]>
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" | |
78c60e3d | 25 | #include "dumpfile.h" |
ebfd146a | 26 | #include "tm.h" |
40e23961 MC |
27 | #include "hash-set.h" |
28 | #include "machmode.h" | |
29 | #include "vec.h" | |
30 | #include "double-int.h" | |
31 | #include "input.h" | |
32 | #include "alias.h" | |
33 | #include "symtab.h" | |
34 | #include "wide-int.h" | |
35 | #include "inchash.h" | |
ebfd146a | 36 | #include "tree.h" |
40e23961 | 37 | #include "fold-const.h" |
d8a2d370 | 38 | #include "stor-layout.h" |
ebfd146a | 39 | #include "target.h" |
60393bbc | 40 | #include "predict.h" |
60393bbc | 41 | #include "hard-reg-set.h" |
60393bbc AM |
42 | #include "function.h" |
43 | #include "dominance.h" | |
44 | #include "cfg.h" | |
ebfd146a | 45 | #include "basic-block.h" |
cf835838 | 46 | #include "gimple-pretty-print.h" |
2fb9a547 AM |
47 | #include "tree-ssa-alias.h" |
48 | #include "internal-fn.h" | |
49 | #include "tree-eh.h" | |
50 | #include "gimple-expr.h" | |
51 | #include "is-a.h" | |
18f429e2 | 52 | #include "gimple.h" |
45b0be94 | 53 | #include "gimplify.h" |
5be5c238 | 54 | #include "gimple-iterator.h" |
18f429e2 | 55 | #include "gimplify-me.h" |
442b4905 AM |
56 | #include "gimple-ssa.h" |
57 | #include "tree-cfg.h" | |
58 | #include "tree-phinodes.h" | |
59 | #include "ssa-iterators.h" | |
d8a2d370 | 60 | #include "stringpool.h" |
442b4905 | 61 | #include "tree-ssanames.h" |
e28030cf | 62 | #include "tree-ssa-loop-manip.h" |
ebfd146a | 63 | #include "cfgloop.h" |
0136f8f0 AH |
64 | #include "tree-ssa-loop.h" |
65 | #include "tree-scalar-evolution.h" | |
36566b39 PK |
66 | #include "hashtab.h" |
67 | #include "rtl.h" | |
68 | #include "flags.h" | |
69 | #include "statistics.h" | |
70 | #include "real.h" | |
71 | #include "fixed-value.h" | |
72 | #include "insn-config.h" | |
73 | #include "expmed.h" | |
74 | #include "dojump.h" | |
75 | #include "explow.h" | |
76 | #include "calls.h" | |
77 | #include "emit-rtl.h" | |
78 | #include "varasm.h" | |
79 | #include "stmt.h" | |
ebfd146a | 80 | #include "expr.h" |
7ee2468b | 81 | #include "recog.h" /* FIXME: for insn_data */ |
b0710fe1 | 82 | #include "insn-codes.h" |
ebfd146a | 83 | #include "optabs.h" |
718f9c0f | 84 | #include "diagnostic-core.h" |
ebfd146a | 85 | #include "tree-vectorizer.h" |
c582198b AM |
86 | #include "hash-map.h" |
87 | #include "plugin-api.h" | |
88 | #include "ipa-ref.h" | |
0136f8f0 | 89 | #include "cgraph.h" |
9b2b7279 | 90 | #include "builtins.h" |
ebfd146a | 91 | |
7ee2468b SB |
92 | /* For lang_hooks.types.type_for_mode. */ |
93 | #include "langhooks.h" | |
ebfd146a | 94 | |
c3e7ee41 BS |
95 | /* Return the vectorized type for the given statement. */ |
96 | ||
97 | tree | |
98 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
99 | { | |
100 | return STMT_VINFO_VECTYPE (stmt_info); | |
101 | } | |
102 | ||
103 | /* Return TRUE iff the given statement is in an inner loop relative to | |
104 | the loop being vectorized. */ | |
105 | bool | |
106 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
107 | { | |
108 | gimple stmt = STMT_VINFO_STMT (stmt_info); | |
109 | basic_block bb = gimple_bb (stmt); | |
110 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
111 | struct loop* loop; | |
112 | ||
113 | if (!loop_vinfo) | |
114 | return false; | |
115 | ||
116 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
117 | ||
118 | return (bb->loop_father == loop->inner); | |
119 | } | |
120 | ||
121 | /* Record the cost of a statement, either by directly informing the | |
122 | target model or by saving it in a vector for later processing. | |
123 | Return a preliminary estimate of the statement's cost. */ | |
124 | ||
125 | unsigned | |
92345349 | 126 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 127 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 128 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 129 | { |
92345349 | 130 | if (body_cost_vec) |
c3e7ee41 | 131 | { |
92345349 BS |
132 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
133 | add_stmt_info_to_vec (body_cost_vec, count, kind, | |
134 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
135 | misalign); | |
c3e7ee41 | 136 | return (unsigned) |
92345349 | 137 | (builtin_vectorization_cost (kind, vectype, misalign) * count); |
c3e7ee41 BS |
138 | |
139 | } | |
140 | else | |
141 | { | |
142 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
143 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
144 | void *target_cost_data; | |
145 | ||
146 | if (loop_vinfo) | |
147 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
148 | else | |
149 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
150 | ||
92345349 BS |
151 | return add_stmt_cost (target_cost_data, count, kind, stmt_info, |
152 | misalign, where); | |
c3e7ee41 BS |
153 | } |
154 | } | |
155 | ||
272c6793 RS |
156 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
157 | ||
158 | static tree | |
159 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
160 | { | |
161 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
162 | "vect_array"); | |
163 | } | |
164 | ||
165 | /* ARRAY is an array of vectors created by create_vector_array. | |
166 | Return an SSA_NAME for the vector in index N. The reference | |
167 | is part of the vectorization of STMT and the vector is associated | |
168 | with scalar destination SCALAR_DEST. */ | |
169 | ||
170 | static tree | |
171 | read_vector_array (gimple stmt, gimple_stmt_iterator *gsi, tree scalar_dest, | |
172 | tree array, unsigned HOST_WIDE_INT n) | |
173 | { | |
174 | tree vect_type, vect, vect_name, array_ref; | |
175 | gimple new_stmt; | |
176 | ||
177 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
178 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
179 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
180 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
181 | build_int_cst (size_type_node, n), | |
182 | NULL_TREE, NULL_TREE); | |
183 | ||
184 | new_stmt = gimple_build_assign (vect, array_ref); | |
185 | vect_name = make_ssa_name (vect, new_stmt); | |
186 | gimple_assign_set_lhs (new_stmt, vect_name); | |
187 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
188 | |
189 | return vect_name; | |
190 | } | |
191 | ||
192 | /* ARRAY is an array of vectors created by create_vector_array. | |
193 | Emit code to store SSA_NAME VECT in index N of the array. | |
194 | The store is part of the vectorization of STMT. */ | |
195 | ||
196 | static void | |
197 | write_vector_array (gimple stmt, gimple_stmt_iterator *gsi, tree vect, | |
198 | tree array, unsigned HOST_WIDE_INT n) | |
199 | { | |
200 | tree array_ref; | |
201 | gimple new_stmt; | |
202 | ||
203 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
204 | build_int_cst (size_type_node, n), | |
205 | NULL_TREE, NULL_TREE); | |
206 | ||
207 | new_stmt = gimple_build_assign (array_ref, vect); | |
208 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
209 | } |
210 | ||
211 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
212 | of *PTR. The memory reference replaces those in FIRST_DR | |
213 | (and its group). */ | |
214 | ||
215 | static tree | |
216 | create_array_ref (tree type, tree ptr, struct data_reference *first_dr) | |
217 | { | |
272c6793 RS |
218 | tree mem_ref, alias_ptr_type; |
219 | ||
220 | alias_ptr_type = reference_alias_ptr_type (DR_REF (first_dr)); | |
221 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); | |
222 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 223 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
224 | return mem_ref; |
225 | } | |
226 | ||
ebfd146a IR |
227 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
228 | ||
229 | /* Function vect_mark_relevant. | |
230 | ||
231 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
232 | ||
233 | static void | |
9771b263 | 234 | vect_mark_relevant (vec<gimple> *worklist, gimple stmt, |
83197f37 IR |
235 | enum vect_relevant relevant, bool live_p, |
236 | bool used_in_pattern) | |
ebfd146a IR |
237 | { |
238 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
239 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
240 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
83197f37 | 241 | gimple pattern_stmt; |
ebfd146a | 242 | |
73fbfcad | 243 | if (dump_enabled_p ()) |
78c60e3d | 244 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 245 | "mark relevant %d, live %d.\n", relevant, live_p); |
ebfd146a | 246 | |
83197f37 IR |
247 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
248 | related pattern stmt instead of the original stmt. However, such stmts | |
249 | may have their own uses that are not in any pattern, in such cases the | |
250 | stmt itself should be marked. */ | |
ebfd146a IR |
251 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
252 | { | |
83197f37 IR |
253 | bool found = false; |
254 | if (!used_in_pattern) | |
255 | { | |
256 | imm_use_iterator imm_iter; | |
257 | use_operand_p use_p; | |
258 | gimple use_stmt; | |
259 | tree lhs; | |
13c931c9 JJ |
260 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
261 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a | 262 | |
83197f37 IR |
263 | if (is_gimple_assign (stmt)) |
264 | lhs = gimple_assign_lhs (stmt); | |
265 | else | |
266 | lhs = gimple_call_lhs (stmt); | |
ebfd146a | 267 | |
83197f37 IR |
268 | /* This use is out of pattern use, if LHS has other uses that are |
269 | pattern uses, we should mark the stmt itself, and not the pattern | |
270 | stmt. */ | |
5ce9450f | 271 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
ab0ef706 JJ |
272 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs) |
273 | { | |
274 | if (is_gimple_debug (USE_STMT (use_p))) | |
275 | continue; | |
276 | use_stmt = USE_STMT (use_p); | |
277 | ||
13c931c9 JJ |
278 | if (!flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) |
279 | continue; | |
280 | ||
ab0ef706 JJ |
281 | if (vinfo_for_stmt (use_stmt) |
282 | && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (use_stmt))) | |
283 | { | |
284 | found = true; | |
285 | break; | |
286 | } | |
287 | } | |
83197f37 IR |
288 | } |
289 | ||
290 | if (!found) | |
291 | { | |
292 | /* This is the last stmt in a sequence that was detected as a | |
293 | pattern that can potentially be vectorized. Don't mark the stmt | |
294 | as relevant/live because it's not going to be vectorized. | |
295 | Instead mark the pattern-stmt that replaces it. */ | |
296 | ||
297 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
298 | ||
73fbfcad | 299 | if (dump_enabled_p ()) |
78c60e3d SS |
300 | dump_printf_loc (MSG_NOTE, vect_location, |
301 | "last stmt in pattern. don't mark" | |
e645e942 | 302 | " relevant/live.\n"); |
83197f37 IR |
303 | stmt_info = vinfo_for_stmt (pattern_stmt); |
304 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
305 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
306 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
307 | stmt = pattern_stmt; | |
308 | } | |
ebfd146a IR |
309 | } |
310 | ||
311 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
312 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
313 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
314 | ||
315 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
316 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
317 | { | |
73fbfcad | 318 | if (dump_enabled_p ()) |
78c60e3d | 319 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 320 | "already marked relevant/live.\n"); |
ebfd146a IR |
321 | return; |
322 | } | |
323 | ||
9771b263 | 324 | worklist->safe_push (stmt); |
ebfd146a IR |
325 | } |
326 | ||
327 | ||
328 | /* Function vect_stmt_relevant_p. | |
329 | ||
330 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
331 | "relevant for vectorization". | |
332 | ||
333 | A stmt is considered "relevant for vectorization" if: | |
334 | - it has uses outside the loop. | |
335 | - it has vdefs (it alters memory). | |
336 | - control stmts in the loop (except for the exit condition). | |
337 | ||
338 | CHECKME: what other side effects would the vectorizer allow? */ | |
339 | ||
340 | static bool | |
341 | vect_stmt_relevant_p (gimple stmt, loop_vec_info loop_vinfo, | |
342 | enum vect_relevant *relevant, bool *live_p) | |
343 | { | |
344 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
345 | ssa_op_iter op_iter; | |
346 | imm_use_iterator imm_iter; | |
347 | use_operand_p use_p; | |
348 | def_operand_p def_p; | |
349 | ||
8644a673 | 350 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
351 | *live_p = false; |
352 | ||
353 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
354 | if (is_ctrl_stmt (stmt) |
355 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
356 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 357 | *relevant = vect_used_in_scope; |
ebfd146a IR |
358 | |
359 | /* changing memory. */ | |
360 | if (gimple_code (stmt) != GIMPLE_PHI) | |
ac6aeab4 RB |
361 | if (gimple_vdef (stmt) |
362 | && !gimple_clobber_p (stmt)) | |
ebfd146a | 363 | { |
73fbfcad | 364 | if (dump_enabled_p ()) |
78c60e3d | 365 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 366 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 367 | *relevant = vect_used_in_scope; |
ebfd146a IR |
368 | } |
369 | ||
370 | /* uses outside the loop. */ | |
371 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
372 | { | |
373 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
374 | { | |
375 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
376 | if (!flow_bb_inside_loop_p (loop, bb)) | |
377 | { | |
73fbfcad | 378 | if (dump_enabled_p ()) |
78c60e3d | 379 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 380 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 381 | |
3157b0c2 AO |
382 | if (is_gimple_debug (USE_STMT (use_p))) |
383 | continue; | |
384 | ||
ebfd146a IR |
385 | /* We expect all such uses to be in the loop exit phis |
386 | (because of loop closed form) */ | |
387 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
388 | gcc_assert (bb == single_exit (loop)->dest); | |
389 | ||
390 | *live_p = true; | |
391 | } | |
392 | } | |
393 | } | |
394 | ||
395 | return (*live_p || *relevant); | |
396 | } | |
397 | ||
398 | ||
b8698a0f | 399 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 400 | |
ff802fa1 | 401 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
402 | used in STMT for anything other than indexing an array. */ |
403 | ||
404 | static bool | |
405 | exist_non_indexing_operands_for_use_p (tree use, gimple stmt) | |
406 | { | |
407 | tree operand; | |
408 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 409 | |
ff802fa1 | 410 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
411 | reference in STMT, then any operand that corresponds to USE |
412 | is not indexing an array. */ | |
413 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
414 | return true; | |
59a05b0c | 415 | |
ebfd146a IR |
416 | /* STMT has a data_ref. FORNOW this means that its of one of |
417 | the following forms: | |
418 | -1- ARRAY_REF = var | |
419 | -2- var = ARRAY_REF | |
420 | (This should have been verified in analyze_data_refs). | |
421 | ||
422 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 423 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
424 | for array indexing. |
425 | ||
426 | Therefore, all we need to check is if STMT falls into the | |
427 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
428 | |
429 | if (!gimple_assign_copy_p (stmt)) | |
5ce9450f JJ |
430 | { |
431 | if (is_gimple_call (stmt) | |
432 | && gimple_call_internal_p (stmt)) | |
433 | switch (gimple_call_internal_fn (stmt)) | |
434 | { | |
435 | case IFN_MASK_STORE: | |
436 | operand = gimple_call_arg (stmt, 3); | |
437 | if (operand == use) | |
438 | return true; | |
439 | /* FALLTHRU */ | |
440 | case IFN_MASK_LOAD: | |
441 | operand = gimple_call_arg (stmt, 2); | |
442 | if (operand == use) | |
443 | return true; | |
444 | break; | |
445 | default: | |
446 | break; | |
447 | } | |
448 | return false; | |
449 | } | |
450 | ||
59a05b0c EB |
451 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
452 | return false; | |
ebfd146a | 453 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
454 | if (TREE_CODE (operand) != SSA_NAME) |
455 | return false; | |
456 | ||
457 | if (operand == use) | |
458 | return true; | |
459 | ||
460 | return false; | |
461 | } | |
462 | ||
463 | ||
b8698a0f | 464 | /* |
ebfd146a IR |
465 | Function process_use. |
466 | ||
467 | Inputs: | |
468 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b8698a0f | 469 | - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt |
ff802fa1 | 470 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 471 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
472 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
473 | be performed. | |
ebfd146a IR |
474 | |
475 | Outputs: | |
476 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
477 | relevance info of the DEF_STMT of this USE: | |
478 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
479 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
480 | Exceptions: | |
481 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 482 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 483 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
484 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
485 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
486 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
487 | be modified accordingly. | |
488 | ||
489 | Return true if everything is as expected. Return false otherwise. */ | |
490 | ||
491 | static bool | |
b8698a0f | 492 | process_use (gimple stmt, tree use, loop_vec_info loop_vinfo, bool live_p, |
9771b263 | 493 | enum vect_relevant relevant, vec<gimple> *worklist, |
aec7ae7d | 494 | bool force) |
ebfd146a IR |
495 | { |
496 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
497 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
498 | stmt_vec_info dstmt_vinfo; | |
499 | basic_block bb, def_bb; | |
500 | tree def; | |
501 | gimple def_stmt; | |
502 | enum vect_def_type dt; | |
503 | ||
b8698a0f | 504 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 505 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 506 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
507 | return true; |
508 | ||
24ee1384 | 509 | if (!vect_is_simple_use (use, stmt, loop_vinfo, NULL, &def_stmt, &def, &dt)) |
b8698a0f | 510 | { |
73fbfcad | 511 | if (dump_enabled_p ()) |
78c60e3d | 512 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 513 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
514 | return false; |
515 | } | |
516 | ||
517 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
518 | return true; | |
519 | ||
520 | def_bb = gimple_bb (def_stmt); | |
521 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
522 | { | |
73fbfcad | 523 | if (dump_enabled_p ()) |
e645e942 | 524 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n"); |
ebfd146a IR |
525 | return true; |
526 | } | |
527 | ||
b8698a0f L |
528 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
529 | DEF_STMT must have already been processed, because this should be the | |
530 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
531 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
532 | check that everything is as expected, and we are done. */ |
533 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
534 | bb = gimple_bb (stmt); | |
535 | if (gimple_code (stmt) == GIMPLE_PHI | |
536 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
537 | && gimple_code (def_stmt) != GIMPLE_PHI | |
538 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
539 | && bb->loop_father == def_bb->loop_father) | |
540 | { | |
73fbfcad | 541 | if (dump_enabled_p ()) |
78c60e3d | 542 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 543 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a IR |
544 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
545 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
546 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 547 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 548 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
549 | return true; |
550 | } | |
551 | ||
552 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
553 | outer-loop-header-bb: | |
554 | d = def_stmt | |
555 | inner-loop: | |
556 | stmt # use (d) | |
557 | outer-loop-tail-bb: | |
558 | ... */ | |
559 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
560 | { | |
73fbfcad | 561 | if (dump_enabled_p ()) |
78c60e3d | 562 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 563 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 564 | |
ebfd146a IR |
565 | switch (relevant) |
566 | { | |
8644a673 | 567 | case vect_unused_in_scope: |
7c5222ff IR |
568 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
569 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 570 | break; |
7c5222ff | 571 | |
ebfd146a | 572 | case vect_used_in_outer_by_reduction: |
7c5222ff | 573 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
574 | relevant = vect_used_by_reduction; |
575 | break; | |
7c5222ff | 576 | |
ebfd146a | 577 | case vect_used_in_outer: |
7c5222ff | 578 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 579 | relevant = vect_used_in_scope; |
ebfd146a | 580 | break; |
7c5222ff | 581 | |
8644a673 | 582 | case vect_used_in_scope: |
ebfd146a IR |
583 | break; |
584 | ||
585 | default: | |
586 | gcc_unreachable (); | |
b8698a0f | 587 | } |
ebfd146a IR |
588 | } |
589 | ||
590 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
591 | outer-loop-header-bb: | |
592 | ... | |
593 | inner-loop: | |
594 | d = def_stmt | |
06066f92 | 595 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
596 | stmt # use (d) */ |
597 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
598 | { | |
73fbfcad | 599 | if (dump_enabled_p ()) |
78c60e3d | 600 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 601 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 602 | |
ebfd146a IR |
603 | switch (relevant) |
604 | { | |
8644a673 | 605 | case vect_unused_in_scope: |
b8698a0f | 606 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 607 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 608 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
609 | break; |
610 | ||
ebfd146a IR |
611 | case vect_used_by_reduction: |
612 | relevant = vect_used_in_outer_by_reduction; | |
613 | break; | |
614 | ||
8644a673 | 615 | case vect_used_in_scope: |
ebfd146a IR |
616 | relevant = vect_used_in_outer; |
617 | break; | |
618 | ||
619 | default: | |
620 | gcc_unreachable (); | |
621 | } | |
622 | } | |
623 | ||
83197f37 IR |
624 | vect_mark_relevant (worklist, def_stmt, relevant, live_p, |
625 | is_pattern_stmt_p (stmt_vinfo)); | |
ebfd146a IR |
626 | return true; |
627 | } | |
628 | ||
629 | ||
630 | /* Function vect_mark_stmts_to_be_vectorized. | |
631 | ||
632 | Not all stmts in the loop need to be vectorized. For example: | |
633 | ||
634 | for i... | |
635 | for j... | |
636 | 1. T0 = i + j | |
637 | 2. T1 = a[T0] | |
638 | ||
639 | 3. j = j + 1 | |
640 | ||
641 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
642 | addressing of vectorized data-refs are handled differently. | |
643 | ||
644 | This pass detects such stmts. */ | |
645 | ||
646 | bool | |
647 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
648 | { | |
ebfd146a IR |
649 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
650 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
651 | unsigned int nbbs = loop->num_nodes; | |
652 | gimple_stmt_iterator si; | |
653 | gimple stmt; | |
654 | unsigned int i; | |
655 | stmt_vec_info stmt_vinfo; | |
656 | basic_block bb; | |
657 | gimple phi; | |
658 | bool live_p; | |
06066f92 IR |
659 | enum vect_relevant relevant, tmp_relevant; |
660 | enum vect_def_type def_type; | |
ebfd146a | 661 | |
73fbfcad | 662 | if (dump_enabled_p ()) |
78c60e3d | 663 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 664 | "=== vect_mark_stmts_to_be_vectorized ===\n"); |
ebfd146a | 665 | |
00f96dc9 | 666 | auto_vec<gimple, 64> worklist; |
ebfd146a IR |
667 | |
668 | /* 1. Init worklist. */ | |
669 | for (i = 0; i < nbbs; i++) | |
670 | { | |
671 | bb = bbs[i]; | |
672 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 673 | { |
ebfd146a | 674 | phi = gsi_stmt (si); |
73fbfcad | 675 | if (dump_enabled_p ()) |
ebfd146a | 676 | { |
78c60e3d SS |
677 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
678 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
679 | } |
680 | ||
681 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 682 | vect_mark_relevant (&worklist, phi, relevant, live_p, false); |
ebfd146a IR |
683 | } |
684 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
685 | { | |
686 | stmt = gsi_stmt (si); | |
73fbfcad | 687 | if (dump_enabled_p ()) |
ebfd146a | 688 | { |
78c60e3d SS |
689 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
690 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
b8698a0f | 691 | } |
ebfd146a IR |
692 | |
693 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 694 | vect_mark_relevant (&worklist, stmt, relevant, live_p, false); |
ebfd146a IR |
695 | } |
696 | } | |
697 | ||
698 | /* 2. Process_worklist */ | |
9771b263 | 699 | while (worklist.length () > 0) |
ebfd146a IR |
700 | { |
701 | use_operand_p use_p; | |
702 | ssa_op_iter iter; | |
703 | ||
9771b263 | 704 | stmt = worklist.pop (); |
73fbfcad | 705 | if (dump_enabled_p ()) |
ebfd146a | 706 | { |
78c60e3d SS |
707 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
708 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
ebfd146a IR |
709 | } |
710 | ||
b8698a0f L |
711 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
712 | (DEF_STMT) as relevant/irrelevant and live/dead according to the | |
ebfd146a IR |
713 | liveness and relevance properties of STMT. */ |
714 | stmt_vinfo = vinfo_for_stmt (stmt); | |
715 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
716 | live_p = STMT_VINFO_LIVE_P (stmt_vinfo); | |
717 | ||
718 | /* Generally, the liveness and relevance properties of STMT are | |
719 | propagated as is to the DEF_STMTs of its USEs: | |
720 | live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO) | |
721 | relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO) | |
722 | ||
723 | One exception is when STMT has been identified as defining a reduction | |
724 | variable; in this case we set the liveness/relevance as follows: | |
725 | live_p = false | |
726 | relevant = vect_used_by_reduction | |
727 | This is because we distinguish between two kinds of relevant stmts - | |
b8698a0f | 728 | those that are used by a reduction computation, and those that are |
ff802fa1 | 729 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 730 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 731 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 732 | |
06066f92 IR |
733 | def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
734 | tmp_relevant = relevant; | |
735 | switch (def_type) | |
ebfd146a | 736 | { |
06066f92 IR |
737 | case vect_reduction_def: |
738 | switch (tmp_relevant) | |
739 | { | |
740 | case vect_unused_in_scope: | |
741 | relevant = vect_used_by_reduction; | |
742 | break; | |
743 | ||
744 | case vect_used_by_reduction: | |
745 | if (gimple_code (stmt) == GIMPLE_PHI) | |
746 | break; | |
747 | /* fall through */ | |
748 | ||
749 | default: | |
73fbfcad | 750 | if (dump_enabled_p ()) |
78c60e3d | 751 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 752 | "unsupported use of reduction.\n"); |
06066f92 IR |
753 | return false; |
754 | } | |
755 | ||
b8698a0f | 756 | live_p = false; |
06066f92 | 757 | break; |
b8698a0f | 758 | |
06066f92 IR |
759 | case vect_nested_cycle: |
760 | if (tmp_relevant != vect_unused_in_scope | |
761 | && tmp_relevant != vect_used_in_outer_by_reduction | |
762 | && tmp_relevant != vect_used_in_outer) | |
763 | { | |
73fbfcad | 764 | if (dump_enabled_p ()) |
78c60e3d | 765 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 766 | "unsupported use of nested cycle.\n"); |
7c5222ff | 767 | |
06066f92 IR |
768 | return false; |
769 | } | |
7c5222ff | 770 | |
b8698a0f L |
771 | live_p = false; |
772 | break; | |
773 | ||
06066f92 IR |
774 | case vect_double_reduction_def: |
775 | if (tmp_relevant != vect_unused_in_scope | |
776 | && tmp_relevant != vect_used_by_reduction) | |
777 | { | |
73fbfcad | 778 | if (dump_enabled_p ()) |
78c60e3d | 779 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 780 | "unsupported use of double reduction.\n"); |
7c5222ff | 781 | |
7c5222ff | 782 | return false; |
06066f92 IR |
783 | } |
784 | ||
785 | live_p = false; | |
b8698a0f | 786 | break; |
7c5222ff | 787 | |
06066f92 IR |
788 | default: |
789 | break; | |
7c5222ff | 790 | } |
b8698a0f | 791 | |
aec7ae7d | 792 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
793 | { |
794 | /* Pattern statements are not inserted into the code, so | |
795 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
796 | have to scan the RHS or function arguments instead. */ | |
797 | if (is_gimple_assign (stmt)) | |
798 | { | |
69d2aade JJ |
799 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
800 | tree op = gimple_assign_rhs1 (stmt); | |
801 | ||
802 | i = 1; | |
803 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
804 | { | |
805 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
aec7ae7d | 806 | live_p, relevant, &worklist, false) |
69d2aade | 807 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
aec7ae7d | 808 | live_p, relevant, &worklist, false)) |
566d377a | 809 | return false; |
69d2aade JJ |
810 | i = 2; |
811 | } | |
812 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 813 | { |
69d2aade | 814 | op = gimple_op (stmt, i); |
9d5e7640 | 815 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, |
aec7ae7d | 816 | &worklist, false)) |
07687835 | 817 | return false; |
9d5e7640 IR |
818 | } |
819 | } | |
820 | else if (is_gimple_call (stmt)) | |
821 | { | |
822 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
823 | { | |
824 | tree arg = gimple_call_arg (stmt, i); | |
825 | if (!process_use (stmt, arg, loop_vinfo, live_p, relevant, | |
aec7ae7d | 826 | &worklist, false)) |
07687835 | 827 | return false; |
9d5e7640 IR |
828 | } |
829 | } | |
830 | } | |
831 | else | |
832 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
833 | { | |
834 | tree op = USE_FROM_PTR (use_p); | |
835 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, | |
aec7ae7d | 836 | &worklist, false)) |
07687835 | 837 | return false; |
9d5e7640 | 838 | } |
aec7ae7d JJ |
839 | |
840 | if (STMT_VINFO_GATHER_P (stmt_vinfo)) | |
841 | { | |
842 | tree off; | |
843 | tree decl = vect_check_gather (stmt, loop_vinfo, NULL, &off, NULL); | |
844 | gcc_assert (decl); | |
845 | if (!process_use (stmt, off, loop_vinfo, live_p, relevant, | |
846 | &worklist, true)) | |
566d377a | 847 | return false; |
aec7ae7d | 848 | } |
ebfd146a IR |
849 | } /* while worklist */ |
850 | ||
ebfd146a IR |
851 | return true; |
852 | } | |
853 | ||
854 | ||
b8698a0f | 855 | /* Function vect_model_simple_cost. |
ebfd146a | 856 | |
b8698a0f | 857 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
858 | single op. Right now, this does not account for multiple insns that could |
859 | be generated for the single vector op. We will handle that shortly. */ | |
860 | ||
861 | void | |
b8698a0f | 862 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 BS |
863 | enum vect_def_type *dt, |
864 | stmt_vector_for_cost *prologue_cost_vec, | |
865 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
866 | { |
867 | int i; | |
92345349 | 868 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
869 | |
870 | /* The SLP costs were already calculated during SLP tree build. */ | |
871 | if (PURE_SLP_STMT (stmt_info)) | |
872 | return; | |
873 | ||
ebfd146a IR |
874 | /* FORNOW: Assuming maximum 2 args per stmts. */ |
875 | for (i = 0; i < 2; i++) | |
92345349 BS |
876 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
877 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, vector_stmt, | |
878 | stmt_info, 0, vect_prologue); | |
c3e7ee41 BS |
879 | |
880 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
92345349 BS |
881 | inside_cost = record_stmt_cost (body_cost_vec, ncopies, vector_stmt, |
882 | stmt_info, 0, vect_body); | |
c3e7ee41 | 883 | |
73fbfcad | 884 | if (dump_enabled_p ()) |
78c60e3d SS |
885 | dump_printf_loc (MSG_NOTE, vect_location, |
886 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 887 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
888 | } |
889 | ||
890 | ||
8bd37302 BS |
891 | /* Model cost for type demotion and promotion operations. PWR is normally |
892 | zero for single-step promotions and demotions. It will be one if | |
893 | two-step promotion/demotion is required, and so on. Each additional | |
894 | step doubles the number of instructions required. */ | |
895 | ||
896 | static void | |
897 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
898 | enum vect_def_type *dt, int pwr) | |
899 | { | |
900 | int i, tmp; | |
92345349 | 901 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 BS |
902 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
903 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
904 | void *target_cost_data; | |
8bd37302 BS |
905 | |
906 | /* The SLP costs were already calculated during SLP tree build. */ | |
907 | if (PURE_SLP_STMT (stmt_info)) | |
908 | return; | |
909 | ||
c3e7ee41 BS |
910 | if (loop_vinfo) |
911 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
912 | else | |
913 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
914 | ||
8bd37302 BS |
915 | for (i = 0; i < pwr + 1; i++) |
916 | { | |
917 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
918 | (i + 1) : i; | |
c3e7ee41 | 919 | inside_cost += add_stmt_cost (target_cost_data, vect_pow2 (tmp), |
92345349 BS |
920 | vec_promote_demote, stmt_info, 0, |
921 | vect_body); | |
8bd37302 BS |
922 | } |
923 | ||
924 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
925 | for (i = 0; i < 2; i++) | |
92345349 BS |
926 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
927 | prologue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt, | |
928 | stmt_info, 0, vect_prologue); | |
8bd37302 | 929 | |
73fbfcad | 930 | if (dump_enabled_p ()) |
78c60e3d SS |
931 | dump_printf_loc (MSG_NOTE, vect_location, |
932 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 933 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
934 | } |
935 | ||
0d0293ac | 936 | /* Function vect_cost_group_size |
b8698a0f | 937 | |
0d0293ac | 938 | For grouped load or store, return the group_size only if it is the first |
ebfd146a IR |
939 | load or store of a group, else return 1. This ensures that group size is |
940 | only returned once per group. */ | |
941 | ||
942 | static int | |
0d0293ac | 943 | vect_cost_group_size (stmt_vec_info stmt_info) |
ebfd146a | 944 | { |
e14c1050 | 945 | gimple first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a IR |
946 | |
947 | if (first_stmt == STMT_VINFO_STMT (stmt_info)) | |
e14c1050 | 948 | return GROUP_SIZE (stmt_info); |
ebfd146a IR |
949 | |
950 | return 1; | |
951 | } | |
952 | ||
953 | ||
954 | /* Function vect_model_store_cost | |
955 | ||
0d0293ac MM |
956 | Models cost for stores. In the case of grouped accesses, one access |
957 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a IR |
958 | |
959 | void | |
b8698a0f | 960 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
272c6793 | 961 | bool store_lanes_p, enum vect_def_type dt, |
92345349 BS |
962 | slp_tree slp_node, |
963 | stmt_vector_for_cost *prologue_cost_vec, | |
964 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
965 | { |
966 | int group_size; | |
92345349 | 967 | unsigned int inside_cost = 0, prologue_cost = 0; |
720f5239 IR |
968 | struct data_reference *first_dr; |
969 | gimple first_stmt; | |
ebfd146a IR |
970 | |
971 | /* The SLP costs were already calculated during SLP tree build. */ | |
972 | if (PURE_SLP_STMT (stmt_info)) | |
973 | return; | |
974 | ||
8644a673 | 975 | if (dt == vect_constant_def || dt == vect_external_def) |
92345349 BS |
976 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
977 | stmt_info, 0, vect_prologue); | |
ebfd146a | 978 | |
0d0293ac MM |
979 | /* Grouped access? */ |
980 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
720f5239 IR |
981 | { |
982 | if (slp_node) | |
983 | { | |
9771b263 | 984 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
720f5239 IR |
985 | group_size = 1; |
986 | } | |
987 | else | |
988 | { | |
e14c1050 | 989 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 990 | group_size = vect_cost_group_size (stmt_info); |
720f5239 IR |
991 | } |
992 | ||
993 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
994 | } | |
0d0293ac | 995 | /* Not a grouped access. */ |
ebfd146a | 996 | else |
720f5239 IR |
997 | { |
998 | group_size = 1; | |
999 | first_dr = STMT_VINFO_DATA_REF (stmt_info); | |
1000 | } | |
ebfd146a | 1001 | |
272c6793 | 1002 | /* We assume that the cost of a single store-lanes instruction is |
0d0293ac | 1003 | equivalent to the cost of GROUP_SIZE separate stores. If a grouped |
272c6793 RS |
1004 | access is instead being provided by a permute-and-store operation, |
1005 | include the cost of the permutes. */ | |
1006 | if (!store_lanes_p && group_size > 1) | |
ebfd146a | 1007 | { |
e1377713 ES |
1008 | /* Uses a high and low interleave or shuffle operations for each |
1009 | needed permute. */ | |
1010 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; | |
92345349 BS |
1011 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
1012 | stmt_info, 0, vect_body); | |
ebfd146a | 1013 | |
73fbfcad | 1014 | if (dump_enabled_p ()) |
78c60e3d | 1015 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1016 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 1017 | group_size); |
ebfd146a IR |
1018 | } |
1019 | ||
1020 | /* Costs of the stores. */ | |
92345349 | 1021 | vect_get_store_cost (first_dr, ncopies, &inside_cost, body_cost_vec); |
ebfd146a | 1022 | |
73fbfcad | 1023 | if (dump_enabled_p ()) |
78c60e3d SS |
1024 | dump_printf_loc (MSG_NOTE, vect_location, |
1025 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 1026 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
1027 | } |
1028 | ||
1029 | ||
720f5239 IR |
1030 | /* Calculate cost of DR's memory access. */ |
1031 | void | |
1032 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1033 | unsigned int *inside_cost, |
92345349 | 1034 | stmt_vector_for_cost *body_cost_vec) |
720f5239 IR |
1035 | { |
1036 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
c3e7ee41 BS |
1037 | gimple stmt = DR_STMT (dr); |
1038 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
720f5239 IR |
1039 | |
1040 | switch (alignment_support_scheme) | |
1041 | { | |
1042 | case dr_aligned: | |
1043 | { | |
92345349 BS |
1044 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
1045 | vector_store, stmt_info, 0, | |
1046 | vect_body); | |
720f5239 | 1047 | |
73fbfcad | 1048 | if (dump_enabled_p ()) |
78c60e3d | 1049 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1050 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
1051 | break; |
1052 | } | |
1053 | ||
1054 | case dr_unaligned_supported: | |
1055 | { | |
720f5239 | 1056 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1057 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1058 | unaligned_store, stmt_info, |
92345349 | 1059 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1060 | if (dump_enabled_p ()) |
78c60e3d SS |
1061 | dump_printf_loc (MSG_NOTE, vect_location, |
1062 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1063 | "hardware.\n"); |
720f5239 IR |
1064 | break; |
1065 | } | |
1066 | ||
38eec4c6 UW |
1067 | case dr_unaligned_unsupported: |
1068 | { | |
1069 | *inside_cost = VECT_MAX_COST; | |
1070 | ||
73fbfcad | 1071 | if (dump_enabled_p ()) |
78c60e3d | 1072 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1073 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1074 | break; |
1075 | } | |
1076 | ||
720f5239 IR |
1077 | default: |
1078 | gcc_unreachable (); | |
1079 | } | |
1080 | } | |
1081 | ||
1082 | ||
ebfd146a IR |
1083 | /* Function vect_model_load_cost |
1084 | ||
0d0293ac MM |
1085 | Models cost for loads. In the case of grouped accesses, the last access |
1086 | has the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1087 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1088 | access scheme chosen. */ |
1089 | ||
1090 | void | |
92345349 BS |
1091 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, |
1092 | bool load_lanes_p, slp_tree slp_node, | |
1093 | stmt_vector_for_cost *prologue_cost_vec, | |
1094 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
1095 | { |
1096 | int group_size; | |
ebfd146a IR |
1097 | gimple first_stmt; |
1098 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; | |
92345349 | 1099 | unsigned int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
1100 | |
1101 | /* The SLP costs were already calculated during SLP tree build. */ | |
1102 | if (PURE_SLP_STMT (stmt_info)) | |
1103 | return; | |
1104 | ||
0d0293ac | 1105 | /* Grouped accesses? */ |
e14c1050 | 1106 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 1107 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && first_stmt && !slp_node) |
ebfd146a | 1108 | { |
0d0293ac | 1109 | group_size = vect_cost_group_size (stmt_info); |
ebfd146a IR |
1110 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
1111 | } | |
0d0293ac | 1112 | /* Not a grouped access. */ |
ebfd146a IR |
1113 | else |
1114 | { | |
1115 | group_size = 1; | |
1116 | first_dr = dr; | |
1117 | } | |
1118 | ||
272c6793 | 1119 | /* We assume that the cost of a single load-lanes instruction is |
0d0293ac | 1120 | equivalent to the cost of GROUP_SIZE separate loads. If a grouped |
272c6793 RS |
1121 | access is instead being provided by a load-and-permute operation, |
1122 | include the cost of the permutes. */ | |
1123 | if (!load_lanes_p && group_size > 1) | |
ebfd146a | 1124 | { |
2c23db6d ES |
1125 | /* Uses an even and odd extract operations or shuffle operations |
1126 | for each needed permute. */ | |
1127 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; | |
1128 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, | |
1129 | stmt_info, 0, vect_body); | |
ebfd146a | 1130 | |
73fbfcad | 1131 | if (dump_enabled_p ()) |
e645e942 TJ |
1132 | dump_printf_loc (MSG_NOTE, vect_location, |
1133 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1134 | group_size); |
ebfd146a IR |
1135 | } |
1136 | ||
1137 | /* The loads themselves. */ | |
a82960aa RG |
1138 | if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
1139 | { | |
a21892ad BS |
1140 | /* N scalar loads plus gathering them into a vector. */ |
1141 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
92345349 | 1142 | inside_cost += record_stmt_cost (body_cost_vec, |
c3e7ee41 | 1143 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), |
92345349 BS |
1144 | scalar_load, stmt_info, 0, vect_body); |
1145 | inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct, | |
1146 | stmt_info, 0, vect_body); | |
a82960aa RG |
1147 | } |
1148 | else | |
1149 | vect_get_load_cost (first_dr, ncopies, | |
1150 | ((!STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1151 | || group_size > 1 || slp_node), | |
92345349 BS |
1152 | &inside_cost, &prologue_cost, |
1153 | prologue_cost_vec, body_cost_vec, true); | |
720f5239 | 1154 | |
73fbfcad | 1155 | if (dump_enabled_p ()) |
78c60e3d SS |
1156 | dump_printf_loc (MSG_NOTE, vect_location, |
1157 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1158 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1159 | } |
1160 | ||
1161 | ||
1162 | /* Calculate cost of DR's memory access. */ | |
1163 | void | |
1164 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1165 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1166 | unsigned int *prologue_cost, |
1167 | stmt_vector_for_cost *prologue_cost_vec, | |
1168 | stmt_vector_for_cost *body_cost_vec, | |
1169 | bool record_prologue_costs) | |
720f5239 IR |
1170 | { |
1171 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
c3e7ee41 BS |
1172 | gimple stmt = DR_STMT (dr); |
1173 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
720f5239 IR |
1174 | |
1175 | switch (alignment_support_scheme) | |
ebfd146a IR |
1176 | { |
1177 | case dr_aligned: | |
1178 | { | |
92345349 BS |
1179 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1180 | stmt_info, 0, vect_body); | |
ebfd146a | 1181 | |
73fbfcad | 1182 | if (dump_enabled_p ()) |
78c60e3d | 1183 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1184 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1185 | |
1186 | break; | |
1187 | } | |
1188 | case dr_unaligned_supported: | |
1189 | { | |
720f5239 | 1190 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1191 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1192 | unaligned_load, stmt_info, |
92345349 | 1193 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1194 | |
73fbfcad | 1195 | if (dump_enabled_p ()) |
78c60e3d SS |
1196 | dump_printf_loc (MSG_NOTE, vect_location, |
1197 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1198 | "hardware.\n"); |
ebfd146a IR |
1199 | |
1200 | break; | |
1201 | } | |
1202 | case dr_explicit_realign: | |
1203 | { | |
92345349 BS |
1204 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1205 | vector_load, stmt_info, 0, vect_body); | |
1206 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1207 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1208 | |
1209 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1210 | the containing loop, the following cost should be added to the | |
92345349 | 1211 | prologue costs. */ |
ebfd146a | 1212 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1213 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1214 | stmt_info, 0, vect_body); | |
ebfd146a | 1215 | |
73fbfcad | 1216 | if (dump_enabled_p ()) |
e645e942 TJ |
1217 | dump_printf_loc (MSG_NOTE, vect_location, |
1218 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1219 | |
ebfd146a IR |
1220 | break; |
1221 | } | |
1222 | case dr_explicit_realign_optimized: | |
1223 | { | |
73fbfcad | 1224 | if (dump_enabled_p ()) |
e645e942 | 1225 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1226 | "vect_model_load_cost: unaligned software " |
e645e942 | 1227 | "pipelined.\n"); |
ebfd146a IR |
1228 | |
1229 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1230 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1231 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1232 | access, then the above cost should only be considered for one |
ff802fa1 | 1233 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1234 | and a realignment op. */ |
1235 | ||
92345349 | 1236 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1237 | { |
92345349 BS |
1238 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1239 | vector_stmt, stmt_info, | |
1240 | 0, vect_prologue); | |
ebfd146a | 1241 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1242 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1243 | vector_stmt, stmt_info, | |
1244 | 0, vect_prologue); | |
ebfd146a IR |
1245 | } |
1246 | ||
92345349 BS |
1247 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1248 | stmt_info, 0, vect_body); | |
1249 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1250 | stmt_info, 0, vect_body); | |
8bd37302 | 1251 | |
73fbfcad | 1252 | if (dump_enabled_p ()) |
78c60e3d | 1253 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1254 | "vect_model_load_cost: explicit realign optimized" |
1255 | "\n"); | |
8bd37302 | 1256 | |
ebfd146a IR |
1257 | break; |
1258 | } | |
1259 | ||
38eec4c6 UW |
1260 | case dr_unaligned_unsupported: |
1261 | { | |
1262 | *inside_cost = VECT_MAX_COST; | |
1263 | ||
73fbfcad | 1264 | if (dump_enabled_p ()) |
78c60e3d | 1265 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1266 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1267 | break; |
1268 | } | |
1269 | ||
ebfd146a IR |
1270 | default: |
1271 | gcc_unreachable (); | |
1272 | } | |
ebfd146a IR |
1273 | } |
1274 | ||
418b7df3 RG |
1275 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1276 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1277 | |
418b7df3 RG |
1278 | static void |
1279 | vect_init_vector_1 (gimple stmt, gimple new_stmt, gimple_stmt_iterator *gsi) | |
ebfd146a | 1280 | { |
ebfd146a | 1281 | if (gsi) |
418b7df3 | 1282 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1283 | else |
1284 | { | |
418b7df3 | 1285 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1286 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1287 | |
a70d6342 IR |
1288 | if (loop_vinfo) |
1289 | { | |
1290 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1291 | basic_block new_bb; |
1292 | edge pe; | |
a70d6342 IR |
1293 | |
1294 | if (nested_in_vect_loop_p (loop, stmt)) | |
1295 | loop = loop->inner; | |
b8698a0f | 1296 | |
a70d6342 | 1297 | pe = loop_preheader_edge (loop); |
418b7df3 | 1298 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1299 | gcc_assert (!new_bb); |
1300 | } | |
1301 | else | |
1302 | { | |
1303 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1304 | basic_block bb; | |
1305 | gimple_stmt_iterator gsi_bb_start; | |
1306 | ||
1307 | gcc_assert (bb_vinfo); | |
1308 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1309 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1310 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1311 | } |
ebfd146a IR |
1312 | } |
1313 | ||
73fbfcad | 1314 | if (dump_enabled_p ()) |
ebfd146a | 1315 | { |
78c60e3d SS |
1316 | dump_printf_loc (MSG_NOTE, vect_location, |
1317 | "created new init_stmt: "); | |
1318 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
ebfd146a | 1319 | } |
418b7df3 RG |
1320 | } |
1321 | ||
1322 | /* Function vect_init_vector. | |
ebfd146a | 1323 | |
5467ee52 RG |
1324 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1325 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1326 | vector type a vector with all elements equal to VAL is created first. | |
1327 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1328 | initialization at the loop preheader. | |
418b7df3 RG |
1329 | Return the DEF of INIT_STMT. |
1330 | It will be used in the vectorization of STMT. */ | |
1331 | ||
1332 | tree | |
5467ee52 | 1333 | vect_init_vector (gimple stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 RG |
1334 | { |
1335 | tree new_var; | |
1336 | gimple init_stmt; | |
1337 | tree vec_oprnd; | |
1338 | tree new_temp; | |
1339 | ||
5467ee52 RG |
1340 | if (TREE_CODE (type) == VECTOR_TYPE |
1341 | && TREE_CODE (TREE_TYPE (val)) != VECTOR_TYPE) | |
418b7df3 | 1342 | { |
5467ee52 | 1343 | if (!types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) |
418b7df3 | 1344 | { |
5467ee52 RG |
1345 | if (CONSTANT_CLASS_P (val)) |
1346 | val = fold_unary (VIEW_CONVERT_EXPR, TREE_TYPE (type), val); | |
418b7df3 RG |
1347 | else |
1348 | { | |
b731b390 | 1349 | new_temp = make_ssa_name (TREE_TYPE (type)); |
0d0e4a03 | 1350 | init_stmt = gimple_build_assign (new_temp, NOP_EXPR, val); |
418b7df3 | 1351 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1352 | val = new_temp; |
418b7df3 RG |
1353 | } |
1354 | } | |
5467ee52 | 1355 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1356 | } |
1357 | ||
5467ee52 | 1358 | new_var = vect_get_new_vect_var (type, vect_simple_var, "cst_"); |
5467ee52 | 1359 | init_stmt = gimple_build_assign (new_var, val); |
418b7df3 RG |
1360 | new_temp = make_ssa_name (new_var, init_stmt); |
1361 | gimple_assign_set_lhs (init_stmt, new_temp); | |
1362 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
ebfd146a IR |
1363 | vec_oprnd = gimple_assign_lhs (init_stmt); |
1364 | return vec_oprnd; | |
1365 | } | |
1366 | ||
a70d6342 | 1367 | |
ebfd146a IR |
1368 | /* Function vect_get_vec_def_for_operand. |
1369 | ||
ff802fa1 | 1370 | OP is an operand in STMT. This function returns a (vector) def that will be |
ebfd146a IR |
1371 | used in the vectorized stmt for STMT. |
1372 | ||
1373 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1374 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1375 | ||
1376 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1377 | needs to be introduced. */ | |
1378 | ||
1379 | tree | |
1380 | vect_get_vec_def_for_operand (tree op, gimple stmt, tree *scalar_def) | |
1381 | { | |
1382 | tree vec_oprnd; | |
1383 | gimple vec_stmt; | |
1384 | gimple def_stmt; | |
1385 | stmt_vec_info def_stmt_info = NULL; | |
1386 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
9dc3f7de | 1387 | unsigned int nunits; |
ebfd146a | 1388 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
ebfd146a | 1389 | tree def; |
ebfd146a IR |
1390 | enum vect_def_type dt; |
1391 | bool is_simple_use; | |
1392 | tree vector_type; | |
1393 | ||
73fbfcad | 1394 | if (dump_enabled_p ()) |
ebfd146a | 1395 | { |
78c60e3d SS |
1396 | dump_printf_loc (MSG_NOTE, vect_location, |
1397 | "vect_get_vec_def_for_operand: "); | |
1398 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
e645e942 | 1399 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1400 | } |
1401 | ||
24ee1384 IR |
1402 | is_simple_use = vect_is_simple_use (op, stmt, loop_vinfo, NULL, |
1403 | &def_stmt, &def, &dt); | |
ebfd146a | 1404 | gcc_assert (is_simple_use); |
73fbfcad | 1405 | if (dump_enabled_p ()) |
ebfd146a | 1406 | { |
78c60e3d | 1407 | int loc_printed = 0; |
ebfd146a IR |
1408 | if (def) |
1409 | { | |
78c60e3d SS |
1410 | dump_printf_loc (MSG_NOTE, vect_location, "def = "); |
1411 | loc_printed = 1; | |
1412 | dump_generic_expr (MSG_NOTE, TDF_SLIM, def); | |
e645e942 | 1413 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1414 | } |
1415 | if (def_stmt) | |
1416 | { | |
78c60e3d SS |
1417 | if (loc_printed) |
1418 | dump_printf (MSG_NOTE, " def_stmt = "); | |
1419 | else | |
1420 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1421 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
ebfd146a IR |
1422 | } |
1423 | } | |
1424 | ||
1425 | switch (dt) | |
1426 | { | |
1427 | /* Case 1: operand is a constant. */ | |
1428 | case vect_constant_def: | |
1429 | { | |
7569a6cc RG |
1430 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); |
1431 | gcc_assert (vector_type); | |
9dc3f7de | 1432 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); |
7569a6cc | 1433 | |
b8698a0f | 1434 | if (scalar_def) |
ebfd146a IR |
1435 | *scalar_def = op; |
1436 | ||
1437 | /* Create 'vect_cst_ = {cst,cst,...,cst}' */ | |
73fbfcad | 1438 | if (dump_enabled_p ()) |
78c60e3d | 1439 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1440 | "Create vector_cst. nunits = %d\n", nunits); |
ebfd146a | 1441 | |
418b7df3 | 1442 | return vect_init_vector (stmt, op, vector_type, NULL); |
ebfd146a IR |
1443 | } |
1444 | ||
1445 | /* Case 2: operand is defined outside the loop - loop invariant. */ | |
8644a673 | 1446 | case vect_external_def: |
ebfd146a IR |
1447 | { |
1448 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (def)); | |
1449 | gcc_assert (vector_type); | |
ebfd146a | 1450 | |
b8698a0f | 1451 | if (scalar_def) |
ebfd146a IR |
1452 | *scalar_def = def; |
1453 | ||
1454 | /* Create 'vec_inv = {inv,inv,..,inv}' */ | |
73fbfcad | 1455 | if (dump_enabled_p ()) |
e645e942 | 1456 | dump_printf_loc (MSG_NOTE, vect_location, "Create vector_inv.\n"); |
ebfd146a | 1457 | |
418b7df3 | 1458 | return vect_init_vector (stmt, def, vector_type, NULL); |
ebfd146a IR |
1459 | } |
1460 | ||
1461 | /* Case 3: operand is defined inside the loop. */ | |
8644a673 | 1462 | case vect_internal_def: |
ebfd146a | 1463 | { |
b8698a0f | 1464 | if (scalar_def) |
ebfd146a IR |
1465 | *scalar_def = NULL/* FIXME tuples: def_stmt*/; |
1466 | ||
1467 | /* Get the def from the vectorized stmt. */ | |
1468 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1469 | |
ebfd146a | 1470 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1471 | /* Get vectorized pattern statement. */ |
1472 | if (!vec_stmt | |
1473 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1474 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1475 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1476 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1477 | gcc_assert (vec_stmt); |
1478 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1479 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1480 | else if (is_gimple_call (vec_stmt)) | |
1481 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1482 | else | |
1483 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1484 | return vec_oprnd; | |
1485 | } | |
1486 | ||
1487 | /* Case 4: operand is defined by a loop header phi - reduction */ | |
1488 | case vect_reduction_def: | |
06066f92 | 1489 | case vect_double_reduction_def: |
7c5222ff | 1490 | case vect_nested_cycle: |
ebfd146a IR |
1491 | { |
1492 | struct loop *loop; | |
1493 | ||
1494 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
b8698a0f | 1495 | loop = (gimple_bb (def_stmt))->loop_father; |
ebfd146a IR |
1496 | |
1497 | /* Get the def before the loop */ | |
1498 | op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop)); | |
1499 | return get_initial_def_for_reduction (stmt, op, scalar_def); | |
1500 | } | |
1501 | ||
1502 | /* Case 5: operand is defined by loop-header phi - induction. */ | |
1503 | case vect_induction_def: | |
1504 | { | |
1505 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1506 | ||
1507 | /* Get the def from the vectorized stmt. */ | |
1508 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1509 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1510 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1511 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1512 | else | |
1513 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1514 | return vec_oprnd; |
1515 | } | |
1516 | ||
1517 | default: | |
1518 | gcc_unreachable (); | |
1519 | } | |
1520 | } | |
1521 | ||
1522 | ||
1523 | /* Function vect_get_vec_def_for_stmt_copy | |
1524 | ||
ff802fa1 | 1525 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1526 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1527 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1528 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1529 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1530 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1531 | DT is the type of the vector def VEC_OPRND. |
1532 | ||
1533 | Context: | |
1534 | In case the vectorization factor (VF) is bigger than the number | |
1535 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1536 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1537 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1538 | smallest data-type determines the VF, and as a result, when vectorizing |
1539 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1540 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1541 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1542 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1543 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1544 | ||
1545 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1546 | |
ebfd146a IR |
1547 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1548 | VS1.1: vx.1 = memref1 VS1.2 | |
1549 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1550 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1551 | |
1552 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1553 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1554 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1555 | VSnew.3: vz3 = vx.3 + ... | |
1556 | ||
1557 | The vectorization of S1 is explained in vectorizable_load. | |
1558 | The vectorization of S2: | |
b8698a0f L |
1559 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1560 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1561 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1562 | returns the vector-def 'vx.0'. |
1563 | ||
b8698a0f L |
1564 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1565 | function is called to get the relevant vector-def for each operand. It is | |
1566 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1567 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1568 | ||
b8698a0f L |
1569 | For example, to obtain the vector-def 'vx.1' in order to create the |
1570 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1571 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1572 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1573 | and return its def ('vx.1'). | |
1574 | Overall, to create the above sequence this function will be called 3 times: | |
1575 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1576 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1577 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1578 | ||
1579 | tree | |
1580 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1581 | { | |
1582 | gimple vec_stmt_for_operand; | |
1583 | stmt_vec_info def_stmt_info; | |
1584 | ||
1585 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1586 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1587 | return vec_oprnd; |
1588 | ||
1589 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1590 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1591 | gcc_assert (def_stmt_info); | |
1592 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1593 | gcc_assert (vec_stmt_for_operand); | |
1594 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1595 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) | |
1596 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1597 | else | |
1598 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1599 | return vec_oprnd; | |
1600 | } | |
1601 | ||
1602 | ||
1603 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1604 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a IR |
1605 | |
1606 | static void | |
b8698a0f | 1607 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1608 | vec<tree> *vec_oprnds0, |
1609 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1610 | { |
9771b263 | 1611 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1612 | |
1613 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1614 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1615 | |
9771b263 | 1616 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1617 | { |
9771b263 | 1618 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1619 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1620 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1621 | } |
1622 | } | |
1623 | ||
1624 | ||
d092494c IR |
1625 | /* Get vectorized definitions for OP0 and OP1. |
1626 | REDUC_INDEX is the index of reduction operand in case of reduction, | |
1627 | and -1 otherwise. */ | |
ebfd146a | 1628 | |
d092494c | 1629 | void |
ebfd146a | 1630 | vect_get_vec_defs (tree op0, tree op1, gimple stmt, |
9771b263 DN |
1631 | vec<tree> *vec_oprnds0, |
1632 | vec<tree> *vec_oprnds1, | |
d092494c | 1633 | slp_tree slp_node, int reduc_index) |
ebfd146a IR |
1634 | { |
1635 | if (slp_node) | |
d092494c IR |
1636 | { |
1637 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1638 | auto_vec<tree> ops (nops); |
1639 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1640 | |
9771b263 | 1641 | ops.quick_push (op0); |
d092494c | 1642 | if (op1) |
9771b263 | 1643 | ops.quick_push (op1); |
d092494c IR |
1644 | |
1645 | vect_get_slp_defs (ops, slp_node, &vec_defs, reduc_index); | |
1646 | ||
37b5ec8f | 1647 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1648 | if (op1) |
37b5ec8f | 1649 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1650 | } |
ebfd146a IR |
1651 | else |
1652 | { | |
1653 | tree vec_oprnd; | |
1654 | ||
9771b263 | 1655 | vec_oprnds0->create (1); |
b8698a0f | 1656 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt, NULL); |
9771b263 | 1657 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1658 | |
1659 | if (op1) | |
1660 | { | |
9771b263 | 1661 | vec_oprnds1->create (1); |
b8698a0f | 1662 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt, NULL); |
9771b263 | 1663 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1664 | } |
1665 | } | |
1666 | } | |
1667 | ||
1668 | ||
1669 | /* Function vect_finish_stmt_generation. | |
1670 | ||
1671 | Insert a new stmt. */ | |
1672 | ||
1673 | void | |
1674 | vect_finish_stmt_generation (gimple stmt, gimple vec_stmt, | |
1675 | gimple_stmt_iterator *gsi) | |
1676 | { | |
1677 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1678 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 1679 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
ebfd146a IR |
1680 | |
1681 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1682 | ||
54e8e2c3 RG |
1683 | if (!gsi_end_p (*gsi) |
1684 | && gimple_has_mem_ops (vec_stmt)) | |
1685 | { | |
1686 | gimple at_stmt = gsi_stmt (*gsi); | |
1687 | tree vuse = gimple_vuse (at_stmt); | |
1688 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1689 | { | |
1690 | tree vdef = gimple_vdef (at_stmt); | |
1691 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1692 | /* If we have an SSA vuse and insert a store, update virtual | |
1693 | SSA form to avoid triggering the renamer. Do so only | |
1694 | if we can easily see all uses - which is what almost always | |
1695 | happens with the way vectorized stmts are inserted. */ | |
1696 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1697 | && ((is_gimple_assign (vec_stmt) | |
1698 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1699 | || (is_gimple_call (vec_stmt) | |
1700 | && !(gimple_call_flags (vec_stmt) | |
1701 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1702 | { | |
1703 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1704 | gimple_set_vdef (vec_stmt, new_vdef); | |
1705 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1706 | } | |
1707 | } | |
1708 | } | |
ebfd146a IR |
1709 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1710 | ||
b8698a0f | 1711 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, loop_vinfo, |
a70d6342 | 1712 | bb_vinfo)); |
ebfd146a | 1713 | |
73fbfcad | 1714 | if (dump_enabled_p ()) |
ebfd146a | 1715 | { |
78c60e3d SS |
1716 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); |
1717 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
ebfd146a IR |
1718 | } |
1719 | ||
ad885386 | 1720 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
8e91d222 JJ |
1721 | |
1722 | /* While EH edges will generally prevent vectorization, stmt might | |
1723 | e.g. be in a must-not-throw region. Ensure newly created stmts | |
1724 | that could throw are part of the same region. */ | |
1725 | int lp_nr = lookup_stmt_eh_lp (stmt); | |
1726 | if (lp_nr != 0 && stmt_could_throw_p (vec_stmt)) | |
1727 | add_stmt_to_eh_lp (vec_stmt, lp_nr); | |
ebfd146a IR |
1728 | } |
1729 | ||
1730 | /* Checks if CALL can be vectorized in type VECTYPE. Returns | |
1731 | a function declaration if the target has a vectorized version | |
1732 | of the function, or NULL_TREE if the function cannot be vectorized. */ | |
1733 | ||
1734 | tree | |
538dd0b7 | 1735 | vectorizable_function (gcall *call, tree vectype_out, tree vectype_in) |
ebfd146a IR |
1736 | { |
1737 | tree fndecl = gimple_call_fndecl (call); | |
ebfd146a IR |
1738 | |
1739 | /* We only handle functions that do not read or clobber memory -- i.e. | |
1740 | const or novops ones. */ | |
1741 | if (!(gimple_call_flags (call) & (ECF_CONST | ECF_NOVOPS))) | |
1742 | return NULL_TREE; | |
1743 | ||
1744 | if (!fndecl | |
1745 | || TREE_CODE (fndecl) != FUNCTION_DECL | |
1746 | || !DECL_BUILT_IN (fndecl)) | |
1747 | return NULL_TREE; | |
1748 | ||
62f7fd21 | 1749 | return targetm.vectorize.builtin_vectorized_function (fndecl, vectype_out, |
ebfd146a IR |
1750 | vectype_in); |
1751 | } | |
1752 | ||
5ce9450f JJ |
1753 | |
1754 | static tree permute_vec_elements (tree, tree, tree, gimple, | |
1755 | gimple_stmt_iterator *); | |
1756 | ||
1757 | ||
1758 | /* Function vectorizable_mask_load_store. | |
1759 | ||
1760 | Check if STMT performs a conditional load or store that can be vectorized. | |
1761 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
1762 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. | |
1763 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1764 | ||
1765 | static bool | |
1766 | vectorizable_mask_load_store (gimple stmt, gimple_stmt_iterator *gsi, | |
1767 | gimple *vec_stmt, slp_tree slp_node) | |
1768 | { | |
1769 | tree vec_dest = NULL; | |
1770 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1771 | stmt_vec_info prev_stmt_info; | |
1772 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1773 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1774 | bool nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
1775 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
1776 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
1777 | tree elem_type; | |
1778 | gimple new_stmt; | |
1779 | tree dummy; | |
1780 | tree dataref_ptr = NULL_TREE; | |
1781 | gimple ptr_incr; | |
1782 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
1783 | int ncopies; | |
1784 | int i, j; | |
1785 | bool inv_p; | |
1786 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; | |
1787 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
1788 | int gather_scale = 1; | |
1789 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
1790 | bool is_store; | |
1791 | tree mask; | |
1792 | gimple def_stmt; | |
1793 | tree def; | |
1794 | enum vect_def_type dt; | |
1795 | ||
1796 | if (slp_node != NULL) | |
1797 | return false; | |
1798 | ||
1799 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
1800 | gcc_assert (ncopies >= 1); | |
1801 | ||
1802 | is_store = gimple_call_internal_fn (stmt) == IFN_MASK_STORE; | |
1803 | mask = gimple_call_arg (stmt, 2); | |
1804 | if (TYPE_PRECISION (TREE_TYPE (mask)) | |
1805 | != GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (vectype)))) | |
1806 | return false; | |
1807 | ||
1808 | /* FORNOW. This restriction should be relaxed. */ | |
1809 | if (nested_in_vect_loop && ncopies > 1) | |
1810 | { | |
1811 | if (dump_enabled_p ()) | |
1812 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1813 | "multiple types in nested loop."); | |
1814 | return false; | |
1815 | } | |
1816 | ||
1817 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) | |
1818 | return false; | |
1819 | ||
1820 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
1821 | return false; | |
1822 | ||
1823 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
1824 | return false; | |
1825 | ||
1826 | elem_type = TREE_TYPE (vectype); | |
1827 | ||
1828 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1829 | return false; | |
1830 | ||
1831 | if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) | |
1832 | return false; | |
1833 | ||
1834 | if (STMT_VINFO_GATHER_P (stmt_info)) | |
1835 | { | |
1836 | gimple def_stmt; | |
1837 | tree def; | |
1838 | gather_decl = vect_check_gather (stmt, loop_vinfo, &gather_base, | |
1839 | &gather_off, &gather_scale); | |
1840 | gcc_assert (gather_decl); | |
1841 | if (!vect_is_simple_use_1 (gather_off, NULL, loop_vinfo, NULL, | |
1842 | &def_stmt, &def, &gather_dt, | |
1843 | &gather_off_vectype)) | |
1844 | { | |
1845 | if (dump_enabled_p ()) | |
1846 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1847 | "gather index use not simple."); | |
1848 | return false; | |
1849 | } | |
03b9e8e4 JJ |
1850 | |
1851 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
1852 | tree masktype | |
1853 | = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (TREE_CHAIN (arglist)))); | |
1854 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
1855 | { | |
1856 | if (dump_enabled_p ()) | |
1857 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1858 | "masked gather with integer mask not supported."); | |
1859 | return false; | |
1860 | } | |
5ce9450f JJ |
1861 | } |
1862 | else if (tree_int_cst_compare (nested_in_vect_loop | |
1863 | ? STMT_VINFO_DR_STEP (stmt_info) | |
1864 | : DR_STEP (dr), size_zero_node) <= 0) | |
1865 | return false; | |
1866 | else if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
1867 | || !can_vec_mask_load_store_p (TYPE_MODE (vectype), !is_store)) | |
1868 | return false; | |
1869 | ||
1870 | if (TREE_CODE (mask) != SSA_NAME) | |
1871 | return false; | |
1872 | ||
1873 | if (!vect_is_simple_use (mask, stmt, loop_vinfo, NULL, | |
1874 | &def_stmt, &def, &dt)) | |
1875 | return false; | |
1876 | ||
1877 | if (is_store) | |
1878 | { | |
1879 | tree rhs = gimple_call_arg (stmt, 3); | |
1880 | if (!vect_is_simple_use (rhs, stmt, loop_vinfo, NULL, | |
1881 | &def_stmt, &def, &dt)) | |
1882 | return false; | |
1883 | } | |
1884 | ||
1885 | if (!vec_stmt) /* transformation not required. */ | |
1886 | { | |
1887 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
1888 | if (is_store) | |
1889 | vect_model_store_cost (stmt_info, ncopies, false, dt, | |
1890 | NULL, NULL, NULL); | |
1891 | else | |
1892 | vect_model_load_cost (stmt_info, ncopies, false, NULL, NULL, NULL); | |
1893 | return true; | |
1894 | } | |
1895 | ||
1896 | /** Transform. **/ | |
1897 | ||
1898 | if (STMT_VINFO_GATHER_P (stmt_info)) | |
1899 | { | |
1900 | tree vec_oprnd0 = NULL_TREE, op; | |
1901 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
1902 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
acdcd61b | 1903 | tree ptr, vec_mask = NULL_TREE, mask_op = NULL_TREE, var, scale; |
5ce9450f | 1904 | tree perm_mask = NULL_TREE, prev_res = NULL_TREE; |
acdcd61b | 1905 | tree mask_perm_mask = NULL_TREE; |
5ce9450f JJ |
1906 | edge pe = loop_preheader_edge (loop); |
1907 | gimple_seq seq; | |
1908 | basic_block new_bb; | |
1909 | enum { NARROW, NONE, WIDEN } modifier; | |
1910 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
1911 | ||
acdcd61b JJ |
1912 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); |
1913 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1914 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1915 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1916 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1917 | scaletype = TREE_VALUE (arglist); | |
1918 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
1919 | && types_compatible_p (srctype, masktype)); | |
1920 | ||
5ce9450f JJ |
1921 | if (nunits == gather_off_nunits) |
1922 | modifier = NONE; | |
1923 | else if (nunits == gather_off_nunits / 2) | |
1924 | { | |
1925 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
1926 | modifier = WIDEN; | |
1927 | ||
1928 | for (i = 0; i < gather_off_nunits; ++i) | |
1929 | sel[i] = i | nunits; | |
1930 | ||
557be5a8 | 1931 | perm_mask = vect_gen_perm_mask_checked (gather_off_vectype, sel); |
5ce9450f JJ |
1932 | } |
1933 | else if (nunits == gather_off_nunits * 2) | |
1934 | { | |
1935 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
1936 | modifier = NARROW; | |
1937 | ||
1938 | for (i = 0; i < nunits; ++i) | |
1939 | sel[i] = i < gather_off_nunits | |
1940 | ? i : i + nunits - gather_off_nunits; | |
1941 | ||
557be5a8 | 1942 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); |
5ce9450f | 1943 | ncopies *= 2; |
acdcd61b JJ |
1944 | for (i = 0; i < nunits; ++i) |
1945 | sel[i] = i | gather_off_nunits; | |
557be5a8 | 1946 | mask_perm_mask = vect_gen_perm_mask_checked (masktype, sel); |
5ce9450f JJ |
1947 | } |
1948 | else | |
1949 | gcc_unreachable (); | |
1950 | ||
5ce9450f JJ |
1951 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); |
1952 | ||
1953 | ptr = fold_convert (ptrtype, gather_base); | |
1954 | if (!is_gimple_min_invariant (ptr)) | |
1955 | { | |
1956 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
1957 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
1958 | gcc_assert (!new_bb); | |
1959 | } | |
1960 | ||
1961 | scale = build_int_cst (scaletype, gather_scale); | |
1962 | ||
1963 | prev_stmt_info = NULL; | |
1964 | for (j = 0; j < ncopies; ++j) | |
1965 | { | |
1966 | if (modifier == WIDEN && (j & 1)) | |
1967 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
1968 | perm_mask, stmt, gsi); | |
1969 | else if (j == 0) | |
1970 | op = vec_oprnd0 | |
1971 | = vect_get_vec_def_for_operand (gather_off, stmt, NULL); | |
1972 | else | |
1973 | op = vec_oprnd0 | |
1974 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
1975 | ||
1976 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
1977 | { | |
1978 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
1979 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
1980 | var = vect_get_new_vect_var (idxtype, vect_simple_var, NULL); | |
b731b390 | 1981 | var = make_ssa_name (var); |
5ce9450f JJ |
1982 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
1983 | new_stmt | |
0d0e4a03 | 1984 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
1985 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1986 | op = var; | |
1987 | } | |
1988 | ||
acdcd61b JJ |
1989 | if (mask_perm_mask && (j & 1)) |
1990 | mask_op = permute_vec_elements (mask_op, mask_op, | |
1991 | mask_perm_mask, stmt, gsi); | |
5ce9450f JJ |
1992 | else |
1993 | { | |
acdcd61b JJ |
1994 | if (j == 0) |
1995 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, NULL); | |
1996 | else | |
1997 | { | |
1998 | vect_is_simple_use (vec_mask, NULL, loop_vinfo, NULL, | |
1999 | &def_stmt, &def, &dt); | |
2000 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); | |
2001 | } | |
5ce9450f | 2002 | |
acdcd61b JJ |
2003 | mask_op = vec_mask; |
2004 | if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask))) | |
2005 | { | |
2006 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)) | |
2007 | == TYPE_VECTOR_SUBPARTS (masktype)); | |
2008 | var = vect_get_new_vect_var (masktype, vect_simple_var, | |
2009 | NULL); | |
b731b390 | 2010 | var = make_ssa_name (var); |
acdcd61b JJ |
2011 | mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op); |
2012 | new_stmt | |
0d0e4a03 | 2013 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, mask_op); |
acdcd61b JJ |
2014 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2015 | mask_op = var; | |
2016 | } | |
5ce9450f JJ |
2017 | } |
2018 | ||
2019 | new_stmt | |
2020 | = gimple_build_call (gather_decl, 5, mask_op, ptr, op, mask_op, | |
2021 | scale); | |
2022 | ||
2023 | if (!useless_type_conversion_p (vectype, rettype)) | |
2024 | { | |
2025 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
2026 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
2027 | var = vect_get_new_vect_var (rettype, vect_simple_var, NULL); | |
2028 | op = make_ssa_name (var, new_stmt); | |
2029 | gimple_call_set_lhs (new_stmt, op); | |
2030 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 2031 | var = make_ssa_name (vec_dest); |
5ce9450f | 2032 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
0d0e4a03 | 2033 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
2034 | } |
2035 | else | |
2036 | { | |
2037 | var = make_ssa_name (vec_dest, new_stmt); | |
2038 | gimple_call_set_lhs (new_stmt, var); | |
2039 | } | |
2040 | ||
2041 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2042 | ||
2043 | if (modifier == NARROW) | |
2044 | { | |
2045 | if ((j & 1) == 0) | |
2046 | { | |
2047 | prev_res = var; | |
2048 | continue; | |
2049 | } | |
2050 | var = permute_vec_elements (prev_res, var, | |
2051 | perm_mask, stmt, gsi); | |
2052 | new_stmt = SSA_NAME_DEF_STMT (var); | |
2053 | } | |
2054 | ||
2055 | if (prev_stmt_info == NULL) | |
2056 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2057 | else | |
2058 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2059 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2060 | } | |
3efe2e2c JJ |
2061 | |
2062 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
2063 | from the IL. */ | |
2064 | tree lhs = gimple_call_lhs (stmt); | |
2065 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
2066 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
2067 | set_vinfo_for_stmt (stmt, NULL); | |
2068 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2069 | gsi_replace (gsi, new_stmt, true); | |
5ce9450f JJ |
2070 | return true; |
2071 | } | |
2072 | else if (is_store) | |
2073 | { | |
2074 | tree vec_rhs = NULL_TREE, vec_mask = NULL_TREE; | |
2075 | prev_stmt_info = NULL; | |
2076 | for (i = 0; i < ncopies; i++) | |
2077 | { | |
2078 | unsigned align, misalign; | |
2079 | ||
2080 | if (i == 0) | |
2081 | { | |
2082 | tree rhs = gimple_call_arg (stmt, 3); | |
2083 | vec_rhs = vect_get_vec_def_for_operand (rhs, stmt, NULL); | |
2084 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, NULL); | |
2085 | /* We should have catched mismatched types earlier. */ | |
2086 | gcc_assert (useless_type_conversion_p (vectype, | |
2087 | TREE_TYPE (vec_rhs))); | |
2088 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, | |
2089 | NULL_TREE, &dummy, gsi, | |
2090 | &ptr_incr, false, &inv_p); | |
2091 | gcc_assert (!inv_p); | |
2092 | } | |
2093 | else | |
2094 | { | |
2095 | vect_is_simple_use (vec_rhs, NULL, loop_vinfo, NULL, &def_stmt, | |
2096 | &def, &dt); | |
2097 | vec_rhs = vect_get_vec_def_for_stmt_copy (dt, vec_rhs); | |
2098 | vect_is_simple_use (vec_mask, NULL, loop_vinfo, NULL, &def_stmt, | |
2099 | &def, &dt); | |
2100 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); | |
2101 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2102 | TYPE_SIZE_UNIT (vectype)); | |
2103 | } | |
2104 | ||
2105 | align = TYPE_ALIGN_UNIT (vectype); | |
2106 | if (aligned_access_p (dr)) | |
2107 | misalign = 0; | |
2108 | else if (DR_MISALIGNMENT (dr) == -1) | |
2109 | { | |
2110 | align = TYPE_ALIGN_UNIT (elem_type); | |
2111 | misalign = 0; | |
2112 | } | |
2113 | else | |
2114 | misalign = DR_MISALIGNMENT (dr); | |
2115 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2116 | misalign); | |
2117 | new_stmt | |
2118 | = gimple_build_call_internal (IFN_MASK_STORE, 4, dataref_ptr, | |
2119 | gimple_call_arg (stmt, 1), | |
2120 | vec_mask, vec_rhs); | |
2121 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2122 | if (i == 0) | |
2123 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2124 | else | |
2125 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2126 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2127 | } | |
2128 | } | |
2129 | else | |
2130 | { | |
2131 | tree vec_mask = NULL_TREE; | |
2132 | prev_stmt_info = NULL; | |
2133 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); | |
2134 | for (i = 0; i < ncopies; i++) | |
2135 | { | |
2136 | unsigned align, misalign; | |
2137 | ||
2138 | if (i == 0) | |
2139 | { | |
2140 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, NULL); | |
2141 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, | |
2142 | NULL_TREE, &dummy, gsi, | |
2143 | &ptr_incr, false, &inv_p); | |
2144 | gcc_assert (!inv_p); | |
2145 | } | |
2146 | else | |
2147 | { | |
2148 | vect_is_simple_use (vec_mask, NULL, loop_vinfo, NULL, &def_stmt, | |
2149 | &def, &dt); | |
2150 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); | |
2151 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2152 | TYPE_SIZE_UNIT (vectype)); | |
2153 | } | |
2154 | ||
2155 | align = TYPE_ALIGN_UNIT (vectype); | |
2156 | if (aligned_access_p (dr)) | |
2157 | misalign = 0; | |
2158 | else if (DR_MISALIGNMENT (dr) == -1) | |
2159 | { | |
2160 | align = TYPE_ALIGN_UNIT (elem_type); | |
2161 | misalign = 0; | |
2162 | } | |
2163 | else | |
2164 | misalign = DR_MISALIGNMENT (dr); | |
2165 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2166 | misalign); | |
2167 | new_stmt | |
2168 | = gimple_build_call_internal (IFN_MASK_LOAD, 3, dataref_ptr, | |
2169 | gimple_call_arg (stmt, 1), | |
2170 | vec_mask); | |
b731b390 | 2171 | gimple_call_set_lhs (new_stmt, make_ssa_name (vec_dest)); |
5ce9450f JJ |
2172 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2173 | if (i == 0) | |
2174 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2175 | else | |
2176 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2177 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2178 | } | |
2179 | } | |
2180 | ||
3efe2e2c JJ |
2181 | if (!is_store) |
2182 | { | |
2183 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
2184 | from the IL. */ | |
2185 | tree lhs = gimple_call_lhs (stmt); | |
2186 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
2187 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
2188 | set_vinfo_for_stmt (stmt, NULL); | |
2189 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2190 | gsi_replace (gsi, new_stmt, true); | |
2191 | } | |
2192 | ||
5ce9450f JJ |
2193 | return true; |
2194 | } | |
2195 | ||
2196 | ||
ebfd146a IR |
2197 | /* Function vectorizable_call. |
2198 | ||
538dd0b7 | 2199 | Check if GS performs a function call that can be vectorized. |
b8698a0f | 2200 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
2201 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2202 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2203 | ||
2204 | static bool | |
538dd0b7 | 2205 | vectorizable_call (gimple gs, gimple_stmt_iterator *gsi, gimple *vec_stmt, |
190c2236 | 2206 | slp_tree slp_node) |
ebfd146a | 2207 | { |
538dd0b7 | 2208 | gcall *stmt; |
ebfd146a IR |
2209 | tree vec_dest; |
2210 | tree scalar_dest; | |
2211 | tree op, type; | |
2212 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
538dd0b7 | 2213 | stmt_vec_info stmt_info = vinfo_for_stmt (gs), prev_stmt_info; |
ebfd146a IR |
2214 | tree vectype_out, vectype_in; |
2215 | int nunits_in; | |
2216 | int nunits_out; | |
2217 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
190c2236 | 2218 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b690cc0f | 2219 | tree fndecl, new_temp, def, rhs_type; |
ebfd146a | 2220 | gimple def_stmt; |
0502fb85 UB |
2221 | enum vect_def_type dt[3] |
2222 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
63827fb8 | 2223 | gimple new_stmt = NULL; |
ebfd146a | 2224 | int ncopies, j; |
6e1aa848 | 2225 | vec<tree> vargs = vNULL; |
ebfd146a IR |
2226 | enum { NARROW, NONE, WIDEN } modifier; |
2227 | size_t i, nargs; | |
9d5e7640 | 2228 | tree lhs; |
ebfd146a | 2229 | |
190c2236 | 2230 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2231 | return false; |
2232 | ||
8644a673 | 2233 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2234 | return false; |
2235 | ||
538dd0b7 DM |
2236 | /* Is GS a vectorizable call? */ |
2237 | stmt = dyn_cast <gcall *> (gs); | |
2238 | if (!stmt) | |
ebfd146a IR |
2239 | return false; |
2240 | ||
5ce9450f JJ |
2241 | if (gimple_call_internal_p (stmt) |
2242 | && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD | |
2243 | || gimple_call_internal_fn (stmt) == IFN_MASK_STORE)) | |
2244 | return vectorizable_mask_load_store (stmt, gsi, vec_stmt, | |
2245 | slp_node); | |
2246 | ||
0136f8f0 AH |
2247 | if (gimple_call_lhs (stmt) == NULL_TREE |
2248 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
ebfd146a IR |
2249 | return false; |
2250 | ||
0136f8f0 | 2251 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); |
5a2c1986 | 2252 | |
b690cc0f RG |
2253 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
2254 | ||
ebfd146a IR |
2255 | /* Process function arguments. */ |
2256 | rhs_type = NULL_TREE; | |
b690cc0f | 2257 | vectype_in = NULL_TREE; |
ebfd146a IR |
2258 | nargs = gimple_call_num_args (stmt); |
2259 | ||
1b1562a5 MM |
2260 | /* Bail out if the function has more than three arguments, we do not have |
2261 | interesting builtin functions to vectorize with more than two arguments | |
2262 | except for fma. No arguments is also not good. */ | |
2263 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
2264 | return false; |
2265 | ||
74bf76ed JJ |
2266 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
2267 | if (gimple_call_internal_p (stmt) | |
2268 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2269 | { | |
2270 | nargs = 0; | |
2271 | rhs_type = unsigned_type_node; | |
2272 | } | |
2273 | ||
ebfd146a IR |
2274 | for (i = 0; i < nargs; i++) |
2275 | { | |
b690cc0f RG |
2276 | tree opvectype; |
2277 | ||
ebfd146a IR |
2278 | op = gimple_call_arg (stmt, i); |
2279 | ||
2280 | /* We can only handle calls with arguments of the same type. */ | |
2281 | if (rhs_type | |
8533c9d8 | 2282 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 2283 | { |
73fbfcad | 2284 | if (dump_enabled_p ()) |
78c60e3d | 2285 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2286 | "argument types differ.\n"); |
ebfd146a IR |
2287 | return false; |
2288 | } | |
b690cc0f RG |
2289 | if (!rhs_type) |
2290 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 2291 | |
24ee1384 | 2292 | if (!vect_is_simple_use_1 (op, stmt, loop_vinfo, bb_vinfo, |
b690cc0f | 2293 | &def_stmt, &def, &dt[i], &opvectype)) |
ebfd146a | 2294 | { |
73fbfcad | 2295 | if (dump_enabled_p ()) |
78c60e3d | 2296 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2297 | "use not simple.\n"); |
ebfd146a IR |
2298 | return false; |
2299 | } | |
ebfd146a | 2300 | |
b690cc0f RG |
2301 | if (!vectype_in) |
2302 | vectype_in = opvectype; | |
2303 | else if (opvectype | |
2304 | && opvectype != vectype_in) | |
2305 | { | |
73fbfcad | 2306 | if (dump_enabled_p ()) |
78c60e3d | 2307 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2308 | "argument vector types differ.\n"); |
b690cc0f RG |
2309 | return false; |
2310 | } | |
2311 | } | |
2312 | /* If all arguments are external or constant defs use a vector type with | |
2313 | the same size as the output vector type. */ | |
ebfd146a | 2314 | if (!vectype_in) |
b690cc0f | 2315 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
2316 | if (vec_stmt) |
2317 | gcc_assert (vectype_in); | |
2318 | if (!vectype_in) | |
2319 | { | |
73fbfcad | 2320 | if (dump_enabled_p ()) |
7d8930a0 | 2321 | { |
78c60e3d SS |
2322 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2323 | "no vectype for scalar type "); | |
2324 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 2325 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
2326 | } |
2327 | ||
2328 | return false; | |
2329 | } | |
ebfd146a IR |
2330 | |
2331 | /* FORNOW */ | |
b690cc0f RG |
2332 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
2333 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
2334 | if (nunits_in == nunits_out / 2) |
2335 | modifier = NARROW; | |
2336 | else if (nunits_out == nunits_in) | |
2337 | modifier = NONE; | |
2338 | else if (nunits_out == nunits_in / 2) | |
2339 | modifier = WIDEN; | |
2340 | else | |
2341 | return false; | |
2342 | ||
2343 | /* For now, we only vectorize functions if a target specific builtin | |
2344 | is available. TODO -- in some cases, it might be profitable to | |
2345 | insert the calls for pieces of the vector, in order to be able | |
2346 | to vectorize other operations in the loop. */ | |
2347 | fndecl = vectorizable_function (stmt, vectype_out, vectype_in); | |
2348 | if (fndecl == NULL_TREE) | |
2349 | { | |
74bf76ed JJ |
2350 | if (gimple_call_internal_p (stmt) |
2351 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE | |
2352 | && !slp_node | |
2353 | && loop_vinfo | |
2354 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2355 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
2356 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2357 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
2358 | { | |
2359 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
2360 | { 0, 1, 2, ... vf - 1 } vector. */ | |
2361 | gcc_assert (nargs == 0); | |
2362 | } | |
2363 | else | |
2364 | { | |
2365 | if (dump_enabled_p ()) | |
2366 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 2367 | "function is not vectorizable.\n"); |
74bf76ed JJ |
2368 | return false; |
2369 | } | |
ebfd146a IR |
2370 | } |
2371 | ||
5006671f | 2372 | gcc_assert (!gimple_vuse (stmt)); |
ebfd146a | 2373 | |
190c2236 JJ |
2374 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
2375 | ncopies = 1; | |
2376 | else if (modifier == NARROW) | |
ebfd146a IR |
2377 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; |
2378 | else | |
2379 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2380 | ||
2381 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
2382 | needs to be generated. */ | |
2383 | gcc_assert (ncopies >= 1); | |
2384 | ||
2385 | if (!vec_stmt) /* transformation not required. */ | |
2386 | { | |
2387 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 2388 | if (dump_enabled_p ()) |
e645e942 TJ |
2389 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ===" |
2390 | "\n"); | |
c3e7ee41 | 2391 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
2392 | return true; |
2393 | } | |
2394 | ||
2395 | /** Transform. **/ | |
2396 | ||
73fbfcad | 2397 | if (dump_enabled_p ()) |
e645e942 | 2398 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
2399 | |
2400 | /* Handle def. */ | |
2401 | scalar_dest = gimple_call_lhs (stmt); | |
2402 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
2403 | ||
2404 | prev_stmt_info = NULL; | |
2405 | switch (modifier) | |
2406 | { | |
2407 | case NONE: | |
2408 | for (j = 0; j < ncopies; ++j) | |
2409 | { | |
2410 | /* Build argument list for the vectorized call. */ | |
2411 | if (j == 0) | |
9771b263 | 2412 | vargs.create (nargs); |
ebfd146a | 2413 | else |
9771b263 | 2414 | vargs.truncate (0); |
ebfd146a | 2415 | |
190c2236 JJ |
2416 | if (slp_node) |
2417 | { | |
ef062b13 | 2418 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2419 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2420 | |
2421 | for (i = 0; i < nargs; i++) | |
9771b263 | 2422 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2423 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2424 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2425 | |
2426 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2427 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
2428 | { |
2429 | size_t k; | |
2430 | for (k = 0; k < nargs; k++) | |
2431 | { | |
37b5ec8f | 2432 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 2433 | vargs[k] = vec_oprndsk[i]; |
190c2236 JJ |
2434 | } |
2435 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2436 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2437 | gimple_call_set_lhs (new_stmt, new_temp); | |
2438 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 2439 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2440 | } |
2441 | ||
2442 | for (i = 0; i < nargs; i++) | |
2443 | { | |
37b5ec8f | 2444 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2445 | vec_oprndsi.release (); |
190c2236 | 2446 | } |
190c2236 JJ |
2447 | continue; |
2448 | } | |
2449 | ||
ebfd146a IR |
2450 | for (i = 0; i < nargs; i++) |
2451 | { | |
2452 | op = gimple_call_arg (stmt, i); | |
2453 | if (j == 0) | |
2454 | vec_oprnd0 | |
2455 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
2456 | else | |
63827fb8 IR |
2457 | { |
2458 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
2459 | vec_oprnd0 | |
2460 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
2461 | } | |
ebfd146a | 2462 | |
9771b263 | 2463 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
2464 | } |
2465 | ||
74bf76ed JJ |
2466 | if (gimple_call_internal_p (stmt) |
2467 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2468 | { | |
2469 | tree *v = XALLOCAVEC (tree, nunits_out); | |
2470 | int k; | |
2471 | for (k = 0; k < nunits_out; ++k) | |
2472 | v[k] = build_int_cst (unsigned_type_node, j * nunits_out + k); | |
2473 | tree cst = build_vector (vectype_out, v); | |
2474 | tree new_var | |
2475 | = vect_get_new_vect_var (vectype_out, vect_simple_var, "cst_"); | |
2476 | gimple init_stmt = gimple_build_assign (new_var, cst); | |
2477 | new_temp = make_ssa_name (new_var, init_stmt); | |
2478 | gimple_assign_set_lhs (init_stmt, new_temp); | |
2479 | vect_init_vector_1 (stmt, init_stmt, NULL); | |
b731b390 | 2480 | new_temp = make_ssa_name (vec_dest); |
74bf76ed JJ |
2481 | new_stmt = gimple_build_assign (new_temp, |
2482 | gimple_assign_lhs (init_stmt)); | |
2483 | } | |
2484 | else | |
2485 | { | |
2486 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2487 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2488 | gimple_call_set_lhs (new_stmt, new_temp); | |
2489 | } | |
ebfd146a IR |
2490 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2491 | ||
2492 | if (j == 0) | |
2493 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2494 | else | |
2495 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2496 | ||
2497 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2498 | } | |
2499 | ||
2500 | break; | |
2501 | ||
2502 | case NARROW: | |
2503 | for (j = 0; j < ncopies; ++j) | |
2504 | { | |
2505 | /* Build argument list for the vectorized call. */ | |
2506 | if (j == 0) | |
9771b263 | 2507 | vargs.create (nargs * 2); |
ebfd146a | 2508 | else |
9771b263 | 2509 | vargs.truncate (0); |
ebfd146a | 2510 | |
190c2236 JJ |
2511 | if (slp_node) |
2512 | { | |
ef062b13 | 2513 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2514 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2515 | |
2516 | for (i = 0; i < nargs; i++) | |
9771b263 | 2517 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2518 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2519 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2520 | |
2521 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2522 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
2523 | { |
2524 | size_t k; | |
9771b263 | 2525 | vargs.truncate (0); |
190c2236 JJ |
2526 | for (k = 0; k < nargs; k++) |
2527 | { | |
37b5ec8f | 2528 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
2529 | vargs.quick_push (vec_oprndsk[i]); |
2530 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 JJ |
2531 | } |
2532 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2533 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2534 | gimple_call_set_lhs (new_stmt, new_temp); | |
2535 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 2536 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2537 | } |
2538 | ||
2539 | for (i = 0; i < nargs; i++) | |
2540 | { | |
37b5ec8f | 2541 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2542 | vec_oprndsi.release (); |
190c2236 | 2543 | } |
190c2236 JJ |
2544 | continue; |
2545 | } | |
2546 | ||
ebfd146a IR |
2547 | for (i = 0; i < nargs; i++) |
2548 | { | |
2549 | op = gimple_call_arg (stmt, i); | |
2550 | if (j == 0) | |
2551 | { | |
2552 | vec_oprnd0 | |
2553 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
2554 | vec_oprnd1 | |
63827fb8 | 2555 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2556 | } |
2557 | else | |
2558 | { | |
336ecb65 | 2559 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 2560 | vec_oprnd0 |
63827fb8 | 2561 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 2562 | vec_oprnd1 |
63827fb8 | 2563 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2564 | } |
2565 | ||
9771b263 DN |
2566 | vargs.quick_push (vec_oprnd0); |
2567 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
2568 | } |
2569 | ||
2570 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2571 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2572 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
2573 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2574 | ||
2575 | if (j == 0) | |
2576 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2577 | else | |
2578 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2579 | ||
2580 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2581 | } | |
2582 | ||
2583 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
2584 | ||
2585 | break; | |
2586 | ||
2587 | case WIDEN: | |
2588 | /* No current target implements this case. */ | |
2589 | return false; | |
2590 | } | |
2591 | ||
9771b263 | 2592 | vargs.release (); |
ebfd146a | 2593 | |
ebfd146a IR |
2594 | /* The call in STMT might prevent it from being removed in dce. |
2595 | We however cannot remove it here, due to the way the ssa name | |
2596 | it defines is mapped to the new definition. So just replace | |
2597 | rhs of the statement with something harmless. */ | |
2598 | ||
dd34c087 JJ |
2599 | if (slp_node) |
2600 | return true; | |
2601 | ||
ebfd146a | 2602 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
2603 | if (is_pattern_stmt_p (stmt_info)) |
2604 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
2605 | else | |
2606 | lhs = gimple_call_lhs (stmt); | |
2607 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
ebfd146a | 2608 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 2609 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
2610 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
2611 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
2612 | |
2613 | return true; | |
2614 | } | |
2615 | ||
2616 | ||
0136f8f0 AH |
2617 | struct simd_call_arg_info |
2618 | { | |
2619 | tree vectype; | |
2620 | tree op; | |
2621 | enum vect_def_type dt; | |
2622 | HOST_WIDE_INT linear_step; | |
2623 | unsigned int align; | |
2624 | }; | |
2625 | ||
2626 | /* Function vectorizable_simd_clone_call. | |
2627 | ||
2628 | Check if STMT performs a function call that can be vectorized | |
2629 | by calling a simd clone of the function. | |
2630 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2631 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2632 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2633 | ||
2634 | static bool | |
2635 | vectorizable_simd_clone_call (gimple stmt, gimple_stmt_iterator *gsi, | |
2636 | gimple *vec_stmt, slp_tree slp_node) | |
2637 | { | |
2638 | tree vec_dest; | |
2639 | tree scalar_dest; | |
2640 | tree op, type; | |
2641 | tree vec_oprnd0 = NULL_TREE; | |
2642 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
2643 | tree vectype; | |
2644 | unsigned int nunits; | |
2645 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2646 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
2647 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; | |
2648 | tree fndecl, new_temp, def; | |
2649 | gimple def_stmt; | |
2650 | gimple new_stmt = NULL; | |
2651 | int ncopies, j; | |
2652 | vec<simd_call_arg_info> arginfo = vNULL; | |
2653 | vec<tree> vargs = vNULL; | |
2654 | size_t i, nargs; | |
2655 | tree lhs, rtype, ratype; | |
2656 | vec<constructor_elt, va_gc> *ret_ctor_elts; | |
2657 | ||
2658 | /* Is STMT a vectorizable call? */ | |
2659 | if (!is_gimple_call (stmt)) | |
2660 | return false; | |
2661 | ||
2662 | fndecl = gimple_call_fndecl (stmt); | |
2663 | if (fndecl == NULL_TREE) | |
2664 | return false; | |
2665 | ||
d52f5295 | 2666 | struct cgraph_node *node = cgraph_node::get (fndecl); |
0136f8f0 AH |
2667 | if (node == NULL || node->simd_clones == NULL) |
2668 | return false; | |
2669 | ||
2670 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
2671 | return false; | |
2672 | ||
2673 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
2674 | return false; | |
2675 | ||
2676 | if (gimple_call_lhs (stmt) | |
2677 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
2678 | return false; | |
2679 | ||
2680 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); | |
2681 | ||
2682 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2683 | ||
2684 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt)) | |
2685 | return false; | |
2686 | ||
2687 | /* FORNOW */ | |
2688 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
2689 | return false; | |
2690 | ||
2691 | /* Process function arguments. */ | |
2692 | nargs = gimple_call_num_args (stmt); | |
2693 | ||
2694 | /* Bail out if the function has zero arguments. */ | |
2695 | if (nargs == 0) | |
2696 | return false; | |
2697 | ||
2698 | arginfo.create (nargs); | |
2699 | ||
2700 | for (i = 0; i < nargs; i++) | |
2701 | { | |
2702 | simd_call_arg_info thisarginfo; | |
2703 | affine_iv iv; | |
2704 | ||
2705 | thisarginfo.linear_step = 0; | |
2706 | thisarginfo.align = 0; | |
2707 | thisarginfo.op = NULL_TREE; | |
2708 | ||
2709 | op = gimple_call_arg (stmt, i); | |
2710 | if (!vect_is_simple_use_1 (op, stmt, loop_vinfo, bb_vinfo, | |
2711 | &def_stmt, &def, &thisarginfo.dt, | |
2712 | &thisarginfo.vectype) | |
2713 | || thisarginfo.dt == vect_uninitialized_def) | |
2714 | { | |
2715 | if (dump_enabled_p ()) | |
2716 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2717 | "use not simple.\n"); | |
2718 | arginfo.release (); | |
2719 | return false; | |
2720 | } | |
2721 | ||
2722 | if (thisarginfo.dt == vect_constant_def | |
2723 | || thisarginfo.dt == vect_external_def) | |
2724 | gcc_assert (thisarginfo.vectype == NULL_TREE); | |
2725 | else | |
2726 | gcc_assert (thisarginfo.vectype != NULL_TREE); | |
2727 | ||
6c9e85fb JJ |
2728 | /* For linear arguments, the analyze phase should have saved |
2729 | the base and step in STMT_VINFO_SIMD_CLONE_INFO. */ | |
2730 | if (i * 2 + 3 <= STMT_VINFO_SIMD_CLONE_INFO (stmt_info).length () | |
2731 | && STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 2 + 2]) | |
2732 | { | |
2733 | gcc_assert (vec_stmt); | |
2734 | thisarginfo.linear_step | |
2735 | = tree_to_shwi (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 2 + 2]); | |
2736 | thisarginfo.op | |
2737 | = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 2 + 1]; | |
2738 | /* If loop has been peeled for alignment, we need to adjust it. */ | |
2739 | tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo); | |
2740 | tree n2 = LOOP_VINFO_NITERS (loop_vinfo); | |
2741 | if (n1 != n2) | |
2742 | { | |
2743 | tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2); | |
2744 | tree step = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 2 + 2]; | |
2745 | tree opt = TREE_TYPE (thisarginfo.op); | |
2746 | bias = fold_convert (TREE_TYPE (step), bias); | |
2747 | bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step); | |
2748 | thisarginfo.op | |
2749 | = fold_build2 (POINTER_TYPE_P (opt) | |
2750 | ? POINTER_PLUS_EXPR : PLUS_EXPR, opt, | |
2751 | thisarginfo.op, bias); | |
2752 | } | |
2753 | } | |
2754 | else if (!vec_stmt | |
2755 | && thisarginfo.dt != vect_constant_def | |
2756 | && thisarginfo.dt != vect_external_def | |
2757 | && loop_vinfo | |
2758 | && TREE_CODE (op) == SSA_NAME | |
2759 | && simple_iv (loop, loop_containing_stmt (stmt), op, | |
2760 | &iv, false) | |
2761 | && tree_fits_shwi_p (iv.step)) | |
0136f8f0 AH |
2762 | { |
2763 | thisarginfo.linear_step = tree_to_shwi (iv.step); | |
2764 | thisarginfo.op = iv.base; | |
2765 | } | |
2766 | else if ((thisarginfo.dt == vect_constant_def | |
2767 | || thisarginfo.dt == vect_external_def) | |
2768 | && POINTER_TYPE_P (TREE_TYPE (op))) | |
2769 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; | |
2770 | ||
2771 | arginfo.quick_push (thisarginfo); | |
2772 | } | |
2773 | ||
2774 | unsigned int badness = 0; | |
2775 | struct cgraph_node *bestn = NULL; | |
6c9e85fb JJ |
2776 | if (STMT_VINFO_SIMD_CLONE_INFO (stmt_info).exists ()) |
2777 | bestn = cgraph_node::get (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[0]); | |
0136f8f0 AH |
2778 | else |
2779 | for (struct cgraph_node *n = node->simd_clones; n != NULL; | |
2780 | n = n->simdclone->next_clone) | |
2781 | { | |
2782 | unsigned int this_badness = 0; | |
2783 | if (n->simdclone->simdlen | |
2784 | > (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
2785 | || n->simdclone->nargs != nargs) | |
2786 | continue; | |
2787 | if (n->simdclone->simdlen | |
2788 | < (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
2789 | this_badness += (exact_log2 (LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
2790 | - exact_log2 (n->simdclone->simdlen)) * 1024; | |
2791 | if (n->simdclone->inbranch) | |
2792 | this_badness += 2048; | |
2793 | int target_badness = targetm.simd_clone.usable (n); | |
2794 | if (target_badness < 0) | |
2795 | continue; | |
2796 | this_badness += target_badness * 512; | |
2797 | /* FORNOW: Have to add code to add the mask argument. */ | |
2798 | if (n->simdclone->inbranch) | |
2799 | continue; | |
2800 | for (i = 0; i < nargs; i++) | |
2801 | { | |
2802 | switch (n->simdclone->args[i].arg_type) | |
2803 | { | |
2804 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
2805 | if (!useless_type_conversion_p | |
2806 | (n->simdclone->args[i].orig_type, | |
2807 | TREE_TYPE (gimple_call_arg (stmt, i)))) | |
2808 | i = -1; | |
2809 | else if (arginfo[i].dt == vect_constant_def | |
2810 | || arginfo[i].dt == vect_external_def | |
2811 | || arginfo[i].linear_step) | |
2812 | this_badness += 64; | |
2813 | break; | |
2814 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
2815 | if (arginfo[i].dt != vect_constant_def | |
2816 | && arginfo[i].dt != vect_external_def) | |
2817 | i = -1; | |
2818 | break; | |
2819 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
2820 | if (arginfo[i].dt == vect_constant_def | |
2821 | || arginfo[i].dt == vect_external_def | |
2822 | || (arginfo[i].linear_step | |
2823 | != n->simdclone->args[i].linear_step)) | |
2824 | i = -1; | |
2825 | break; | |
2826 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
2827 | /* FORNOW */ | |
2828 | i = -1; | |
2829 | break; | |
2830 | case SIMD_CLONE_ARG_TYPE_MASK: | |
2831 | gcc_unreachable (); | |
2832 | } | |
2833 | if (i == (size_t) -1) | |
2834 | break; | |
2835 | if (n->simdclone->args[i].alignment > arginfo[i].align) | |
2836 | { | |
2837 | i = -1; | |
2838 | break; | |
2839 | } | |
2840 | if (arginfo[i].align) | |
2841 | this_badness += (exact_log2 (arginfo[i].align) | |
2842 | - exact_log2 (n->simdclone->args[i].alignment)); | |
2843 | } | |
2844 | if (i == (size_t) -1) | |
2845 | continue; | |
2846 | if (bestn == NULL || this_badness < badness) | |
2847 | { | |
2848 | bestn = n; | |
2849 | badness = this_badness; | |
2850 | } | |
2851 | } | |
2852 | ||
2853 | if (bestn == NULL) | |
2854 | { | |
2855 | arginfo.release (); | |
2856 | return false; | |
2857 | } | |
2858 | ||
2859 | for (i = 0; i < nargs; i++) | |
2860 | if ((arginfo[i].dt == vect_constant_def | |
2861 | || arginfo[i].dt == vect_external_def) | |
2862 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
2863 | { | |
2864 | arginfo[i].vectype | |
2865 | = get_vectype_for_scalar_type (TREE_TYPE (gimple_call_arg (stmt, | |
2866 | i))); | |
2867 | if (arginfo[i].vectype == NULL | |
2868 | || (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
2869 | > bestn->simdclone->simdlen)) | |
2870 | { | |
2871 | arginfo.release (); | |
2872 | return false; | |
2873 | } | |
2874 | } | |
2875 | ||
2876 | fndecl = bestn->decl; | |
2877 | nunits = bestn->simdclone->simdlen; | |
2878 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
2879 | ||
2880 | /* If the function isn't const, only allow it in simd loops where user | |
2881 | has asserted that at least nunits consecutive iterations can be | |
2882 | performed using SIMD instructions. */ | |
2883 | if ((loop == NULL || (unsigned) loop->safelen < nunits) | |
2884 | && gimple_vuse (stmt)) | |
2885 | { | |
2886 | arginfo.release (); | |
2887 | return false; | |
2888 | } | |
2889 | ||
2890 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
2891 | needs to be generated. */ | |
2892 | gcc_assert (ncopies >= 1); | |
2893 | ||
2894 | if (!vec_stmt) /* transformation not required. */ | |
2895 | { | |
6c9e85fb JJ |
2896 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (bestn->decl); |
2897 | for (i = 0; i < nargs; i++) | |
2898 | if (bestn->simdclone->args[i].arg_type | |
2899 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
2900 | { | |
2901 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_grow_cleared (i * 2 | |
2902 | + 1); | |
2903 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (arginfo[i].op); | |
2904 | tree lst = POINTER_TYPE_P (TREE_TYPE (arginfo[i].op)) | |
2905 | ? size_type_node : TREE_TYPE (arginfo[i].op); | |
2906 | tree ls = build_int_cst (lst, arginfo[i].linear_step); | |
2907 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (ls); | |
2908 | } | |
0136f8f0 AH |
2909 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; |
2910 | if (dump_enabled_p ()) | |
2911 | dump_printf_loc (MSG_NOTE, vect_location, | |
2912 | "=== vectorizable_simd_clone_call ===\n"); | |
2913 | /* vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); */ | |
2914 | arginfo.release (); | |
2915 | return true; | |
2916 | } | |
2917 | ||
2918 | /** Transform. **/ | |
2919 | ||
2920 | if (dump_enabled_p ()) | |
2921 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); | |
2922 | ||
2923 | /* Handle def. */ | |
2924 | scalar_dest = gimple_call_lhs (stmt); | |
2925 | vec_dest = NULL_TREE; | |
2926 | rtype = NULL_TREE; | |
2927 | ratype = NULL_TREE; | |
2928 | if (scalar_dest) | |
2929 | { | |
2930 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
2931 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); | |
2932 | if (TREE_CODE (rtype) == ARRAY_TYPE) | |
2933 | { | |
2934 | ratype = rtype; | |
2935 | rtype = TREE_TYPE (ratype); | |
2936 | } | |
2937 | } | |
2938 | ||
2939 | prev_stmt_info = NULL; | |
2940 | for (j = 0; j < ncopies; ++j) | |
2941 | { | |
2942 | /* Build argument list for the vectorized call. */ | |
2943 | if (j == 0) | |
2944 | vargs.create (nargs); | |
2945 | else | |
2946 | vargs.truncate (0); | |
2947 | ||
2948 | for (i = 0; i < nargs; i++) | |
2949 | { | |
2950 | unsigned int k, l, m, o; | |
2951 | tree atype; | |
2952 | op = gimple_call_arg (stmt, i); | |
2953 | switch (bestn->simdclone->args[i].arg_type) | |
2954 | { | |
2955 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
2956 | atype = bestn->simdclone->args[i].vector_type; | |
2957 | o = nunits / TYPE_VECTOR_SUBPARTS (atype); | |
2958 | for (m = j * o; m < (j + 1) * o; m++) | |
2959 | { | |
2960 | if (TYPE_VECTOR_SUBPARTS (atype) | |
2961 | < TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)) | |
2962 | { | |
2963 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); | |
2964 | k = (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
2965 | / TYPE_VECTOR_SUBPARTS (atype)); | |
2966 | gcc_assert ((k & (k - 1)) == 0); | |
2967 | if (m == 0) | |
2968 | vec_oprnd0 | |
2969 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
2970 | else | |
2971 | { | |
2972 | vec_oprnd0 = arginfo[i].op; | |
2973 | if ((m & (k - 1)) == 0) | |
2974 | vec_oprnd0 | |
2975 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
2976 | vec_oprnd0); | |
2977 | } | |
2978 | arginfo[i].op = vec_oprnd0; | |
2979 | vec_oprnd0 | |
2980 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, | |
2981 | size_int (prec), | |
2982 | bitsize_int ((m & (k - 1)) * prec)); | |
2983 | new_stmt | |
b731b390 | 2984 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
2985 | vec_oprnd0); |
2986 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2987 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
2988 | } | |
2989 | else | |
2990 | { | |
2991 | k = (TYPE_VECTOR_SUBPARTS (atype) | |
2992 | / TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)); | |
2993 | gcc_assert ((k & (k - 1)) == 0); | |
2994 | vec<constructor_elt, va_gc> *ctor_elts; | |
2995 | if (k != 1) | |
2996 | vec_alloc (ctor_elts, k); | |
2997 | else | |
2998 | ctor_elts = NULL; | |
2999 | for (l = 0; l < k; l++) | |
3000 | { | |
3001 | if (m == 0 && l == 0) | |
3002 | vec_oprnd0 | |
3003 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
3004 | else | |
3005 | vec_oprnd0 | |
3006 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3007 | arginfo[i].op); | |
3008 | arginfo[i].op = vec_oprnd0; | |
3009 | if (k == 1) | |
3010 | break; | |
3011 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, | |
3012 | vec_oprnd0); | |
3013 | } | |
3014 | if (k == 1) | |
3015 | vargs.safe_push (vec_oprnd0); | |
3016 | else | |
3017 | { | |
3018 | vec_oprnd0 = build_constructor (atype, ctor_elts); | |
3019 | new_stmt | |
b731b390 | 3020 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3021 | vec_oprnd0); |
3022 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3023 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3024 | } | |
3025 | } | |
3026 | } | |
3027 | break; | |
3028 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3029 | vargs.safe_push (op); | |
3030 | break; | |
3031 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
3032 | if (j == 0) | |
3033 | { | |
3034 | gimple_seq stmts; | |
3035 | arginfo[i].op | |
3036 | = force_gimple_operand (arginfo[i].op, &stmts, true, | |
3037 | NULL_TREE); | |
3038 | if (stmts != NULL) | |
3039 | { | |
3040 | basic_block new_bb; | |
3041 | edge pe = loop_preheader_edge (loop); | |
3042 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
3043 | gcc_assert (!new_bb); | |
3044 | } | |
b731b390 | 3045 | tree phi_res = copy_ssa_name (op); |
538dd0b7 | 3046 | gphi *new_phi = create_phi_node (phi_res, loop->header); |
0136f8f0 AH |
3047 | set_vinfo_for_stmt (new_phi, |
3048 | new_stmt_vec_info (new_phi, loop_vinfo, | |
3049 | NULL)); | |
3050 | add_phi_arg (new_phi, arginfo[i].op, | |
3051 | loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
3052 | enum tree_code code | |
3053 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3054 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3055 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3056 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3057 | widest_int cst |
3058 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3059 | ncopies * nunits); | |
3060 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3061 | tree phi_arg = copy_ssa_name (op); |
0d0e4a03 JJ |
3062 | new_stmt |
3063 | = gimple_build_assign (phi_arg, code, phi_res, tcst); | |
0136f8f0 AH |
3064 | gimple_stmt_iterator si = gsi_after_labels (loop->header); |
3065 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); | |
3066 | set_vinfo_for_stmt (new_stmt, | |
3067 | new_stmt_vec_info (new_stmt, loop_vinfo, | |
3068 | NULL)); | |
3069 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), | |
3070 | UNKNOWN_LOCATION); | |
3071 | arginfo[i].op = phi_res; | |
3072 | vargs.safe_push (phi_res); | |
3073 | } | |
3074 | else | |
3075 | { | |
3076 | enum tree_code code | |
3077 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3078 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3079 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3080 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3081 | widest_int cst |
3082 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3083 | j * nunits); | |
3084 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3085 | new_temp = make_ssa_name (TREE_TYPE (op)); |
0d0e4a03 JJ |
3086 | new_stmt = gimple_build_assign (new_temp, code, |
3087 | arginfo[i].op, tcst); | |
0136f8f0 AH |
3088 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3089 | vargs.safe_push (new_temp); | |
3090 | } | |
3091 | break; | |
3092 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
3093 | default: | |
3094 | gcc_unreachable (); | |
3095 | } | |
3096 | } | |
3097 | ||
3098 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
3099 | if (vec_dest) | |
3100 | { | |
3101 | gcc_assert (ratype || TYPE_VECTOR_SUBPARTS (rtype) == nunits); | |
3102 | if (ratype) | |
b731b390 | 3103 | new_temp = create_tmp_var (ratype); |
0136f8f0 AH |
3104 | else if (TYPE_VECTOR_SUBPARTS (vectype) |
3105 | == TYPE_VECTOR_SUBPARTS (rtype)) | |
3106 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3107 | else | |
3108 | new_temp = make_ssa_name (rtype, new_stmt); | |
3109 | gimple_call_set_lhs (new_stmt, new_temp); | |
3110 | } | |
3111 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3112 | ||
3113 | if (vec_dest) | |
3114 | { | |
3115 | if (TYPE_VECTOR_SUBPARTS (vectype) < nunits) | |
3116 | { | |
3117 | unsigned int k, l; | |
3118 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); | |
3119 | k = nunits / TYPE_VECTOR_SUBPARTS (vectype); | |
3120 | gcc_assert ((k & (k - 1)) == 0); | |
3121 | for (l = 0; l < k; l++) | |
3122 | { | |
3123 | tree t; | |
3124 | if (ratype) | |
3125 | { | |
3126 | t = build_fold_addr_expr (new_temp); | |
3127 | t = build2 (MEM_REF, vectype, t, | |
3128 | build_int_cst (TREE_TYPE (t), | |
3129 | l * prec / BITS_PER_UNIT)); | |
3130 | } | |
3131 | else | |
3132 | t = build3 (BIT_FIELD_REF, vectype, new_temp, | |
3133 | size_int (prec), bitsize_int (l * prec)); | |
3134 | new_stmt | |
b731b390 | 3135 | = gimple_build_assign (make_ssa_name (vectype), t); |
0136f8f0 AH |
3136 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3137 | if (j == 0 && l == 0) | |
3138 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3139 | else | |
3140 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3141 | ||
3142 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3143 | } | |
3144 | ||
3145 | if (ratype) | |
3146 | { | |
3147 | tree clobber = build_constructor (ratype, NULL); | |
3148 | TREE_THIS_VOLATILE (clobber) = 1; | |
3149 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3150 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3151 | } | |
3152 | continue; | |
3153 | } | |
3154 | else if (TYPE_VECTOR_SUBPARTS (vectype) > nunits) | |
3155 | { | |
3156 | unsigned int k = (TYPE_VECTOR_SUBPARTS (vectype) | |
3157 | / TYPE_VECTOR_SUBPARTS (rtype)); | |
3158 | gcc_assert ((k & (k - 1)) == 0); | |
3159 | if ((j & (k - 1)) == 0) | |
3160 | vec_alloc (ret_ctor_elts, k); | |
3161 | if (ratype) | |
3162 | { | |
3163 | unsigned int m, o = nunits / TYPE_VECTOR_SUBPARTS (rtype); | |
3164 | for (m = 0; m < o; m++) | |
3165 | { | |
3166 | tree tem = build4 (ARRAY_REF, rtype, new_temp, | |
3167 | size_int (m), NULL_TREE, NULL_TREE); | |
3168 | new_stmt | |
b731b390 | 3169 | = gimple_build_assign (make_ssa_name (rtype), tem); |
0136f8f0 AH |
3170 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3171 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, | |
3172 | gimple_assign_lhs (new_stmt)); | |
3173 | } | |
3174 | tree clobber = build_constructor (ratype, NULL); | |
3175 | TREE_THIS_VOLATILE (clobber) = 1; | |
3176 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3177 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3178 | } | |
3179 | else | |
3180 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); | |
3181 | if ((j & (k - 1)) != k - 1) | |
3182 | continue; | |
3183 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); | |
3184 | new_stmt | |
b731b390 | 3185 | = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0); |
0136f8f0 AH |
3186 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3187 | ||
3188 | if ((unsigned) j == k - 1) | |
3189 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3190 | else | |
3191 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3192 | ||
3193 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3194 | continue; | |
3195 | } | |
3196 | else if (ratype) | |
3197 | { | |
3198 | tree t = build_fold_addr_expr (new_temp); | |
3199 | t = build2 (MEM_REF, vectype, t, | |
3200 | build_int_cst (TREE_TYPE (t), 0)); | |
3201 | new_stmt | |
b731b390 | 3202 | = gimple_build_assign (make_ssa_name (vec_dest), t); |
0136f8f0 AH |
3203 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3204 | tree clobber = build_constructor (ratype, NULL); | |
3205 | TREE_THIS_VOLATILE (clobber) = 1; | |
3206 | vect_finish_stmt_generation (stmt, | |
3207 | gimple_build_assign (new_temp, | |
3208 | clobber), gsi); | |
3209 | } | |
3210 | } | |
3211 | ||
3212 | if (j == 0) | |
3213 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3214 | else | |
3215 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3216 | ||
3217 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3218 | } | |
3219 | ||
3220 | vargs.release (); | |
3221 | ||
3222 | /* The call in STMT might prevent it from being removed in dce. | |
3223 | We however cannot remove it here, due to the way the ssa name | |
3224 | it defines is mapped to the new definition. So just replace | |
3225 | rhs of the statement with something harmless. */ | |
3226 | ||
3227 | if (slp_node) | |
3228 | return true; | |
3229 | ||
3230 | if (scalar_dest) | |
3231 | { | |
3232 | type = TREE_TYPE (scalar_dest); | |
3233 | if (is_pattern_stmt_p (stmt_info)) | |
3234 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
3235 | else | |
3236 | lhs = gimple_call_lhs (stmt); | |
3237 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
3238 | } | |
3239 | else | |
3240 | new_stmt = gimple_build_nop (); | |
3241 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
3242 | set_vinfo_for_stmt (stmt, NULL); | |
3243 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2865f32a | 3244 | gsi_replace (gsi, new_stmt, true); |
0136f8f0 AH |
3245 | unlink_stmt_vdef (stmt); |
3246 | ||
3247 | return true; | |
3248 | } | |
3249 | ||
3250 | ||
ebfd146a IR |
3251 | /* Function vect_gen_widened_results_half |
3252 | ||
3253 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 3254 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 3255 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
3256 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
3257 | needs to be created (DECL is a function-decl of a target-builtin). | |
3258 | STMT is the original scalar stmt that we are vectorizing. */ | |
3259 | ||
3260 | static gimple | |
3261 | vect_gen_widened_results_half (enum tree_code code, | |
3262 | tree decl, | |
3263 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
3264 | tree vec_dest, gimple_stmt_iterator *gsi, | |
3265 | gimple stmt) | |
b8698a0f | 3266 | { |
ebfd146a | 3267 | gimple new_stmt; |
b8698a0f L |
3268 | tree new_temp; |
3269 | ||
3270 | /* Generate half of the widened result: */ | |
3271 | if (code == CALL_EXPR) | |
3272 | { | |
3273 | /* Target specific support */ | |
ebfd146a IR |
3274 | if (op_type == binary_op) |
3275 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
3276 | else | |
3277 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
3278 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3279 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
3280 | } |
3281 | else | |
ebfd146a | 3282 | { |
b8698a0f L |
3283 | /* Generic support */ |
3284 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
3285 | if (op_type != binary_op) |
3286 | vec_oprnd1 = NULL; | |
0d0e4a03 | 3287 | new_stmt = gimple_build_assign (vec_dest, code, vec_oprnd0, vec_oprnd1); |
ebfd146a IR |
3288 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3289 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 3290 | } |
ebfd146a IR |
3291 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3292 | ||
ebfd146a IR |
3293 | return new_stmt; |
3294 | } | |
3295 | ||
4a00c761 JJ |
3296 | |
3297 | /* Get vectorized definitions for loop-based vectorization. For the first | |
3298 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
3299 | scalar operand), and for the rest we get a copy with | |
3300 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
3301 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
3302 | The vectors are collected into VEC_OPRNDS. */ | |
3303 | ||
3304 | static void | |
3305 | vect_get_loop_based_defs (tree *oprnd, gimple stmt, enum vect_def_type dt, | |
9771b263 | 3306 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
3307 | { |
3308 | tree vec_oprnd; | |
3309 | ||
3310 | /* Get first vector operand. */ | |
3311 | /* All the vector operands except the very first one (that is scalar oprnd) | |
3312 | are stmt copies. */ | |
3313 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
3314 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt, NULL); | |
3315 | else | |
3316 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
3317 | ||
9771b263 | 3318 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3319 | |
3320 | /* Get second vector operand. */ | |
3321 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 3322 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3323 | |
3324 | *oprnd = vec_oprnd; | |
3325 | ||
3326 | /* For conversion in multiple steps, continue to get operands | |
3327 | recursively. */ | |
3328 | if (multi_step_cvt) | |
3329 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
3330 | } | |
3331 | ||
3332 | ||
3333 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
3334 | For multi-step conversions store the resulting vectors and call the function | |
3335 | recursively. */ | |
3336 | ||
3337 | static void | |
9771b263 | 3338 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
4a00c761 | 3339 | int multi_step_cvt, gimple stmt, |
9771b263 | 3340 | vec<tree> vec_dsts, |
4a00c761 JJ |
3341 | gimple_stmt_iterator *gsi, |
3342 | slp_tree slp_node, enum tree_code code, | |
3343 | stmt_vec_info *prev_stmt_info) | |
3344 | { | |
3345 | unsigned int i; | |
3346 | tree vop0, vop1, new_tmp, vec_dest; | |
3347 | gimple new_stmt; | |
3348 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3349 | ||
9771b263 | 3350 | vec_dest = vec_dsts.pop (); |
4a00c761 | 3351 | |
9771b263 | 3352 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
3353 | { |
3354 | /* Create demotion operation. */ | |
9771b263 DN |
3355 | vop0 = (*vec_oprnds)[i]; |
3356 | vop1 = (*vec_oprnds)[i + 1]; | |
0d0e4a03 | 3357 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
4a00c761 JJ |
3358 | new_tmp = make_ssa_name (vec_dest, new_stmt); |
3359 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
3360 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3361 | ||
3362 | if (multi_step_cvt) | |
3363 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 3364 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
3365 | else |
3366 | { | |
3367 | /* This is the last step of the conversion sequence. Store the | |
3368 | vectors in SLP_NODE or in vector info of the scalar statement | |
3369 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
3370 | if (slp_node) | |
9771b263 | 3371 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
3372 | else |
3373 | { | |
3374 | if (!*prev_stmt_info) | |
3375 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3376 | else | |
3377 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
3378 | ||
3379 | *prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3380 | } | |
3381 | } | |
3382 | } | |
3383 | ||
3384 | /* For multi-step demotion operations we first generate demotion operations | |
3385 | from the source type to the intermediate types, and then combine the | |
3386 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
3387 | type. */ | |
3388 | if (multi_step_cvt) | |
3389 | { | |
3390 | /* At each level of recursion we have half of the operands we had at the | |
3391 | previous level. */ | |
9771b263 | 3392 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
3393 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
3394 | stmt, vec_dsts, gsi, slp_node, | |
3395 | VEC_PACK_TRUNC_EXPR, | |
3396 | prev_stmt_info); | |
3397 | } | |
3398 | ||
9771b263 | 3399 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3400 | } |
3401 | ||
3402 | ||
3403 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
3404 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
3405 | the resulting vectors and call the function recursively. */ | |
3406 | ||
3407 | static void | |
9771b263 DN |
3408 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
3409 | vec<tree> *vec_oprnds1, | |
4a00c761 JJ |
3410 | gimple stmt, tree vec_dest, |
3411 | gimple_stmt_iterator *gsi, | |
3412 | enum tree_code code1, | |
3413 | enum tree_code code2, tree decl1, | |
3414 | tree decl2, int op_type) | |
3415 | { | |
3416 | int i; | |
3417 | tree vop0, vop1, new_tmp1, new_tmp2; | |
3418 | gimple new_stmt1, new_stmt2; | |
6e1aa848 | 3419 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 3420 | |
9771b263 DN |
3421 | vec_tmp.create (vec_oprnds0->length () * 2); |
3422 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
3423 | { |
3424 | if (op_type == binary_op) | |
9771b263 | 3425 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
3426 | else |
3427 | vop1 = NULL_TREE; | |
3428 | ||
3429 | /* Generate the two halves of promotion operation. */ | |
3430 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
3431 | op_type, vec_dest, gsi, stmt); | |
3432 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
3433 | op_type, vec_dest, gsi, stmt); | |
3434 | if (is_gimple_call (new_stmt1)) | |
3435 | { | |
3436 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
3437 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
3438 | } | |
3439 | else | |
3440 | { | |
3441 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
3442 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
3443 | } | |
3444 | ||
3445 | /* Store the results for the next step. */ | |
9771b263 DN |
3446 | vec_tmp.quick_push (new_tmp1); |
3447 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
3448 | } |
3449 | ||
689eaba3 | 3450 | vec_oprnds0->release (); |
4a00c761 JJ |
3451 | *vec_oprnds0 = vec_tmp; |
3452 | } | |
3453 | ||
3454 | ||
b8698a0f L |
3455 | /* Check if STMT performs a conversion operation, that can be vectorized. |
3456 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 3457 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
3458 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
3459 | ||
3460 | static bool | |
3461 | vectorizable_conversion (gimple stmt, gimple_stmt_iterator *gsi, | |
3462 | gimple *vec_stmt, slp_tree slp_node) | |
3463 | { | |
3464 | tree vec_dest; | |
3465 | tree scalar_dest; | |
4a00c761 | 3466 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
3467 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
3468 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3469 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3470 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 3471 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
3472 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
3473 | tree new_temp; | |
3474 | tree def; | |
3475 | gimple def_stmt; | |
3476 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
3477 | gimple new_stmt = NULL; | |
3478 | stmt_vec_info prev_stmt_info; | |
3479 | int nunits_in; | |
3480 | int nunits_out; | |
3481 | tree vectype_out, vectype_in; | |
4a00c761 JJ |
3482 | int ncopies, i, j; |
3483 | tree lhs_type, rhs_type; | |
ebfd146a | 3484 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
3485 | vec<tree> vec_oprnds0 = vNULL; |
3486 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 3487 | tree vop0; |
4a00c761 JJ |
3488 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3489 | int multi_step_cvt = 0; | |
6e1aa848 DN |
3490 | vec<tree> vec_dsts = vNULL; |
3491 | vec<tree> interm_types = vNULL; | |
4a00c761 JJ |
3492 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
3493 | int op_type; | |
ef4bddc2 | 3494 | machine_mode rhs_mode; |
4a00c761 | 3495 | unsigned short fltsz; |
ebfd146a IR |
3496 | |
3497 | /* Is STMT a vectorizable conversion? */ | |
3498 | ||
4a00c761 | 3499 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3500 | return false; |
3501 | ||
8644a673 | 3502 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3503 | return false; |
3504 | ||
3505 | if (!is_gimple_assign (stmt)) | |
3506 | return false; | |
3507 | ||
3508 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3509 | return false; | |
3510 | ||
3511 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
3512 | if (!CONVERT_EXPR_CODE_P (code) |
3513 | && code != FIX_TRUNC_EXPR | |
3514 | && code != FLOAT_EXPR | |
3515 | && code != WIDEN_MULT_EXPR | |
3516 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
3517 | return false; |
3518 | ||
4a00c761 JJ |
3519 | op_type = TREE_CODE_LENGTH (code); |
3520 | ||
ebfd146a | 3521 | /* Check types of lhs and rhs. */ |
b690cc0f | 3522 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 3523 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
3524 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
3525 | ||
ebfd146a IR |
3526 | op0 = gimple_assign_rhs1 (stmt); |
3527 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
3528 | |
3529 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
3530 | && !((INTEGRAL_TYPE_P (lhs_type) | |
3531 | && INTEGRAL_TYPE_P (rhs_type)) | |
3532 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
3533 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
3534 | return false; | |
3535 | ||
3536 | if ((INTEGRAL_TYPE_P (lhs_type) | |
3537 | && (TYPE_PRECISION (lhs_type) | |
3538 | != GET_MODE_PRECISION (TYPE_MODE (lhs_type)))) | |
3539 | || (INTEGRAL_TYPE_P (rhs_type) | |
3540 | && (TYPE_PRECISION (rhs_type) | |
3541 | != GET_MODE_PRECISION (TYPE_MODE (rhs_type))))) | |
3542 | { | |
73fbfcad | 3543 | if (dump_enabled_p ()) |
78c60e3d | 3544 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
3545 | "type conversion to/from bit-precision unsupported." |
3546 | "\n"); | |
4a00c761 JJ |
3547 | return false; |
3548 | } | |
3549 | ||
b690cc0f | 3550 | /* Check the operands of the operation. */ |
24ee1384 | 3551 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
b690cc0f RG |
3552 | &def_stmt, &def, &dt[0], &vectype_in)) |
3553 | { | |
73fbfcad | 3554 | if (dump_enabled_p ()) |
78c60e3d | 3555 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3556 | "use not simple.\n"); |
b690cc0f RG |
3557 | return false; |
3558 | } | |
4a00c761 JJ |
3559 | if (op_type == binary_op) |
3560 | { | |
3561 | bool ok; | |
3562 | ||
3563 | op1 = gimple_assign_rhs2 (stmt); | |
3564 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
3565 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
3566 | OP1. */ | |
3567 | if (CONSTANT_CLASS_P (op0)) | |
f5709183 | 3568 | ok = vect_is_simple_use_1 (op1, stmt, loop_vinfo, bb_vinfo, |
4a00c761 JJ |
3569 | &def_stmt, &def, &dt[1], &vectype_in); |
3570 | else | |
f5709183 | 3571 | ok = vect_is_simple_use (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
24ee1384 | 3572 | &def, &dt[1]); |
4a00c761 JJ |
3573 | |
3574 | if (!ok) | |
3575 | { | |
73fbfcad | 3576 | if (dump_enabled_p ()) |
78c60e3d | 3577 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3578 | "use not simple.\n"); |
4a00c761 JJ |
3579 | return false; |
3580 | } | |
3581 | } | |
3582 | ||
b690cc0f RG |
3583 | /* If op0 is an external or constant defs use a vector type of |
3584 | the same size as the output vector type. */ | |
ebfd146a | 3585 | if (!vectype_in) |
b690cc0f | 3586 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
3587 | if (vec_stmt) |
3588 | gcc_assert (vectype_in); | |
3589 | if (!vectype_in) | |
3590 | { | |
73fbfcad | 3591 | if (dump_enabled_p ()) |
4a00c761 | 3592 | { |
78c60e3d SS |
3593 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3594 | "no vectype for scalar type "); | |
3595 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 3596 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 3597 | } |
7d8930a0 IR |
3598 | |
3599 | return false; | |
3600 | } | |
ebfd146a | 3601 | |
b690cc0f RG |
3602 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
3603 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4a00c761 | 3604 | if (nunits_in < nunits_out) |
ebfd146a IR |
3605 | modifier = NARROW; |
3606 | else if (nunits_out == nunits_in) | |
3607 | modifier = NONE; | |
ebfd146a | 3608 | else |
4a00c761 | 3609 | modifier = WIDEN; |
ebfd146a | 3610 | |
ff802fa1 IR |
3611 | /* Multiple types in SLP are handled by creating the appropriate number of |
3612 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3613 | case of SLP. */ | |
437f4a00 | 3614 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a | 3615 | ncopies = 1; |
4a00c761 JJ |
3616 | else if (modifier == NARROW) |
3617 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
3618 | else | |
3619 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
b8698a0f | 3620 | |
ebfd146a IR |
3621 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
3622 | needs to be generated. */ | |
3623 | gcc_assert (ncopies >= 1); | |
3624 | ||
ebfd146a | 3625 | /* Supportable by target? */ |
4a00c761 | 3626 | switch (modifier) |
ebfd146a | 3627 | { |
4a00c761 JJ |
3628 | case NONE: |
3629 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
3630 | return false; | |
3631 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
3632 | &decl1, &code1)) | |
3633 | break; | |
3634 | /* FALLTHRU */ | |
3635 | unsupported: | |
73fbfcad | 3636 | if (dump_enabled_p ()) |
78c60e3d | 3637 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3638 | "conversion not supported by target.\n"); |
ebfd146a | 3639 | return false; |
ebfd146a | 3640 | |
4a00c761 JJ |
3641 | case WIDEN: |
3642 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
3643 | &code1, &code2, &multi_step_cvt, |
3644 | &interm_types)) | |
4a00c761 JJ |
3645 | { |
3646 | /* Binary widening operation can only be supported directly by the | |
3647 | architecture. */ | |
3648 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
3649 | break; | |
3650 | } | |
3651 | ||
3652 | if (code != FLOAT_EXPR | |
3653 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
3654 | <= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
3655 | goto unsupported; | |
3656 | ||
3657 | rhs_mode = TYPE_MODE (rhs_type); | |
3658 | fltsz = GET_MODE_SIZE (TYPE_MODE (lhs_type)); | |
3659 | for (rhs_mode = GET_MODE_2XWIDER_MODE (TYPE_MODE (rhs_type)); | |
3660 | rhs_mode != VOIDmode && GET_MODE_SIZE (rhs_mode) <= fltsz; | |
3661 | rhs_mode = GET_MODE_2XWIDER_MODE (rhs_mode)) | |
3662 | { | |
3663 | cvt_type | |
3664 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
3665 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
3666 | if (cvt_type == NULL_TREE) | |
3667 | goto unsupported; | |
3668 | ||
3669 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
3670 | { | |
3671 | if (!supportable_convert_operation (code, vectype_out, | |
3672 | cvt_type, &decl1, &codecvt1)) | |
3673 | goto unsupported; | |
3674 | } | |
3675 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
3676 | cvt_type, &codecvt1, |
3677 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
3678 | &interm_types)) |
3679 | continue; | |
3680 | else | |
3681 | gcc_assert (multi_step_cvt == 0); | |
3682 | ||
3683 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
3684 | vectype_in, &code1, &code2, |
3685 | &multi_step_cvt, &interm_types)) | |
4a00c761 JJ |
3686 | break; |
3687 | } | |
3688 | ||
3689 | if (rhs_mode == VOIDmode || GET_MODE_SIZE (rhs_mode) > fltsz) | |
3690 | goto unsupported; | |
3691 | ||
3692 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
3693 | codecvt2 = ERROR_MARK; | |
3694 | else | |
3695 | { | |
3696 | multi_step_cvt++; | |
9771b263 | 3697 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
3698 | cvt_type = NULL_TREE; |
3699 | } | |
3700 | break; | |
3701 | ||
3702 | case NARROW: | |
3703 | gcc_assert (op_type == unary_op); | |
3704 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
3705 | &code1, &multi_step_cvt, | |
3706 | &interm_types)) | |
3707 | break; | |
3708 | ||
3709 | if (code != FIX_TRUNC_EXPR | |
3710 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
3711 | >= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
3712 | goto unsupported; | |
3713 | ||
3714 | rhs_mode = TYPE_MODE (rhs_type); | |
3715 | cvt_type | |
3716 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
3717 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
3718 | if (cvt_type == NULL_TREE) | |
3719 | goto unsupported; | |
3720 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
3721 | &decl1, &codecvt1)) | |
3722 | goto unsupported; | |
3723 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
3724 | &code1, &multi_step_cvt, | |
3725 | &interm_types)) | |
3726 | break; | |
3727 | goto unsupported; | |
3728 | ||
3729 | default: | |
3730 | gcc_unreachable (); | |
ebfd146a IR |
3731 | } |
3732 | ||
3733 | if (!vec_stmt) /* transformation not required. */ | |
3734 | { | |
73fbfcad | 3735 | if (dump_enabled_p ()) |
78c60e3d | 3736 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3737 | "=== vectorizable_conversion ===\n"); |
4a00c761 | 3738 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
3739 | { |
3740 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
c3e7ee41 | 3741 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
8bd37302 | 3742 | } |
4a00c761 JJ |
3743 | else if (modifier == NARROW) |
3744 | { | |
3745 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
8bd37302 | 3746 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 JJ |
3747 | } |
3748 | else | |
3749 | { | |
3750 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
8bd37302 | 3751 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 | 3752 | } |
9771b263 | 3753 | interm_types.release (); |
ebfd146a IR |
3754 | return true; |
3755 | } | |
3756 | ||
3757 | /** Transform. **/ | |
73fbfcad | 3758 | if (dump_enabled_p ()) |
78c60e3d | 3759 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3760 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 3761 | |
4a00c761 JJ |
3762 | if (op_type == binary_op) |
3763 | { | |
3764 | if (CONSTANT_CLASS_P (op0)) | |
3765 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
3766 | else if (CONSTANT_CLASS_P (op1)) | |
3767 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
3768 | } | |
3769 | ||
3770 | /* In case of multi-step conversion, we first generate conversion operations | |
3771 | to the intermediate types, and then from that types to the final one. | |
3772 | We create vector destinations for the intermediate type (TYPES) received | |
3773 | from supportable_*_operation, and store them in the correct order | |
3774 | for future use in vect_create_vectorized_*_stmts (). */ | |
9771b263 | 3775 | vec_dsts.create (multi_step_cvt + 1); |
82294ec1 JJ |
3776 | vec_dest = vect_create_destination_var (scalar_dest, |
3777 | (cvt_type && modifier == WIDEN) | |
3778 | ? cvt_type : vectype_out); | |
9771b263 | 3779 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3780 | |
3781 | if (multi_step_cvt) | |
3782 | { | |
9771b263 DN |
3783 | for (i = interm_types.length () - 1; |
3784 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
3785 | { |
3786 | vec_dest = vect_create_destination_var (scalar_dest, | |
3787 | intermediate_type); | |
9771b263 | 3788 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3789 | } |
3790 | } | |
ebfd146a | 3791 | |
4a00c761 | 3792 | if (cvt_type) |
82294ec1 JJ |
3793 | vec_dest = vect_create_destination_var (scalar_dest, |
3794 | modifier == WIDEN | |
3795 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
3796 | |
3797 | if (!slp_node) | |
3798 | { | |
30862efc | 3799 | if (modifier == WIDEN) |
4a00c761 | 3800 | { |
c3284718 | 3801 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 3802 | if (op_type == binary_op) |
9771b263 | 3803 | vec_oprnds1.create (1); |
4a00c761 | 3804 | } |
30862efc | 3805 | else if (modifier == NARROW) |
9771b263 DN |
3806 | vec_oprnds0.create ( |
3807 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
3808 | } |
3809 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 3810 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 3811 | |
4a00c761 | 3812 | last_oprnd = op0; |
ebfd146a IR |
3813 | prev_stmt_info = NULL; |
3814 | switch (modifier) | |
3815 | { | |
3816 | case NONE: | |
3817 | for (j = 0; j < ncopies; j++) | |
3818 | { | |
ebfd146a | 3819 | if (j == 0) |
d092494c IR |
3820 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node, |
3821 | -1); | |
ebfd146a IR |
3822 | else |
3823 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
3824 | ||
9771b263 | 3825 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
3826 | { |
3827 | /* Arguments are ready, create the new vector stmt. */ | |
3828 | if (code1 == CALL_EXPR) | |
3829 | { | |
3830 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
3831 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3832 | gimple_call_set_lhs (new_stmt, new_temp); | |
3833 | } | |
3834 | else | |
3835 | { | |
3836 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
0d0e4a03 | 3837 | new_stmt = gimple_build_assign (vec_dest, code1, vop0); |
4a00c761 JJ |
3838 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3839 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3840 | } | |
3841 | ||
3842 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3843 | if (slp_node) | |
9771b263 | 3844 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
3845 | } |
3846 | ||
ebfd146a IR |
3847 | if (j == 0) |
3848 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3849 | else | |
3850 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3851 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3852 | } | |
3853 | break; | |
3854 | ||
3855 | case WIDEN: | |
3856 | /* In case the vectorization factor (VF) is bigger than the number | |
3857 | of elements that we can fit in a vectype (nunits), we have to | |
3858 | generate more than one vector stmt - i.e - we need to "unroll" | |
3859 | the vector stmt by a factor VF/nunits. */ | |
3860 | for (j = 0; j < ncopies; j++) | |
3861 | { | |
4a00c761 | 3862 | /* Handle uses. */ |
ebfd146a | 3863 | if (j == 0) |
4a00c761 JJ |
3864 | { |
3865 | if (slp_node) | |
3866 | { | |
3867 | if (code == WIDEN_LSHIFT_EXPR) | |
3868 | { | |
3869 | unsigned int k; | |
ebfd146a | 3870 | |
4a00c761 JJ |
3871 | vec_oprnd1 = op1; |
3872 | /* Store vec_oprnd1 for every vector stmt to be created | |
3873 | for SLP_NODE. We check during the analysis that all | |
3874 | the shift arguments are the same. */ | |
3875 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 3876 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
3877 | |
3878 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
3879 | slp_node, -1); | |
3880 | } | |
3881 | else | |
3882 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
3883 | &vec_oprnds1, slp_node, -1); | |
3884 | } | |
3885 | else | |
3886 | { | |
3887 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
9771b263 | 3888 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
3889 | if (op_type == binary_op) |
3890 | { | |
3891 | if (code == WIDEN_LSHIFT_EXPR) | |
3892 | vec_oprnd1 = op1; | |
3893 | else | |
3894 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, | |
3895 | NULL); | |
9771b263 | 3896 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
3897 | } |
3898 | } | |
3899 | } | |
ebfd146a | 3900 | else |
4a00c761 JJ |
3901 | { |
3902 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
3903 | vec_oprnds0.truncate (0); |
3904 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
3905 | if (op_type == binary_op) |
3906 | { | |
3907 | if (code == WIDEN_LSHIFT_EXPR) | |
3908 | vec_oprnd1 = op1; | |
3909 | else | |
3910 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
3911 | vec_oprnd1); | |
9771b263 DN |
3912 | vec_oprnds1.truncate (0); |
3913 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
3914 | } |
3915 | } | |
ebfd146a | 3916 | |
4a00c761 JJ |
3917 | /* Arguments are ready. Create the new vector stmts. */ |
3918 | for (i = multi_step_cvt; i >= 0; i--) | |
3919 | { | |
9771b263 | 3920 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
3921 | enum tree_code c1 = code1, c2 = code2; |
3922 | if (i == 0 && codecvt2 != ERROR_MARK) | |
3923 | { | |
3924 | c1 = codecvt1; | |
3925 | c2 = codecvt2; | |
3926 | } | |
3927 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
3928 | &vec_oprnds1, | |
3929 | stmt, this_dest, gsi, | |
3930 | c1, c2, decl1, decl2, | |
3931 | op_type); | |
3932 | } | |
3933 | ||
9771b263 | 3934 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
3935 | { |
3936 | if (cvt_type) | |
3937 | { | |
3938 | if (codecvt1 == CALL_EXPR) | |
3939 | { | |
3940 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
3941 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3942 | gimple_call_set_lhs (new_stmt, new_temp); | |
3943 | } | |
3944 | else | |
3945 | { | |
3946 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 3947 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
3948 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
3949 | vop0); | |
4a00c761 JJ |
3950 | } |
3951 | ||
3952 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3953 | } | |
3954 | else | |
3955 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
3956 | ||
3957 | if (slp_node) | |
9771b263 | 3958 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
3959 | else |
3960 | { | |
3961 | if (!prev_stmt_info) | |
3962 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3963 | else | |
3964 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3965 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3966 | } | |
3967 | } | |
ebfd146a | 3968 | } |
4a00c761 JJ |
3969 | |
3970 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
3971 | break; |
3972 | ||
3973 | case NARROW: | |
3974 | /* In case the vectorization factor (VF) is bigger than the number | |
3975 | of elements that we can fit in a vectype (nunits), we have to | |
3976 | generate more than one vector stmt - i.e - we need to "unroll" | |
3977 | the vector stmt by a factor VF/nunits. */ | |
3978 | for (j = 0; j < ncopies; j++) | |
3979 | { | |
3980 | /* Handle uses. */ | |
4a00c761 JJ |
3981 | if (slp_node) |
3982 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
3983 | slp_node, -1); | |
ebfd146a IR |
3984 | else |
3985 | { | |
9771b263 | 3986 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
3987 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
3988 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
3989 | } |
3990 | ||
4a00c761 JJ |
3991 | /* Arguments are ready. Create the new vector stmts. */ |
3992 | if (cvt_type) | |
9771b263 | 3993 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
3994 | { |
3995 | if (codecvt1 == CALL_EXPR) | |
3996 | { | |
3997 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
3998 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3999 | gimple_call_set_lhs (new_stmt, new_temp); | |
4000 | } | |
4001 | else | |
4002 | { | |
4003 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4004 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4005 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4006 | vop0); | |
4a00c761 | 4007 | } |
ebfd146a | 4008 | |
4a00c761 | 4009 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 4010 | vec_oprnds0[i] = new_temp; |
4a00c761 | 4011 | } |
ebfd146a | 4012 | |
4a00c761 JJ |
4013 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
4014 | stmt, vec_dsts, gsi, | |
4015 | slp_node, code1, | |
4016 | &prev_stmt_info); | |
ebfd146a IR |
4017 | } |
4018 | ||
4019 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 4020 | break; |
ebfd146a IR |
4021 | } |
4022 | ||
9771b263 DN |
4023 | vec_oprnds0.release (); |
4024 | vec_oprnds1.release (); | |
4025 | vec_dsts.release (); | |
4026 | interm_types.release (); | |
ebfd146a IR |
4027 | |
4028 | return true; | |
4029 | } | |
ff802fa1 IR |
4030 | |
4031 | ||
ebfd146a IR |
4032 | /* Function vectorizable_assignment. |
4033 | ||
b8698a0f L |
4034 | Check if STMT performs an assignment (copy) that can be vectorized. |
4035 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4036 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4037 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4038 | ||
4039 | static bool | |
4040 | vectorizable_assignment (gimple stmt, gimple_stmt_iterator *gsi, | |
4041 | gimple *vec_stmt, slp_tree slp_node) | |
4042 | { | |
4043 | tree vec_dest; | |
4044 | tree scalar_dest; | |
4045 | tree op; | |
4046 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4047 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
4048 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4049 | tree new_temp; | |
4050 | tree def; | |
4051 | gimple def_stmt; | |
4052 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
fde9c428 | 4053 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
ebfd146a | 4054 | int ncopies; |
f18b55bd | 4055 | int i, j; |
6e1aa848 | 4056 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 4057 | tree vop; |
a70d6342 | 4058 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
f18b55bd IR |
4059 | gimple new_stmt = NULL; |
4060 | stmt_vec_info prev_stmt_info = NULL; | |
fde9c428 RG |
4061 | enum tree_code code; |
4062 | tree vectype_in; | |
ebfd146a IR |
4063 | |
4064 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4065 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4066 | case of SLP. */ | |
437f4a00 | 4067 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
4068 | ncopies = 1; |
4069 | else | |
4070 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4071 | ||
4072 | gcc_assert (ncopies >= 1); | |
ebfd146a | 4073 | |
a70d6342 | 4074 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4075 | return false; |
4076 | ||
8644a673 | 4077 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
4078 | return false; |
4079 | ||
4080 | /* Is vectorizable assignment? */ | |
4081 | if (!is_gimple_assign (stmt)) | |
4082 | return false; | |
4083 | ||
4084 | scalar_dest = gimple_assign_lhs (stmt); | |
4085 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4086 | return false; | |
4087 | ||
fde9c428 | 4088 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 4089 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
4090 | || code == PAREN_EXPR |
4091 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
4092 | op = gimple_assign_rhs1 (stmt); |
4093 | else | |
4094 | return false; | |
4095 | ||
7b7ec6c5 RG |
4096 | if (code == VIEW_CONVERT_EXPR) |
4097 | op = TREE_OPERAND (op, 0); | |
4098 | ||
24ee1384 | 4099 | if (!vect_is_simple_use_1 (op, stmt, loop_vinfo, bb_vinfo, |
fde9c428 | 4100 | &def_stmt, &def, &dt[0], &vectype_in)) |
ebfd146a | 4101 | { |
73fbfcad | 4102 | if (dump_enabled_p ()) |
78c60e3d | 4103 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4104 | "use not simple.\n"); |
ebfd146a IR |
4105 | return false; |
4106 | } | |
4107 | ||
fde9c428 RG |
4108 | /* We can handle NOP_EXPR conversions that do not change the number |
4109 | of elements or the vector size. */ | |
7b7ec6c5 RG |
4110 | if ((CONVERT_EXPR_CODE_P (code) |
4111 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 RG |
4112 | && (!vectype_in |
4113 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
4114 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
4115 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
4116 | return false; | |
4117 | ||
7b7b1813 RG |
4118 | /* We do not handle bit-precision changes. */ |
4119 | if ((CONVERT_EXPR_CODE_P (code) | |
4120 | || code == VIEW_CONVERT_EXPR) | |
4121 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
4122 | && ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4123 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4124 | || ((TYPE_PRECISION (TREE_TYPE (op)) | |
4125 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))) | |
4126 | /* But a conversion that does not change the bit-pattern is ok. */ | |
4127 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4128 | > TYPE_PRECISION (TREE_TYPE (op))) | |
4129 | && TYPE_UNSIGNED (TREE_TYPE (op)))) | |
4130 | { | |
73fbfcad | 4131 | if (dump_enabled_p ()) |
78c60e3d SS |
4132 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4133 | "type conversion to/from bit-precision " | |
e645e942 | 4134 | "unsupported.\n"); |
7b7b1813 RG |
4135 | return false; |
4136 | } | |
4137 | ||
ebfd146a IR |
4138 | if (!vec_stmt) /* transformation not required. */ |
4139 | { | |
4140 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 4141 | if (dump_enabled_p ()) |
78c60e3d | 4142 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4143 | "=== vectorizable_assignment ===\n"); |
c3e7ee41 | 4144 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
4145 | return true; |
4146 | } | |
4147 | ||
4148 | /** Transform. **/ | |
73fbfcad | 4149 | if (dump_enabled_p ()) |
e645e942 | 4150 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
4151 | |
4152 | /* Handle def. */ | |
4153 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4154 | ||
4155 | /* Handle use. */ | |
f18b55bd | 4156 | for (j = 0; j < ncopies; j++) |
ebfd146a | 4157 | { |
f18b55bd IR |
4158 | /* Handle uses. */ |
4159 | if (j == 0) | |
d092494c | 4160 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node, -1); |
f18b55bd IR |
4161 | else |
4162 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
4163 | ||
4164 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 4165 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 4166 | { |
7b7ec6c5 RG |
4167 | if (CONVERT_EXPR_CODE_P (code) |
4168 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 4169 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
4170 | new_stmt = gimple_build_assign (vec_dest, vop); |
4171 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4172 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4173 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4174 | if (slp_node) | |
9771b263 | 4175 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 4176 | } |
ebfd146a IR |
4177 | |
4178 | if (slp_node) | |
f18b55bd IR |
4179 | continue; |
4180 | ||
4181 | if (j == 0) | |
4182 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4183 | else | |
4184 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4185 | ||
4186 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4187 | } | |
b8698a0f | 4188 | |
9771b263 | 4189 | vec_oprnds.release (); |
ebfd146a IR |
4190 | return true; |
4191 | } | |
4192 | ||
9dc3f7de | 4193 | |
1107f3ae IR |
4194 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
4195 | either as shift by a scalar or by a vector. */ | |
4196 | ||
4197 | bool | |
4198 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
4199 | { | |
4200 | ||
ef4bddc2 | 4201 | machine_mode vec_mode; |
1107f3ae IR |
4202 | optab optab; |
4203 | int icode; | |
4204 | tree vectype; | |
4205 | ||
4206 | vectype = get_vectype_for_scalar_type (scalar_type); | |
4207 | if (!vectype) | |
4208 | return false; | |
4209 | ||
4210 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4211 | if (!optab | |
4212 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
4213 | { | |
4214 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4215 | if (!optab | |
4216 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
4217 | == CODE_FOR_nothing)) | |
4218 | return false; | |
4219 | } | |
4220 | ||
4221 | vec_mode = TYPE_MODE (vectype); | |
4222 | icode = (int) optab_handler (optab, vec_mode); | |
4223 | if (icode == CODE_FOR_nothing) | |
4224 | return false; | |
4225 | ||
4226 | return true; | |
4227 | } | |
4228 | ||
4229 | ||
9dc3f7de IR |
4230 | /* Function vectorizable_shift. |
4231 | ||
4232 | Check if STMT performs a shift operation that can be vectorized. | |
4233 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4234 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
4235 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4236 | ||
4237 | static bool | |
4238 | vectorizable_shift (gimple stmt, gimple_stmt_iterator *gsi, | |
4239 | gimple *vec_stmt, slp_tree slp_node) | |
4240 | { | |
4241 | tree vec_dest; | |
4242 | tree scalar_dest; | |
4243 | tree op0, op1 = NULL; | |
4244 | tree vec_oprnd1 = NULL_TREE; | |
4245 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4246 | tree vectype; | |
4247 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4248 | enum tree_code code; | |
ef4bddc2 | 4249 | machine_mode vec_mode; |
9dc3f7de IR |
4250 | tree new_temp; |
4251 | optab optab; | |
4252 | int icode; | |
ef4bddc2 | 4253 | machine_mode optab_op2_mode; |
9dc3f7de IR |
4254 | tree def; |
4255 | gimple def_stmt; | |
4256 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
4257 | gimple new_stmt = NULL; | |
4258 | stmt_vec_info prev_stmt_info; | |
4259 | int nunits_in; | |
4260 | int nunits_out; | |
4261 | tree vectype_out; | |
cede2577 | 4262 | tree op1_vectype; |
9dc3f7de IR |
4263 | int ncopies; |
4264 | int j, i; | |
6e1aa848 DN |
4265 | vec<tree> vec_oprnds0 = vNULL; |
4266 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
4267 | tree vop0, vop1; |
4268 | unsigned int k; | |
49eab32e | 4269 | bool scalar_shift_arg = true; |
9dc3f7de IR |
4270 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
4271 | int vf; | |
4272 | ||
4273 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
4274 | return false; | |
4275 | ||
4276 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
4277 | return false; | |
4278 | ||
4279 | /* Is STMT a vectorizable binary/unary operation? */ | |
4280 | if (!is_gimple_assign (stmt)) | |
4281 | return false; | |
4282 | ||
4283 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4284 | return false; | |
4285 | ||
4286 | code = gimple_assign_rhs_code (stmt); | |
4287 | ||
4288 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
4289 | || code == RROTATE_EXPR)) | |
4290 | return false; | |
4291 | ||
4292 | scalar_dest = gimple_assign_lhs (stmt); | |
4293 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
7b7b1813 RG |
4294 | if (TYPE_PRECISION (TREE_TYPE (scalar_dest)) |
4295 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4296 | { | |
73fbfcad | 4297 | if (dump_enabled_p ()) |
78c60e3d | 4298 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4299 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
4300 | return false; |
4301 | } | |
9dc3f7de IR |
4302 | |
4303 | op0 = gimple_assign_rhs1 (stmt); | |
24ee1384 | 4304 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
9dc3f7de IR |
4305 | &def_stmt, &def, &dt[0], &vectype)) |
4306 | { | |
73fbfcad | 4307 | if (dump_enabled_p ()) |
78c60e3d | 4308 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4309 | "use not simple.\n"); |
9dc3f7de IR |
4310 | return false; |
4311 | } | |
4312 | /* If op0 is an external or constant def use a vector type with | |
4313 | the same size as the output vector type. */ | |
4314 | if (!vectype) | |
4315 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
4316 | if (vec_stmt) | |
4317 | gcc_assert (vectype); | |
4318 | if (!vectype) | |
4319 | { | |
73fbfcad | 4320 | if (dump_enabled_p ()) |
78c60e3d | 4321 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4322 | "no vectype for scalar type\n"); |
9dc3f7de IR |
4323 | return false; |
4324 | } | |
4325 | ||
4326 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4327 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4328 | if (nunits_out != nunits_in) | |
4329 | return false; | |
4330 | ||
4331 | op1 = gimple_assign_rhs2 (stmt); | |
24ee1384 IR |
4332 | if (!vect_is_simple_use_1 (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
4333 | &def, &dt[1], &op1_vectype)) | |
9dc3f7de | 4334 | { |
73fbfcad | 4335 | if (dump_enabled_p ()) |
78c60e3d | 4336 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4337 | "use not simple.\n"); |
9dc3f7de IR |
4338 | return false; |
4339 | } | |
4340 | ||
4341 | if (loop_vinfo) | |
4342 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4343 | else | |
4344 | vf = 1; | |
4345 | ||
4346 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4347 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4348 | case of SLP. */ | |
437f4a00 | 4349 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
9dc3f7de IR |
4350 | ncopies = 1; |
4351 | else | |
4352 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
4353 | ||
4354 | gcc_assert (ncopies >= 1); | |
4355 | ||
4356 | /* Determine whether the shift amount is a vector, or scalar. If the | |
4357 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
4358 | ||
49eab32e JJ |
4359 | if (dt[1] == vect_internal_def && !slp_node) |
4360 | scalar_shift_arg = false; | |
4361 | else if (dt[1] == vect_constant_def | |
4362 | || dt[1] == vect_external_def | |
4363 | || dt[1] == vect_internal_def) | |
4364 | { | |
4365 | /* In SLP, need to check whether the shift count is the same, | |
4366 | in loops if it is a constant or invariant, it is always | |
4367 | a scalar shift. */ | |
4368 | if (slp_node) | |
4369 | { | |
9771b263 | 4370 | vec<gimple> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
49eab32e JJ |
4371 | gimple slpstmt; |
4372 | ||
9771b263 | 4373 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
4374 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
4375 | scalar_shift_arg = false; | |
4376 | } | |
4377 | } | |
4378 | else | |
4379 | { | |
73fbfcad | 4380 | if (dump_enabled_p ()) |
78c60e3d | 4381 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4382 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
4383 | return false; |
4384 | } | |
4385 | ||
9dc3f7de | 4386 | /* Vector shifted by vector. */ |
49eab32e | 4387 | if (!scalar_shift_arg) |
9dc3f7de IR |
4388 | { |
4389 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 4390 | if (dump_enabled_p ()) |
78c60e3d | 4391 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4392 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 4393 | |
aa948027 JJ |
4394 | if (!op1_vectype) |
4395 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
4396 | if (op1_vectype == NULL_TREE | |
4397 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 4398 | { |
73fbfcad | 4399 | if (dump_enabled_p ()) |
78c60e3d SS |
4400 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4401 | "unusable type for last operand in" | |
e645e942 | 4402 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
4403 | return false; |
4404 | } | |
9dc3f7de IR |
4405 | } |
4406 | /* See if the machine has a vector shifted by scalar insn and if not | |
4407 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 4408 | else |
9dc3f7de IR |
4409 | { |
4410 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4411 | if (optab | |
4412 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
4413 | { | |
73fbfcad | 4414 | if (dump_enabled_p ()) |
78c60e3d | 4415 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4416 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
4417 | } |
4418 | else | |
4419 | { | |
4420 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4421 | if (optab | |
4422 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
4423 | != CODE_FOR_nothing)) | |
4424 | { | |
49eab32e JJ |
4425 | scalar_shift_arg = false; |
4426 | ||
73fbfcad | 4427 | if (dump_enabled_p ()) |
78c60e3d | 4428 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4429 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
4430 | |
4431 | /* Unlike the other binary operators, shifts/rotates have | |
4432 | the rhs being int, instead of the same type as the lhs, | |
4433 | so make sure the scalar is the right type if we are | |
aa948027 | 4434 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
4435 | if (dt[1] == vect_constant_def) |
4436 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
4437 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
4438 | TREE_TYPE (op1))) | |
4439 | { | |
4440 | if (slp_node | |
4441 | && TYPE_MODE (TREE_TYPE (vectype)) | |
4442 | != TYPE_MODE (TREE_TYPE (op1))) | |
4443 | { | |
73fbfcad | 4444 | if (dump_enabled_p ()) |
78c60e3d SS |
4445 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4446 | "unusable type for last operand in" | |
e645e942 | 4447 | " vector/vector shift/rotate.\n"); |
aa948027 JJ |
4448 | return false; |
4449 | } | |
4450 | if (vec_stmt && !slp_node) | |
4451 | { | |
4452 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
4453 | op1 = vect_init_vector (stmt, op1, | |
4454 | TREE_TYPE (vectype), NULL); | |
4455 | } | |
4456 | } | |
9dc3f7de IR |
4457 | } |
4458 | } | |
4459 | } | |
9dc3f7de IR |
4460 | |
4461 | /* Supportable by target? */ | |
4462 | if (!optab) | |
4463 | { | |
73fbfcad | 4464 | if (dump_enabled_p ()) |
78c60e3d | 4465 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4466 | "no optab.\n"); |
9dc3f7de IR |
4467 | return false; |
4468 | } | |
4469 | vec_mode = TYPE_MODE (vectype); | |
4470 | icode = (int) optab_handler (optab, vec_mode); | |
4471 | if (icode == CODE_FOR_nothing) | |
4472 | { | |
73fbfcad | 4473 | if (dump_enabled_p ()) |
78c60e3d | 4474 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4475 | "op not supported by target.\n"); |
9dc3f7de IR |
4476 | /* Check only during analysis. */ |
4477 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
4478 | || (vf < vect_min_worthwhile_factor (code) | |
4479 | && !vec_stmt)) | |
4480 | return false; | |
73fbfcad | 4481 | if (dump_enabled_p ()) |
e645e942 TJ |
4482 | dump_printf_loc (MSG_NOTE, vect_location, |
4483 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
4484 | } |
4485 | ||
4486 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
4487 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
4488 | && vf < vect_min_worthwhile_factor (code) | |
4489 | && !vec_stmt) | |
4490 | { | |
73fbfcad | 4491 | if (dump_enabled_p ()) |
78c60e3d | 4492 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4493 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
4494 | return false; |
4495 | } | |
4496 | ||
4497 | if (!vec_stmt) /* transformation not required. */ | |
4498 | { | |
4499 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 4500 | if (dump_enabled_p ()) |
e645e942 TJ |
4501 | dump_printf_loc (MSG_NOTE, vect_location, |
4502 | "=== vectorizable_shift ===\n"); | |
c3e7ee41 | 4503 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
9dc3f7de IR |
4504 | return true; |
4505 | } | |
4506 | ||
4507 | /** Transform. **/ | |
4508 | ||
73fbfcad | 4509 | if (dump_enabled_p ()) |
78c60e3d | 4510 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4511 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
4512 | |
4513 | /* Handle def. */ | |
4514 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4515 | ||
9dc3f7de IR |
4516 | prev_stmt_info = NULL; |
4517 | for (j = 0; j < ncopies; j++) | |
4518 | { | |
4519 | /* Handle uses. */ | |
4520 | if (j == 0) | |
4521 | { | |
4522 | if (scalar_shift_arg) | |
4523 | { | |
4524 | /* Vector shl and shr insn patterns can be defined with scalar | |
4525 | operand 2 (shift operand). In this case, use constant or loop | |
4526 | invariant op1 directly, without extending it to vector mode | |
4527 | first. */ | |
4528 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
4529 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
4530 | { | |
73fbfcad | 4531 | if (dump_enabled_p ()) |
78c60e3d | 4532 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4533 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 4534 | vec_oprnd1 = op1; |
8930f723 | 4535 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 4536 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
4537 | if (slp_node) |
4538 | { | |
4539 | /* Store vec_oprnd1 for every vector stmt to be created | |
4540 | for SLP_NODE. We check during the analysis that all | |
4541 | the shift arguments are the same. | |
4542 | TODO: Allow different constants for different vector | |
4543 | stmts generated for an SLP instance. */ | |
4544 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4545 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
4546 | } |
4547 | } | |
4548 | } | |
4549 | ||
4550 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
4551 | (a special case for certain kind of vector shifts); otherwise, | |
4552 | operand 1 should be of a vector type (the usual case). */ | |
4553 | if (vec_oprnd1) | |
4554 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
d092494c | 4555 | slp_node, -1); |
9dc3f7de IR |
4556 | else |
4557 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
d092494c | 4558 | slp_node, -1); |
9dc3f7de IR |
4559 | } |
4560 | else | |
4561 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
4562 | ||
4563 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 4564 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 4565 | { |
9771b263 | 4566 | vop1 = vec_oprnds1[i]; |
0d0e4a03 | 4567 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
9dc3f7de IR |
4568 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4569 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4570 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4571 | if (slp_node) | |
9771b263 | 4572 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
4573 | } |
4574 | ||
4575 | if (slp_node) | |
4576 | continue; | |
4577 | ||
4578 | if (j == 0) | |
4579 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4580 | else | |
4581 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4582 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4583 | } | |
4584 | ||
9771b263 DN |
4585 | vec_oprnds0.release (); |
4586 | vec_oprnds1.release (); | |
9dc3f7de IR |
4587 | |
4588 | return true; | |
4589 | } | |
4590 | ||
4591 | ||
ebfd146a IR |
4592 | /* Function vectorizable_operation. |
4593 | ||
16949072 RG |
4594 | Check if STMT performs a binary, unary or ternary operation that can |
4595 | be vectorized. | |
b8698a0f | 4596 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
4597 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4598 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4599 | ||
4600 | static bool | |
4601 | vectorizable_operation (gimple stmt, gimple_stmt_iterator *gsi, | |
4602 | gimple *vec_stmt, slp_tree slp_node) | |
4603 | { | |
00f07b86 | 4604 | tree vec_dest; |
ebfd146a | 4605 | tree scalar_dest; |
16949072 | 4606 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 4607 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 4608 | tree vectype; |
ebfd146a IR |
4609 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4610 | enum tree_code code; | |
ef4bddc2 | 4611 | machine_mode vec_mode; |
ebfd146a IR |
4612 | tree new_temp; |
4613 | int op_type; | |
00f07b86 | 4614 | optab optab; |
ebfd146a | 4615 | int icode; |
ebfd146a IR |
4616 | tree def; |
4617 | gimple def_stmt; | |
16949072 RG |
4618 | enum vect_def_type dt[3] |
4619 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
ebfd146a IR |
4620 | gimple new_stmt = NULL; |
4621 | stmt_vec_info prev_stmt_info; | |
b690cc0f | 4622 | int nunits_in; |
ebfd146a IR |
4623 | int nunits_out; |
4624 | tree vectype_out; | |
4625 | int ncopies; | |
4626 | int j, i; | |
6e1aa848 DN |
4627 | vec<tree> vec_oprnds0 = vNULL; |
4628 | vec<tree> vec_oprnds1 = vNULL; | |
4629 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 4630 | tree vop0, vop1, vop2; |
a70d6342 IR |
4631 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
4632 | int vf; | |
4633 | ||
a70d6342 | 4634 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4635 | return false; |
4636 | ||
8644a673 | 4637 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
4638 | return false; |
4639 | ||
4640 | /* Is STMT a vectorizable binary/unary operation? */ | |
4641 | if (!is_gimple_assign (stmt)) | |
4642 | return false; | |
4643 | ||
4644 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4645 | return false; | |
4646 | ||
ebfd146a IR |
4647 | code = gimple_assign_rhs_code (stmt); |
4648 | ||
4649 | /* For pointer addition, we should use the normal plus for | |
4650 | the vector addition. */ | |
4651 | if (code == POINTER_PLUS_EXPR) | |
4652 | code = PLUS_EXPR; | |
4653 | ||
4654 | /* Support only unary or binary operations. */ | |
4655 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 4656 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 4657 | { |
73fbfcad | 4658 | if (dump_enabled_p ()) |
78c60e3d | 4659 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4660 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 4661 | op_type); |
ebfd146a IR |
4662 | return false; |
4663 | } | |
4664 | ||
b690cc0f RG |
4665 | scalar_dest = gimple_assign_lhs (stmt); |
4666 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
4667 | ||
7b7b1813 RG |
4668 | /* Most operations cannot handle bit-precision types without extra |
4669 | truncations. */ | |
4670 | if ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4671 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4672 | /* Exception are bitwise binary operations. */ | |
4673 | && code != BIT_IOR_EXPR | |
4674 | && code != BIT_XOR_EXPR | |
4675 | && code != BIT_AND_EXPR) | |
4676 | { | |
73fbfcad | 4677 | if (dump_enabled_p ()) |
78c60e3d | 4678 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4679 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
4680 | return false; |
4681 | } | |
4682 | ||
ebfd146a | 4683 | op0 = gimple_assign_rhs1 (stmt); |
24ee1384 | 4684 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
b690cc0f | 4685 | &def_stmt, &def, &dt[0], &vectype)) |
ebfd146a | 4686 | { |
73fbfcad | 4687 | if (dump_enabled_p ()) |
78c60e3d | 4688 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4689 | "use not simple.\n"); |
ebfd146a IR |
4690 | return false; |
4691 | } | |
b690cc0f RG |
4692 | /* If op0 is an external or constant def use a vector type with |
4693 | the same size as the output vector type. */ | |
4694 | if (!vectype) | |
4695 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
4696 | if (vec_stmt) |
4697 | gcc_assert (vectype); | |
4698 | if (!vectype) | |
4699 | { | |
73fbfcad | 4700 | if (dump_enabled_p ()) |
7d8930a0 | 4701 | { |
78c60e3d SS |
4702 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4703 | "no vectype for scalar type "); | |
4704 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
4705 | TREE_TYPE (op0)); | |
e645e942 | 4706 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
4707 | } |
4708 | ||
4709 | return false; | |
4710 | } | |
b690cc0f RG |
4711 | |
4712 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4713 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4714 | if (nunits_out != nunits_in) | |
4715 | return false; | |
ebfd146a | 4716 | |
16949072 | 4717 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
4718 | { |
4719 | op1 = gimple_assign_rhs2 (stmt); | |
24ee1384 IR |
4720 | if (!vect_is_simple_use (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
4721 | &def, &dt[1])) | |
ebfd146a | 4722 | { |
73fbfcad | 4723 | if (dump_enabled_p ()) |
78c60e3d | 4724 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4725 | "use not simple.\n"); |
ebfd146a IR |
4726 | return false; |
4727 | } | |
4728 | } | |
16949072 RG |
4729 | if (op_type == ternary_op) |
4730 | { | |
4731 | op2 = gimple_assign_rhs3 (stmt); | |
24ee1384 IR |
4732 | if (!vect_is_simple_use (op2, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
4733 | &def, &dt[2])) | |
16949072 | 4734 | { |
73fbfcad | 4735 | if (dump_enabled_p ()) |
78c60e3d | 4736 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4737 | "use not simple.\n"); |
16949072 RG |
4738 | return false; |
4739 | } | |
4740 | } | |
ebfd146a | 4741 | |
b690cc0f RG |
4742 | if (loop_vinfo) |
4743 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4744 | else | |
4745 | vf = 1; | |
4746 | ||
4747 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 4748 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 4749 | case of SLP. */ |
437f4a00 | 4750 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
b690cc0f RG |
4751 | ncopies = 1; |
4752 | else | |
4753 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
4754 | ||
4755 | gcc_assert (ncopies >= 1); | |
4756 | ||
9dc3f7de | 4757 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
4758 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
4759 | || code == RROTATE_EXPR) | |
9dc3f7de | 4760 | return false; |
ebfd146a | 4761 | |
ebfd146a | 4762 | /* Supportable by target? */ |
00f07b86 RH |
4763 | |
4764 | vec_mode = TYPE_MODE (vectype); | |
4765 | if (code == MULT_HIGHPART_EXPR) | |
ebfd146a | 4766 | { |
00f07b86 | 4767 | if (can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype))) |
dee54b6e | 4768 | icode = LAST_INSN_CODE; |
00f07b86 RH |
4769 | else |
4770 | icode = CODE_FOR_nothing; | |
ebfd146a | 4771 | } |
00f07b86 RH |
4772 | else |
4773 | { | |
4774 | optab = optab_for_tree_code (code, vectype, optab_default); | |
4775 | if (!optab) | |
5deb57cb | 4776 | { |
73fbfcad | 4777 | if (dump_enabled_p ()) |
78c60e3d | 4778 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4779 | "no optab.\n"); |
00f07b86 | 4780 | return false; |
5deb57cb | 4781 | } |
00f07b86 | 4782 | icode = (int) optab_handler (optab, vec_mode); |
5deb57cb JJ |
4783 | } |
4784 | ||
ebfd146a IR |
4785 | if (icode == CODE_FOR_nothing) |
4786 | { | |
73fbfcad | 4787 | if (dump_enabled_p ()) |
78c60e3d | 4788 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4789 | "op not supported by target.\n"); |
ebfd146a IR |
4790 | /* Check only during analysis. */ |
4791 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
5deb57cb | 4792 | || (!vec_stmt && vf < vect_min_worthwhile_factor (code))) |
ebfd146a | 4793 | return false; |
73fbfcad | 4794 | if (dump_enabled_p ()) |
e645e942 TJ |
4795 | dump_printf_loc (MSG_NOTE, vect_location, |
4796 | "proceeding using word mode.\n"); | |
383d9c83 IR |
4797 | } |
4798 | ||
4a00c761 | 4799 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
4800 | if (!VECTOR_MODE_P (vec_mode) |
4801 | && !vec_stmt | |
4802 | && vf < vect_min_worthwhile_factor (code)) | |
7d8930a0 | 4803 | { |
73fbfcad | 4804 | if (dump_enabled_p ()) |
78c60e3d | 4805 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4806 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 4807 | return false; |
7d8930a0 | 4808 | } |
ebfd146a | 4809 | |
ebfd146a IR |
4810 | if (!vec_stmt) /* transformation not required. */ |
4811 | { | |
4a00c761 | 4812 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 4813 | if (dump_enabled_p ()) |
78c60e3d | 4814 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4815 | "=== vectorizable_operation ===\n"); |
c3e7ee41 | 4816 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
4817 | return true; |
4818 | } | |
4819 | ||
4820 | /** Transform. **/ | |
4821 | ||
73fbfcad | 4822 | if (dump_enabled_p ()) |
78c60e3d | 4823 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4824 | "transform binary/unary operation.\n"); |
383d9c83 | 4825 | |
ebfd146a | 4826 | /* Handle def. */ |
00f07b86 | 4827 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 4828 | |
ebfd146a IR |
4829 | /* In case the vectorization factor (VF) is bigger than the number |
4830 | of elements that we can fit in a vectype (nunits), we have to generate | |
4831 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
4832 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
4833 | from one copy of the vector stmt to the next, in the field | |
4834 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
4835 | stages to find the correct vector defs to be used when vectorizing | |
4836 | stmts that use the defs of the current stmt. The example below | |
4837 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
4838 | we need to create 4 vectorized stmts): | |
4839 | ||
4840 | before vectorization: | |
4841 | RELATED_STMT VEC_STMT | |
4842 | S1: x = memref - - | |
4843 | S2: z = x + 1 - - | |
4844 | ||
4845 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
4846 | there): | |
4847 | RELATED_STMT VEC_STMT | |
4848 | VS1_0: vx0 = memref0 VS1_1 - | |
4849 | VS1_1: vx1 = memref1 VS1_2 - | |
4850 | VS1_2: vx2 = memref2 VS1_3 - | |
4851 | VS1_3: vx3 = memref3 - - | |
4852 | S1: x = load - VS1_0 | |
4853 | S2: z = x + 1 - - | |
4854 | ||
4855 | step2: vectorize stmt S2 (done here): | |
4856 | To vectorize stmt S2 we first need to find the relevant vector | |
4857 | def for the first operand 'x'. This is, as usual, obtained from | |
4858 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
4859 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
4860 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
4861 | the vector stmt VS2_0, and as usual, record it in the | |
4862 | STMT_VINFO_VEC_STMT of stmt S2. | |
4863 | When creating the second copy (VS2_1), we obtain the relevant vector | |
4864 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
4865 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
4866 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
4867 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
4868 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
4869 | chain of stmts and pointers: | |
4870 | RELATED_STMT VEC_STMT | |
4871 | VS1_0: vx0 = memref0 VS1_1 - | |
4872 | VS1_1: vx1 = memref1 VS1_2 - | |
4873 | VS1_2: vx2 = memref2 VS1_3 - | |
4874 | VS1_3: vx3 = memref3 - - | |
4875 | S1: x = load - VS1_0 | |
4876 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
4877 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
4878 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
4879 | VS2_3: vz3 = vx3 + v1 - - | |
4880 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
4881 | |
4882 | prev_stmt_info = NULL; | |
4883 | for (j = 0; j < ncopies; j++) | |
4884 | { | |
4885 | /* Handle uses. */ | |
4886 | if (j == 0) | |
4a00c761 JJ |
4887 | { |
4888 | if (op_type == binary_op || op_type == ternary_op) | |
4889 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
4890 | slp_node, -1); | |
4891 | else | |
4892 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
4893 | slp_node, -1); | |
4894 | if (op_type == ternary_op) | |
36ba4aae | 4895 | { |
9771b263 DN |
4896 | vec_oprnds2.create (1); |
4897 | vec_oprnds2.quick_push (vect_get_vec_def_for_operand (op2, | |
4898 | stmt, | |
4899 | NULL)); | |
36ba4aae | 4900 | } |
4a00c761 | 4901 | } |
ebfd146a | 4902 | else |
4a00c761 JJ |
4903 | { |
4904 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
4905 | if (op_type == ternary_op) | |
4906 | { | |
9771b263 DN |
4907 | tree vec_oprnd = vec_oprnds2.pop (); |
4908 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
4909 | vec_oprnd)); | |
4a00c761 JJ |
4910 | } |
4911 | } | |
4912 | ||
4913 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 4914 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 4915 | { |
4a00c761 | 4916 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 4917 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 4918 | vop2 = ((op_type == ternary_op) |
9771b263 | 4919 | ? vec_oprnds2[i] : NULL_TREE); |
0d0e4a03 | 4920 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1, vop2); |
4a00c761 JJ |
4921 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4922 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4923 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4924 | if (slp_node) | |
9771b263 | 4925 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
4926 | } |
4927 | ||
4a00c761 JJ |
4928 | if (slp_node) |
4929 | continue; | |
4930 | ||
4931 | if (j == 0) | |
4932 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4933 | else | |
4934 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4935 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
4936 | } |
4937 | ||
9771b263 DN |
4938 | vec_oprnds0.release (); |
4939 | vec_oprnds1.release (); | |
4940 | vec_oprnds2.release (); | |
ebfd146a | 4941 | |
ebfd146a IR |
4942 | return true; |
4943 | } | |
4944 | ||
c716e67f XDL |
4945 | /* A helper function to ensure data reference DR's base alignment |
4946 | for STMT_INFO. */ | |
4947 | ||
4948 | static void | |
4949 | ensure_base_align (stmt_vec_info stmt_info, struct data_reference *dr) | |
4950 | { | |
4951 | if (!dr->aux) | |
4952 | return; | |
4953 | ||
4954 | if (((dataref_aux *)dr->aux)->base_misaligned) | |
4955 | { | |
4956 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
4957 | tree base_decl = ((dataref_aux *)dr->aux)->base_decl; | |
4958 | ||
428f0c67 JH |
4959 | if (decl_in_symtab_p (base_decl)) |
4960 | symtab_node::get (base_decl)->increase_alignment (TYPE_ALIGN (vectype)); | |
4961 | else | |
4962 | { | |
4963 | DECL_ALIGN (base_decl) = TYPE_ALIGN (vectype); | |
4964 | DECL_USER_ALIGN (base_decl) = 1; | |
4965 | } | |
c716e67f XDL |
4966 | ((dataref_aux *)dr->aux)->base_misaligned = false; |
4967 | } | |
4968 | } | |
4969 | ||
ebfd146a | 4970 | |
09dfa495 BM |
4971 | /* Given a vector type VECTYPE returns the VECTOR_CST mask that implements |
4972 | reversal of the vector elements. If that is impossible to do, | |
4973 | returns NULL. */ | |
4974 | ||
4975 | static tree | |
4976 | perm_mask_for_reverse (tree vectype) | |
4977 | { | |
4978 | int i, nunits; | |
4979 | unsigned char *sel; | |
4980 | ||
4981 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4982 | sel = XALLOCAVEC (unsigned char, nunits); | |
4983 | ||
4984 | for (i = 0; i < nunits; ++i) | |
4985 | sel[i] = nunits - 1 - i; | |
4986 | ||
557be5a8 AL |
4987 | if (!can_vec_perm_p (TYPE_MODE (vectype), false, sel)) |
4988 | return NULL_TREE; | |
4989 | return vect_gen_perm_mask_checked (vectype, sel); | |
09dfa495 BM |
4990 | } |
4991 | ||
ebfd146a IR |
4992 | /* Function vectorizable_store. |
4993 | ||
b8698a0f L |
4994 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
4995 | can be vectorized. | |
4996 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4997 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4998 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4999 | ||
5000 | static bool | |
5001 | vectorizable_store (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
c716e67f | 5002 | slp_tree slp_node) |
ebfd146a IR |
5003 | { |
5004 | tree scalar_dest; | |
5005 | tree data_ref; | |
5006 | tree op; | |
5007 | tree vec_oprnd = NULL_TREE; | |
5008 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5009 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
5010 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
272c6793 | 5011 | tree elem_type; |
ebfd146a | 5012 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 5013 | struct loop *loop = NULL; |
ef4bddc2 | 5014 | machine_mode vec_mode; |
ebfd146a IR |
5015 | tree dummy; |
5016 | enum dr_alignment_support alignment_support_scheme; | |
5017 | tree def; | |
5018 | gimple def_stmt; | |
5019 | enum vect_def_type dt; | |
5020 | stmt_vec_info prev_stmt_info = NULL; | |
5021 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 5022 | tree dataref_offset = NULL_TREE; |
fef4d2b3 | 5023 | gimple ptr_incr = NULL; |
ebfd146a IR |
5024 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
5025 | int ncopies; | |
5026 | int j; | |
5027 | gimple next_stmt, first_stmt = NULL; | |
0d0293ac | 5028 | bool grouped_store = false; |
272c6793 | 5029 | bool store_lanes_p = false; |
ebfd146a | 5030 | unsigned int group_size, i; |
6e1aa848 DN |
5031 | vec<tree> dr_chain = vNULL; |
5032 | vec<tree> oprnds = vNULL; | |
5033 | vec<tree> result_chain = vNULL; | |
ebfd146a | 5034 | bool inv_p; |
09dfa495 BM |
5035 | bool negative = false; |
5036 | tree offset = NULL_TREE; | |
6e1aa848 | 5037 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 5038 | bool slp = (slp_node != NULL); |
ebfd146a | 5039 | unsigned int vec_num; |
a70d6342 | 5040 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
272c6793 | 5041 | tree aggr_type; |
a70d6342 IR |
5042 | |
5043 | if (loop_vinfo) | |
5044 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
5045 | |
5046 | /* Multiple types in SLP are handled by creating the appropriate number of | |
5047 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5048 | case of SLP. */ | |
437f4a00 | 5049 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
5050 | ncopies = 1; |
5051 | else | |
5052 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
5053 | ||
5054 | gcc_assert (ncopies >= 1); | |
5055 | ||
5056 | /* FORNOW. This restriction should be relaxed. */ | |
a70d6342 | 5057 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) |
ebfd146a | 5058 | { |
73fbfcad | 5059 | if (dump_enabled_p ()) |
78c60e3d | 5060 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5061 | "multiple types in nested loop.\n"); |
ebfd146a IR |
5062 | return false; |
5063 | } | |
5064 | ||
a70d6342 | 5065 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5066 | return false; |
5067 | ||
8644a673 | 5068 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
5069 | return false; |
5070 | ||
5071 | /* Is vectorizable store? */ | |
5072 | ||
5073 | if (!is_gimple_assign (stmt)) | |
5074 | return false; | |
5075 | ||
5076 | scalar_dest = gimple_assign_lhs (stmt); | |
ab0ef706 JJ |
5077 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
5078 | && is_pattern_stmt_p (stmt_info)) | |
5079 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
ebfd146a | 5080 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
38000232 | 5081 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF |
ebfd146a | 5082 | && TREE_CODE (scalar_dest) != INDIRECT_REF |
e9dbe7bb IR |
5083 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
5084 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
5085 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
5086 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
5087 | return false; |
5088 | ||
5089 | gcc_assert (gimple_assign_single_p (stmt)); | |
5090 | op = gimple_assign_rhs1 (stmt); | |
24ee1384 IR |
5091 | if (!vect_is_simple_use (op, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
5092 | &def, &dt)) | |
ebfd146a | 5093 | { |
73fbfcad | 5094 | if (dump_enabled_p ()) |
78c60e3d | 5095 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5096 | "use not simple.\n"); |
ebfd146a IR |
5097 | return false; |
5098 | } | |
5099 | ||
272c6793 | 5100 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 5101 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 5102 | |
ebfd146a IR |
5103 | /* FORNOW. In some cases can vectorize even if data-type not supported |
5104 | (e.g. - array initialization with 0). */ | |
947131ba | 5105 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
5106 | return false; |
5107 | ||
5108 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
5109 | return false; | |
5110 | ||
09dfa495 BM |
5111 | negative = |
5112 | tree_int_cst_compare (loop && nested_in_vect_loop_p (loop, stmt) | |
5113 | ? STMT_VINFO_DR_STEP (stmt_info) : DR_STEP (dr), | |
5114 | size_zero_node) < 0; | |
5115 | if (negative && ncopies > 1) | |
a1e53f3f | 5116 | { |
73fbfcad | 5117 | if (dump_enabled_p ()) |
78c60e3d | 5118 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
f234d260 | 5119 | "multiple types with negative step.\n"); |
a1e53f3f L |
5120 | return false; |
5121 | } | |
5122 | ||
09dfa495 BM |
5123 | if (negative) |
5124 | { | |
5125 | gcc_assert (!grouped_store); | |
5126 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
5127 | if (alignment_support_scheme != dr_aligned | |
5128 | && alignment_support_scheme != dr_unaligned_supported) | |
5129 | { | |
5130 | if (dump_enabled_p ()) | |
5131 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
f234d260 | 5132 | "negative step but alignment required.\n"); |
09dfa495 BM |
5133 | return false; |
5134 | } | |
f234d260 BM |
5135 | if (dt != vect_constant_def |
5136 | && dt != vect_external_def | |
5137 | && !perm_mask_for_reverse (vectype)) | |
09dfa495 BM |
5138 | { |
5139 | if (dump_enabled_p ()) | |
5140 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
f234d260 | 5141 | "negative step and reversing not supported.\n"); |
09dfa495 BM |
5142 | return false; |
5143 | } | |
5144 | } | |
5145 | ||
0d0293ac | 5146 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 5147 | { |
0d0293ac | 5148 | grouped_store = true; |
e14c1050 | 5149 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
b602d918 RS |
5150 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
5151 | { | |
e14c1050 | 5152 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
5153 | if (vect_store_lanes_supported (vectype, group_size)) |
5154 | store_lanes_p = true; | |
0d0293ac | 5155 | else if (!vect_grouped_store_supported (vectype, group_size)) |
b602d918 RS |
5156 | return false; |
5157 | } | |
b8698a0f | 5158 | |
ebfd146a IR |
5159 | if (first_stmt == stmt) |
5160 | { | |
5161 | /* STMT is the leader of the group. Check the operands of all the | |
5162 | stmts of the group. */ | |
e14c1050 | 5163 | next_stmt = GROUP_NEXT_ELEMENT (stmt_info); |
ebfd146a IR |
5164 | while (next_stmt) |
5165 | { | |
5166 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
5167 | op = gimple_assign_rhs1 (next_stmt); | |
24ee1384 IR |
5168 | if (!vect_is_simple_use (op, next_stmt, loop_vinfo, bb_vinfo, |
5169 | &def_stmt, &def, &dt)) | |
ebfd146a | 5170 | { |
73fbfcad | 5171 | if (dump_enabled_p ()) |
78c60e3d | 5172 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5173 | "use not simple.\n"); |
ebfd146a IR |
5174 | return false; |
5175 | } | |
e14c1050 | 5176 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
5177 | } |
5178 | } | |
5179 | } | |
5180 | ||
5181 | if (!vec_stmt) /* transformation not required. */ | |
5182 | { | |
5183 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
92345349 BS |
5184 | vect_model_store_cost (stmt_info, ncopies, store_lanes_p, dt, |
5185 | NULL, NULL, NULL); | |
ebfd146a IR |
5186 | return true; |
5187 | } | |
5188 | ||
5189 | /** Transform. **/ | |
5190 | ||
c716e67f XDL |
5191 | ensure_base_align (stmt_info, dr); |
5192 | ||
0d0293ac | 5193 | if (grouped_store) |
ebfd146a IR |
5194 | { |
5195 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 5196 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 5197 | |
e14c1050 | 5198 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
5199 | |
5200 | /* FORNOW */ | |
a70d6342 | 5201 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
5202 | |
5203 | /* We vectorize all the stmts of the interleaving group when we | |
5204 | reach the last stmt in the group. */ | |
e14c1050 IR |
5205 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
5206 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
5207 | && !slp) |
5208 | { | |
5209 | *vec_stmt = NULL; | |
5210 | return true; | |
5211 | } | |
5212 | ||
5213 | if (slp) | |
4b5caab7 | 5214 | { |
0d0293ac | 5215 | grouped_store = false; |
4b5caab7 IR |
5216 | /* VEC_NUM is the number of vect stmts to be created for this |
5217 | group. */ | |
5218 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 5219 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
4b5caab7 | 5220 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
d092494c | 5221 | op = gimple_assign_rhs1 (first_stmt); |
4b5caab7 | 5222 | } |
ebfd146a | 5223 | else |
4b5caab7 IR |
5224 | /* VEC_NUM is the number of vect stmts to be created for this |
5225 | group. */ | |
ebfd146a IR |
5226 | vec_num = group_size; |
5227 | } | |
b8698a0f | 5228 | else |
ebfd146a IR |
5229 | { |
5230 | first_stmt = stmt; | |
5231 | first_dr = dr; | |
5232 | group_size = vec_num = 1; | |
ebfd146a | 5233 | } |
b8698a0f | 5234 | |
73fbfcad | 5235 | if (dump_enabled_p ()) |
78c60e3d | 5236 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5237 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 5238 | |
9771b263 DN |
5239 | dr_chain.create (group_size); |
5240 | oprnds.create (group_size); | |
ebfd146a | 5241 | |
720f5239 | 5242 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 5243 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
5244 | /* Targets with store-lane instructions must not require explicit |
5245 | realignment. */ | |
5246 | gcc_assert (!store_lanes_p | |
5247 | || alignment_support_scheme == dr_aligned | |
5248 | || alignment_support_scheme == dr_unaligned_supported); | |
5249 | ||
09dfa495 BM |
5250 | if (negative) |
5251 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
5252 | ||
272c6793 RS |
5253 | if (store_lanes_p) |
5254 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
5255 | else | |
5256 | aggr_type = vectype; | |
ebfd146a IR |
5257 | |
5258 | /* In case the vectorization factor (VF) is bigger than the number | |
5259 | of elements that we can fit in a vectype (nunits), we have to generate | |
5260 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 5261 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
5262 | vect_get_vec_def_for_copy_stmt. */ |
5263 | ||
0d0293ac | 5264 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
5265 | |
5266 | S1: &base + 2 = x2 | |
5267 | S2: &base = x0 | |
5268 | S3: &base + 1 = x1 | |
5269 | S4: &base + 3 = x3 | |
5270 | ||
5271 | We create vectorized stores starting from base address (the access of the | |
5272 | first stmt in the chain (S2 in the above example), when the last store stmt | |
5273 | of the chain (S4) is reached: | |
5274 | ||
5275 | VS1: &base = vx2 | |
5276 | VS2: &base + vec_size*1 = vx0 | |
5277 | VS3: &base + vec_size*2 = vx1 | |
5278 | VS4: &base + vec_size*3 = vx3 | |
5279 | ||
5280 | Then permutation statements are generated: | |
5281 | ||
3fcc1b55 JJ |
5282 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
5283 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 5284 | ... |
b8698a0f | 5285 | |
ebfd146a IR |
5286 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
5287 | (the order of the data-refs in the output of vect_permute_store_chain | |
5288 | corresponds to the order of scalar stmts in the interleaving chain - see | |
5289 | the documentation of vect_permute_store_chain()). | |
5290 | ||
5291 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 5292 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 5293 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 5294 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
5295 | */ |
5296 | ||
5297 | prev_stmt_info = NULL; | |
5298 | for (j = 0; j < ncopies; j++) | |
5299 | { | |
5300 | gimple new_stmt; | |
ebfd146a IR |
5301 | |
5302 | if (j == 0) | |
5303 | { | |
5304 | if (slp) | |
5305 | { | |
5306 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c IR |
5307 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
5308 | NULL, slp_node, -1); | |
ebfd146a | 5309 | |
9771b263 | 5310 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
5311 | } |
5312 | else | |
5313 | { | |
b8698a0f L |
5314 | /* For interleaved stores we collect vectorized defs for all the |
5315 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
5316 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
5317 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
5318 | ||
0d0293ac | 5319 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 5320 | OPRNDS are of size 1. */ |
b8698a0f | 5321 | next_stmt = first_stmt; |
ebfd146a IR |
5322 | for (i = 0; i < group_size; i++) |
5323 | { | |
b8698a0f L |
5324 | /* Since gaps are not supported for interleaved stores, |
5325 | GROUP_SIZE is the exact number of stmts in the chain. | |
5326 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
5327 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
5328 | iteration of the loop will be executed. */ |
5329 | gcc_assert (next_stmt | |
5330 | && gimple_assign_single_p (next_stmt)); | |
5331 | op = gimple_assign_rhs1 (next_stmt); | |
5332 | ||
b8698a0f | 5333 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt, |
ebfd146a | 5334 | NULL); |
9771b263 DN |
5335 | dr_chain.quick_push (vec_oprnd); |
5336 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 5337 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
5338 | } |
5339 | } | |
5340 | ||
5341 | /* We should have catched mismatched types earlier. */ | |
5342 | gcc_assert (useless_type_conversion_p (vectype, | |
5343 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
5344 | bool simd_lane_access_p |
5345 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
5346 | if (simd_lane_access_p | |
5347 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
5348 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
5349 | && integer_zerop (DR_OFFSET (first_dr)) | |
5350 | && integer_zerop (DR_INIT (first_dr)) | |
5351 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
5352 | get_alias_set (DR_REF (first_dr)))) | |
5353 | { | |
5354 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
5355 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
5356 | (DR_REF (first_dr)), 0); | |
8928eff3 | 5357 | inv_p = false; |
74bf76ed JJ |
5358 | } |
5359 | else | |
5360 | dataref_ptr | |
5361 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
5362 | simd_lane_access_p ? loop : NULL, | |
09dfa495 | 5363 | offset, &dummy, gsi, &ptr_incr, |
74bf76ed | 5364 | simd_lane_access_p, &inv_p); |
a70d6342 | 5365 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 5366 | } |
b8698a0f | 5367 | else |
ebfd146a | 5368 | { |
b8698a0f L |
5369 | /* For interleaved stores we created vectorized defs for all the |
5370 | defs stored in OPRNDS in the previous iteration (previous copy). | |
5371 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
5372 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
5373 | next copy. | |
0d0293ac | 5374 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
5375 | OPRNDS are of size 1. */ |
5376 | for (i = 0; i < group_size; i++) | |
5377 | { | |
9771b263 | 5378 | op = oprnds[i]; |
24ee1384 IR |
5379 | vect_is_simple_use (op, NULL, loop_vinfo, bb_vinfo, &def_stmt, |
5380 | &def, &dt); | |
b8698a0f | 5381 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
9771b263 DN |
5382 | dr_chain[i] = vec_oprnd; |
5383 | oprnds[i] = vec_oprnd; | |
ebfd146a | 5384 | } |
74bf76ed JJ |
5385 | if (dataref_offset) |
5386 | dataref_offset | |
5387 | = int_const_binop (PLUS_EXPR, dataref_offset, | |
5388 | TYPE_SIZE_UNIT (aggr_type)); | |
5389 | else | |
5390 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
5391 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
5392 | } |
5393 | ||
272c6793 | 5394 | if (store_lanes_p) |
ebfd146a | 5395 | { |
272c6793 | 5396 | tree vec_array; |
267d3070 | 5397 | |
272c6793 RS |
5398 | /* Combine all the vectors into an array. */ |
5399 | vec_array = create_vector_array (vectype, vec_num); | |
5400 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 5401 | { |
9771b263 | 5402 | vec_oprnd = dr_chain[i]; |
272c6793 | 5403 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 5404 | } |
b8698a0f | 5405 | |
272c6793 RS |
5406 | /* Emit: |
5407 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
5408 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
5409 | new_stmt = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array); | |
5410 | gimple_call_set_lhs (new_stmt, data_ref); | |
267d3070 | 5411 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
5412 | } |
5413 | else | |
5414 | { | |
5415 | new_stmt = NULL; | |
0d0293ac | 5416 | if (grouped_store) |
272c6793 | 5417 | { |
b6b9227d JJ |
5418 | if (j == 0) |
5419 | result_chain.create (group_size); | |
272c6793 RS |
5420 | /* Permute. */ |
5421 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
5422 | &result_chain); | |
5423 | } | |
c2d7ab2a | 5424 | |
272c6793 RS |
5425 | next_stmt = first_stmt; |
5426 | for (i = 0; i < vec_num; i++) | |
5427 | { | |
644ffefd | 5428 | unsigned align, misalign; |
272c6793 RS |
5429 | |
5430 | if (i > 0) | |
5431 | /* Bump the vector pointer. */ | |
5432 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
5433 | stmt, NULL_TREE); | |
5434 | ||
5435 | if (slp) | |
9771b263 | 5436 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
5437 | else if (grouped_store) |
5438 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 5439 | vect_permute_store_chain(). */ |
9771b263 | 5440 | vec_oprnd = result_chain[i]; |
272c6793 RS |
5441 | |
5442 | data_ref = build2 (MEM_REF, TREE_TYPE (vec_oprnd), dataref_ptr, | |
74bf76ed JJ |
5443 | dataref_offset |
5444 | ? dataref_offset | |
5445 | : build_int_cst (reference_alias_ptr_type | |
5446 | (DR_REF (first_dr)), 0)); | |
644ffefd | 5447 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 | 5448 | if (aligned_access_p (first_dr)) |
644ffefd | 5449 | misalign = 0; |
272c6793 RS |
5450 | else if (DR_MISALIGNMENT (first_dr) == -1) |
5451 | { | |
5452 | TREE_TYPE (data_ref) | |
5453 | = build_aligned_type (TREE_TYPE (data_ref), | |
5454 | TYPE_ALIGN (elem_type)); | |
644ffefd MJ |
5455 | align = TYPE_ALIGN_UNIT (elem_type); |
5456 | misalign = 0; | |
272c6793 RS |
5457 | } |
5458 | else | |
5459 | { | |
5460 | TREE_TYPE (data_ref) | |
5461 | = build_aligned_type (TREE_TYPE (data_ref), | |
5462 | TYPE_ALIGN (elem_type)); | |
644ffefd | 5463 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 5464 | } |
74bf76ed JJ |
5465 | if (dataref_offset == NULL_TREE) |
5466 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
5467 | misalign); | |
c2d7ab2a | 5468 | |
f234d260 BM |
5469 | if (negative |
5470 | && dt != vect_constant_def | |
5471 | && dt != vect_external_def) | |
09dfa495 BM |
5472 | { |
5473 | tree perm_mask = perm_mask_for_reverse (vectype); | |
5474 | tree perm_dest | |
5475 | = vect_create_destination_var (gimple_assign_rhs1 (stmt), | |
5476 | vectype); | |
b731b390 | 5477 | tree new_temp = make_ssa_name (perm_dest); |
09dfa495 BM |
5478 | |
5479 | /* Generate the permute statement. */ | |
5480 | gimple perm_stmt | |
0d0e4a03 JJ |
5481 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
5482 | vec_oprnd, perm_mask); | |
09dfa495 BM |
5483 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
5484 | ||
5485 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); | |
5486 | vec_oprnd = new_temp; | |
5487 | } | |
5488 | ||
272c6793 RS |
5489 | /* Arguments are ready. Create the new vector stmt. */ |
5490 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
5491 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
5492 | |
5493 | if (slp) | |
5494 | continue; | |
5495 | ||
e14c1050 | 5496 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
5497 | if (!next_stmt) |
5498 | break; | |
5499 | } | |
ebfd146a | 5500 | } |
1da0876c RS |
5501 | if (!slp) |
5502 | { | |
5503 | if (j == 0) | |
5504 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5505 | else | |
5506 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5507 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5508 | } | |
ebfd146a IR |
5509 | } |
5510 | ||
9771b263 DN |
5511 | dr_chain.release (); |
5512 | oprnds.release (); | |
5513 | result_chain.release (); | |
5514 | vec_oprnds.release (); | |
ebfd146a IR |
5515 | |
5516 | return true; | |
5517 | } | |
5518 | ||
557be5a8 AL |
5519 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
5520 | VECTOR_CST mask. No checks are made that the target platform supports the | |
5521 | mask, so callers may wish to test can_vec_perm_p separately, or use | |
5522 | vect_gen_perm_mask_checked. */ | |
a1e53f3f | 5523 | |
3fcc1b55 | 5524 | tree |
557be5a8 | 5525 | vect_gen_perm_mask_any (tree vectype, const unsigned char *sel) |
a1e53f3f | 5526 | { |
d2a12ae7 | 5527 | tree mask_elt_type, mask_type, mask_vec, *mask_elts; |
2635892a | 5528 | int i, nunits; |
a1e53f3f | 5529 | |
22e4dee7 | 5530 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
22e4dee7 | 5531 | |
96f9265a RG |
5532 | mask_elt_type = lang_hooks.types.type_for_mode |
5533 | (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype))), 1); | |
22e4dee7 | 5534 | mask_type = get_vectype_for_scalar_type (mask_elt_type); |
a1e53f3f | 5535 | |
d2a12ae7 | 5536 | mask_elts = XALLOCAVEC (tree, nunits); |
aec7ae7d | 5537 | for (i = nunits - 1; i >= 0; i--) |
d2a12ae7 RG |
5538 | mask_elts[i] = build_int_cst (mask_elt_type, sel[i]); |
5539 | mask_vec = build_vector (mask_type, mask_elts); | |
a1e53f3f | 5540 | |
2635892a | 5541 | return mask_vec; |
a1e53f3f L |
5542 | } |
5543 | ||
cf7aa6a3 AL |
5544 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_p, |
5545 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ | |
557be5a8 AL |
5546 | |
5547 | tree | |
5548 | vect_gen_perm_mask_checked (tree vectype, const unsigned char *sel) | |
5549 | { | |
5550 | gcc_assert (can_vec_perm_p (TYPE_MODE (vectype), false, sel)); | |
5551 | return vect_gen_perm_mask_any (vectype, sel); | |
5552 | } | |
5553 | ||
aec7ae7d JJ |
5554 | /* Given a vector variable X and Y, that was generated for the scalar |
5555 | STMT, generate instructions to permute the vector elements of X and Y | |
5556 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
5557 | permuted vector variable. */ | |
a1e53f3f L |
5558 | |
5559 | static tree | |
aec7ae7d JJ |
5560 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple stmt, |
5561 | gimple_stmt_iterator *gsi) | |
a1e53f3f L |
5562 | { |
5563 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 5564 | tree perm_dest, data_ref; |
a1e53f3f L |
5565 | gimple perm_stmt; |
5566 | ||
acdcd61b | 5567 | perm_dest = vect_create_destination_var (gimple_get_lhs (stmt), vectype); |
b731b390 | 5568 | data_ref = make_ssa_name (perm_dest); |
a1e53f3f L |
5569 | |
5570 | /* Generate the permute statement. */ | |
0d0e4a03 | 5571 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
a1e53f3f L |
5572 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
5573 | ||
5574 | return data_ref; | |
5575 | } | |
5576 | ||
6b916b36 RB |
5577 | /* Hoist the definitions of all SSA uses on STMT out of the loop LOOP, |
5578 | inserting them on the loops preheader edge. Returns true if we | |
5579 | were successful in doing so (and thus STMT can be moved then), | |
5580 | otherwise returns false. */ | |
5581 | ||
5582 | static bool | |
5583 | hoist_defs_of_uses (gimple stmt, struct loop *loop) | |
5584 | { | |
5585 | ssa_op_iter i; | |
5586 | tree op; | |
5587 | bool any = false; | |
5588 | ||
5589 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
5590 | { | |
5591 | gimple def_stmt = SSA_NAME_DEF_STMT (op); | |
5592 | if (!gimple_nop_p (def_stmt) | |
5593 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
5594 | { | |
5595 | /* Make sure we don't need to recurse. While we could do | |
5596 | so in simple cases when there are more complex use webs | |
5597 | we don't have an easy way to preserve stmt order to fulfil | |
5598 | dependencies within them. */ | |
5599 | tree op2; | |
5600 | ssa_op_iter i2; | |
d1417442 JJ |
5601 | if (gimple_code (def_stmt) == GIMPLE_PHI) |
5602 | return false; | |
6b916b36 RB |
5603 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
5604 | { | |
5605 | gimple def_stmt2 = SSA_NAME_DEF_STMT (op2); | |
5606 | if (!gimple_nop_p (def_stmt2) | |
5607 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2))) | |
5608 | return false; | |
5609 | } | |
5610 | any = true; | |
5611 | } | |
5612 | } | |
5613 | ||
5614 | if (!any) | |
5615 | return true; | |
5616 | ||
5617 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
5618 | { | |
5619 | gimple def_stmt = SSA_NAME_DEF_STMT (op); | |
5620 | if (!gimple_nop_p (def_stmt) | |
5621 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
5622 | { | |
5623 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); | |
5624 | gsi_remove (&gsi, false); | |
5625 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); | |
5626 | } | |
5627 | } | |
5628 | ||
5629 | return true; | |
5630 | } | |
5631 | ||
ebfd146a IR |
5632 | /* vectorizable_load. |
5633 | ||
b8698a0f L |
5634 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
5635 | can be vectorized. | |
5636 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
5637 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5638 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5639 | ||
5640 | static bool | |
5641 | vectorizable_load (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
c716e67f | 5642 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a IR |
5643 | { |
5644 | tree scalar_dest; | |
5645 | tree vec_dest = NULL; | |
5646 | tree data_ref = NULL; | |
5647 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 5648 | stmt_vec_info prev_stmt_info; |
ebfd146a | 5649 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 5650 | struct loop *loop = NULL; |
ebfd146a | 5651 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 5652 | bool nested_in_vect_loop = false; |
c716e67f | 5653 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
ebfd146a | 5654 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
272c6793 | 5655 | tree elem_type; |
ebfd146a | 5656 | tree new_temp; |
ef4bddc2 | 5657 | machine_mode mode; |
ebfd146a IR |
5658 | gimple new_stmt = NULL; |
5659 | tree dummy; | |
5660 | enum dr_alignment_support alignment_support_scheme; | |
5661 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 5662 | tree dataref_offset = NULL_TREE; |
fef4d2b3 | 5663 | gimple ptr_incr = NULL; |
ebfd146a IR |
5664 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
5665 | int ncopies; | |
a64b9c26 | 5666 | int i, j, group_size, group_gap; |
ebfd146a IR |
5667 | tree msq = NULL_TREE, lsq; |
5668 | tree offset = NULL_TREE; | |
356bbc4c | 5669 | tree byte_offset = NULL_TREE; |
ebfd146a | 5670 | tree realignment_token = NULL_TREE; |
538dd0b7 | 5671 | gphi *phi = NULL; |
6e1aa848 | 5672 | vec<tree> dr_chain = vNULL; |
0d0293ac | 5673 | bool grouped_load = false; |
272c6793 | 5674 | bool load_lanes_p = false; |
ebfd146a | 5675 | gimple first_stmt; |
ebfd146a | 5676 | bool inv_p; |
319e6439 | 5677 | bool negative = false; |
ebfd146a IR |
5678 | bool compute_in_loop = false; |
5679 | struct loop *at_loop; | |
5680 | int vec_num; | |
5681 | bool slp = (slp_node != NULL); | |
5682 | bool slp_perm = false; | |
5683 | enum tree_code code; | |
a70d6342 IR |
5684 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
5685 | int vf; | |
272c6793 | 5686 | tree aggr_type; |
aec7ae7d JJ |
5687 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; |
5688 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
5689 | int gather_scale = 1; | |
5690 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
a70d6342 IR |
5691 | |
5692 | if (loop_vinfo) | |
5693 | { | |
5694 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
5695 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
5696 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
5697 | } | |
5698 | else | |
3533e503 | 5699 | vf = 1; |
ebfd146a IR |
5700 | |
5701 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 5702 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 5703 | case of SLP. */ |
437f4a00 | 5704 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
5705 | ncopies = 1; |
5706 | else | |
5707 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
5708 | ||
5709 | gcc_assert (ncopies >= 1); | |
5710 | ||
5711 | /* FORNOW. This restriction should be relaxed. */ | |
5712 | if (nested_in_vect_loop && ncopies > 1) | |
5713 | { | |
73fbfcad | 5714 | if (dump_enabled_p ()) |
78c60e3d | 5715 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5716 | "multiple types in nested loop.\n"); |
ebfd146a IR |
5717 | return false; |
5718 | } | |
5719 | ||
f2556b68 RB |
5720 | /* Invalidate assumptions made by dependence analysis when vectorization |
5721 | on the unrolled body effectively re-orders stmts. */ | |
5722 | if (ncopies > 1 | |
5723 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
5724 | && ((unsigned)LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
5725 | > STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
5726 | { | |
5727 | if (dump_enabled_p ()) | |
5728 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5729 | "cannot perform implicit CSE when unrolling " | |
5730 | "with negative dependence distance\n"); | |
5731 | return false; | |
5732 | } | |
5733 | ||
a70d6342 | 5734 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5735 | return false; |
5736 | ||
8644a673 | 5737 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
5738 | return false; |
5739 | ||
5740 | /* Is vectorizable load? */ | |
5741 | if (!is_gimple_assign (stmt)) | |
5742 | return false; | |
5743 | ||
5744 | scalar_dest = gimple_assign_lhs (stmt); | |
5745 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
5746 | return false; | |
5747 | ||
5748 | code = gimple_assign_rhs_code (stmt); | |
5749 | if (code != ARRAY_REF | |
38000232 | 5750 | && code != BIT_FIELD_REF |
ebfd146a | 5751 | && code != INDIRECT_REF |
e9dbe7bb IR |
5752 | && code != COMPONENT_REF |
5753 | && code != IMAGPART_EXPR | |
70f34814 | 5754 | && code != REALPART_EXPR |
42373e0b RG |
5755 | && code != MEM_REF |
5756 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
ebfd146a IR |
5757 | return false; |
5758 | ||
5759 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
5760 | return false; | |
5761 | ||
7b7b1813 | 5762 | elem_type = TREE_TYPE (vectype); |
947131ba | 5763 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
5764 | |
5765 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
5766 | (e.g. - data copies). */ | |
947131ba | 5767 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 5768 | { |
73fbfcad | 5769 | if (dump_enabled_p ()) |
78c60e3d | 5770 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5771 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
5772 | return false; |
5773 | } | |
5774 | ||
ebfd146a | 5775 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 5776 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 5777 | { |
0d0293ac | 5778 | grouped_load = true; |
ebfd146a | 5779 | /* FORNOW */ |
aec7ae7d | 5780 | gcc_assert (! nested_in_vect_loop && !STMT_VINFO_GATHER_P (stmt_info)); |
ebfd146a | 5781 | |
e14c1050 | 5782 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
d5f035ea RB |
5783 | |
5784 | /* If this is single-element interleaving with an element distance | |
5785 | that leaves unused vector loads around punt - we at least create | |
5786 | very sub-optimal code in that case (and blow up memory, | |
5787 | see PR65518). */ | |
5788 | if (first_stmt == stmt | |
5789 | && !GROUP_NEXT_ELEMENT (stmt_info) | |
5790 | && GROUP_SIZE (stmt_info) > TYPE_VECTOR_SUBPARTS (vectype)) | |
5791 | { | |
5792 | if (dump_enabled_p ()) | |
5793 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5794 | "single-element interleaving not supported " | |
5795 | "for not adjacent vector loads\n"); | |
5796 | return false; | |
5797 | } | |
5798 | ||
b602d918 RS |
5799 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
5800 | { | |
e14c1050 | 5801 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
5802 | if (vect_load_lanes_supported (vectype, group_size)) |
5803 | load_lanes_p = true; | |
0d0293ac | 5804 | else if (!vect_grouped_load_supported (vectype, group_size)) |
b602d918 RS |
5805 | return false; |
5806 | } | |
f2556b68 RB |
5807 | |
5808 | /* Invalidate assumptions made by dependence analysis when vectorization | |
5809 | on the unrolled body effectively re-orders stmts. */ | |
5810 | if (!PURE_SLP_STMT (stmt_info) | |
5811 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
5812 | && ((unsigned)LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
5813 | > STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
5814 | { | |
5815 | if (dump_enabled_p ()) | |
5816 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5817 | "cannot perform implicit CSE when performing " | |
5818 | "group loads with negative dependence distance\n"); | |
5819 | return false; | |
5820 | } | |
96bb56b2 RB |
5821 | |
5822 | /* Similarly when the stmt is a load that is both part of a SLP | |
5823 | instance and a loop vectorized stmt via the same-dr mechanism | |
5824 | we have to give up. */ | |
5825 | if (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info) | |
5826 | && (STMT_SLP_TYPE (stmt_info) | |
5827 | != STMT_SLP_TYPE (vinfo_for_stmt | |
5828 | (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info))))) | |
5829 | { | |
5830 | if (dump_enabled_p ()) | |
5831 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5832 | "conflicting SLP types for CSEd load\n"); | |
5833 | return false; | |
5834 | } | |
ebfd146a IR |
5835 | } |
5836 | ||
a1e53f3f | 5837 | |
aec7ae7d JJ |
5838 | if (STMT_VINFO_GATHER_P (stmt_info)) |
5839 | { | |
5840 | gimple def_stmt; | |
5841 | tree def; | |
5842 | gather_decl = vect_check_gather (stmt, loop_vinfo, &gather_base, | |
5843 | &gather_off, &gather_scale); | |
5844 | gcc_assert (gather_decl); | |
24ee1384 | 5845 | if (!vect_is_simple_use_1 (gather_off, NULL, loop_vinfo, bb_vinfo, |
aec7ae7d JJ |
5846 | &def_stmt, &def, &gather_dt, |
5847 | &gather_off_vectype)) | |
5848 | { | |
73fbfcad | 5849 | if (dump_enabled_p ()) |
78c60e3d | 5850 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5851 | "gather index use not simple.\n"); |
aec7ae7d JJ |
5852 | return false; |
5853 | } | |
5854 | } | |
7d75abc8 | 5855 | else if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
14ac6aa2 | 5856 | ; |
319e6439 RG |
5857 | else |
5858 | { | |
5859 | negative = tree_int_cst_compare (nested_in_vect_loop | |
5860 | ? STMT_VINFO_DR_STEP (stmt_info) | |
5861 | : DR_STEP (dr), | |
5862 | size_zero_node) < 0; | |
5863 | if (negative && ncopies > 1) | |
5864 | { | |
73fbfcad | 5865 | if (dump_enabled_p ()) |
78c60e3d | 5866 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5867 | "multiple types with negative step.\n"); |
319e6439 RG |
5868 | return false; |
5869 | } | |
5870 | ||
5871 | if (negative) | |
5872 | { | |
08940f33 RB |
5873 | if (grouped_load) |
5874 | { | |
5875 | if (dump_enabled_p ()) | |
5876 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 TJ |
5877 | "negative step for group load not supported" |
5878 | "\n"); | |
08940f33 RB |
5879 | return false; |
5880 | } | |
319e6439 RG |
5881 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
5882 | if (alignment_support_scheme != dr_aligned | |
5883 | && alignment_support_scheme != dr_unaligned_supported) | |
5884 | { | |
73fbfcad | 5885 | if (dump_enabled_p ()) |
78c60e3d | 5886 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5887 | "negative step but alignment required.\n"); |
319e6439 RG |
5888 | return false; |
5889 | } | |
5890 | if (!perm_mask_for_reverse (vectype)) | |
5891 | { | |
73fbfcad | 5892 | if (dump_enabled_p ()) |
78c60e3d | 5893 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
5894 | "negative step and reversing not supported." |
5895 | "\n"); | |
319e6439 RG |
5896 | return false; |
5897 | } | |
5898 | } | |
7d75abc8 | 5899 | } |
aec7ae7d | 5900 | |
ebfd146a IR |
5901 | if (!vec_stmt) /* transformation not required. */ |
5902 | { | |
5903 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
92345349 | 5904 | vect_model_load_cost (stmt_info, ncopies, load_lanes_p, NULL, NULL, NULL); |
ebfd146a IR |
5905 | return true; |
5906 | } | |
5907 | ||
73fbfcad | 5908 | if (dump_enabled_p ()) |
78c60e3d | 5909 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5910 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a IR |
5911 | |
5912 | /** Transform. **/ | |
5913 | ||
c716e67f XDL |
5914 | ensure_base_align (stmt_info, dr); |
5915 | ||
aec7ae7d JJ |
5916 | if (STMT_VINFO_GATHER_P (stmt_info)) |
5917 | { | |
5918 | tree vec_oprnd0 = NULL_TREE, op; | |
5919 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
5920 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
d3c2fee0 | 5921 | tree ptr, mask, var, scale, merge, perm_mask = NULL_TREE, prev_res = NULL_TREE; |
aec7ae7d JJ |
5922 | edge pe = loop_preheader_edge (loop); |
5923 | gimple_seq seq; | |
5924 | basic_block new_bb; | |
5925 | enum { NARROW, NONE, WIDEN } modifier; | |
5926 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
5927 | ||
5928 | if (nunits == gather_off_nunits) | |
5929 | modifier = NONE; | |
5930 | else if (nunits == gather_off_nunits / 2) | |
5931 | { | |
5932 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
5933 | modifier = WIDEN; | |
5934 | ||
5935 | for (i = 0; i < gather_off_nunits; ++i) | |
5936 | sel[i] = i | nunits; | |
5937 | ||
557be5a8 | 5938 | perm_mask = vect_gen_perm_mask_checked (gather_off_vectype, sel); |
aec7ae7d JJ |
5939 | } |
5940 | else if (nunits == gather_off_nunits * 2) | |
5941 | { | |
5942 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
5943 | modifier = NARROW; | |
5944 | ||
5945 | for (i = 0; i < nunits; ++i) | |
5946 | sel[i] = i < gather_off_nunits | |
5947 | ? i : i + nunits - gather_off_nunits; | |
5948 | ||
557be5a8 | 5949 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); |
aec7ae7d JJ |
5950 | ncopies *= 2; |
5951 | } | |
5952 | else | |
5953 | gcc_unreachable (); | |
5954 | ||
5955 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); | |
5956 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5957 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5958 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5959 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5960 | scaletype = TREE_VALUE (arglist); | |
d3c2fee0 | 5961 | gcc_checking_assert (types_compatible_p (srctype, rettype)); |
aec7ae7d JJ |
5962 | |
5963 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5964 | ||
5965 | ptr = fold_convert (ptrtype, gather_base); | |
5966 | if (!is_gimple_min_invariant (ptr)) | |
5967 | { | |
5968 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
5969 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
5970 | gcc_assert (!new_bb); | |
5971 | } | |
5972 | ||
5973 | /* Currently we support only unconditional gather loads, | |
5974 | so mask should be all ones. */ | |
d3c2fee0 AI |
5975 | if (TREE_CODE (masktype) == INTEGER_TYPE) |
5976 | mask = build_int_cst (masktype, -1); | |
5977 | else if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
5978 | { | |
5979 | mask = build_int_cst (TREE_TYPE (masktype), -1); | |
5980 | mask = build_vector_from_val (masktype, mask); | |
03b9e8e4 | 5981 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
d3c2fee0 | 5982 | } |
aec7ae7d JJ |
5983 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) |
5984 | { | |
5985 | REAL_VALUE_TYPE r; | |
5986 | long tmp[6]; | |
5987 | for (j = 0; j < 6; ++j) | |
5988 | tmp[j] = -1; | |
5989 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
5990 | mask = build_real (TREE_TYPE (masktype), r); | |
d3c2fee0 | 5991 | mask = build_vector_from_val (masktype, mask); |
03b9e8e4 | 5992 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
aec7ae7d JJ |
5993 | } |
5994 | else | |
5995 | gcc_unreachable (); | |
aec7ae7d JJ |
5996 | |
5997 | scale = build_int_cst (scaletype, gather_scale); | |
5998 | ||
d3c2fee0 AI |
5999 | if (TREE_CODE (TREE_TYPE (rettype)) == INTEGER_TYPE) |
6000 | merge = build_int_cst (TREE_TYPE (rettype), 0); | |
6001 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (rettype))) | |
6002 | { | |
6003 | REAL_VALUE_TYPE r; | |
6004 | long tmp[6]; | |
6005 | for (j = 0; j < 6; ++j) | |
6006 | tmp[j] = 0; | |
6007 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (rettype))); | |
6008 | merge = build_real (TREE_TYPE (rettype), r); | |
6009 | } | |
6010 | else | |
6011 | gcc_unreachable (); | |
6012 | merge = build_vector_from_val (rettype, merge); | |
6013 | merge = vect_init_vector (stmt, merge, rettype, NULL); | |
6014 | ||
aec7ae7d JJ |
6015 | prev_stmt_info = NULL; |
6016 | for (j = 0; j < ncopies; ++j) | |
6017 | { | |
6018 | if (modifier == WIDEN && (j & 1)) | |
6019 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
6020 | perm_mask, stmt, gsi); | |
6021 | else if (j == 0) | |
6022 | op = vec_oprnd0 | |
6023 | = vect_get_vec_def_for_operand (gather_off, stmt, NULL); | |
6024 | else | |
6025 | op = vec_oprnd0 | |
6026 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
6027 | ||
6028 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
6029 | { | |
6030 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
6031 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
6032 | var = vect_get_new_vect_var (idxtype, vect_simple_var, NULL); | |
b731b390 | 6033 | var = make_ssa_name (var); |
aec7ae7d JJ |
6034 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
6035 | new_stmt | |
0d0e4a03 | 6036 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
6037 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6038 | op = var; | |
6039 | } | |
6040 | ||
6041 | new_stmt | |
d3c2fee0 | 6042 | = gimple_build_call (gather_decl, 5, merge, ptr, op, mask, scale); |
aec7ae7d JJ |
6043 | |
6044 | if (!useless_type_conversion_p (vectype, rettype)) | |
6045 | { | |
6046 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
6047 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
6048 | var = vect_get_new_vect_var (rettype, vect_simple_var, NULL); | |
aec7ae7d JJ |
6049 | op = make_ssa_name (var, new_stmt); |
6050 | gimple_call_set_lhs (new_stmt, op); | |
6051 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 6052 | var = make_ssa_name (vec_dest); |
aec7ae7d JJ |
6053 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
6054 | new_stmt | |
0d0e4a03 | 6055 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
6056 | } |
6057 | else | |
6058 | { | |
6059 | var = make_ssa_name (vec_dest, new_stmt); | |
6060 | gimple_call_set_lhs (new_stmt, var); | |
6061 | } | |
6062 | ||
6063 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6064 | ||
6065 | if (modifier == NARROW) | |
6066 | { | |
6067 | if ((j & 1) == 0) | |
6068 | { | |
6069 | prev_res = var; | |
6070 | continue; | |
6071 | } | |
6072 | var = permute_vec_elements (prev_res, var, | |
6073 | perm_mask, stmt, gsi); | |
6074 | new_stmt = SSA_NAME_DEF_STMT (var); | |
6075 | } | |
6076 | ||
6077 | if (prev_stmt_info == NULL) | |
6078 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6079 | else | |
6080 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6081 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6082 | } | |
6083 | return true; | |
6084 | } | |
7d75abc8 MM |
6085 | else if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
6086 | { | |
6087 | gimple_stmt_iterator incr_gsi; | |
6088 | bool insert_after; | |
6089 | gimple incr; | |
6090 | tree offvar; | |
7d75abc8 MM |
6091 | tree ivstep; |
6092 | tree running_off; | |
9771b263 | 6093 | vec<constructor_elt, va_gc> *v = NULL; |
7d75abc8 | 6094 | gimple_seq stmts = NULL; |
14ac6aa2 RB |
6095 | tree stride_base, stride_step, alias_off; |
6096 | ||
6097 | gcc_assert (!nested_in_vect_loop); | |
7d75abc8 | 6098 | |
14ac6aa2 RB |
6099 | stride_base |
6100 | = fold_build_pointer_plus | |
6101 | (unshare_expr (DR_BASE_ADDRESS (dr)), | |
6102 | size_binop (PLUS_EXPR, | |
6103 | convert_to_ptrofftype (unshare_expr (DR_OFFSET (dr))), | |
c3284718 | 6104 | convert_to_ptrofftype (DR_INIT (dr)))); |
14ac6aa2 | 6105 | stride_step = fold_convert (sizetype, unshare_expr (DR_STEP (dr))); |
7d75abc8 MM |
6106 | |
6107 | /* For a load with loop-invariant (but other than power-of-2) | |
6108 | stride (i.e. not a grouped access) like so: | |
6109 | ||
6110 | for (i = 0; i < n; i += stride) | |
6111 | ... = array[i]; | |
6112 | ||
6113 | we generate a new induction variable and new accesses to | |
6114 | form a new vector (or vectors, depending on ncopies): | |
6115 | ||
6116 | for (j = 0; ; j += VF*stride) | |
6117 | tmp1 = array[j]; | |
6118 | tmp2 = array[j + stride]; | |
6119 | ... | |
6120 | vectemp = {tmp1, tmp2, ...} | |
6121 | */ | |
6122 | ||
6123 | ivstep = stride_step; | |
6124 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
6125 | build_int_cst (TREE_TYPE (ivstep), vf)); | |
6126 | ||
6127 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
6128 | ||
6129 | create_iv (stride_base, ivstep, NULL, | |
6130 | loop, &incr_gsi, insert_after, | |
6131 | &offvar, NULL); | |
6132 | incr = gsi_stmt (incr_gsi); | |
6133 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo, NULL)); | |
6134 | ||
6135 | stride_step = force_gimple_operand (stride_step, &stmts, true, NULL_TREE); | |
6136 | if (stmts) | |
6137 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
6138 | ||
6139 | prev_stmt_info = NULL; | |
6140 | running_off = offvar; | |
14ac6aa2 | 6141 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (dr)), 0); |
7d75abc8 MM |
6142 | for (j = 0; j < ncopies; j++) |
6143 | { | |
6144 | tree vec_inv; | |
6145 | ||
9771b263 | 6146 | vec_alloc (v, nunits); |
7d75abc8 MM |
6147 | for (i = 0; i < nunits; i++) |
6148 | { | |
6149 | tree newref, newoff; | |
6150 | gimple incr; | |
14ac6aa2 RB |
6151 | newref = build2 (MEM_REF, TREE_TYPE (vectype), |
6152 | running_off, alias_off); | |
7d75abc8 MM |
6153 | |
6154 | newref = force_gimple_operand_gsi (gsi, newref, true, | |
6155 | NULL_TREE, true, | |
6156 | GSI_SAME_STMT); | |
6157 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, newref); | |
b731b390 | 6158 | newoff = copy_ssa_name (running_off); |
0d0e4a03 JJ |
6159 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
6160 | running_off, stride_step); | |
7d75abc8 MM |
6161 | vect_finish_stmt_generation (stmt, incr, gsi); |
6162 | ||
6163 | running_off = newoff; | |
6164 | } | |
6165 | ||
6166 | vec_inv = build_constructor (vectype, v); | |
6167 | new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi); | |
6168 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
7d75abc8 MM |
6169 | |
6170 | if (j == 0) | |
6171 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6172 | else | |
6173 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6174 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6175 | } | |
6176 | return true; | |
6177 | } | |
aec7ae7d | 6178 | |
0d0293ac | 6179 | if (grouped_load) |
ebfd146a | 6180 | { |
e14c1050 | 6181 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
6aa904c4 | 6182 | if (slp |
01d8bf07 | 6183 | && !SLP_TREE_LOAD_PERMUTATION (slp_node).exists () |
9771b263 DN |
6184 | && first_stmt != SLP_TREE_SCALAR_STMTS (slp_node)[0]) |
6185 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 6186 | |
ebfd146a | 6187 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
6188 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
6189 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
6190 | ??? But we can only do so if there is exactly one | |
6191 | as we have no way to get at the rest. Leave the CSE | |
6192 | opportunity alone. | |
6193 | ??? With the group load eventually participating | |
6194 | in multiple different permutations (having multiple | |
6195 | slp nodes which refer to the same group) the CSE | |
6196 | is even wrong code. See PR56270. */ | |
6197 | && !slp) | |
ebfd146a IR |
6198 | { |
6199 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
6200 | return true; | |
6201 | } | |
6202 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 6203 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a IR |
6204 | |
6205 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
6206 | if (slp) | |
6207 | { | |
0d0293ac | 6208 | grouped_load = false; |
ebfd146a | 6209 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
01d8bf07 | 6210 | if (SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
a70d6342 | 6211 | slp_perm = true; |
a64b9c26 | 6212 | group_gap = GROUP_GAP (vinfo_for_stmt (first_stmt)); |
a70d6342 | 6213 | } |
ebfd146a | 6214 | else |
a64b9c26 RB |
6215 | { |
6216 | vec_num = group_size; | |
6217 | group_gap = 0; | |
6218 | } | |
ebfd146a IR |
6219 | } |
6220 | else | |
6221 | { | |
6222 | first_stmt = stmt; | |
6223 | first_dr = dr; | |
6224 | group_size = vec_num = 1; | |
a64b9c26 | 6225 | group_gap = 0; |
ebfd146a IR |
6226 | } |
6227 | ||
720f5239 | 6228 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 6229 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
6230 | /* Targets with load-lane instructions must not require explicit |
6231 | realignment. */ | |
6232 | gcc_assert (!load_lanes_p | |
6233 | || alignment_support_scheme == dr_aligned | |
6234 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
6235 | |
6236 | /* In case the vectorization factor (VF) is bigger than the number | |
6237 | of elements that we can fit in a vectype (nunits), we have to generate | |
6238 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 6239 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 6240 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 6241 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 6242 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
6243 | stmts that use the defs of the current stmt. The example below |
6244 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
6245 | need to create 4 vectorized stmts): | |
ebfd146a IR |
6246 | |
6247 | before vectorization: | |
6248 | RELATED_STMT VEC_STMT | |
6249 | S1: x = memref - - | |
6250 | S2: z = x + 1 - - | |
6251 | ||
6252 | step 1: vectorize stmt S1: | |
6253 | We first create the vector stmt VS1_0, and, as usual, record a | |
6254 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
6255 | Next, we create the vector stmt VS1_1, and record a pointer to | |
6256 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 6257 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
6258 | stmts and pointers: |
6259 | RELATED_STMT VEC_STMT | |
6260 | VS1_0: vx0 = memref0 VS1_1 - | |
6261 | VS1_1: vx1 = memref1 VS1_2 - | |
6262 | VS1_2: vx2 = memref2 VS1_3 - | |
6263 | VS1_3: vx3 = memref3 - - | |
6264 | S1: x = load - VS1_0 | |
6265 | S2: z = x + 1 - - | |
6266 | ||
b8698a0f L |
6267 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
6268 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
6269 | stmt S2. */ |
6270 | ||
0d0293ac | 6271 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
6272 | |
6273 | S1: x2 = &base + 2 | |
6274 | S2: x0 = &base | |
6275 | S3: x1 = &base + 1 | |
6276 | S4: x3 = &base + 3 | |
6277 | ||
b8698a0f | 6278 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
6279 | starting from the access of the first stmt of the chain: |
6280 | ||
6281 | VS1: vx0 = &base | |
6282 | VS2: vx1 = &base + vec_size*1 | |
6283 | VS3: vx3 = &base + vec_size*2 | |
6284 | VS4: vx4 = &base + vec_size*3 | |
6285 | ||
6286 | Then permutation statements are generated: | |
6287 | ||
e2c83630 RH |
6288 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
6289 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
6290 | ... |
6291 | ||
6292 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
6293 | (the order of the data-refs in the output of vect_permute_load_chain | |
6294 | corresponds to the order of scalar stmts in the interleaving chain - see | |
6295 | the documentation of vect_permute_load_chain()). | |
6296 | The generation of permutation stmts and recording them in | |
0d0293ac | 6297 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 6298 | |
b8698a0f | 6299 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
6300 | permutation stmts above are created for every copy. The result vector |
6301 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
6302 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
6303 | |
6304 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
6305 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
6306 | we generate the following code: | |
6307 | p = initial_addr; | |
6308 | indx = 0; | |
6309 | loop { | |
6310 | p = p + indx * vectype_size; | |
6311 | vec_dest = *(p); | |
6312 | indx = indx + 1; | |
6313 | } | |
6314 | ||
6315 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 6316 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
6317 | then generate the following code, in which the data in each iteration is |
6318 | obtained by two vector loads, one from the previous iteration, and one | |
6319 | from the current iteration: | |
6320 | p1 = initial_addr; | |
6321 | msq_init = *(floor(p1)) | |
6322 | p2 = initial_addr + VS - 1; | |
6323 | realignment_token = call target_builtin; | |
6324 | indx = 0; | |
6325 | loop { | |
6326 | p2 = p2 + indx * vectype_size | |
6327 | lsq = *(floor(p2)) | |
6328 | vec_dest = realign_load (msq, lsq, realignment_token) | |
6329 | indx = indx + 1; | |
6330 | msq = lsq; | |
6331 | } */ | |
6332 | ||
6333 | /* If the misalignment remains the same throughout the execution of the | |
6334 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 6335 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
6336 | This can only occur when vectorizing memory accesses in the inner-loop |
6337 | nested within an outer-loop that is being vectorized. */ | |
6338 | ||
d1e4b493 | 6339 | if (nested_in_vect_loop |
211bea38 | 6340 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
ebfd146a IR |
6341 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) |
6342 | { | |
6343 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
6344 | compute_in_loop = true; | |
6345 | } | |
6346 | ||
6347 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
6348 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 6349 | && !compute_in_loop) |
ebfd146a IR |
6350 | { |
6351 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
6352 | alignment_support_scheme, NULL_TREE, | |
6353 | &at_loop); | |
6354 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
6355 | { | |
538dd0b7 | 6356 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (msq)); |
356bbc4c JJ |
6357 | byte_offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
6358 | size_one_node); | |
ebfd146a IR |
6359 | } |
6360 | } | |
6361 | else | |
6362 | at_loop = loop; | |
6363 | ||
a1e53f3f L |
6364 | if (negative) |
6365 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
6366 | ||
272c6793 RS |
6367 | if (load_lanes_p) |
6368 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
6369 | else | |
6370 | aggr_type = vectype; | |
6371 | ||
ebfd146a IR |
6372 | prev_stmt_info = NULL; |
6373 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 6374 | { |
272c6793 | 6375 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 6376 | if (j == 0) |
74bf76ed JJ |
6377 | { |
6378 | bool simd_lane_access_p | |
6379 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
6380 | if (simd_lane_access_p | |
6381 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
6382 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
6383 | && integer_zerop (DR_OFFSET (first_dr)) | |
6384 | && integer_zerop (DR_INIT (first_dr)) | |
6385 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
6386 | get_alias_set (DR_REF (first_dr))) | |
6387 | && (alignment_support_scheme == dr_aligned | |
6388 | || alignment_support_scheme == dr_unaligned_supported)) | |
6389 | { | |
6390 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
6391 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
6392 | (DR_REF (first_dr)), 0); | |
8928eff3 | 6393 | inv_p = false; |
74bf76ed JJ |
6394 | } |
6395 | else | |
6396 | dataref_ptr | |
6397 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
6398 | offset, &dummy, gsi, &ptr_incr, | |
356bbc4c JJ |
6399 | simd_lane_access_p, &inv_p, |
6400 | byte_offset); | |
74bf76ed JJ |
6401 | } |
6402 | else if (dataref_offset) | |
6403 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
6404 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 6405 | else |
272c6793 RS |
6406 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
6407 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 6408 | |
0d0293ac | 6409 | if (grouped_load || slp_perm) |
9771b263 | 6410 | dr_chain.create (vec_num); |
5ce1ee7f | 6411 | |
272c6793 | 6412 | if (load_lanes_p) |
ebfd146a | 6413 | { |
272c6793 RS |
6414 | tree vec_array; |
6415 | ||
6416 | vec_array = create_vector_array (vectype, vec_num); | |
6417 | ||
6418 | /* Emit: | |
6419 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
6420 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
6421 | new_stmt = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
6422 | gimple_call_set_lhs (new_stmt, vec_array); | |
6423 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a | 6424 | |
272c6793 RS |
6425 | /* Extract each vector into an SSA_NAME. */ |
6426 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 6427 | { |
272c6793 RS |
6428 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
6429 | vec_array, i); | |
9771b263 | 6430 | dr_chain.quick_push (new_temp); |
272c6793 RS |
6431 | } |
6432 | ||
6433 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 6434 | vect_record_grouped_load_vectors (stmt, dr_chain); |
272c6793 RS |
6435 | } |
6436 | else | |
6437 | { | |
6438 | for (i = 0; i < vec_num; i++) | |
6439 | { | |
6440 | if (i > 0) | |
6441 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
6442 | stmt, NULL_TREE); | |
6443 | ||
6444 | /* 2. Create the vector-load in the loop. */ | |
6445 | switch (alignment_support_scheme) | |
6446 | { | |
6447 | case dr_aligned: | |
6448 | case dr_unaligned_supported: | |
be1ac4ec | 6449 | { |
644ffefd MJ |
6450 | unsigned int align, misalign; |
6451 | ||
272c6793 RS |
6452 | data_ref |
6453 | = build2 (MEM_REF, vectype, dataref_ptr, | |
74bf76ed JJ |
6454 | dataref_offset |
6455 | ? dataref_offset | |
6456 | : build_int_cst (reference_alias_ptr_type | |
6457 | (DR_REF (first_dr)), 0)); | |
644ffefd | 6458 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 RS |
6459 | if (alignment_support_scheme == dr_aligned) |
6460 | { | |
6461 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 6462 | misalign = 0; |
272c6793 RS |
6463 | } |
6464 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
6465 | { | |
6466 | TREE_TYPE (data_ref) | |
6467 | = build_aligned_type (TREE_TYPE (data_ref), | |
6468 | TYPE_ALIGN (elem_type)); | |
644ffefd MJ |
6469 | align = TYPE_ALIGN_UNIT (elem_type); |
6470 | misalign = 0; | |
272c6793 RS |
6471 | } |
6472 | else | |
6473 | { | |
6474 | TREE_TYPE (data_ref) | |
6475 | = build_aligned_type (TREE_TYPE (data_ref), | |
6476 | TYPE_ALIGN (elem_type)); | |
644ffefd | 6477 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 6478 | } |
74bf76ed JJ |
6479 | if (dataref_offset == NULL_TREE) |
6480 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), | |
6481 | align, misalign); | |
272c6793 | 6482 | break; |
be1ac4ec | 6483 | } |
272c6793 | 6484 | case dr_explicit_realign: |
267d3070 | 6485 | { |
272c6793 | 6486 | tree ptr, bump; |
272c6793 | 6487 | |
d88981fc | 6488 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
272c6793 RS |
6489 | |
6490 | if (compute_in_loop) | |
6491 | msq = vect_setup_realignment (first_stmt, gsi, | |
6492 | &realignment_token, | |
6493 | dr_explicit_realign, | |
6494 | dataref_ptr, NULL); | |
6495 | ||
b731b390 | 6496 | ptr = copy_ssa_name (dataref_ptr); |
0d0e4a03 JJ |
6497 | new_stmt = gimple_build_assign |
6498 | (ptr, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
6499 | build_int_cst |
6500 | (TREE_TYPE (dataref_ptr), | |
6501 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
6502 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6503 | data_ref | |
6504 | = build2 (MEM_REF, vectype, ptr, | |
6505 | build_int_cst (reference_alias_ptr_type | |
6506 | (DR_REF (first_dr)), 0)); | |
6507 | vec_dest = vect_create_destination_var (scalar_dest, | |
6508 | vectype); | |
6509 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
6510 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
6511 | gimple_assign_set_lhs (new_stmt, new_temp); | |
6512 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
6513 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
6514 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6515 | msq = new_temp; | |
6516 | ||
d88981fc | 6517 | bump = size_binop (MULT_EXPR, vs, |
7b7b1813 | 6518 | TYPE_SIZE_UNIT (elem_type)); |
d88981fc | 6519 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
272c6793 | 6520 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
0d0e4a03 JJ |
6521 | new_stmt = gimple_build_assign |
6522 | (NULL_TREE, BIT_AND_EXPR, ptr, | |
272c6793 RS |
6523 | build_int_cst |
6524 | (TREE_TYPE (ptr), | |
6525 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
070ecdfd | 6526 | ptr = copy_ssa_name (dataref_ptr, new_stmt); |
272c6793 RS |
6527 | gimple_assign_set_lhs (new_stmt, ptr); |
6528 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6529 | data_ref | |
6530 | = build2 (MEM_REF, vectype, ptr, | |
6531 | build_int_cst (reference_alias_ptr_type | |
6532 | (DR_REF (first_dr)), 0)); | |
6533 | break; | |
267d3070 | 6534 | } |
272c6793 | 6535 | case dr_explicit_realign_optimized: |
b731b390 | 6536 | new_temp = copy_ssa_name (dataref_ptr); |
0d0e4a03 JJ |
6537 | new_stmt = gimple_build_assign |
6538 | (new_temp, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
6539 | build_int_cst |
6540 | (TREE_TYPE (dataref_ptr), | |
6541 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
6542 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6543 | data_ref | |
6544 | = build2 (MEM_REF, vectype, new_temp, | |
6545 | build_int_cst (reference_alias_ptr_type | |
6546 | (DR_REF (first_dr)), 0)); | |
6547 | break; | |
6548 | default: | |
6549 | gcc_unreachable (); | |
6550 | } | |
ebfd146a | 6551 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 6552 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
6553 | new_temp = make_ssa_name (vec_dest, new_stmt); |
6554 | gimple_assign_set_lhs (new_stmt, new_temp); | |
6555 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6556 | ||
272c6793 RS |
6557 | /* 3. Handle explicit realignment if necessary/supported. |
6558 | Create in loop: | |
6559 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
6560 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
6561 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 6562 | { |
272c6793 RS |
6563 | lsq = gimple_assign_lhs (new_stmt); |
6564 | if (!realignment_token) | |
6565 | realignment_token = dataref_ptr; | |
6566 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
0d0e4a03 JJ |
6567 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, |
6568 | msq, lsq, realignment_token); | |
272c6793 RS |
6569 | new_temp = make_ssa_name (vec_dest, new_stmt); |
6570 | gimple_assign_set_lhs (new_stmt, new_temp); | |
6571 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6572 | ||
6573 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
6574 | { | |
6575 | gcc_assert (phi); | |
6576 | if (i == vec_num - 1 && j == ncopies - 1) | |
6577 | add_phi_arg (phi, lsq, | |
6578 | loop_latch_edge (containing_loop), | |
9e227d60 | 6579 | UNKNOWN_LOCATION); |
272c6793 RS |
6580 | msq = lsq; |
6581 | } | |
ebfd146a | 6582 | } |
ebfd146a | 6583 | |
59fd17e3 RB |
6584 | /* 4. Handle invariant-load. */ |
6585 | if (inv_p && !bb_vinfo) | |
6586 | { | |
59fd17e3 | 6587 | gcc_assert (!grouped_load); |
d1417442 JJ |
6588 | /* If we have versioned for aliasing or the loop doesn't |
6589 | have any data dependencies that would preclude this, | |
6590 | then we are sure this is a loop invariant load and | |
6591 | thus we can insert it on the preheader edge. */ | |
6592 | if (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) | |
6593 | && !nested_in_vect_loop | |
6b916b36 | 6594 | && hoist_defs_of_uses (stmt, loop)) |
a0e35eb0 RB |
6595 | { |
6596 | if (dump_enabled_p ()) | |
6597 | { | |
6598 | dump_printf_loc (MSG_NOTE, vect_location, | |
6599 | "hoisting out of the vectorized " | |
6600 | "loop: "); | |
6601 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
a0e35eb0 | 6602 | } |
b731b390 | 6603 | tree tem = copy_ssa_name (scalar_dest); |
a0e35eb0 RB |
6604 | gsi_insert_on_edge_immediate |
6605 | (loop_preheader_edge (loop), | |
6606 | gimple_build_assign (tem, | |
6607 | unshare_expr | |
6608 | (gimple_assign_rhs1 (stmt)))); | |
6609 | new_temp = vect_init_vector (stmt, tem, vectype, NULL); | |
6610 | } | |
6611 | else | |
6612 | { | |
6613 | gimple_stmt_iterator gsi2 = *gsi; | |
6614 | gsi_next (&gsi2); | |
6615 | new_temp = vect_init_vector (stmt, scalar_dest, | |
6616 | vectype, &gsi2); | |
6617 | } | |
59fd17e3 | 6618 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
a0e35eb0 RB |
6619 | set_vinfo_for_stmt (new_stmt, |
6620 | new_stmt_vec_info (new_stmt, loop_vinfo, | |
6621 | bb_vinfo)); | |
59fd17e3 RB |
6622 | } |
6623 | ||
272c6793 RS |
6624 | if (negative) |
6625 | { | |
aec7ae7d JJ |
6626 | tree perm_mask = perm_mask_for_reverse (vectype); |
6627 | new_temp = permute_vec_elements (new_temp, new_temp, | |
6628 | perm_mask, stmt, gsi); | |
ebfd146a IR |
6629 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
6630 | } | |
267d3070 | 6631 | |
272c6793 | 6632 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
6633 | vect_transform_grouped_load (). */ |
6634 | if (grouped_load || slp_perm) | |
9771b263 | 6635 | dr_chain.quick_push (new_temp); |
267d3070 | 6636 | |
272c6793 RS |
6637 | /* Store vector loads in the corresponding SLP_NODE. */ |
6638 | if (slp && !slp_perm) | |
9771b263 | 6639 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
272c6793 | 6640 | } |
a64b9c26 RB |
6641 | /* Bump the vector pointer to account for a gap. */ |
6642 | if (slp && group_gap != 0) | |
6643 | { | |
6644 | tree bump = size_binop (MULT_EXPR, | |
6645 | TYPE_SIZE_UNIT (elem_type), | |
6646 | size_int (group_gap)); | |
6647 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
6648 | stmt, bump); | |
6649 | } | |
ebfd146a IR |
6650 | } |
6651 | ||
6652 | if (slp && !slp_perm) | |
6653 | continue; | |
6654 | ||
6655 | if (slp_perm) | |
6656 | { | |
01d8bf07 | 6657 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
ebfd146a IR |
6658 | slp_node_instance, false)) |
6659 | { | |
9771b263 | 6660 | dr_chain.release (); |
ebfd146a IR |
6661 | return false; |
6662 | } | |
6663 | } | |
6664 | else | |
6665 | { | |
0d0293ac | 6666 | if (grouped_load) |
ebfd146a | 6667 | { |
272c6793 | 6668 | if (!load_lanes_p) |
0d0293ac | 6669 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 6670 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
6671 | } |
6672 | else | |
6673 | { | |
6674 | if (j == 0) | |
6675 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6676 | else | |
6677 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6678 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6679 | } | |
6680 | } | |
9771b263 | 6681 | dr_chain.release (); |
ebfd146a IR |
6682 | } |
6683 | ||
ebfd146a IR |
6684 | return true; |
6685 | } | |
6686 | ||
6687 | /* Function vect_is_simple_cond. | |
b8698a0f | 6688 | |
ebfd146a IR |
6689 | Input: |
6690 | LOOP - the loop that is being vectorized. | |
6691 | COND - Condition that is checked for simple use. | |
6692 | ||
e9e1d143 RG |
6693 | Output: |
6694 | *COMP_VECTYPE - the vector type for the comparison. | |
6695 | ||
ebfd146a IR |
6696 | Returns whether a COND can be vectorized. Checks whether |
6697 | condition operands are supportable using vec_is_simple_use. */ | |
6698 | ||
87aab9b2 | 6699 | static bool |
24ee1384 IR |
6700 | vect_is_simple_cond (tree cond, gimple stmt, loop_vec_info loop_vinfo, |
6701 | bb_vec_info bb_vinfo, tree *comp_vectype) | |
ebfd146a IR |
6702 | { |
6703 | tree lhs, rhs; | |
6704 | tree def; | |
6705 | enum vect_def_type dt; | |
e9e1d143 | 6706 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a IR |
6707 | |
6708 | if (!COMPARISON_CLASS_P (cond)) | |
6709 | return false; | |
6710 | ||
6711 | lhs = TREE_OPERAND (cond, 0); | |
6712 | rhs = TREE_OPERAND (cond, 1); | |
6713 | ||
6714 | if (TREE_CODE (lhs) == SSA_NAME) | |
6715 | { | |
6716 | gimple lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); | |
24ee1384 IR |
6717 | if (!vect_is_simple_use_1 (lhs, stmt, loop_vinfo, bb_vinfo, |
6718 | &lhs_def_stmt, &def, &dt, &vectype1)) | |
ebfd146a IR |
6719 | return false; |
6720 | } | |
6721 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
6722 | && TREE_CODE (lhs) != FIXED_CST) | |
6723 | return false; | |
6724 | ||
6725 | if (TREE_CODE (rhs) == SSA_NAME) | |
6726 | { | |
6727 | gimple rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); | |
24ee1384 IR |
6728 | if (!vect_is_simple_use_1 (rhs, stmt, loop_vinfo, bb_vinfo, |
6729 | &rhs_def_stmt, &def, &dt, &vectype2)) | |
ebfd146a IR |
6730 | return false; |
6731 | } | |
f7e531cf | 6732 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST |
ebfd146a IR |
6733 | && TREE_CODE (rhs) != FIXED_CST) |
6734 | return false; | |
6735 | ||
e9e1d143 | 6736 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
ebfd146a IR |
6737 | return true; |
6738 | } | |
6739 | ||
6740 | /* vectorizable_condition. | |
6741 | ||
b8698a0f L |
6742 | Check if STMT is conditional modify expression that can be vectorized. |
6743 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
6744 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
6745 | at GSI. |
6746 | ||
6747 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
6748 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
6749 | else caluse if it is 2). | |
ebfd146a IR |
6750 | |
6751 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
6752 | ||
4bbe8262 | 6753 | bool |
ebfd146a | 6754 | vectorizable_condition (gimple stmt, gimple_stmt_iterator *gsi, |
f7e531cf IR |
6755 | gimple *vec_stmt, tree reduc_def, int reduc_index, |
6756 | slp_tree slp_node) | |
ebfd146a IR |
6757 | { |
6758 | tree scalar_dest = NULL_TREE; | |
6759 | tree vec_dest = NULL_TREE; | |
ebfd146a IR |
6760 | tree cond_expr, then_clause, else_clause; |
6761 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
6762 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
df11cc78 | 6763 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
6764 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
6765 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
ebfd146a IR |
6766 | tree vec_compare, vec_cond_expr; |
6767 | tree new_temp; | |
6768 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
ebfd146a | 6769 | tree def; |
a855b1b1 | 6770 | enum vect_def_type dt, dts[4]; |
ebfd146a | 6771 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
f7e531cf | 6772 | int ncopies; |
ebfd146a | 6773 | enum tree_code code; |
a855b1b1 | 6774 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
6775 | int i, j; |
6776 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
6777 | vec<tree> vec_oprnds0 = vNULL; |
6778 | vec<tree> vec_oprnds1 = vNULL; | |
6779 | vec<tree> vec_oprnds2 = vNULL; | |
6780 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 6781 | tree vec_cmp_type; |
b8698a0f | 6782 | |
f7e531cf IR |
6783 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
6784 | ncopies = 1; | |
6785 | else | |
6786 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
437f4a00 | 6787 | |
ebfd146a | 6788 | gcc_assert (ncopies >= 1); |
a855b1b1 | 6789 | if (reduc_index && ncopies > 1) |
ebfd146a IR |
6790 | return false; /* FORNOW */ |
6791 | ||
f7e531cf IR |
6792 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
6793 | return false; | |
6794 | ||
6795 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
ebfd146a IR |
6796 | return false; |
6797 | ||
4bbe8262 IR |
6798 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
6799 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
6800 | && reduc_def)) | |
ebfd146a IR |
6801 | return false; |
6802 | ||
ebfd146a | 6803 | /* FORNOW: not yet supported. */ |
b8698a0f | 6804 | if (STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 6805 | { |
73fbfcad | 6806 | if (dump_enabled_p ()) |
78c60e3d | 6807 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6808 | "value used after loop.\n"); |
ebfd146a IR |
6809 | return false; |
6810 | } | |
6811 | ||
6812 | /* Is vectorizable conditional operation? */ | |
6813 | if (!is_gimple_assign (stmt)) | |
6814 | return false; | |
6815 | ||
6816 | code = gimple_assign_rhs_code (stmt); | |
6817 | ||
6818 | if (code != COND_EXPR) | |
6819 | return false; | |
6820 | ||
4e71066d RG |
6821 | cond_expr = gimple_assign_rhs1 (stmt); |
6822 | then_clause = gimple_assign_rhs2 (stmt); | |
6823 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 6824 | |
24ee1384 IR |
6825 | if (!vect_is_simple_cond (cond_expr, stmt, loop_vinfo, bb_vinfo, |
6826 | &comp_vectype) | |
e9e1d143 | 6827 | || !comp_vectype) |
ebfd146a IR |
6828 | return false; |
6829 | ||
6830 | if (TREE_CODE (then_clause) == SSA_NAME) | |
6831 | { | |
6832 | gimple then_def_stmt = SSA_NAME_DEF_STMT (then_clause); | |
24ee1384 | 6833 | if (!vect_is_simple_use (then_clause, stmt, loop_vinfo, bb_vinfo, |
ebfd146a IR |
6834 | &then_def_stmt, &def, &dt)) |
6835 | return false; | |
6836 | } | |
b8698a0f | 6837 | else if (TREE_CODE (then_clause) != INTEGER_CST |
ebfd146a IR |
6838 | && TREE_CODE (then_clause) != REAL_CST |
6839 | && TREE_CODE (then_clause) != FIXED_CST) | |
6840 | return false; | |
6841 | ||
6842 | if (TREE_CODE (else_clause) == SSA_NAME) | |
6843 | { | |
6844 | gimple else_def_stmt = SSA_NAME_DEF_STMT (else_clause); | |
24ee1384 | 6845 | if (!vect_is_simple_use (else_clause, stmt, loop_vinfo, bb_vinfo, |
ebfd146a IR |
6846 | &else_def_stmt, &def, &dt)) |
6847 | return false; | |
6848 | } | |
b8698a0f | 6849 | else if (TREE_CODE (else_clause) != INTEGER_CST |
ebfd146a IR |
6850 | && TREE_CODE (else_clause) != REAL_CST |
6851 | && TREE_CODE (else_clause) != FIXED_CST) | |
6852 | return false; | |
6853 | ||
74946978 MP |
6854 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (vectype))); |
6855 | /* The result of a vector comparison should be signed type. */ | |
6856 | tree cmp_type = build_nonstandard_integer_type (prec, 0); | |
6857 | vec_cmp_type = get_same_sized_vectype (cmp_type, vectype); | |
6858 | if (vec_cmp_type == NULL_TREE) | |
6859 | return false; | |
784fb9b3 | 6860 | |
b8698a0f | 6861 | if (!vec_stmt) |
ebfd146a IR |
6862 | { |
6863 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
e9e1d143 | 6864 | return expand_vec_cond_expr_p (vectype, comp_vectype); |
ebfd146a IR |
6865 | } |
6866 | ||
f7e531cf IR |
6867 | /* Transform. */ |
6868 | ||
6869 | if (!slp_node) | |
6870 | { | |
9771b263 DN |
6871 | vec_oprnds0.create (1); |
6872 | vec_oprnds1.create (1); | |
6873 | vec_oprnds2.create (1); | |
6874 | vec_oprnds3.create (1); | |
f7e531cf | 6875 | } |
ebfd146a IR |
6876 | |
6877 | /* Handle def. */ | |
6878 | scalar_dest = gimple_assign_lhs (stmt); | |
6879 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
6880 | ||
6881 | /* Handle cond expr. */ | |
a855b1b1 MM |
6882 | for (j = 0; j < ncopies; j++) |
6883 | { | |
538dd0b7 | 6884 | gassign *new_stmt = NULL; |
a855b1b1 MM |
6885 | if (j == 0) |
6886 | { | |
f7e531cf IR |
6887 | if (slp_node) |
6888 | { | |
00f96dc9 TS |
6889 | auto_vec<tree, 4> ops; |
6890 | auto_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 6891 | |
9771b263 DN |
6892 | ops.safe_push (TREE_OPERAND (cond_expr, 0)); |
6893 | ops.safe_push (TREE_OPERAND (cond_expr, 1)); | |
6894 | ops.safe_push (then_clause); | |
6895 | ops.safe_push (else_clause); | |
f7e531cf | 6896 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); |
37b5ec8f JJ |
6897 | vec_oprnds3 = vec_defs.pop (); |
6898 | vec_oprnds2 = vec_defs.pop (); | |
6899 | vec_oprnds1 = vec_defs.pop (); | |
6900 | vec_oprnds0 = vec_defs.pop (); | |
f7e531cf | 6901 | |
9771b263 DN |
6902 | ops.release (); |
6903 | vec_defs.release (); | |
f7e531cf IR |
6904 | } |
6905 | else | |
6906 | { | |
6907 | gimple gtemp; | |
6908 | vec_cond_lhs = | |
a855b1b1 MM |
6909 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), |
6910 | stmt, NULL); | |
24ee1384 IR |
6911 | vect_is_simple_use (TREE_OPERAND (cond_expr, 0), stmt, |
6912 | loop_vinfo, NULL, >emp, &def, &dts[0]); | |
f7e531cf IR |
6913 | |
6914 | vec_cond_rhs = | |
6915 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), | |
6916 | stmt, NULL); | |
24ee1384 IR |
6917 | vect_is_simple_use (TREE_OPERAND (cond_expr, 1), stmt, |
6918 | loop_vinfo, NULL, >emp, &def, &dts[1]); | |
f7e531cf IR |
6919 | if (reduc_index == 1) |
6920 | vec_then_clause = reduc_def; | |
6921 | else | |
6922 | { | |
6923 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
6924 | stmt, NULL); | |
24ee1384 | 6925 | vect_is_simple_use (then_clause, stmt, loop_vinfo, |
f7e531cf IR |
6926 | NULL, >emp, &def, &dts[2]); |
6927 | } | |
6928 | if (reduc_index == 2) | |
6929 | vec_else_clause = reduc_def; | |
6930 | else | |
6931 | { | |
6932 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
a855b1b1 | 6933 | stmt, NULL); |
24ee1384 | 6934 | vect_is_simple_use (else_clause, stmt, loop_vinfo, |
a855b1b1 | 6935 | NULL, >emp, &def, &dts[3]); |
f7e531cf | 6936 | } |
a855b1b1 MM |
6937 | } |
6938 | } | |
6939 | else | |
6940 | { | |
f7e531cf | 6941 | vec_cond_lhs = vect_get_vec_def_for_stmt_copy (dts[0], |
9771b263 | 6942 | vec_oprnds0.pop ()); |
f7e531cf | 6943 | vec_cond_rhs = vect_get_vec_def_for_stmt_copy (dts[1], |
9771b263 | 6944 | vec_oprnds1.pop ()); |
a855b1b1 | 6945 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 6946 | vec_oprnds2.pop ()); |
a855b1b1 | 6947 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 6948 | vec_oprnds3.pop ()); |
f7e531cf IR |
6949 | } |
6950 | ||
6951 | if (!slp_node) | |
6952 | { | |
9771b263 DN |
6953 | vec_oprnds0.quick_push (vec_cond_lhs); |
6954 | vec_oprnds1.quick_push (vec_cond_rhs); | |
6955 | vec_oprnds2.quick_push (vec_then_clause); | |
6956 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
6957 | } |
6958 | ||
9dc3f7de | 6959 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 6960 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 6961 | { |
9771b263 DN |
6962 | vec_cond_rhs = vec_oprnds1[i]; |
6963 | vec_then_clause = vec_oprnds2[i]; | |
6964 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 6965 | |
784fb9b3 JJ |
6966 | vec_compare = build2 (TREE_CODE (cond_expr), vec_cmp_type, |
6967 | vec_cond_lhs, vec_cond_rhs); | |
f7e531cf IR |
6968 | vec_cond_expr = build3 (VEC_COND_EXPR, vectype, |
6969 | vec_compare, vec_then_clause, vec_else_clause); | |
a855b1b1 | 6970 | |
f7e531cf IR |
6971 | new_stmt = gimple_build_assign (vec_dest, vec_cond_expr); |
6972 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
6973 | gimple_assign_set_lhs (new_stmt, new_temp); | |
6974 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6975 | if (slp_node) | |
9771b263 | 6976 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
6977 | } |
6978 | ||
6979 | if (slp_node) | |
6980 | continue; | |
6981 | ||
6982 | if (j == 0) | |
6983 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6984 | else | |
6985 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6986 | ||
6987 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 6988 | } |
b8698a0f | 6989 | |
9771b263 DN |
6990 | vec_oprnds0.release (); |
6991 | vec_oprnds1.release (); | |
6992 | vec_oprnds2.release (); | |
6993 | vec_oprnds3.release (); | |
f7e531cf | 6994 | |
ebfd146a IR |
6995 | return true; |
6996 | } | |
6997 | ||
6998 | ||
8644a673 | 6999 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
7000 | |
7001 | bool | |
a70d6342 | 7002 | vect_analyze_stmt (gimple stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 7003 | { |
8644a673 | 7004 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 7005 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 7006 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 7007 | bool ok; |
a70d6342 | 7008 | tree scalar_type, vectype; |
363477c0 JJ |
7009 | gimple pattern_stmt; |
7010 | gimple_seq pattern_def_seq; | |
ebfd146a | 7011 | |
73fbfcad | 7012 | if (dump_enabled_p ()) |
ebfd146a | 7013 | { |
78c60e3d SS |
7014 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
7015 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
8644a673 | 7016 | } |
ebfd146a | 7017 | |
1825a1f3 | 7018 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 7019 | { |
73fbfcad | 7020 | if (dump_enabled_p ()) |
78c60e3d | 7021 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7022 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
7023 | |
7024 | return false; | |
7025 | } | |
b8698a0f L |
7026 | |
7027 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
7028 | to include: |
7029 | - the COND_EXPR which is the loop exit condition | |
7030 | - any LABEL_EXPRs in the loop | |
b8698a0f | 7031 | - computations that are used only for array indexing or loop control. |
8644a673 | 7032 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 7033 | instance, therefore, all the statements are relevant. |
ebfd146a | 7034 | |
d092494c | 7035 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 7036 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
7037 | statements. In basic blocks we are called from some SLP instance |
7038 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
7039 | already will be part of SLP instance. */ | |
83197f37 IR |
7040 | |
7041 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 7042 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 7043 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 7044 | { |
9d5e7640 | 7045 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 7046 | && pattern_stmt |
9d5e7640 IR |
7047 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
7048 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
7049 | { | |
83197f37 | 7050 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
7051 | stmt = pattern_stmt; |
7052 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 7053 | if (dump_enabled_p ()) |
9d5e7640 | 7054 | { |
78c60e3d SS |
7055 | dump_printf_loc (MSG_NOTE, vect_location, |
7056 | "==> examining pattern statement: "); | |
7057 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
9d5e7640 IR |
7058 | } |
7059 | } | |
7060 | else | |
7061 | { | |
73fbfcad | 7062 | if (dump_enabled_p ()) |
e645e942 | 7063 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 7064 | |
9d5e7640 IR |
7065 | return true; |
7066 | } | |
8644a673 | 7067 | } |
83197f37 | 7068 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 7069 | && node == NULL |
83197f37 IR |
7070 | && pattern_stmt |
7071 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
7072 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
7073 | { | |
7074 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 7075 | if (dump_enabled_p ()) |
83197f37 | 7076 | { |
78c60e3d SS |
7077 | dump_printf_loc (MSG_NOTE, vect_location, |
7078 | "==> examining pattern statement: "); | |
7079 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
83197f37 IR |
7080 | } |
7081 | ||
7082 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node)) | |
7083 | return false; | |
7084 | } | |
ebfd146a | 7085 | |
1107f3ae | 7086 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 7087 | && node == NULL |
363477c0 | 7088 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 7089 | { |
363477c0 | 7090 | gimple_stmt_iterator si; |
1107f3ae | 7091 | |
363477c0 JJ |
7092 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
7093 | { | |
7094 | gimple pattern_def_stmt = gsi_stmt (si); | |
7095 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) | |
7096 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
7097 | { | |
7098 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 7099 | if (dump_enabled_p ()) |
363477c0 | 7100 | { |
78c60e3d SS |
7101 | dump_printf_loc (MSG_NOTE, vect_location, |
7102 | "==> examining pattern def statement: "); | |
7103 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
363477c0 | 7104 | } |
1107f3ae | 7105 | |
363477c0 JJ |
7106 | if (!vect_analyze_stmt (pattern_def_stmt, |
7107 | need_to_vectorize, node)) | |
7108 | return false; | |
7109 | } | |
7110 | } | |
7111 | } | |
1107f3ae | 7112 | |
8644a673 IR |
7113 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
7114 | { | |
7115 | case vect_internal_def: | |
7116 | break; | |
ebfd146a | 7117 | |
8644a673 | 7118 | case vect_reduction_def: |
7c5222ff | 7119 | case vect_nested_cycle: |
a70d6342 | 7120 | gcc_assert (!bb_vinfo && (relevance == vect_used_in_outer |
8644a673 | 7121 | || relevance == vect_used_in_outer_by_reduction |
a70d6342 | 7122 | || relevance == vect_unused_in_scope)); |
8644a673 IR |
7123 | break; |
7124 | ||
7125 | case vect_induction_def: | |
7126 | case vect_constant_def: | |
7127 | case vect_external_def: | |
7128 | case vect_unknown_def_type: | |
7129 | default: | |
7130 | gcc_unreachable (); | |
7131 | } | |
ebfd146a | 7132 | |
a70d6342 IR |
7133 | if (bb_vinfo) |
7134 | { | |
7135 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
7136 | ||
b690cc0f | 7137 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
73fbfcad | 7138 | if (dump_enabled_p ()) |
a70d6342 | 7139 | { |
78c60e3d SS |
7140 | dump_printf_loc (MSG_NOTE, vect_location, |
7141 | "get vectype for scalar type: "); | |
7142 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
e645e942 | 7143 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
7144 | } |
7145 | ||
7146 | vectype = get_vectype_for_scalar_type (scalar_type); | |
7147 | if (!vectype) | |
7148 | { | |
73fbfcad | 7149 | if (dump_enabled_p ()) |
a70d6342 | 7150 | { |
78c60e3d SS |
7151 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
7152 | "not SLPed: unsupported data-type "); | |
7153 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
7154 | scalar_type); | |
e645e942 | 7155 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
a70d6342 IR |
7156 | } |
7157 | return false; | |
7158 | } | |
7159 | ||
73fbfcad | 7160 | if (dump_enabled_p ()) |
a70d6342 | 7161 | { |
78c60e3d SS |
7162 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); |
7163 | dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); | |
e645e942 | 7164 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
7165 | } |
7166 | ||
7167 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
7168 | } | |
7169 | ||
8644a673 | 7170 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 7171 | { |
8644a673 | 7172 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
0136f8f0 AH |
7173 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
7174 | || (is_gimple_call (stmt) | |
7175 | && gimple_call_lhs (stmt) == NULL_TREE)); | |
8644a673 | 7176 | *need_to_vectorize = true; |
ebfd146a IR |
7177 | } |
7178 | ||
8644a673 | 7179 | ok = true; |
b8698a0f | 7180 | if (!bb_vinfo |
a70d6342 IR |
7181 | && (STMT_VINFO_RELEVANT_P (stmt_info) |
7182 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
0136f8f0 AH |
7183 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, NULL) |
7184 | || vectorizable_conversion (stmt, NULL, NULL, NULL) | |
9dc3f7de | 7185 | || vectorizable_shift (stmt, NULL, NULL, NULL) |
8644a673 IR |
7186 | || vectorizable_operation (stmt, NULL, NULL, NULL) |
7187 | || vectorizable_assignment (stmt, NULL, NULL, NULL) | |
7188 | || vectorizable_load (stmt, NULL, NULL, NULL, NULL) | |
190c2236 | 7189 | || vectorizable_call (stmt, NULL, NULL, NULL) |
8644a673 | 7190 | || vectorizable_store (stmt, NULL, NULL, NULL) |
b5aeb3bb | 7191 | || vectorizable_reduction (stmt, NULL, NULL, NULL) |
f7e531cf | 7192 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, NULL)); |
a70d6342 IR |
7193 | else |
7194 | { | |
7195 | if (bb_vinfo) | |
0136f8f0 AH |
7196 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) |
7197 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
4a00c761 | 7198 | || vectorizable_shift (stmt, NULL, NULL, node) |
9dc3f7de | 7199 | || vectorizable_operation (stmt, NULL, NULL, node) |
a70d6342 IR |
7200 | || vectorizable_assignment (stmt, NULL, NULL, node) |
7201 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
190c2236 | 7202 | || vectorizable_call (stmt, NULL, NULL, node) |
f7e531cf IR |
7203 | || vectorizable_store (stmt, NULL, NULL, node) |
7204 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node)); | |
b8698a0f | 7205 | } |
8644a673 IR |
7206 | |
7207 | if (!ok) | |
ebfd146a | 7208 | { |
73fbfcad | 7209 | if (dump_enabled_p ()) |
8644a673 | 7210 | { |
78c60e3d SS |
7211 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
7212 | "not vectorized: relevant stmt not "); | |
7213 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
7214 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 7215 | } |
b8698a0f | 7216 | |
ebfd146a IR |
7217 | return false; |
7218 | } | |
7219 | ||
a70d6342 IR |
7220 | if (bb_vinfo) |
7221 | return true; | |
7222 | ||
8644a673 IR |
7223 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
7224 | need extra handling, except for vectorizable reductions. */ | |
7225 | if (STMT_VINFO_LIVE_P (stmt_info) | |
7226 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
7227 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 7228 | |
8644a673 | 7229 | if (!ok) |
ebfd146a | 7230 | { |
73fbfcad | 7231 | if (dump_enabled_p ()) |
8644a673 | 7232 | { |
78c60e3d SS |
7233 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
7234 | "not vectorized: live stmt not "); | |
7235 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
7236 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 7237 | } |
b8698a0f | 7238 | |
8644a673 | 7239 | return false; |
ebfd146a IR |
7240 | } |
7241 | ||
ebfd146a IR |
7242 | return true; |
7243 | } | |
7244 | ||
7245 | ||
7246 | /* Function vect_transform_stmt. | |
7247 | ||
7248 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
7249 | ||
7250 | bool | |
7251 | vect_transform_stmt (gimple stmt, gimple_stmt_iterator *gsi, | |
0d0293ac | 7252 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
7253 | slp_instance slp_node_instance) |
7254 | { | |
7255 | bool is_store = false; | |
7256 | gimple vec_stmt = NULL; | |
7257 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a | 7258 | bool done; |
ebfd146a IR |
7259 | |
7260 | switch (STMT_VINFO_TYPE (stmt_info)) | |
7261 | { | |
7262 | case type_demotion_vec_info_type: | |
ebfd146a | 7263 | case type_promotion_vec_info_type: |
ebfd146a IR |
7264 | case type_conversion_vec_info_type: |
7265 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
7266 | gcc_assert (done); | |
7267 | break; | |
7268 | ||
7269 | case induc_vec_info_type: | |
7270 | gcc_assert (!slp_node); | |
7271 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
7272 | gcc_assert (done); | |
7273 | break; | |
7274 | ||
9dc3f7de IR |
7275 | case shift_vec_info_type: |
7276 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
7277 | gcc_assert (done); | |
7278 | break; | |
7279 | ||
ebfd146a IR |
7280 | case op_vec_info_type: |
7281 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
7282 | gcc_assert (done); | |
7283 | break; | |
7284 | ||
7285 | case assignment_vec_info_type: | |
7286 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
7287 | gcc_assert (done); | |
7288 | break; | |
7289 | ||
7290 | case load_vec_info_type: | |
b8698a0f | 7291 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
7292 | slp_node_instance); |
7293 | gcc_assert (done); | |
7294 | break; | |
7295 | ||
7296 | case store_vec_info_type: | |
7297 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
7298 | gcc_assert (done); | |
0d0293ac | 7299 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
7300 | { |
7301 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 7302 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
7303 | one are skipped, and there vec_stmt_info shouldn't be freed |
7304 | meanwhile. */ | |
0d0293ac | 7305 | *grouped_store = true; |
ebfd146a IR |
7306 | if (STMT_VINFO_VEC_STMT (stmt_info)) |
7307 | is_store = true; | |
7308 | } | |
7309 | else | |
7310 | is_store = true; | |
7311 | break; | |
7312 | ||
7313 | case condition_vec_info_type: | |
f7e531cf | 7314 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
7315 | gcc_assert (done); |
7316 | break; | |
7317 | ||
7318 | case call_vec_info_type: | |
190c2236 | 7319 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 7320 | stmt = gsi_stmt (*gsi); |
5ce9450f JJ |
7321 | if (is_gimple_call (stmt) |
7322 | && gimple_call_internal_p (stmt) | |
7323 | && gimple_call_internal_fn (stmt) == IFN_MASK_STORE) | |
7324 | is_store = true; | |
ebfd146a IR |
7325 | break; |
7326 | ||
0136f8f0 AH |
7327 | case call_simd_clone_vec_info_type: |
7328 | done = vectorizable_simd_clone_call (stmt, gsi, &vec_stmt, slp_node); | |
7329 | stmt = gsi_stmt (*gsi); | |
7330 | break; | |
7331 | ||
ebfd146a | 7332 | case reduc_vec_info_type: |
b5aeb3bb | 7333 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
7334 | gcc_assert (done); |
7335 | break; | |
7336 | ||
7337 | default: | |
7338 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
7339 | { | |
73fbfcad | 7340 | if (dump_enabled_p ()) |
78c60e3d | 7341 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7342 | "stmt not supported.\n"); |
ebfd146a IR |
7343 | gcc_unreachable (); |
7344 | } | |
7345 | } | |
7346 | ||
7347 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that | |
7348 | is being vectorized, but outside the immediately enclosing loop. */ | |
7349 | if (vec_stmt | |
a70d6342 IR |
7350 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
7351 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
7352 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
7353 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
7354 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 7355 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 7356 | vect_used_in_outer_by_reduction)) |
ebfd146a | 7357 | { |
a70d6342 IR |
7358 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
7359 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
7360 | imm_use_iterator imm_iter; |
7361 | use_operand_p use_p; | |
7362 | tree scalar_dest; | |
7363 | gimple exit_phi; | |
7364 | ||
73fbfcad | 7365 | if (dump_enabled_p ()) |
78c60e3d | 7366 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 7367 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
7368 | |
7369 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
7370 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
7371 | STMT). */ | |
7372 | if (gimple_code (stmt) == GIMPLE_PHI) | |
7373 | scalar_dest = PHI_RESULT (stmt); | |
7374 | else | |
7375 | scalar_dest = gimple_assign_lhs (stmt); | |
7376 | ||
7377 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
7378 | { | |
7379 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
7380 | { | |
7381 | exit_phi = USE_STMT (use_p); | |
7382 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
7383 | } | |
7384 | } | |
7385 | } | |
7386 | ||
7387 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
7388 | being vectorized. */ | |
7389 | if (STMT_VINFO_LIVE_P (stmt_info) | |
7390 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
7391 | { | |
7392 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
7393 | gcc_assert (done); | |
7394 | } | |
7395 | ||
7396 | if (vec_stmt) | |
83197f37 | 7397 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 7398 | |
b8698a0f | 7399 | return is_store; |
ebfd146a IR |
7400 | } |
7401 | ||
7402 | ||
b8698a0f | 7403 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
7404 | stmt_vec_info. */ |
7405 | ||
7406 | void | |
7407 | vect_remove_stores (gimple first_stmt) | |
7408 | { | |
7409 | gimple next = first_stmt; | |
7410 | gimple tmp; | |
7411 | gimple_stmt_iterator next_si; | |
7412 | ||
7413 | while (next) | |
7414 | { | |
78048b1c JJ |
7415 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
7416 | ||
7417 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
7418 | if (is_pattern_stmt_p (stmt_info)) | |
7419 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
7420 | /* Free the attached stmt_vec_info and remove the stmt. */ |
7421 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 7422 | unlink_stmt_vdef (next); |
ebfd146a | 7423 | gsi_remove (&next_si, true); |
3d3f2249 | 7424 | release_defs (next); |
ebfd146a IR |
7425 | free_stmt_vec_info (next); |
7426 | next = tmp; | |
7427 | } | |
7428 | } | |
7429 | ||
7430 | ||
7431 | /* Function new_stmt_vec_info. | |
7432 | ||
7433 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
7434 | ||
7435 | stmt_vec_info | |
b8698a0f | 7436 | new_stmt_vec_info (gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 7437 | bb_vec_info bb_vinfo) |
ebfd146a IR |
7438 | { |
7439 | stmt_vec_info res; | |
7440 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
7441 | ||
7442 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
7443 | STMT_VINFO_STMT (res) = stmt; | |
7444 | STMT_VINFO_LOOP_VINFO (res) = loop_vinfo; | |
a70d6342 | 7445 | STMT_VINFO_BB_VINFO (res) = bb_vinfo; |
8644a673 | 7446 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
7447 | STMT_VINFO_LIVE_P (res) = false; |
7448 | STMT_VINFO_VECTYPE (res) = NULL; | |
7449 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 7450 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
7451 | STMT_VINFO_IN_PATTERN_P (res) = false; |
7452 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 7453 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a IR |
7454 | STMT_VINFO_DATA_REF (res) = NULL; |
7455 | ||
7456 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
7457 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
7458 | STMT_VINFO_DR_INIT (res) = NULL; | |
7459 | STMT_VINFO_DR_STEP (res) = NULL; | |
7460 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
7461 | ||
7462 | if (gimple_code (stmt) == GIMPLE_PHI | |
7463 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
7464 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
7465 | else | |
8644a673 IR |
7466 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
7467 | ||
9771b263 | 7468 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 7469 | STMT_SLP_TYPE (res) = loop_vect; |
e14c1050 IR |
7470 | GROUP_FIRST_ELEMENT (res) = NULL; |
7471 | GROUP_NEXT_ELEMENT (res) = NULL; | |
7472 | GROUP_SIZE (res) = 0; | |
7473 | GROUP_STORE_COUNT (res) = 0; | |
7474 | GROUP_GAP (res) = 0; | |
7475 | GROUP_SAME_DR_STMT (res) = NULL; | |
ebfd146a IR |
7476 | |
7477 | return res; | |
7478 | } | |
7479 | ||
7480 | ||
7481 | /* Create a hash table for stmt_vec_info. */ | |
7482 | ||
7483 | void | |
7484 | init_stmt_vec_info_vec (void) | |
7485 | { | |
9771b263 DN |
7486 | gcc_assert (!stmt_vec_info_vec.exists ()); |
7487 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
7488 | } |
7489 | ||
7490 | ||
7491 | /* Free hash table for stmt_vec_info. */ | |
7492 | ||
7493 | void | |
7494 | free_stmt_vec_info_vec (void) | |
7495 | { | |
93675444 JJ |
7496 | unsigned int i; |
7497 | vec_void_p info; | |
7498 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) | |
7499 | if (info != NULL) | |
7500 | free_stmt_vec_info (STMT_VINFO_STMT ((stmt_vec_info) info)); | |
9771b263 DN |
7501 | gcc_assert (stmt_vec_info_vec.exists ()); |
7502 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
7503 | } |
7504 | ||
7505 | ||
7506 | /* Free stmt vectorization related info. */ | |
7507 | ||
7508 | void | |
7509 | free_stmt_vec_info (gimple stmt) | |
7510 | { | |
7511 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
7512 | ||
7513 | if (!stmt_info) | |
7514 | return; | |
7515 | ||
78048b1c JJ |
7516 | /* Check if this statement has a related "pattern stmt" |
7517 | (introduced by the vectorizer during the pattern recognition | |
7518 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
7519 | too. */ | |
7520 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
7521 | { | |
7522 | stmt_vec_info patt_info | |
7523 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
7524 | if (patt_info) | |
7525 | { | |
363477c0 | 7526 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
f0281fde RB |
7527 | gimple patt_stmt = STMT_VINFO_STMT (patt_info); |
7528 | gimple_set_bb (patt_stmt, NULL); | |
7529 | tree lhs = gimple_get_lhs (patt_stmt); | |
7530 | if (TREE_CODE (lhs) == SSA_NAME) | |
7531 | release_ssa_name (lhs); | |
363477c0 JJ |
7532 | if (seq) |
7533 | { | |
7534 | gimple_stmt_iterator si; | |
7535 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
f0281fde RB |
7536 | { |
7537 | gimple seq_stmt = gsi_stmt (si); | |
7538 | gimple_set_bb (seq_stmt, NULL); | |
7539 | lhs = gimple_get_lhs (patt_stmt); | |
7540 | if (TREE_CODE (lhs) == SSA_NAME) | |
7541 | release_ssa_name (lhs); | |
7542 | free_stmt_vec_info (seq_stmt); | |
7543 | } | |
363477c0 | 7544 | } |
f0281fde | 7545 | free_stmt_vec_info (patt_stmt); |
78048b1c JJ |
7546 | } |
7547 | } | |
7548 | ||
9771b263 | 7549 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
6c9e85fb | 7550 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).release (); |
ebfd146a IR |
7551 | set_vinfo_for_stmt (stmt, NULL); |
7552 | free (stmt_info); | |
7553 | } | |
7554 | ||
7555 | ||
bb67d9c7 | 7556 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 7557 | |
bb67d9c7 | 7558 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
7559 | by the target. */ |
7560 | ||
bb67d9c7 RG |
7561 | static tree |
7562 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a | 7563 | { |
ef4bddc2 RS |
7564 | machine_mode inner_mode = TYPE_MODE (scalar_type); |
7565 | machine_mode simd_mode; | |
2f816591 | 7566 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
7567 | int nunits; |
7568 | tree vectype; | |
7569 | ||
cc4b5170 | 7570 | if (nbytes == 0) |
ebfd146a IR |
7571 | return NULL_TREE; |
7572 | ||
48f2e373 RB |
7573 | if (GET_MODE_CLASS (inner_mode) != MODE_INT |
7574 | && GET_MODE_CLASS (inner_mode) != MODE_FLOAT) | |
7575 | return NULL_TREE; | |
7576 | ||
7b7b1813 RG |
7577 | /* For vector types of elements whose mode precision doesn't |
7578 | match their types precision we use a element type of mode | |
7579 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
7580 | they support the proper result truncation/extension. |
7581 | We also make sure to build vector types with INTEGER_TYPE | |
7582 | component type only. */ | |
6d7971b8 | 7583 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
7584 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
7585 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
7586 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
7587 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 7588 | |
ccbf5bb4 RG |
7589 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
7590 | When the component mode passes the above test simply use a type | |
7591 | corresponding to that mode. The theory is that any use that | |
7592 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 7593 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 7594 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
7595 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
7596 | ||
7597 | /* We can't build a vector type of elements with alignment bigger than | |
7598 | their size. */ | |
dfc2e2ac | 7599 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
7600 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
7601 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 7602 | |
dfc2e2ac RB |
7603 | /* If we felt back to using the mode fail if there was |
7604 | no scalar type for it. */ | |
7605 | if (scalar_type == NULL_TREE) | |
7606 | return NULL_TREE; | |
7607 | ||
bb67d9c7 RG |
7608 | /* If no size was supplied use the mode the target prefers. Otherwise |
7609 | lookup a vector mode of the specified size. */ | |
7610 | if (size == 0) | |
7611 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
7612 | else | |
7613 | simd_mode = mode_for_vector (inner_mode, size / nbytes); | |
cc4b5170 RG |
7614 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
7615 | if (nunits <= 1) | |
7616 | return NULL_TREE; | |
ebfd146a IR |
7617 | |
7618 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
7619 | |
7620 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
7621 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 7622 | return NULL_TREE; |
ebfd146a IR |
7623 | |
7624 | return vectype; | |
7625 | } | |
7626 | ||
bb67d9c7 RG |
7627 | unsigned int current_vector_size; |
7628 | ||
7629 | /* Function get_vectype_for_scalar_type. | |
7630 | ||
7631 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
7632 | by the target. */ | |
7633 | ||
7634 | tree | |
7635 | get_vectype_for_scalar_type (tree scalar_type) | |
7636 | { | |
7637 | tree vectype; | |
7638 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
7639 | current_vector_size); | |
7640 | if (vectype | |
7641 | && current_vector_size == 0) | |
7642 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
7643 | return vectype; | |
7644 | } | |
7645 | ||
b690cc0f RG |
7646 | /* Function get_same_sized_vectype |
7647 | ||
7648 | Returns a vector type corresponding to SCALAR_TYPE of size | |
7649 | VECTOR_TYPE if supported by the target. */ | |
7650 | ||
7651 | tree | |
bb67d9c7 | 7652 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 7653 | { |
bb67d9c7 RG |
7654 | return get_vectype_for_scalar_type_and_size |
7655 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
7656 | } |
7657 | ||
ebfd146a IR |
7658 | /* Function vect_is_simple_use. |
7659 | ||
7660 | Input: | |
a70d6342 IR |
7661 | LOOP_VINFO - the vect info of the loop that is being vectorized. |
7662 | BB_VINFO - the vect info of the basic block that is being vectorized. | |
24ee1384 | 7663 | OPERAND - operand of STMT in the loop or bb. |
ebfd146a IR |
7664 | DEF - the defining stmt in case OPERAND is an SSA_NAME. |
7665 | ||
7666 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 7667 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 7668 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 7669 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
7670 | is the case in reduction/induction computations). |
7671 | For basic blocks, supportable operands are constants and bb invariants. | |
7672 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
7673 | |
7674 | bool | |
24ee1384 | 7675 | vect_is_simple_use (tree operand, gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 7676 | bb_vec_info bb_vinfo, gimple *def_stmt, |
ebfd146a | 7677 | tree *def, enum vect_def_type *dt) |
b8698a0f | 7678 | { |
ebfd146a IR |
7679 | basic_block bb; |
7680 | stmt_vec_info stmt_vinfo; | |
a70d6342 | 7681 | struct loop *loop = NULL; |
b8698a0f | 7682 | |
a70d6342 IR |
7683 | if (loop_vinfo) |
7684 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
7685 | |
7686 | *def_stmt = NULL; | |
7687 | *def = NULL_TREE; | |
b8698a0f | 7688 | |
73fbfcad | 7689 | if (dump_enabled_p ()) |
ebfd146a | 7690 | { |
78c60e3d SS |
7691 | dump_printf_loc (MSG_NOTE, vect_location, |
7692 | "vect_is_simple_use: operand "); | |
7693 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
e645e942 | 7694 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 7695 | } |
b8698a0f | 7696 | |
b758f602 | 7697 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
7698 | { |
7699 | *dt = vect_constant_def; | |
7700 | return true; | |
7701 | } | |
b8698a0f | 7702 | |
ebfd146a IR |
7703 | if (is_gimple_min_invariant (operand)) |
7704 | { | |
7705 | *def = operand; | |
8644a673 | 7706 | *dt = vect_external_def; |
ebfd146a IR |
7707 | return true; |
7708 | } | |
7709 | ||
7710 | if (TREE_CODE (operand) == PAREN_EXPR) | |
7711 | { | |
73fbfcad | 7712 | if (dump_enabled_p ()) |
e645e942 | 7713 | dump_printf_loc (MSG_NOTE, vect_location, "non-associatable copy.\n"); |
ebfd146a IR |
7714 | operand = TREE_OPERAND (operand, 0); |
7715 | } | |
b8698a0f | 7716 | |
ebfd146a IR |
7717 | if (TREE_CODE (operand) != SSA_NAME) |
7718 | { | |
73fbfcad | 7719 | if (dump_enabled_p ()) |
78c60e3d | 7720 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7721 | "not ssa-name.\n"); |
ebfd146a IR |
7722 | return false; |
7723 | } | |
b8698a0f | 7724 | |
ebfd146a IR |
7725 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
7726 | if (*def_stmt == NULL) | |
7727 | { | |
73fbfcad | 7728 | if (dump_enabled_p ()) |
78c60e3d | 7729 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7730 | "no def_stmt.\n"); |
ebfd146a IR |
7731 | return false; |
7732 | } | |
7733 | ||
73fbfcad | 7734 | if (dump_enabled_p ()) |
ebfd146a | 7735 | { |
78c60e3d SS |
7736 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
7737 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
ebfd146a IR |
7738 | } |
7739 | ||
8644a673 | 7740 | /* Empty stmt is expected only in case of a function argument. |
ebfd146a IR |
7741 | (Otherwise - we expect a phi_node or a GIMPLE_ASSIGN). */ |
7742 | if (gimple_nop_p (*def_stmt)) | |
7743 | { | |
7744 | *def = operand; | |
8644a673 | 7745 | *dt = vect_external_def; |
ebfd146a IR |
7746 | return true; |
7747 | } | |
7748 | ||
7749 | bb = gimple_bb (*def_stmt); | |
a70d6342 IR |
7750 | |
7751 | if ((loop && !flow_bb_inside_loop_p (loop, bb)) | |
7752 | || (!loop && bb != BB_VINFO_BB (bb_vinfo)) | |
b8698a0f | 7753 | || (!loop && gimple_code (*def_stmt) == GIMPLE_PHI)) |
8644a673 | 7754 | *dt = vect_external_def; |
ebfd146a IR |
7755 | else |
7756 | { | |
7757 | stmt_vinfo = vinfo_for_stmt (*def_stmt); | |
7758 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
7759 | } | |
7760 | ||
24ee1384 IR |
7761 | if (*dt == vect_unknown_def_type |
7762 | || (stmt | |
7763 | && *dt == vect_double_reduction_def | |
7764 | && gimple_code (stmt) != GIMPLE_PHI)) | |
ebfd146a | 7765 | { |
73fbfcad | 7766 | if (dump_enabled_p ()) |
78c60e3d | 7767 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7768 | "Unsupported pattern.\n"); |
ebfd146a IR |
7769 | return false; |
7770 | } | |
7771 | ||
73fbfcad | 7772 | if (dump_enabled_p ()) |
e645e942 | 7773 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: %d.\n", *dt); |
ebfd146a IR |
7774 | |
7775 | switch (gimple_code (*def_stmt)) | |
7776 | { | |
7777 | case GIMPLE_PHI: | |
7778 | *def = gimple_phi_result (*def_stmt); | |
7779 | break; | |
7780 | ||
7781 | case GIMPLE_ASSIGN: | |
7782 | *def = gimple_assign_lhs (*def_stmt); | |
7783 | break; | |
7784 | ||
7785 | case GIMPLE_CALL: | |
7786 | *def = gimple_call_lhs (*def_stmt); | |
7787 | if (*def != NULL) | |
7788 | break; | |
7789 | /* FALLTHRU */ | |
7790 | default: | |
73fbfcad | 7791 | if (dump_enabled_p ()) |
78c60e3d | 7792 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7793 | "unsupported defining stmt:\n"); |
ebfd146a IR |
7794 | return false; |
7795 | } | |
7796 | ||
7797 | return true; | |
7798 | } | |
7799 | ||
b690cc0f RG |
7800 | /* Function vect_is_simple_use_1. |
7801 | ||
7802 | Same as vect_is_simple_use_1 but also determines the vector operand | |
7803 | type of OPERAND and stores it to *VECTYPE. If the definition of | |
7804 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
7805 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
7806 | is responsible to compute the best suited vector type for the | |
7807 | scalar operand. */ | |
7808 | ||
7809 | bool | |
24ee1384 | 7810 | vect_is_simple_use_1 (tree operand, gimple stmt, loop_vec_info loop_vinfo, |
b690cc0f RG |
7811 | bb_vec_info bb_vinfo, gimple *def_stmt, |
7812 | tree *def, enum vect_def_type *dt, tree *vectype) | |
7813 | { | |
24ee1384 IR |
7814 | if (!vect_is_simple_use (operand, stmt, loop_vinfo, bb_vinfo, def_stmt, |
7815 | def, dt)) | |
b690cc0f RG |
7816 | return false; |
7817 | ||
7818 | /* Now get a vector type if the def is internal, otherwise supply | |
7819 | NULL_TREE and leave it up to the caller to figure out a proper | |
7820 | type for the use stmt. */ | |
7821 | if (*dt == vect_internal_def | |
7822 | || *dt == vect_induction_def | |
7823 | || *dt == vect_reduction_def | |
7824 | || *dt == vect_double_reduction_def | |
7825 | || *dt == vect_nested_cycle) | |
7826 | { | |
7827 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
7828 | |
7829 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
7830 | && !STMT_VINFO_RELEVANT (stmt_info) | |
7831 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 7832 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 7833 | |
b690cc0f RG |
7834 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
7835 | gcc_assert (*vectype != NULL_TREE); | |
7836 | } | |
7837 | else if (*dt == vect_uninitialized_def | |
7838 | || *dt == vect_constant_def | |
7839 | || *dt == vect_external_def) | |
7840 | *vectype = NULL_TREE; | |
7841 | else | |
7842 | gcc_unreachable (); | |
7843 | ||
7844 | return true; | |
7845 | } | |
7846 | ||
ebfd146a IR |
7847 | |
7848 | /* Function supportable_widening_operation | |
7849 | ||
b8698a0f L |
7850 | Check whether an operation represented by the code CODE is a |
7851 | widening operation that is supported by the target platform in | |
b690cc0f RG |
7852 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
7853 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 7854 | |
ebfd146a IR |
7855 | Widening operations we currently support are NOP (CONVERT), FLOAT |
7856 | and WIDEN_MULT. This function checks if these operations are supported | |
7857 | by the target platform either directly (via vector tree-codes), or via | |
7858 | target builtins. | |
7859 | ||
7860 | Output: | |
b8698a0f L |
7861 | - CODE1 and CODE2 are codes of vector operations to be used when |
7862 | vectorizing the operation, if available. | |
ebfd146a IR |
7863 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
7864 | case of multi-step conversion (like char->short->int - in that case | |
7865 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
7866 | - INTERM_TYPES contains the intermediate type required to perform the |
7867 | widening operation (short in the above example). */ | |
ebfd146a IR |
7868 | |
7869 | bool | |
b690cc0f RG |
7870 | supportable_widening_operation (enum tree_code code, gimple stmt, |
7871 | tree vectype_out, tree vectype_in, | |
ebfd146a IR |
7872 | enum tree_code *code1, enum tree_code *code2, |
7873 | int *multi_step_cvt, | |
9771b263 | 7874 | vec<tree> *interm_types) |
ebfd146a IR |
7875 | { |
7876 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
7877 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 7878 | struct loop *vect_loop = NULL; |
ef4bddc2 | 7879 | machine_mode vec_mode; |
81f40b79 | 7880 | enum insn_code icode1, icode2; |
ebfd146a | 7881 | optab optab1, optab2; |
b690cc0f RG |
7882 | tree vectype = vectype_in; |
7883 | tree wide_vectype = vectype_out; | |
ebfd146a | 7884 | enum tree_code c1, c2; |
4a00c761 JJ |
7885 | int i; |
7886 | tree prev_type, intermediate_type; | |
ef4bddc2 | 7887 | machine_mode intermediate_mode, prev_mode; |
4a00c761 | 7888 | optab optab3, optab4; |
ebfd146a | 7889 | |
4a00c761 | 7890 | *multi_step_cvt = 0; |
4ef69dfc IR |
7891 | if (loop_info) |
7892 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
7893 | ||
ebfd146a IR |
7894 | switch (code) |
7895 | { | |
7896 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
7897 | /* The result of a vectorized widening operation usually requires |
7898 | two vectors (because the widened results do not fit into one vector). | |
7899 | The generated vector results would normally be expected to be | |
7900 | generated in the same order as in the original scalar computation, | |
7901 | i.e. if 8 results are generated in each vector iteration, they are | |
7902 | to be organized as follows: | |
7903 | vect1: [res1,res2,res3,res4], | |
7904 | vect2: [res5,res6,res7,res8]. | |
7905 | ||
7906 | However, in the special case that the result of the widening | |
7907 | operation is used in a reduction computation only, the order doesn't | |
7908 | matter (because when vectorizing a reduction we change the order of | |
7909 | the computation). Some targets can take advantage of this and | |
7910 | generate more efficient code. For example, targets like Altivec, | |
7911 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
7912 | generate the following vectors: | |
7913 | vect1: [res1,res3,res5,res7], | |
7914 | vect2: [res2,res4,res6,res8]. | |
7915 | ||
7916 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
7917 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
7918 | iterations in parallel). We therefore don't allow to change the | |
7919 | order of the computation in the inner-loop during outer-loop | |
7920 | vectorization. */ | |
7921 | /* TODO: Another case in which order doesn't *really* matter is when we | |
7922 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
7923 | Normally, pack_trunc performs an even/odd permute, whereas the | |
7924 | repack from an even/odd expansion would be an interleave, which | |
7925 | would be significantly simpler for e.g. AVX2. */ | |
7926 | /* In any case, in order to avoid duplicating the code below, recurse | |
7927 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
7928 | are properly set up for the caller. If we fail, we'll continue with | |
7929 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
7930 | if (vect_loop | |
7931 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
7932 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
7933 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
7934 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
7935 | code1, code2, multi_step_cvt, |
7936 | interm_types)) | |
ebc047a2 CH |
7937 | { |
7938 | /* Elements in a vector with vect_used_by_reduction property cannot | |
7939 | be reordered if the use chain with this property does not have the | |
7940 | same operation. One such an example is s += a * b, where elements | |
7941 | in a and b cannot be reordered. Here we check if the vector defined | |
7942 | by STMT is only directly used in the reduction statement. */ | |
7943 | tree lhs = gimple_assign_lhs (stmt); | |
7944 | use_operand_p dummy; | |
7945 | gimple use_stmt; | |
7946 | stmt_vec_info use_stmt_info = NULL; | |
7947 | if (single_imm_use (lhs, &dummy, &use_stmt) | |
7948 | && (use_stmt_info = vinfo_for_stmt (use_stmt)) | |
7949 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) | |
7950 | return true; | |
7951 | } | |
4a00c761 JJ |
7952 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
7953 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
7954 | break; |
7955 | ||
6ae6116f RH |
7956 | case VEC_WIDEN_MULT_EVEN_EXPR: |
7957 | /* Support the recursion induced just above. */ | |
7958 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
7959 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
7960 | break; | |
7961 | ||
36ba4aae | 7962 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
7963 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
7964 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
7965 | break; |
7966 | ||
ebfd146a | 7967 | CASE_CONVERT: |
4a00c761 JJ |
7968 | c1 = VEC_UNPACK_LO_EXPR; |
7969 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
7970 | break; |
7971 | ||
7972 | case FLOAT_EXPR: | |
4a00c761 JJ |
7973 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
7974 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
7975 | break; |
7976 | ||
7977 | case FIX_TRUNC_EXPR: | |
7978 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
7979 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
7980 | computing the operation. */ | |
7981 | return false; | |
7982 | ||
7983 | default: | |
7984 | gcc_unreachable (); | |
7985 | } | |
7986 | ||
6ae6116f | 7987 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
4a00c761 JJ |
7988 | { |
7989 | enum tree_code ctmp = c1; | |
7990 | c1 = c2; | |
7991 | c2 = ctmp; | |
7992 | } | |
7993 | ||
ebfd146a IR |
7994 | if (code == FIX_TRUNC_EXPR) |
7995 | { | |
7996 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
7997 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
7998 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
7999 | } |
8000 | else | |
8001 | { | |
8002 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
8003 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
8004 | } | |
8005 | ||
8006 | if (!optab1 || !optab2) | |
8007 | return false; | |
8008 | ||
8009 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
8010 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
8011 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
8012 | return false; |
8013 | ||
4a00c761 JJ |
8014 | *code1 = c1; |
8015 | *code2 = c2; | |
8016 | ||
8017 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
8018 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
8019 | return true; | |
8020 | ||
b8698a0f | 8021 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 8022 | types. */ |
ebfd146a | 8023 | |
4a00c761 JJ |
8024 | prev_type = vectype; |
8025 | prev_mode = vec_mode; | |
b8698a0f | 8026 | |
4a00c761 JJ |
8027 | if (!CONVERT_EXPR_CODE_P (code)) |
8028 | return false; | |
b8698a0f | 8029 | |
4a00c761 JJ |
8030 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
8031 | intermediate steps in promotion sequence. We try | |
8032 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
8033 | not. */ | |
9771b263 | 8034 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
8035 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
8036 | { | |
8037 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
8038 | intermediate_type | |
8039 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
8040 | TYPE_UNSIGNED (prev_type)); | |
8041 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); | |
8042 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
8043 | ||
8044 | if (!optab3 || !optab4 | |
8045 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
8046 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
8047 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
8048 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
8049 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
8050 | == CODE_FOR_nothing) | |
8051 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
8052 | == CODE_FOR_nothing)) | |
8053 | break; | |
ebfd146a | 8054 | |
9771b263 | 8055 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
8056 | (*multi_step_cvt)++; |
8057 | ||
8058 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
8059 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
8060 | return true; | |
8061 | ||
8062 | prev_type = intermediate_type; | |
8063 | prev_mode = intermediate_mode; | |
ebfd146a IR |
8064 | } |
8065 | ||
9771b263 | 8066 | interm_types->release (); |
4a00c761 | 8067 | return false; |
ebfd146a IR |
8068 | } |
8069 | ||
8070 | ||
8071 | /* Function supportable_narrowing_operation | |
8072 | ||
b8698a0f L |
8073 | Check whether an operation represented by the code CODE is a |
8074 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
8075 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
8076 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 8077 | |
ebfd146a | 8078 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 8079 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
8080 | the target platform directly via vector tree-codes. |
8081 | ||
8082 | Output: | |
b8698a0f L |
8083 | - CODE1 is the code of a vector operation to be used when |
8084 | vectorizing the operation, if available. | |
ebfd146a IR |
8085 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
8086 | case of multi-step conversion (like int->short->char - in that case | |
8087 | MULTI_STEP_CVT will be 1). | |
8088 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 8089 | narrowing operation (short in the above example). */ |
ebfd146a IR |
8090 | |
8091 | bool | |
8092 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 8093 | tree vectype_out, tree vectype_in, |
ebfd146a | 8094 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 8095 | vec<tree> *interm_types) |
ebfd146a | 8096 | { |
ef4bddc2 | 8097 | machine_mode vec_mode; |
ebfd146a IR |
8098 | enum insn_code icode1; |
8099 | optab optab1, interm_optab; | |
b690cc0f RG |
8100 | tree vectype = vectype_in; |
8101 | tree narrow_vectype = vectype_out; | |
ebfd146a | 8102 | enum tree_code c1; |
4a00c761 | 8103 | tree intermediate_type; |
ef4bddc2 | 8104 | machine_mode intermediate_mode, prev_mode; |
ebfd146a | 8105 | int i; |
4a00c761 | 8106 | bool uns; |
ebfd146a | 8107 | |
4a00c761 | 8108 | *multi_step_cvt = 0; |
ebfd146a IR |
8109 | switch (code) |
8110 | { | |
8111 | CASE_CONVERT: | |
8112 | c1 = VEC_PACK_TRUNC_EXPR; | |
8113 | break; | |
8114 | ||
8115 | case FIX_TRUNC_EXPR: | |
8116 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
8117 | break; | |
8118 | ||
8119 | case FLOAT_EXPR: | |
8120 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
8121 | tree code and optabs used for computing the operation. */ | |
8122 | return false; | |
8123 | ||
8124 | default: | |
8125 | gcc_unreachable (); | |
8126 | } | |
8127 | ||
8128 | if (code == FIX_TRUNC_EXPR) | |
8129 | /* The signedness is determined from output operand. */ | |
b690cc0f | 8130 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
8131 | else |
8132 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
8133 | ||
8134 | if (!optab1) | |
8135 | return false; | |
8136 | ||
8137 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 8138 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
8139 | return false; |
8140 | ||
4a00c761 JJ |
8141 | *code1 = c1; |
8142 | ||
8143 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
8144 | return true; | |
8145 | ||
ebfd146a IR |
8146 | /* Check if it's a multi-step conversion that can be done using intermediate |
8147 | types. */ | |
4a00c761 JJ |
8148 | prev_mode = vec_mode; |
8149 | if (code == FIX_TRUNC_EXPR) | |
8150 | uns = TYPE_UNSIGNED (vectype_out); | |
8151 | else | |
8152 | uns = TYPE_UNSIGNED (vectype); | |
8153 | ||
8154 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
8155 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
8156 | costly than signed. */ | |
8157 | if (code == FIX_TRUNC_EXPR && uns) | |
8158 | { | |
8159 | enum insn_code icode2; | |
8160 | ||
8161 | intermediate_type | |
8162 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
8163 | interm_optab | |
8164 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 8165 | if (interm_optab != unknown_optab |
4a00c761 JJ |
8166 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
8167 | && insn_data[icode1].operand[0].mode | |
8168 | == insn_data[icode2].operand[0].mode) | |
8169 | { | |
8170 | uns = false; | |
8171 | optab1 = interm_optab; | |
8172 | icode1 = icode2; | |
8173 | } | |
8174 | } | |
ebfd146a | 8175 | |
4a00c761 JJ |
8176 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
8177 | intermediate steps in promotion sequence. We try | |
8178 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 8179 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
8180 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
8181 | { | |
8182 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
8183 | intermediate_type | |
8184 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
8185 | interm_optab | |
8186 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
8187 | optab_default); | |
8188 | if (!interm_optab | |
8189 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
8190 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
8191 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
8192 | == CODE_FOR_nothing)) | |
8193 | break; | |
8194 | ||
9771b263 | 8195 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
8196 | (*multi_step_cvt)++; |
8197 | ||
8198 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
8199 | return true; | |
8200 | ||
8201 | prev_mode = intermediate_mode; | |
8202 | optab1 = interm_optab; | |
ebfd146a IR |
8203 | } |
8204 | ||
9771b263 | 8205 | interm_types->release (); |
4a00c761 | 8206 | return false; |
ebfd146a | 8207 | } |