmt)) | |
8724 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) | |
8725 | return true; | |
8726 | } | |
4a00c761 JJ |
8727 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
8728 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
8729 | break; |
8730 | ||
81c40241 RB |
8731 | case DOT_PROD_EXPR: |
8732 | c1 = DOT_PROD_EXPR; | |
8733 | c2 = DOT_PROD_EXPR; | |
8734 | break; | |
8735 | ||
8736 | case SAD_EXPR: | |
8737 | c1 = SAD_EXPR; | |
8738 | c2 = SAD_EXPR; | |
8739 | break; | |
8740 | ||
6ae6116f RH |
8741 | case VEC_WIDEN_MULT_EVEN_EXPR: |
8742 | /* Support the recursion induced just above. */ | |
8743 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
8744 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
8745 | break; | |
8746 | ||
36ba4aae | 8747 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
8748 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
8749 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
8750 | break; |
8751 | ||
ebfd146a | 8752 | CASE_CONVERT: |
4a00c761 JJ |
8753 | c1 = VEC_UNPACK_LO_EXPR; |
8754 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
8755 | break; |
8756 | ||
8757 | case FLOAT_EXPR: | |
4a00c761 JJ |
8758 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
8759 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
8760 | break; |
8761 | ||
8762 | case FIX_TRUNC_EXPR: | |
8763 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
8764 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
8765 | computing the operation. */ | |
8766 | return false; | |
8767 | ||
8768 | default: | |
8769 | gcc_unreachable (); | |
8770 | } | |
8771 | ||
6ae6116f | 8772 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
6b4db501 | 8773 | std::swap (c1, c2); |
4a00c761 | 8774 | |
ebfd146a IR |
8775 | if (code == FIX_TRUNC_EXPR) |
8776 | { | |
8777 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
8778 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
8779 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
8780 | } |
8781 | else | |
8782 | { | |
8783 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
8784 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
8785 | } | |
8786 | ||
8787 | if (!optab1 || !optab2) | |
8788 | return false; | |
8789 | ||
8790 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
8791 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
8792 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
8793 | return false; |
8794 | ||
4a00c761 JJ |
8795 | *code1 = c1; |
8796 | *code2 = c2; | |
8797 | ||
8798 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
8799 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
8800 | return true; | |
8801 | ||
b8698a0f | 8802 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 8803 | types. */ |
ebfd146a | 8804 | |
4a00c761 JJ |
8805 | prev_type = vectype; |
8806 | prev_mode = vec_mode; | |
b8698a0f | 8807 | |
4a00c761 JJ |
8808 | if (!CONVERT_EXPR_CODE_P (code)) |
8809 | return false; | |
b8698a0f | 8810 | |
4a00c761 JJ |
8811 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
8812 | intermediate steps in promotion sequence. We try | |
8813 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
8814 | not. */ | |
9771b263 | 8815 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
8816 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
8817 | { | |
8818 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
8819 | intermediate_type | |
8820 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
8821 | TYPE_UNSIGNED (prev_type)); | |
8822 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); | |
8823 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
8824 | ||
8825 | if (!optab3 || !optab4 | |
8826 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
8827 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
8828 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
8829 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
8830 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
8831 | == CODE_FOR_nothing) | |
8832 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
8833 | == CODE_FOR_nothing)) | |
8834 | break; | |
ebfd146a | 8835 | |
9771b263 | 8836 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
8837 | (*multi_step_cvt)++; |
8838 | ||
8839 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
8840 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
8841 | return true; | |
8842 | ||
8843 | prev_type = intermediate_type; | |
8844 | prev_mode = intermediate_mode; | |
ebfd146a IR |
8845 | } |
8846 | ||
9771b263 | 8847 | interm_types->release (); |
4a00c761 | 8848 | return false; |
ebfd146a IR |
8849 | } |
8850 | ||
8851 | ||
8852 | /* Function supportable_narrowing_operation | |
8853 | ||
b8698a0f L |
8854 | Check whether an operation represented by the code CODE is a |
8855 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
8856 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
8857 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 8858 | |
ebfd146a | 8859 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 8860 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
8861 | the target platform directly via vector tree-codes. |
8862 | ||
8863 | Output: | |
b8698a0f L |
8864 | - CODE1 is the code of a vector operation to be used when |
8865 | vectorizing the operation, if available. | |
ebfd146a IR |
8866 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
8867 | case of multi-step conversion (like int->short->char - in that case | |
8868 | MULTI_STEP_CVT will be 1). | |
8869 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 8870 | narrowing operation (short in the above example). */ |
ebfd146a IR |
8871 | |
8872 | bool | |
8873 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 8874 | tree vectype_out, tree vectype_in, |
ebfd146a | 8875 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 8876 | vec<tree> *interm_types) |
ebfd146a | 8877 | { |
ef4bddc2 | 8878 | machine_mode vec_mode; |
ebfd146a IR |
8879 | enum insn_code icode1; |
8880 | optab optab1, interm_optab; | |
b690cc0f RG |
8881 | tree vectype = vectype_in; |
8882 | tree narrow_vectype = vectype_out; | |
ebfd146a | 8883 | enum tree_code c1; |
4a00c761 | 8884 | tree intermediate_type; |
ef4bddc2 | 8885 | machine_mode intermediate_mode, prev_mode; |
ebfd146a | 8886 | int i; |
4a00c761 | 8887 | bool uns; |
ebfd146a | 8888 | |
4a00c761 | 8889 | *multi_step_cvt = 0; |
ebfd146a IR |
8890 | switch (code) |
8891 | { | |
8892 | CASE_CONVERT: | |
8893 | c1 = VEC_PACK_TRUNC_EXPR; | |
8894 | break; | |
8895 | ||
8896 | case FIX_TRUNC_EXPR: | |
8897 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
8898 | break; | |
8899 | ||
8900 | case FLOAT_EXPR: | |
8901 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
8902 | tree code and optabs used for computing the operation. */ | |
8903 | return false; | |
8904 | ||
8905 | default: | |
8906 | gcc_unreachable (); | |
8907 | } | |
8908 | ||
8909 | if (code == FIX_TRUNC_EXPR) | |
8910 | /* The signedness is determined from output operand. */ | |
b690cc0f | 8911 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
8912 | else |
8913 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
8914 | ||
8915 | if (!optab1) | |
8916 | return false; | |
8917 | ||
8918 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 8919 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
8920 | return false; |
8921 | ||
4a00c761 JJ |
8922 | *code1 = c1; |
8923 | ||
8924 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
8925 | return true; | |
8926 | ||
ebfd146a IR |
8927 | /* Check if it's a multi-step conversion that can be done using intermediate |
8928 | types. */ | |
4a00c761 JJ |
8929 | prev_mode = vec_mode; |
8930 | if (code == FIX_TRUNC_EXPR) | |
8931 | uns = TYPE_UNSIGNED (vectype_out); | |
8932 | else | |
8933 | uns = TYPE_UNSIGNED (vectype); | |
8934 | ||
8935 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
8936 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
8937 | costly than signed. */ | |
8938 | if (code == FIX_TRUNC_EXPR && uns) | |
8939 | { | |
8940 | enum insn_code icode2; | |
8941 | ||
8942 | intermediate_type | |
8943 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
8944 | interm_optab | |
8945 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 8946 | if (interm_optab != unknown_optab |
4a00c761 JJ |
8947 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
8948 | && insn_data[icode1].operand[0].mode | |
8949 | == insn_data[icode2].operand[0].mode) | |
8950 | { | |
8951 | uns = false; | |
8952 | optab1 = interm_optab; | |
8953 | icode1 = icode2; | |
8954 | } | |
8955 | } | |
ebfd146a | 8956 | |
4a00c761 JJ |
8957 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
8958 | intermediate steps in promotion sequence. We try | |
8959 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 8960 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
8961 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
8962 | { | |
8963 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
8964 | intermediate_type | |
8965 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
8966 | interm_optab | |
8967 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
8968 | optab_default); | |
8969 | if (!interm_optab | |
8970 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
8971 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
8972 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
8973 | == CODE_FOR_nothing)) | |
8974 | break; | |
8975 | ||
9771b263 | 8976 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
8977 | (*multi_step_cvt)++; |
8978 | ||
8979 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
8980 | return true; | |
8981 | ||
8982 | prev_mode = intermediate_mode; | |
8983 | optab1 = interm_optab; | |
ebfd146a IR |
8984 | } |
8985 | ||
9771b263 | 8986 | interm_types->release (); |
4a00c761 | 8987 | return false; |
ebfd146a | 8988 | } |