]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
d1e082c2 | 2 | Copyright (C) 2003-2013 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 IR |
26 | #include "tm.h" |
27 | #include "ggc.h" | |
28 | #include "tree.h" | |
d8a2d370 | 29 | #include "stor-layout.h" |
ebfd146a IR |
30 | #include "target.h" |
31 | #include "basic-block.h" | |
cf835838 | 32 | #include "gimple-pretty-print.h" |
18f429e2 | 33 | #include "gimple.h" |
45b0be94 | 34 | #include "gimplify.h" |
5be5c238 | 35 | #include "gimple-iterator.h" |
18f429e2 | 36 | #include "gimplify-me.h" |
442b4905 AM |
37 | #include "gimple-ssa.h" |
38 | #include "tree-cfg.h" | |
39 | #include "tree-phinodes.h" | |
40 | #include "ssa-iterators.h" | |
d8a2d370 | 41 | #include "stringpool.h" |
442b4905 | 42 | #include "tree-ssanames.h" |
e28030cf | 43 | #include "tree-ssa-loop-manip.h" |
ebfd146a | 44 | #include "cfgloop.h" |
ebfd146a | 45 | #include "expr.h" |
7ee2468b | 46 | #include "recog.h" /* FIXME: for insn_data */ |
ebfd146a | 47 | #include "optabs.h" |
718f9c0f | 48 | #include "diagnostic-core.h" |
ebfd146a | 49 | #include "tree-vectorizer.h" |
7ee2468b | 50 | #include "dumpfile.h" |
ebfd146a | 51 | |
7ee2468b SB |
52 | /* For lang_hooks.types.type_for_mode. */ |
53 | #include "langhooks.h" | |
ebfd146a | 54 | |
c3e7ee41 BS |
55 | /* Return the vectorized type for the given statement. */ |
56 | ||
57 | tree | |
58 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
59 | { | |
60 | return STMT_VINFO_VECTYPE (stmt_info); | |
61 | } | |
62 | ||
63 | /* Return TRUE iff the given statement is in an inner loop relative to | |
64 | the loop being vectorized. */ | |
65 | bool | |
66 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
67 | { | |
68 | gimple stmt = STMT_VINFO_STMT (stmt_info); | |
69 | basic_block bb = gimple_bb (stmt); | |
70 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
71 | struct loop* loop; | |
72 | ||
73 | if (!loop_vinfo) | |
74 | return false; | |
75 | ||
76 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
77 | ||
78 | return (bb->loop_father == loop->inner); | |
79 | } | |
80 | ||
81 | /* Record the cost of a statement, either by directly informing the | |
82 | target model or by saving it in a vector for later processing. | |
83 | Return a preliminary estimate of the statement's cost. */ | |
84 | ||
85 | unsigned | |
92345349 | 86 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 87 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 88 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 89 | { |
92345349 | 90 | if (body_cost_vec) |
c3e7ee41 | 91 | { |
92345349 BS |
92 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
93 | add_stmt_info_to_vec (body_cost_vec, count, kind, | |
94 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
95 | misalign); | |
c3e7ee41 | 96 | return (unsigned) |
92345349 | 97 | (builtin_vectorization_cost (kind, vectype, misalign) * count); |
c3e7ee41 BS |
98 | |
99 | } | |
100 | else | |
101 | { | |
102 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
103 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
104 | void *target_cost_data; | |
105 | ||
106 | if (loop_vinfo) | |
107 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
108 | else | |
109 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
110 | ||
92345349 BS |
111 | return add_stmt_cost (target_cost_data, count, kind, stmt_info, |
112 | misalign, where); | |
c3e7ee41 BS |
113 | } |
114 | } | |
115 | ||
272c6793 RS |
116 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
117 | ||
118 | static tree | |
119 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
120 | { | |
121 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
122 | "vect_array"); | |
123 | } | |
124 | ||
125 | /* ARRAY is an array of vectors created by create_vector_array. | |
126 | Return an SSA_NAME for the vector in index N. The reference | |
127 | is part of the vectorization of STMT and the vector is associated | |
128 | with scalar destination SCALAR_DEST. */ | |
129 | ||
130 | static tree | |
131 | read_vector_array (gimple stmt, gimple_stmt_iterator *gsi, tree scalar_dest, | |
132 | tree array, unsigned HOST_WIDE_INT n) | |
133 | { | |
134 | tree vect_type, vect, vect_name, array_ref; | |
135 | gimple new_stmt; | |
136 | ||
137 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
138 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
139 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
140 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
141 | build_int_cst (size_type_node, n), | |
142 | NULL_TREE, NULL_TREE); | |
143 | ||
144 | new_stmt = gimple_build_assign (vect, array_ref); | |
145 | vect_name = make_ssa_name (vect, new_stmt); | |
146 | gimple_assign_set_lhs (new_stmt, vect_name); | |
147 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
148 | |
149 | return vect_name; | |
150 | } | |
151 | ||
152 | /* ARRAY is an array of vectors created by create_vector_array. | |
153 | Emit code to store SSA_NAME VECT in index N of the array. | |
154 | The store is part of the vectorization of STMT. */ | |
155 | ||
156 | static void | |
157 | write_vector_array (gimple stmt, gimple_stmt_iterator *gsi, tree vect, | |
158 | tree array, unsigned HOST_WIDE_INT n) | |
159 | { | |
160 | tree array_ref; | |
161 | gimple new_stmt; | |
162 | ||
163 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
164 | build_int_cst (size_type_node, n), | |
165 | NULL_TREE, NULL_TREE); | |
166 | ||
167 | new_stmt = gimple_build_assign (array_ref, vect); | |
168 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
169 | } |
170 | ||
171 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
172 | of *PTR. The memory reference replaces those in FIRST_DR | |
173 | (and its group). */ | |
174 | ||
175 | static tree | |
176 | create_array_ref (tree type, tree ptr, struct data_reference *first_dr) | |
177 | { | |
272c6793 RS |
178 | tree mem_ref, alias_ptr_type; |
179 | ||
180 | alias_ptr_type = reference_alias_ptr_type (DR_REF (first_dr)); | |
181 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); | |
182 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 183 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
184 | return mem_ref; |
185 | } | |
186 | ||
ebfd146a IR |
187 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
188 | ||
189 | /* Function vect_mark_relevant. | |
190 | ||
191 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
192 | ||
193 | static void | |
9771b263 | 194 | vect_mark_relevant (vec<gimple> *worklist, gimple stmt, |
83197f37 IR |
195 | enum vect_relevant relevant, bool live_p, |
196 | bool used_in_pattern) | |
ebfd146a IR |
197 | { |
198 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
199 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
200 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
83197f37 | 201 | gimple pattern_stmt; |
ebfd146a | 202 | |
73fbfcad | 203 | if (dump_enabled_p ()) |
78c60e3d | 204 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 205 | "mark relevant %d, live %d.\n", relevant, live_p); |
ebfd146a | 206 | |
83197f37 IR |
207 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
208 | related pattern stmt instead of the original stmt. However, such stmts | |
209 | may have their own uses that are not in any pattern, in such cases the | |
210 | stmt itself should be marked. */ | |
ebfd146a IR |
211 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
212 | { | |
83197f37 IR |
213 | bool found = false; |
214 | if (!used_in_pattern) | |
215 | { | |
216 | imm_use_iterator imm_iter; | |
217 | use_operand_p use_p; | |
218 | gimple use_stmt; | |
219 | tree lhs; | |
13c931c9 JJ |
220 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
221 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a | 222 | |
83197f37 IR |
223 | if (is_gimple_assign (stmt)) |
224 | lhs = gimple_assign_lhs (stmt); | |
225 | else | |
226 | lhs = gimple_call_lhs (stmt); | |
ebfd146a | 227 | |
83197f37 IR |
228 | /* This use is out of pattern use, if LHS has other uses that are |
229 | pattern uses, we should mark the stmt itself, and not the pattern | |
230 | stmt. */ | |
ab0ef706 JJ |
231 | if (TREE_CODE (lhs) == SSA_NAME) |
232 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs) | |
233 | { | |
234 | if (is_gimple_debug (USE_STMT (use_p))) | |
235 | continue; | |
236 | use_stmt = USE_STMT (use_p); | |
237 | ||
13c931c9 JJ |
238 | if (!flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) |
239 | continue; | |
240 | ||
ab0ef706 JJ |
241 | if (vinfo_for_stmt (use_stmt) |
242 | && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (use_stmt))) | |
243 | { | |
244 | found = true; | |
245 | break; | |
246 | } | |
247 | } | |
83197f37 IR |
248 | } |
249 | ||
250 | if (!found) | |
251 | { | |
252 | /* This is the last stmt in a sequence that was detected as a | |
253 | pattern that can potentially be vectorized. Don't mark the stmt | |
254 | as relevant/live because it's not going to be vectorized. | |
255 | Instead mark the pattern-stmt that replaces it. */ | |
256 | ||
257 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
258 | ||
73fbfcad | 259 | if (dump_enabled_p ()) |
78c60e3d SS |
260 | dump_printf_loc (MSG_NOTE, vect_location, |
261 | "last stmt in pattern. don't mark" | |
e645e942 | 262 | " relevant/live.\n"); |
83197f37 IR |
263 | stmt_info = vinfo_for_stmt (pattern_stmt); |
264 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
265 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
266 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
267 | stmt = pattern_stmt; | |
268 | } | |
ebfd146a IR |
269 | } |
270 | ||
271 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
272 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
273 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
274 | ||
275 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
276 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
277 | { | |
73fbfcad | 278 | if (dump_enabled_p ()) |
78c60e3d | 279 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 280 | "already marked relevant/live.\n"); |
ebfd146a IR |
281 | return; |
282 | } | |
283 | ||
9771b263 | 284 | worklist->safe_push (stmt); |
ebfd146a IR |
285 | } |
286 | ||
287 | ||
288 | /* Function vect_stmt_relevant_p. | |
289 | ||
290 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
291 | "relevant for vectorization". | |
292 | ||
293 | A stmt is considered "relevant for vectorization" if: | |
294 | - it has uses outside the loop. | |
295 | - it has vdefs (it alters memory). | |
296 | - control stmts in the loop (except for the exit condition). | |
297 | ||
298 | CHECKME: what other side effects would the vectorizer allow? */ | |
299 | ||
300 | static bool | |
301 | vect_stmt_relevant_p (gimple stmt, loop_vec_info loop_vinfo, | |
302 | enum vect_relevant *relevant, bool *live_p) | |
303 | { | |
304 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
305 | ssa_op_iter op_iter; | |
306 | imm_use_iterator imm_iter; | |
307 | use_operand_p use_p; | |
308 | def_operand_p def_p; | |
309 | ||
8644a673 | 310 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
311 | *live_p = false; |
312 | ||
313 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
314 | if (is_ctrl_stmt (stmt) |
315 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
316 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 317 | *relevant = vect_used_in_scope; |
ebfd146a IR |
318 | |
319 | /* changing memory. */ | |
320 | if (gimple_code (stmt) != GIMPLE_PHI) | |
5006671f | 321 | if (gimple_vdef (stmt)) |
ebfd146a | 322 | { |
73fbfcad | 323 | if (dump_enabled_p ()) |
78c60e3d | 324 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 325 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 326 | *relevant = vect_used_in_scope; |
ebfd146a IR |
327 | } |
328 | ||
329 | /* uses outside the loop. */ | |
330 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
331 | { | |
332 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
333 | { | |
334 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
335 | if (!flow_bb_inside_loop_p (loop, bb)) | |
336 | { | |
73fbfcad | 337 | if (dump_enabled_p ()) |
78c60e3d | 338 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 339 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 340 | |
3157b0c2 AO |
341 | if (is_gimple_debug (USE_STMT (use_p))) |
342 | continue; | |
343 | ||
ebfd146a IR |
344 | /* We expect all such uses to be in the loop exit phis |
345 | (because of loop closed form) */ | |
346 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
347 | gcc_assert (bb == single_exit (loop)->dest); | |
348 | ||
349 | *live_p = true; | |
350 | } | |
351 | } | |
352 | } | |
353 | ||
354 | return (*live_p || *relevant); | |
355 | } | |
356 | ||
357 | ||
b8698a0f | 358 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 359 | |
ff802fa1 | 360 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
361 | used in STMT for anything other than indexing an array. */ |
362 | ||
363 | static bool | |
364 | exist_non_indexing_operands_for_use_p (tree use, gimple stmt) | |
365 | { | |
366 | tree operand; | |
367 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 368 | |
ff802fa1 | 369 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
370 | reference in STMT, then any operand that corresponds to USE |
371 | is not indexing an array. */ | |
372 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
373 | return true; | |
59a05b0c | 374 | |
ebfd146a IR |
375 | /* STMT has a data_ref. FORNOW this means that its of one of |
376 | the following forms: | |
377 | -1- ARRAY_REF = var | |
378 | -2- var = ARRAY_REF | |
379 | (This should have been verified in analyze_data_refs). | |
380 | ||
381 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 382 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
383 | for array indexing. |
384 | ||
385 | Therefore, all we need to check is if STMT falls into the | |
386 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
387 | |
388 | if (!gimple_assign_copy_p (stmt)) | |
389 | return false; | |
59a05b0c EB |
390 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
391 | return false; | |
ebfd146a | 392 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
393 | if (TREE_CODE (operand) != SSA_NAME) |
394 | return false; | |
395 | ||
396 | if (operand == use) | |
397 | return true; | |
398 | ||
399 | return false; | |
400 | } | |
401 | ||
402 | ||
b8698a0f | 403 | /* |
ebfd146a IR |
404 | Function process_use. |
405 | ||
406 | Inputs: | |
407 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b8698a0f | 408 | - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt |
ff802fa1 | 409 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 410 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
411 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
412 | be performed. | |
ebfd146a IR |
413 | |
414 | Outputs: | |
415 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
416 | relevance info of the DEF_STMT of this USE: | |
417 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
418 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
419 | Exceptions: | |
420 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 421 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 422 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
423 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
424 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
425 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
426 | be modified accordingly. | |
427 | ||
428 | Return true if everything is as expected. Return false otherwise. */ | |
429 | ||
430 | static bool | |
b8698a0f | 431 | process_use (gimple stmt, tree use, loop_vec_info loop_vinfo, bool live_p, |
9771b263 | 432 | enum vect_relevant relevant, vec<gimple> *worklist, |
aec7ae7d | 433 | bool force) |
ebfd146a IR |
434 | { |
435 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
436 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
437 | stmt_vec_info dstmt_vinfo; | |
438 | basic_block bb, def_bb; | |
439 | tree def; | |
440 | gimple def_stmt; | |
441 | enum vect_def_type dt; | |
442 | ||
b8698a0f | 443 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 444 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 445 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
446 | return true; |
447 | ||
24ee1384 | 448 | if (!vect_is_simple_use (use, stmt, loop_vinfo, NULL, &def_stmt, &def, &dt)) |
b8698a0f | 449 | { |
73fbfcad | 450 | if (dump_enabled_p ()) |
78c60e3d | 451 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 452 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
453 | return false; |
454 | } | |
455 | ||
456 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
457 | return true; | |
458 | ||
459 | def_bb = gimple_bb (def_stmt); | |
460 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
461 | { | |
73fbfcad | 462 | if (dump_enabled_p ()) |
e645e942 | 463 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n"); |
ebfd146a IR |
464 | return true; |
465 | } | |
466 | ||
b8698a0f L |
467 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
468 | DEF_STMT must have already been processed, because this should be the | |
469 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
470 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
471 | check that everything is as expected, and we are done. */ |
472 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
473 | bb = gimple_bb (stmt); | |
474 | if (gimple_code (stmt) == GIMPLE_PHI | |
475 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
476 | && gimple_code (def_stmt) != GIMPLE_PHI | |
477 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
478 | && bb->loop_father == def_bb->loop_father) | |
479 | { | |
73fbfcad | 480 | if (dump_enabled_p ()) |
78c60e3d | 481 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 482 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a IR |
483 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
484 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
485 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 486 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 487 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
488 | return true; |
489 | } | |
490 | ||
491 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
492 | outer-loop-header-bb: | |
493 | d = def_stmt | |
494 | inner-loop: | |
495 | stmt # use (d) | |
496 | outer-loop-tail-bb: | |
497 | ... */ | |
498 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
499 | { | |
73fbfcad | 500 | if (dump_enabled_p ()) |
78c60e3d | 501 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 502 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 503 | |
ebfd146a IR |
504 | switch (relevant) |
505 | { | |
8644a673 | 506 | case vect_unused_in_scope: |
7c5222ff IR |
507 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
508 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 509 | break; |
7c5222ff | 510 | |
ebfd146a | 511 | case vect_used_in_outer_by_reduction: |
7c5222ff | 512 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
513 | relevant = vect_used_by_reduction; |
514 | break; | |
7c5222ff | 515 | |
ebfd146a | 516 | case vect_used_in_outer: |
7c5222ff | 517 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 518 | relevant = vect_used_in_scope; |
ebfd146a | 519 | break; |
7c5222ff | 520 | |
8644a673 | 521 | case vect_used_in_scope: |
ebfd146a IR |
522 | break; |
523 | ||
524 | default: | |
525 | gcc_unreachable (); | |
b8698a0f | 526 | } |
ebfd146a IR |
527 | } |
528 | ||
529 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
530 | outer-loop-header-bb: | |
531 | ... | |
532 | inner-loop: | |
533 | d = def_stmt | |
06066f92 | 534 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
535 | stmt # use (d) */ |
536 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
537 | { | |
73fbfcad | 538 | if (dump_enabled_p ()) |
78c60e3d | 539 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 540 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 541 | |
ebfd146a IR |
542 | switch (relevant) |
543 | { | |
8644a673 | 544 | case vect_unused_in_scope: |
b8698a0f | 545 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 546 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 547 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
548 | break; |
549 | ||
ebfd146a IR |
550 | case vect_used_by_reduction: |
551 | relevant = vect_used_in_outer_by_reduction; | |
552 | break; | |
553 | ||
8644a673 | 554 | case vect_used_in_scope: |
ebfd146a IR |
555 | relevant = vect_used_in_outer; |
556 | break; | |
557 | ||
558 | default: | |
559 | gcc_unreachable (); | |
560 | } | |
561 | } | |
562 | ||
83197f37 IR |
563 | vect_mark_relevant (worklist, def_stmt, relevant, live_p, |
564 | is_pattern_stmt_p (stmt_vinfo)); | |
ebfd146a IR |
565 | return true; |
566 | } | |
567 | ||
568 | ||
569 | /* Function vect_mark_stmts_to_be_vectorized. | |
570 | ||
571 | Not all stmts in the loop need to be vectorized. For example: | |
572 | ||
573 | for i... | |
574 | for j... | |
575 | 1. T0 = i + j | |
576 | 2. T1 = a[T0] | |
577 | ||
578 | 3. j = j + 1 | |
579 | ||
580 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
581 | addressing of vectorized data-refs are handled differently. | |
582 | ||
583 | This pass detects such stmts. */ | |
584 | ||
585 | bool | |
586 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
587 | { | |
ebfd146a IR |
588 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
589 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
590 | unsigned int nbbs = loop->num_nodes; | |
591 | gimple_stmt_iterator si; | |
592 | gimple stmt; | |
593 | unsigned int i; | |
594 | stmt_vec_info stmt_vinfo; | |
595 | basic_block bb; | |
596 | gimple phi; | |
597 | bool live_p; | |
06066f92 IR |
598 | enum vect_relevant relevant, tmp_relevant; |
599 | enum vect_def_type def_type; | |
ebfd146a | 600 | |
73fbfcad | 601 | if (dump_enabled_p ()) |
78c60e3d | 602 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 603 | "=== vect_mark_stmts_to_be_vectorized ===\n"); |
ebfd146a | 604 | |
07687835 | 605 | stack_vec<gimple, 64> worklist; |
ebfd146a IR |
606 | |
607 | /* 1. Init worklist. */ | |
608 | for (i = 0; i < nbbs; i++) | |
609 | { | |
610 | bb = bbs[i]; | |
611 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 612 | { |
ebfd146a | 613 | phi = gsi_stmt (si); |
73fbfcad | 614 | if (dump_enabled_p ()) |
ebfd146a | 615 | { |
78c60e3d SS |
616 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
617 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
e645e942 | 618 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
619 | } |
620 | ||
621 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 622 | vect_mark_relevant (&worklist, phi, relevant, live_p, false); |
ebfd146a IR |
623 | } |
624 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
625 | { | |
626 | stmt = gsi_stmt (si); | |
73fbfcad | 627 | if (dump_enabled_p ()) |
ebfd146a | 628 | { |
78c60e3d SS |
629 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
630 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 631 | dump_printf (MSG_NOTE, "\n"); |
b8698a0f | 632 | } |
ebfd146a IR |
633 | |
634 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 635 | vect_mark_relevant (&worklist, stmt, relevant, live_p, false); |
ebfd146a IR |
636 | } |
637 | } | |
638 | ||
639 | /* 2. Process_worklist */ | |
9771b263 | 640 | while (worklist.length () > 0) |
ebfd146a IR |
641 | { |
642 | use_operand_p use_p; | |
643 | ssa_op_iter iter; | |
644 | ||
9771b263 | 645 | stmt = worklist.pop (); |
73fbfcad | 646 | if (dump_enabled_p ()) |
ebfd146a | 647 | { |
78c60e3d SS |
648 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
649 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 650 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
651 | } |
652 | ||
b8698a0f L |
653 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
654 | (DEF_STMT) as relevant/irrelevant and live/dead according to the | |
ebfd146a IR |
655 | liveness and relevance properties of STMT. */ |
656 | stmt_vinfo = vinfo_for_stmt (stmt); | |
657 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
658 | live_p = STMT_VINFO_LIVE_P (stmt_vinfo); | |
659 | ||
660 | /* Generally, the liveness and relevance properties of STMT are | |
661 | propagated as is to the DEF_STMTs of its USEs: | |
662 | live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO) | |
663 | relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO) | |
664 | ||
665 | One exception is when STMT has been identified as defining a reduction | |
666 | variable; in this case we set the liveness/relevance as follows: | |
667 | live_p = false | |
668 | relevant = vect_used_by_reduction | |
669 | This is because we distinguish between two kinds of relevant stmts - | |
b8698a0f | 670 | those that are used by a reduction computation, and those that are |
ff802fa1 | 671 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 672 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 673 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 674 | |
06066f92 IR |
675 | def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
676 | tmp_relevant = relevant; | |
677 | switch (def_type) | |
ebfd146a | 678 | { |
06066f92 IR |
679 | case vect_reduction_def: |
680 | switch (tmp_relevant) | |
681 | { | |
682 | case vect_unused_in_scope: | |
683 | relevant = vect_used_by_reduction; | |
684 | break; | |
685 | ||
686 | case vect_used_by_reduction: | |
687 | if (gimple_code (stmt) == GIMPLE_PHI) | |
688 | break; | |
689 | /* fall through */ | |
690 | ||
691 | default: | |
73fbfcad | 692 | if (dump_enabled_p ()) |
78c60e3d | 693 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 694 | "unsupported use of reduction.\n"); |
06066f92 IR |
695 | return false; |
696 | } | |
697 | ||
b8698a0f | 698 | live_p = false; |
06066f92 | 699 | break; |
b8698a0f | 700 | |
06066f92 IR |
701 | case vect_nested_cycle: |
702 | if (tmp_relevant != vect_unused_in_scope | |
703 | && tmp_relevant != vect_used_in_outer_by_reduction | |
704 | && tmp_relevant != vect_used_in_outer) | |
705 | { | |
73fbfcad | 706 | if (dump_enabled_p ()) |
78c60e3d | 707 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 708 | "unsupported use of nested cycle.\n"); |
7c5222ff | 709 | |
06066f92 IR |
710 | return false; |
711 | } | |
7c5222ff | 712 | |
b8698a0f L |
713 | live_p = false; |
714 | break; | |
715 | ||
06066f92 IR |
716 | case vect_double_reduction_def: |
717 | if (tmp_relevant != vect_unused_in_scope | |
718 | && tmp_relevant != vect_used_by_reduction) | |
719 | { | |
73fbfcad | 720 | if (dump_enabled_p ()) |
78c60e3d | 721 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 722 | "unsupported use of double reduction.\n"); |
7c5222ff | 723 | |
7c5222ff | 724 | return false; |
06066f92 IR |
725 | } |
726 | ||
727 | live_p = false; | |
b8698a0f | 728 | break; |
7c5222ff | 729 | |
06066f92 IR |
730 | default: |
731 | break; | |
7c5222ff | 732 | } |
b8698a0f | 733 | |
aec7ae7d | 734 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
735 | { |
736 | /* Pattern statements are not inserted into the code, so | |
737 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
738 | have to scan the RHS or function arguments instead. */ | |
739 | if (is_gimple_assign (stmt)) | |
740 | { | |
69d2aade JJ |
741 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
742 | tree op = gimple_assign_rhs1 (stmt); | |
743 | ||
744 | i = 1; | |
745 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
746 | { | |
747 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
aec7ae7d | 748 | live_p, relevant, &worklist, false) |
69d2aade | 749 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
aec7ae7d | 750 | live_p, relevant, &worklist, false)) |
07687835 | 751 | return false; |
69d2aade JJ |
752 | i = 2; |
753 | } | |
754 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 755 | { |
69d2aade | 756 | op = gimple_op (stmt, i); |
9d5e7640 | 757 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, |
aec7ae7d | 758 | &worklist, false)) |
07687835 | 759 | return false; |
9d5e7640 IR |
760 | } |
761 | } | |
762 | else if (is_gimple_call (stmt)) | |
763 | { | |
764 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
765 | { | |
766 | tree arg = gimple_call_arg (stmt, i); | |
767 | if (!process_use (stmt, arg, loop_vinfo, live_p, relevant, | |
aec7ae7d | 768 | &worklist, false)) |
07687835 | 769 | return false; |
9d5e7640 IR |
770 | } |
771 | } | |
772 | } | |
773 | else | |
774 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
775 | { | |
776 | tree op = USE_FROM_PTR (use_p); | |
777 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, | |
aec7ae7d | 778 | &worklist, false)) |
07687835 | 779 | return false; |
9d5e7640 | 780 | } |
aec7ae7d JJ |
781 | |
782 | if (STMT_VINFO_GATHER_P (stmt_vinfo)) | |
783 | { | |
784 | tree off; | |
785 | tree decl = vect_check_gather (stmt, loop_vinfo, NULL, &off, NULL); | |
786 | gcc_assert (decl); | |
787 | if (!process_use (stmt, off, loop_vinfo, live_p, relevant, | |
788 | &worklist, true)) | |
07687835 | 789 | return false; |
aec7ae7d | 790 | } |
ebfd146a IR |
791 | } /* while worklist */ |
792 | ||
ebfd146a IR |
793 | return true; |
794 | } | |
795 | ||
796 | ||
b8698a0f | 797 | /* Function vect_model_simple_cost. |
ebfd146a | 798 | |
b8698a0f | 799 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
800 | single op. Right now, this does not account for multiple insns that could |
801 | be generated for the single vector op. We will handle that shortly. */ | |
802 | ||
803 | void | |
b8698a0f | 804 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 BS |
805 | enum vect_def_type *dt, |
806 | stmt_vector_for_cost *prologue_cost_vec, | |
807 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
808 | { |
809 | int i; | |
92345349 | 810 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
811 | |
812 | /* The SLP costs were already calculated during SLP tree build. */ | |
813 | if (PURE_SLP_STMT (stmt_info)) | |
814 | return; | |
815 | ||
ebfd146a IR |
816 | /* FORNOW: Assuming maximum 2 args per stmts. */ |
817 | for (i = 0; i < 2; i++) | |
92345349 BS |
818 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
819 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, vector_stmt, | |
820 | stmt_info, 0, vect_prologue); | |
c3e7ee41 BS |
821 | |
822 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
92345349 BS |
823 | inside_cost = record_stmt_cost (body_cost_vec, ncopies, vector_stmt, |
824 | stmt_info, 0, vect_body); | |
c3e7ee41 | 825 | |
73fbfcad | 826 | if (dump_enabled_p ()) |
78c60e3d SS |
827 | dump_printf_loc (MSG_NOTE, vect_location, |
828 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 829 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
830 | } |
831 | ||
832 | ||
8bd37302 BS |
833 | /* Model cost for type demotion and promotion operations. PWR is normally |
834 | zero for single-step promotions and demotions. It will be one if | |
835 | two-step promotion/demotion is required, and so on. Each additional | |
836 | step doubles the number of instructions required. */ | |
837 | ||
838 | static void | |
839 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
840 | enum vect_def_type *dt, int pwr) | |
841 | { | |
842 | int i, tmp; | |
92345349 | 843 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 BS |
844 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
845 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
846 | void *target_cost_data; | |
8bd37302 BS |
847 | |
848 | /* The SLP costs were already calculated during SLP tree build. */ | |
849 | if (PURE_SLP_STMT (stmt_info)) | |
850 | return; | |
851 | ||
c3e7ee41 BS |
852 | if (loop_vinfo) |
853 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
854 | else | |
855 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
856 | ||
8bd37302 BS |
857 | for (i = 0; i < pwr + 1; i++) |
858 | { | |
859 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
860 | (i + 1) : i; | |
c3e7ee41 | 861 | inside_cost += add_stmt_cost (target_cost_data, vect_pow2 (tmp), |
92345349 BS |
862 | vec_promote_demote, stmt_info, 0, |
863 | vect_body); | |
8bd37302 BS |
864 | } |
865 | ||
866 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
867 | for (i = 0; i < 2; i++) | |
92345349 BS |
868 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
869 | prologue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt, | |
870 | stmt_info, 0, vect_prologue); | |
8bd37302 | 871 | |
73fbfcad | 872 | if (dump_enabled_p ()) |
78c60e3d SS |
873 | dump_printf_loc (MSG_NOTE, vect_location, |
874 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 875 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
876 | } |
877 | ||
0d0293ac | 878 | /* Function vect_cost_group_size |
b8698a0f | 879 | |
0d0293ac | 880 | For grouped load or store, return the group_size only if it is the first |
ebfd146a IR |
881 | load or store of a group, else return 1. This ensures that group size is |
882 | only returned once per group. */ | |
883 | ||
884 | static int | |
0d0293ac | 885 | vect_cost_group_size (stmt_vec_info stmt_info) |
ebfd146a | 886 | { |
e14c1050 | 887 | gimple first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a IR |
888 | |
889 | if (first_stmt == STMT_VINFO_STMT (stmt_info)) | |
e14c1050 | 890 | return GROUP_SIZE (stmt_info); |
ebfd146a IR |
891 | |
892 | return 1; | |
893 | } | |
894 | ||
895 | ||
896 | /* Function vect_model_store_cost | |
897 | ||
0d0293ac MM |
898 | Models cost for stores. In the case of grouped accesses, one access |
899 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a IR |
900 | |
901 | void | |
b8698a0f | 902 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
272c6793 | 903 | bool store_lanes_p, enum vect_def_type dt, |
92345349 BS |
904 | slp_tree slp_node, |
905 | stmt_vector_for_cost *prologue_cost_vec, | |
906 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
907 | { |
908 | int group_size; | |
92345349 | 909 | unsigned int inside_cost = 0, prologue_cost = 0; |
720f5239 IR |
910 | struct data_reference *first_dr; |
911 | gimple first_stmt; | |
ebfd146a IR |
912 | |
913 | /* The SLP costs were already calculated during SLP tree build. */ | |
914 | if (PURE_SLP_STMT (stmt_info)) | |
915 | return; | |
916 | ||
8644a673 | 917 | if (dt == vect_constant_def || dt == vect_external_def) |
92345349 BS |
918 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
919 | stmt_info, 0, vect_prologue); | |
ebfd146a | 920 | |
0d0293ac MM |
921 | /* Grouped access? */ |
922 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
720f5239 IR |
923 | { |
924 | if (slp_node) | |
925 | { | |
9771b263 | 926 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
720f5239 IR |
927 | group_size = 1; |
928 | } | |
929 | else | |
930 | { | |
e14c1050 | 931 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 932 | group_size = vect_cost_group_size (stmt_info); |
720f5239 IR |
933 | } |
934 | ||
935 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
936 | } | |
0d0293ac | 937 | /* Not a grouped access. */ |
ebfd146a | 938 | else |
720f5239 IR |
939 | { |
940 | group_size = 1; | |
941 | first_dr = STMT_VINFO_DATA_REF (stmt_info); | |
942 | } | |
ebfd146a | 943 | |
272c6793 | 944 | /* We assume that the cost of a single store-lanes instruction is |
0d0293ac | 945 | equivalent to the cost of GROUP_SIZE separate stores. If a grouped |
272c6793 RS |
946 | access is instead being provided by a permute-and-store operation, |
947 | include the cost of the permutes. */ | |
948 | if (!store_lanes_p && group_size > 1) | |
ebfd146a IR |
949 | { |
950 | /* Uses a high and low interleave operation for each needed permute. */ | |
c3e7ee41 BS |
951 | |
952 | int nstmts = ncopies * exact_log2 (group_size) * group_size; | |
92345349 BS |
953 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
954 | stmt_info, 0, vect_body); | |
ebfd146a | 955 | |
73fbfcad | 956 | if (dump_enabled_p ()) |
78c60e3d | 957 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 958 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 959 | group_size); |
ebfd146a IR |
960 | } |
961 | ||
962 | /* Costs of the stores. */ | |
92345349 | 963 | vect_get_store_cost (first_dr, ncopies, &inside_cost, body_cost_vec); |
ebfd146a | 964 | |
73fbfcad | 965 | if (dump_enabled_p ()) |
78c60e3d SS |
966 | dump_printf_loc (MSG_NOTE, vect_location, |
967 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 968 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
969 | } |
970 | ||
971 | ||
720f5239 IR |
972 | /* Calculate cost of DR's memory access. */ |
973 | void | |
974 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 975 | unsigned int *inside_cost, |
92345349 | 976 | stmt_vector_for_cost *body_cost_vec) |
720f5239 IR |
977 | { |
978 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
c3e7ee41 BS |
979 | gimple stmt = DR_STMT (dr); |
980 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
720f5239 IR |
981 | |
982 | switch (alignment_support_scheme) | |
983 | { | |
984 | case dr_aligned: | |
985 | { | |
92345349 BS |
986 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
987 | vector_store, stmt_info, 0, | |
988 | vect_body); | |
720f5239 | 989 | |
73fbfcad | 990 | if (dump_enabled_p ()) |
78c60e3d | 991 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 992 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
993 | break; |
994 | } | |
995 | ||
996 | case dr_unaligned_supported: | |
997 | { | |
720f5239 | 998 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 999 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1000 | unaligned_store, stmt_info, |
92345349 | 1001 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1002 | if (dump_enabled_p ()) |
78c60e3d SS |
1003 | dump_printf_loc (MSG_NOTE, vect_location, |
1004 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1005 | "hardware.\n"); |
720f5239 IR |
1006 | break; |
1007 | } | |
1008 | ||
38eec4c6 UW |
1009 | case dr_unaligned_unsupported: |
1010 | { | |
1011 | *inside_cost = VECT_MAX_COST; | |
1012 | ||
73fbfcad | 1013 | if (dump_enabled_p ()) |
78c60e3d | 1014 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1015 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1016 | break; |
1017 | } | |
1018 | ||
720f5239 IR |
1019 | default: |
1020 | gcc_unreachable (); | |
1021 | } | |
1022 | } | |
1023 | ||
1024 | ||
ebfd146a IR |
1025 | /* Function vect_model_load_cost |
1026 | ||
0d0293ac MM |
1027 | Models cost for loads. In the case of grouped accesses, the last access |
1028 | has the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1029 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1030 | access scheme chosen. */ |
1031 | ||
1032 | void | |
92345349 BS |
1033 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, |
1034 | bool load_lanes_p, slp_tree slp_node, | |
1035 | stmt_vector_for_cost *prologue_cost_vec, | |
1036 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
1037 | { |
1038 | int group_size; | |
ebfd146a IR |
1039 | gimple first_stmt; |
1040 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; | |
92345349 | 1041 | unsigned int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
1042 | |
1043 | /* The SLP costs were already calculated during SLP tree build. */ | |
1044 | if (PURE_SLP_STMT (stmt_info)) | |
1045 | return; | |
1046 | ||
0d0293ac | 1047 | /* Grouped accesses? */ |
e14c1050 | 1048 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 1049 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && first_stmt && !slp_node) |
ebfd146a | 1050 | { |
0d0293ac | 1051 | group_size = vect_cost_group_size (stmt_info); |
ebfd146a IR |
1052 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
1053 | } | |
0d0293ac | 1054 | /* Not a grouped access. */ |
ebfd146a IR |
1055 | else |
1056 | { | |
1057 | group_size = 1; | |
1058 | first_dr = dr; | |
1059 | } | |
1060 | ||
272c6793 | 1061 | /* We assume that the cost of a single load-lanes instruction is |
0d0293ac | 1062 | equivalent to the cost of GROUP_SIZE separate loads. If a grouped |
272c6793 RS |
1063 | access is instead being provided by a load-and-permute operation, |
1064 | include the cost of the permutes. */ | |
1065 | if (!load_lanes_p && group_size > 1) | |
ebfd146a IR |
1066 | { |
1067 | /* Uses an even and odd extract operations for each needed permute. */ | |
c3e7ee41 | 1068 | int nstmts = ncopies * exact_log2 (group_size) * group_size; |
92345349 BS |
1069 | inside_cost += record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
1070 | stmt_info, 0, vect_body); | |
ebfd146a | 1071 | |
73fbfcad | 1072 | if (dump_enabled_p ()) |
e645e942 TJ |
1073 | dump_printf_loc (MSG_NOTE, vect_location, |
1074 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1075 | group_size); |
ebfd146a IR |
1076 | } |
1077 | ||
1078 | /* The loads themselves. */ | |
a82960aa RG |
1079 | if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
1080 | { | |
a21892ad BS |
1081 | /* N scalar loads plus gathering them into a vector. */ |
1082 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
92345349 | 1083 | inside_cost += record_stmt_cost (body_cost_vec, |
c3e7ee41 | 1084 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), |
92345349 BS |
1085 | scalar_load, stmt_info, 0, vect_body); |
1086 | inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct, | |
1087 | stmt_info, 0, vect_body); | |
a82960aa RG |
1088 | } |
1089 | else | |
1090 | vect_get_load_cost (first_dr, ncopies, | |
1091 | ((!STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1092 | || group_size > 1 || slp_node), | |
92345349 BS |
1093 | &inside_cost, &prologue_cost, |
1094 | prologue_cost_vec, body_cost_vec, true); | |
720f5239 | 1095 | |
73fbfcad | 1096 | if (dump_enabled_p ()) |
78c60e3d SS |
1097 | dump_printf_loc (MSG_NOTE, vect_location, |
1098 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1099 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1100 | } |
1101 | ||
1102 | ||
1103 | /* Calculate cost of DR's memory access. */ | |
1104 | void | |
1105 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1106 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1107 | unsigned int *prologue_cost, |
1108 | stmt_vector_for_cost *prologue_cost_vec, | |
1109 | stmt_vector_for_cost *body_cost_vec, | |
1110 | bool record_prologue_costs) | |
720f5239 IR |
1111 | { |
1112 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
c3e7ee41 BS |
1113 | gimple stmt = DR_STMT (dr); |
1114 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
720f5239 IR |
1115 | |
1116 | switch (alignment_support_scheme) | |
ebfd146a IR |
1117 | { |
1118 | case dr_aligned: | |
1119 | { | |
92345349 BS |
1120 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1121 | stmt_info, 0, vect_body); | |
ebfd146a | 1122 | |
73fbfcad | 1123 | if (dump_enabled_p ()) |
78c60e3d | 1124 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1125 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1126 | |
1127 | break; | |
1128 | } | |
1129 | case dr_unaligned_supported: | |
1130 | { | |
720f5239 | 1131 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1132 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1133 | unaligned_load, stmt_info, |
92345349 | 1134 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1135 | |
73fbfcad | 1136 | if (dump_enabled_p ()) |
78c60e3d SS |
1137 | dump_printf_loc (MSG_NOTE, vect_location, |
1138 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1139 | "hardware.\n"); |
ebfd146a IR |
1140 | |
1141 | break; | |
1142 | } | |
1143 | case dr_explicit_realign: | |
1144 | { | |
92345349 BS |
1145 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1146 | vector_load, stmt_info, 0, vect_body); | |
1147 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1148 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1149 | |
1150 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1151 | the containing loop, the following cost should be added to the | |
92345349 | 1152 | prologue costs. */ |
ebfd146a | 1153 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1154 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1155 | stmt_info, 0, vect_body); | |
ebfd146a | 1156 | |
73fbfcad | 1157 | if (dump_enabled_p ()) |
e645e942 TJ |
1158 | dump_printf_loc (MSG_NOTE, vect_location, |
1159 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1160 | |
ebfd146a IR |
1161 | break; |
1162 | } | |
1163 | case dr_explicit_realign_optimized: | |
1164 | { | |
73fbfcad | 1165 | if (dump_enabled_p ()) |
e645e942 | 1166 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1167 | "vect_model_load_cost: unaligned software " |
e645e942 | 1168 | "pipelined.\n"); |
ebfd146a IR |
1169 | |
1170 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1171 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1172 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1173 | access, then the above cost should only be considered for one |
ff802fa1 | 1174 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1175 | and a realignment op. */ |
1176 | ||
92345349 | 1177 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1178 | { |
92345349 BS |
1179 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1180 | vector_stmt, stmt_info, | |
1181 | 0, vect_prologue); | |
ebfd146a | 1182 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1183 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1184 | vector_stmt, stmt_info, | |
1185 | 0, vect_prologue); | |
ebfd146a IR |
1186 | } |
1187 | ||
92345349 BS |
1188 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1189 | stmt_info, 0, vect_body); | |
1190 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1191 | stmt_info, 0, vect_body); | |
8bd37302 | 1192 | |
73fbfcad | 1193 | if (dump_enabled_p ()) |
78c60e3d | 1194 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1195 | "vect_model_load_cost: explicit realign optimized" |
1196 | "\n"); | |
8bd37302 | 1197 | |
ebfd146a IR |
1198 | break; |
1199 | } | |
1200 | ||
38eec4c6 UW |
1201 | case dr_unaligned_unsupported: |
1202 | { | |
1203 | *inside_cost = VECT_MAX_COST; | |
1204 | ||
73fbfcad | 1205 | if (dump_enabled_p ()) |
78c60e3d | 1206 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1207 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1208 | break; |
1209 | } | |
1210 | ||
ebfd146a IR |
1211 | default: |
1212 | gcc_unreachable (); | |
1213 | } | |
ebfd146a IR |
1214 | } |
1215 | ||
418b7df3 RG |
1216 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1217 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1218 | |
418b7df3 RG |
1219 | static void |
1220 | vect_init_vector_1 (gimple stmt, gimple new_stmt, gimple_stmt_iterator *gsi) | |
ebfd146a | 1221 | { |
ebfd146a | 1222 | if (gsi) |
418b7df3 | 1223 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1224 | else |
1225 | { | |
418b7df3 | 1226 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1227 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1228 | |
a70d6342 IR |
1229 | if (loop_vinfo) |
1230 | { | |
1231 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1232 | basic_block new_bb; |
1233 | edge pe; | |
a70d6342 IR |
1234 | |
1235 | if (nested_in_vect_loop_p (loop, stmt)) | |
1236 | loop = loop->inner; | |
b8698a0f | 1237 | |
a70d6342 | 1238 | pe = loop_preheader_edge (loop); |
418b7df3 | 1239 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1240 | gcc_assert (!new_bb); |
1241 | } | |
1242 | else | |
1243 | { | |
1244 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1245 | basic_block bb; | |
1246 | gimple_stmt_iterator gsi_bb_start; | |
1247 | ||
1248 | gcc_assert (bb_vinfo); | |
1249 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1250 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1251 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1252 | } |
ebfd146a IR |
1253 | } |
1254 | ||
73fbfcad | 1255 | if (dump_enabled_p ()) |
ebfd146a | 1256 | { |
78c60e3d SS |
1257 | dump_printf_loc (MSG_NOTE, vect_location, |
1258 | "created new init_stmt: "); | |
1259 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
e645e942 | 1260 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 1261 | } |
418b7df3 RG |
1262 | } |
1263 | ||
1264 | /* Function vect_init_vector. | |
ebfd146a | 1265 | |
5467ee52 RG |
1266 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1267 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1268 | vector type a vector with all elements equal to VAL is created first. | |
1269 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1270 | initialization at the loop preheader. | |
418b7df3 RG |
1271 | Return the DEF of INIT_STMT. |
1272 | It will be used in the vectorization of STMT. */ | |
1273 | ||
1274 | tree | |
5467ee52 | 1275 | vect_init_vector (gimple stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 RG |
1276 | { |
1277 | tree new_var; | |
1278 | gimple init_stmt; | |
1279 | tree vec_oprnd; | |
1280 | tree new_temp; | |
1281 | ||
5467ee52 RG |
1282 | if (TREE_CODE (type) == VECTOR_TYPE |
1283 | && TREE_CODE (TREE_TYPE (val)) != VECTOR_TYPE) | |
418b7df3 | 1284 | { |
5467ee52 | 1285 | if (!types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) |
418b7df3 | 1286 | { |
5467ee52 RG |
1287 | if (CONSTANT_CLASS_P (val)) |
1288 | val = fold_unary (VIEW_CONVERT_EXPR, TREE_TYPE (type), val); | |
418b7df3 RG |
1289 | else |
1290 | { | |
83d5977e | 1291 | new_temp = make_ssa_name (TREE_TYPE (type), NULL); |
418b7df3 | 1292 | init_stmt = gimple_build_assign_with_ops (NOP_EXPR, |
83d5977e | 1293 | new_temp, val, |
418b7df3 | 1294 | NULL_TREE); |
418b7df3 | 1295 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1296 | val = new_temp; |
418b7df3 RG |
1297 | } |
1298 | } | |
5467ee52 | 1299 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1300 | } |
1301 | ||
5467ee52 | 1302 | new_var = vect_get_new_vect_var (type, vect_simple_var, "cst_"); |
5467ee52 | 1303 | init_stmt = gimple_build_assign (new_var, val); |
418b7df3 RG |
1304 | new_temp = make_ssa_name (new_var, init_stmt); |
1305 | gimple_assign_set_lhs (init_stmt, new_temp); | |
1306 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
ebfd146a IR |
1307 | vec_oprnd = gimple_assign_lhs (init_stmt); |
1308 | return vec_oprnd; | |
1309 | } | |
1310 | ||
a70d6342 | 1311 | |
ebfd146a IR |
1312 | /* Function vect_get_vec_def_for_operand. |
1313 | ||
ff802fa1 | 1314 | OP is an operand in STMT. This function returns a (vector) def that will be |
ebfd146a IR |
1315 | used in the vectorized stmt for STMT. |
1316 | ||
1317 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1318 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1319 | ||
1320 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1321 | needs to be introduced. */ | |
1322 | ||
1323 | tree | |
1324 | vect_get_vec_def_for_operand (tree op, gimple stmt, tree *scalar_def) | |
1325 | { | |
1326 | tree vec_oprnd; | |
1327 | gimple vec_stmt; | |
1328 | gimple def_stmt; | |
1329 | stmt_vec_info def_stmt_info = NULL; | |
1330 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
9dc3f7de | 1331 | unsigned int nunits; |
ebfd146a | 1332 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
ebfd146a | 1333 | tree def; |
ebfd146a IR |
1334 | enum vect_def_type dt; |
1335 | bool is_simple_use; | |
1336 | tree vector_type; | |
1337 | ||
73fbfcad | 1338 | if (dump_enabled_p ()) |
ebfd146a | 1339 | { |
78c60e3d SS |
1340 | dump_printf_loc (MSG_NOTE, vect_location, |
1341 | "vect_get_vec_def_for_operand: "); | |
1342 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
e645e942 | 1343 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1344 | } |
1345 | ||
24ee1384 IR |
1346 | is_simple_use = vect_is_simple_use (op, stmt, loop_vinfo, NULL, |
1347 | &def_stmt, &def, &dt); | |
ebfd146a | 1348 | gcc_assert (is_simple_use); |
73fbfcad | 1349 | if (dump_enabled_p ()) |
ebfd146a | 1350 | { |
78c60e3d | 1351 | int loc_printed = 0; |
ebfd146a IR |
1352 | if (def) |
1353 | { | |
78c60e3d SS |
1354 | dump_printf_loc (MSG_NOTE, vect_location, "def = "); |
1355 | loc_printed = 1; | |
1356 | dump_generic_expr (MSG_NOTE, TDF_SLIM, def); | |
e645e942 | 1357 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1358 | } |
1359 | if (def_stmt) | |
1360 | { | |
78c60e3d SS |
1361 | if (loc_printed) |
1362 | dump_printf (MSG_NOTE, " def_stmt = "); | |
1363 | else | |
1364 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1365 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
e645e942 | 1366 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1367 | } |
1368 | } | |
1369 | ||
1370 | switch (dt) | |
1371 | { | |
1372 | /* Case 1: operand is a constant. */ | |
1373 | case vect_constant_def: | |
1374 | { | |
7569a6cc RG |
1375 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); |
1376 | gcc_assert (vector_type); | |
9dc3f7de | 1377 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); |
7569a6cc | 1378 | |
b8698a0f | 1379 | if (scalar_def) |
ebfd146a IR |
1380 | *scalar_def = op; |
1381 | ||
1382 | /* Create 'vect_cst_ = {cst,cst,...,cst}' */ | |
73fbfcad | 1383 | if (dump_enabled_p ()) |
78c60e3d | 1384 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1385 | "Create vector_cst. nunits = %d\n", nunits); |
ebfd146a | 1386 | |
418b7df3 | 1387 | return vect_init_vector (stmt, op, vector_type, NULL); |
ebfd146a IR |
1388 | } |
1389 | ||
1390 | /* Case 2: operand is defined outside the loop - loop invariant. */ | |
8644a673 | 1391 | case vect_external_def: |
ebfd146a IR |
1392 | { |
1393 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (def)); | |
1394 | gcc_assert (vector_type); | |
ebfd146a | 1395 | |
b8698a0f | 1396 | if (scalar_def) |
ebfd146a IR |
1397 | *scalar_def = def; |
1398 | ||
1399 | /* Create 'vec_inv = {inv,inv,..,inv}' */ | |
73fbfcad | 1400 | if (dump_enabled_p ()) |
e645e942 | 1401 | dump_printf_loc (MSG_NOTE, vect_location, "Create vector_inv.\n"); |
ebfd146a | 1402 | |
418b7df3 | 1403 | return vect_init_vector (stmt, def, vector_type, NULL); |
ebfd146a IR |
1404 | } |
1405 | ||
1406 | /* Case 3: operand is defined inside the loop. */ | |
8644a673 | 1407 | case vect_internal_def: |
ebfd146a | 1408 | { |
b8698a0f | 1409 | if (scalar_def) |
ebfd146a IR |
1410 | *scalar_def = NULL/* FIXME tuples: def_stmt*/; |
1411 | ||
1412 | /* Get the def from the vectorized stmt. */ | |
1413 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1414 | |
ebfd146a | 1415 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1416 | /* Get vectorized pattern statement. */ |
1417 | if (!vec_stmt | |
1418 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1419 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1420 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1421 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1422 | gcc_assert (vec_stmt); |
1423 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1424 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1425 | else if (is_gimple_call (vec_stmt)) | |
1426 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1427 | else | |
1428 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1429 | return vec_oprnd; | |
1430 | } | |
1431 | ||
1432 | /* Case 4: operand is defined by a loop header phi - reduction */ | |
1433 | case vect_reduction_def: | |
06066f92 | 1434 | case vect_double_reduction_def: |
7c5222ff | 1435 | case vect_nested_cycle: |
ebfd146a IR |
1436 | { |
1437 | struct loop *loop; | |
1438 | ||
1439 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
b8698a0f | 1440 | loop = (gimple_bb (def_stmt))->loop_father; |
ebfd146a IR |
1441 | |
1442 | /* Get the def before the loop */ | |
1443 | op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop)); | |
1444 | return get_initial_def_for_reduction (stmt, op, scalar_def); | |
1445 | } | |
1446 | ||
1447 | /* Case 5: operand is defined by loop-header phi - induction. */ | |
1448 | case vect_induction_def: | |
1449 | { | |
1450 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1451 | ||
1452 | /* Get the def from the vectorized stmt. */ | |
1453 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1454 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1455 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1456 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1457 | else | |
1458 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1459 | return vec_oprnd; |
1460 | } | |
1461 | ||
1462 | default: | |
1463 | gcc_unreachable (); | |
1464 | } | |
1465 | } | |
1466 | ||
1467 | ||
1468 | /* Function vect_get_vec_def_for_stmt_copy | |
1469 | ||
ff802fa1 | 1470 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1471 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1472 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1473 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1474 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1475 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1476 | DT is the type of the vector def VEC_OPRND. |
1477 | ||
1478 | Context: | |
1479 | In case the vectorization factor (VF) is bigger than the number | |
1480 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1481 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1482 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1483 | smallest data-type determines the VF, and as a result, when vectorizing |
1484 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1485 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1486 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1487 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1488 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1489 | ||
1490 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1491 | |
ebfd146a IR |
1492 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1493 | VS1.1: vx.1 = memref1 VS1.2 | |
1494 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1495 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1496 | |
1497 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1498 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1499 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1500 | VSnew.3: vz3 = vx.3 + ... | |
1501 | ||
1502 | The vectorization of S1 is explained in vectorizable_load. | |
1503 | The vectorization of S2: | |
b8698a0f L |
1504 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1505 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1506 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1507 | returns the vector-def 'vx.0'. |
1508 | ||
b8698a0f L |
1509 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1510 | function is called to get the relevant vector-def for each operand. It is | |
1511 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1512 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1513 | ||
b8698a0f L |
1514 | For example, to obtain the vector-def 'vx.1' in order to create the |
1515 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1516 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1517 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1518 | and return its def ('vx.1'). | |
1519 | Overall, to create the above sequence this function will be called 3 times: | |
1520 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1521 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1522 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1523 | ||
1524 | tree | |
1525 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1526 | { | |
1527 | gimple vec_stmt_for_operand; | |
1528 | stmt_vec_info def_stmt_info; | |
1529 | ||
1530 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1531 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1532 | return vec_oprnd; |
1533 | ||
1534 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1535 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1536 | gcc_assert (def_stmt_info); | |
1537 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1538 | gcc_assert (vec_stmt_for_operand); | |
1539 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1540 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) | |
1541 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1542 | else | |
1543 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1544 | return vec_oprnd; | |
1545 | } | |
1546 | ||
1547 | ||
1548 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1549 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a IR |
1550 | |
1551 | static void | |
b8698a0f | 1552 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1553 | vec<tree> *vec_oprnds0, |
1554 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1555 | { |
9771b263 | 1556 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1557 | |
1558 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1559 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1560 | |
9771b263 | 1561 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1562 | { |
9771b263 | 1563 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1564 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1565 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1566 | } |
1567 | } | |
1568 | ||
1569 | ||
d092494c IR |
1570 | /* Get vectorized definitions for OP0 and OP1. |
1571 | REDUC_INDEX is the index of reduction operand in case of reduction, | |
1572 | and -1 otherwise. */ | |
ebfd146a | 1573 | |
d092494c | 1574 | void |
ebfd146a | 1575 | vect_get_vec_defs (tree op0, tree op1, gimple stmt, |
9771b263 DN |
1576 | vec<tree> *vec_oprnds0, |
1577 | vec<tree> *vec_oprnds1, | |
d092494c | 1578 | slp_tree slp_node, int reduc_index) |
ebfd146a IR |
1579 | { |
1580 | if (slp_node) | |
d092494c IR |
1581 | { |
1582 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
9771b263 DN |
1583 | vec<tree> ops; |
1584 | ops.create (nops); | |
37b5ec8f | 1585 | vec<vec<tree> > vec_defs; |
9771b263 | 1586 | vec_defs.create (nops); |
d092494c | 1587 | |
9771b263 | 1588 | ops.quick_push (op0); |
d092494c | 1589 | if (op1) |
9771b263 | 1590 | ops.quick_push (op1); |
d092494c IR |
1591 | |
1592 | vect_get_slp_defs (ops, slp_node, &vec_defs, reduc_index); | |
1593 | ||
37b5ec8f | 1594 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1595 | if (op1) |
37b5ec8f | 1596 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1597 | |
9771b263 DN |
1598 | ops.release (); |
1599 | vec_defs.release (); | |
d092494c | 1600 | } |
ebfd146a IR |
1601 | else |
1602 | { | |
1603 | tree vec_oprnd; | |
1604 | ||
9771b263 | 1605 | vec_oprnds0->create (1); |
b8698a0f | 1606 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt, NULL); |
9771b263 | 1607 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1608 | |
1609 | if (op1) | |
1610 | { | |
9771b263 | 1611 | vec_oprnds1->create (1); |
b8698a0f | 1612 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt, NULL); |
9771b263 | 1613 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1614 | } |
1615 | } | |
1616 | } | |
1617 | ||
1618 | ||
1619 | /* Function vect_finish_stmt_generation. | |
1620 | ||
1621 | Insert a new stmt. */ | |
1622 | ||
1623 | void | |
1624 | vect_finish_stmt_generation (gimple stmt, gimple vec_stmt, | |
1625 | gimple_stmt_iterator *gsi) | |
1626 | { | |
1627 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1628 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 1629 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
ebfd146a IR |
1630 | |
1631 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1632 | ||
54e8e2c3 RG |
1633 | if (!gsi_end_p (*gsi) |
1634 | && gimple_has_mem_ops (vec_stmt)) | |
1635 | { | |
1636 | gimple at_stmt = gsi_stmt (*gsi); | |
1637 | tree vuse = gimple_vuse (at_stmt); | |
1638 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1639 | { | |
1640 | tree vdef = gimple_vdef (at_stmt); | |
1641 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1642 | /* If we have an SSA vuse and insert a store, update virtual | |
1643 | SSA form to avoid triggering the renamer. Do so only | |
1644 | if we can easily see all uses - which is what almost always | |
1645 | happens with the way vectorized stmts are inserted. */ | |
1646 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1647 | && ((is_gimple_assign (vec_stmt) | |
1648 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1649 | || (is_gimple_call (vec_stmt) | |
1650 | && !(gimple_call_flags (vec_stmt) | |
1651 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1652 | { | |
1653 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1654 | gimple_set_vdef (vec_stmt, new_vdef); | |
1655 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1656 | } | |
1657 | } | |
1658 | } | |
ebfd146a IR |
1659 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1660 | ||
b8698a0f | 1661 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, loop_vinfo, |
a70d6342 | 1662 | bb_vinfo)); |
ebfd146a | 1663 | |
73fbfcad | 1664 | if (dump_enabled_p ()) |
ebfd146a | 1665 | { |
78c60e3d SS |
1666 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); |
1667 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
e645e942 | 1668 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1669 | } |
1670 | ||
ad885386 | 1671 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
ebfd146a IR |
1672 | } |
1673 | ||
1674 | /* Checks if CALL can be vectorized in type VECTYPE. Returns | |
1675 | a function declaration if the target has a vectorized version | |
1676 | of the function, or NULL_TREE if the function cannot be vectorized. */ | |
1677 | ||
1678 | tree | |
1679 | vectorizable_function (gimple call, tree vectype_out, tree vectype_in) | |
1680 | { | |
1681 | tree fndecl = gimple_call_fndecl (call); | |
ebfd146a IR |
1682 | |
1683 | /* We only handle functions that do not read or clobber memory -- i.e. | |
1684 | const or novops ones. */ | |
1685 | if (!(gimple_call_flags (call) & (ECF_CONST | ECF_NOVOPS))) | |
1686 | return NULL_TREE; | |
1687 | ||
1688 | if (!fndecl | |
1689 | || TREE_CODE (fndecl) != FUNCTION_DECL | |
1690 | || !DECL_BUILT_IN (fndecl)) | |
1691 | return NULL_TREE; | |
1692 | ||
62f7fd21 | 1693 | return targetm.vectorize.builtin_vectorized_function (fndecl, vectype_out, |
ebfd146a IR |
1694 | vectype_in); |
1695 | } | |
1696 | ||
1697 | /* Function vectorizable_call. | |
1698 | ||
b8698a0f L |
1699 | Check if STMT performs a function call that can be vectorized. |
1700 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1701 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1702 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1703 | ||
1704 | static bool | |
190c2236 JJ |
1705 | vectorizable_call (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, |
1706 | slp_tree slp_node) | |
ebfd146a IR |
1707 | { |
1708 | tree vec_dest; | |
1709 | tree scalar_dest; | |
1710 | tree op, type; | |
1711 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
1712 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
1713 | tree vectype_out, vectype_in; | |
1714 | int nunits_in; | |
1715 | int nunits_out; | |
1716 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
190c2236 | 1717 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b690cc0f | 1718 | tree fndecl, new_temp, def, rhs_type; |
ebfd146a | 1719 | gimple def_stmt; |
0502fb85 UB |
1720 | enum vect_def_type dt[3] |
1721 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
63827fb8 | 1722 | gimple new_stmt = NULL; |
ebfd146a | 1723 | int ncopies, j; |
6e1aa848 | 1724 | vec<tree> vargs = vNULL; |
ebfd146a IR |
1725 | enum { NARROW, NONE, WIDEN } modifier; |
1726 | size_t i, nargs; | |
9d5e7640 | 1727 | tree lhs; |
ebfd146a | 1728 | |
190c2236 | 1729 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
1730 | return false; |
1731 | ||
8644a673 | 1732 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1733 | return false; |
1734 | ||
ebfd146a IR |
1735 | /* Is STMT a vectorizable call? */ |
1736 | if (!is_gimple_call (stmt)) | |
1737 | return false; | |
1738 | ||
1739 | if (TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
1740 | return false; | |
1741 | ||
822ba6d7 | 1742 | if (stmt_can_throw_internal (stmt)) |
5a2c1986 IR |
1743 | return false; |
1744 | ||
b690cc0f RG |
1745 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
1746 | ||
ebfd146a IR |
1747 | /* Process function arguments. */ |
1748 | rhs_type = NULL_TREE; | |
b690cc0f | 1749 | vectype_in = NULL_TREE; |
ebfd146a IR |
1750 | nargs = gimple_call_num_args (stmt); |
1751 | ||
1b1562a5 MM |
1752 | /* Bail out if the function has more than three arguments, we do not have |
1753 | interesting builtin functions to vectorize with more than two arguments | |
1754 | except for fma. No arguments is also not good. */ | |
1755 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
1756 | return false; |
1757 | ||
74bf76ed JJ |
1758 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
1759 | if (gimple_call_internal_p (stmt) | |
1760 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
1761 | { | |
1762 | nargs = 0; | |
1763 | rhs_type = unsigned_type_node; | |
1764 | } | |
1765 | ||
ebfd146a IR |
1766 | for (i = 0; i < nargs; i++) |
1767 | { | |
b690cc0f RG |
1768 | tree opvectype; |
1769 | ||
ebfd146a IR |
1770 | op = gimple_call_arg (stmt, i); |
1771 | ||
1772 | /* We can only handle calls with arguments of the same type. */ | |
1773 | if (rhs_type | |
8533c9d8 | 1774 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 1775 | { |
73fbfcad | 1776 | if (dump_enabled_p ()) |
78c60e3d | 1777 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1778 | "argument types differ.\n"); |
ebfd146a IR |
1779 | return false; |
1780 | } | |
b690cc0f RG |
1781 | if (!rhs_type) |
1782 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 1783 | |
24ee1384 | 1784 | if (!vect_is_simple_use_1 (op, stmt, loop_vinfo, bb_vinfo, |
b690cc0f | 1785 | &def_stmt, &def, &dt[i], &opvectype)) |
ebfd146a | 1786 | { |
73fbfcad | 1787 | if (dump_enabled_p ()) |
78c60e3d | 1788 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1789 | "use not simple.\n"); |
ebfd146a IR |
1790 | return false; |
1791 | } | |
ebfd146a | 1792 | |
b690cc0f RG |
1793 | if (!vectype_in) |
1794 | vectype_in = opvectype; | |
1795 | else if (opvectype | |
1796 | && opvectype != vectype_in) | |
1797 | { | |
73fbfcad | 1798 | if (dump_enabled_p ()) |
78c60e3d | 1799 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1800 | "argument vector types differ.\n"); |
b690cc0f RG |
1801 | return false; |
1802 | } | |
1803 | } | |
1804 | /* If all arguments are external or constant defs use a vector type with | |
1805 | the same size as the output vector type. */ | |
ebfd146a | 1806 | if (!vectype_in) |
b690cc0f | 1807 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
1808 | if (vec_stmt) |
1809 | gcc_assert (vectype_in); | |
1810 | if (!vectype_in) | |
1811 | { | |
73fbfcad | 1812 | if (dump_enabled_p ()) |
7d8930a0 | 1813 | { |
78c60e3d SS |
1814 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1815 | "no vectype for scalar type "); | |
1816 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 1817 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
1818 | } |
1819 | ||
1820 | return false; | |
1821 | } | |
ebfd146a IR |
1822 | |
1823 | /* FORNOW */ | |
b690cc0f RG |
1824 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
1825 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
1826 | if (nunits_in == nunits_out / 2) |
1827 | modifier = NARROW; | |
1828 | else if (nunits_out == nunits_in) | |
1829 | modifier = NONE; | |
1830 | else if (nunits_out == nunits_in / 2) | |
1831 | modifier = WIDEN; | |
1832 | else | |
1833 | return false; | |
1834 | ||
1835 | /* For now, we only vectorize functions if a target specific builtin | |
1836 | is available. TODO -- in some cases, it might be profitable to | |
1837 | insert the calls for pieces of the vector, in order to be able | |
1838 | to vectorize other operations in the loop. */ | |
1839 | fndecl = vectorizable_function (stmt, vectype_out, vectype_in); | |
1840 | if (fndecl == NULL_TREE) | |
1841 | { | |
74bf76ed JJ |
1842 | if (gimple_call_internal_p (stmt) |
1843 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE | |
1844 | && !slp_node | |
1845 | && loop_vinfo | |
1846 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
1847 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
1848 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
1849 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
1850 | { | |
1851 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
1852 | { 0, 1, 2, ... vf - 1 } vector. */ | |
1853 | gcc_assert (nargs == 0); | |
1854 | } | |
1855 | else | |
1856 | { | |
1857 | if (dump_enabled_p ()) | |
1858 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 1859 | "function is not vectorizable.\n"); |
74bf76ed JJ |
1860 | return false; |
1861 | } | |
ebfd146a IR |
1862 | } |
1863 | ||
5006671f | 1864 | gcc_assert (!gimple_vuse (stmt)); |
ebfd146a | 1865 | |
190c2236 JJ |
1866 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
1867 | ncopies = 1; | |
1868 | else if (modifier == NARROW) | |
ebfd146a IR |
1869 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; |
1870 | else | |
1871 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1872 | ||
1873 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
1874 | needs to be generated. */ | |
1875 | gcc_assert (ncopies >= 1); | |
1876 | ||
1877 | if (!vec_stmt) /* transformation not required. */ | |
1878 | { | |
1879 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 1880 | if (dump_enabled_p ()) |
e645e942 TJ |
1881 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ===" |
1882 | "\n"); | |
c3e7ee41 | 1883 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
1884 | return true; |
1885 | } | |
1886 | ||
1887 | /** Transform. **/ | |
1888 | ||
73fbfcad | 1889 | if (dump_enabled_p ()) |
e645e942 | 1890 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
1891 | |
1892 | /* Handle def. */ | |
1893 | scalar_dest = gimple_call_lhs (stmt); | |
1894 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
1895 | ||
1896 | prev_stmt_info = NULL; | |
1897 | switch (modifier) | |
1898 | { | |
1899 | case NONE: | |
1900 | for (j = 0; j < ncopies; ++j) | |
1901 | { | |
1902 | /* Build argument list for the vectorized call. */ | |
1903 | if (j == 0) | |
9771b263 | 1904 | vargs.create (nargs); |
ebfd146a | 1905 | else |
9771b263 | 1906 | vargs.truncate (0); |
ebfd146a | 1907 | |
190c2236 JJ |
1908 | if (slp_node) |
1909 | { | |
37b5ec8f | 1910 | vec<vec<tree> > vec_defs; |
9771b263 DN |
1911 | vec_defs.create (nargs); |
1912 | vec<tree> vec_oprnds0; | |
190c2236 JJ |
1913 | |
1914 | for (i = 0; i < nargs; i++) | |
9771b263 | 1915 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 1916 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 1917 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
1918 | |
1919 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 1920 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
1921 | { |
1922 | size_t k; | |
1923 | for (k = 0; k < nargs; k++) | |
1924 | { | |
37b5ec8f | 1925 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 1926 | vargs[k] = vec_oprndsk[i]; |
190c2236 JJ |
1927 | } |
1928 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1929 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1930 | gimple_call_set_lhs (new_stmt, new_temp); | |
1931 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 1932 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
1933 | } |
1934 | ||
1935 | for (i = 0; i < nargs; i++) | |
1936 | { | |
37b5ec8f | 1937 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 1938 | vec_oprndsi.release (); |
190c2236 | 1939 | } |
9771b263 | 1940 | vec_defs.release (); |
190c2236 JJ |
1941 | continue; |
1942 | } | |
1943 | ||
ebfd146a IR |
1944 | for (i = 0; i < nargs; i++) |
1945 | { | |
1946 | op = gimple_call_arg (stmt, i); | |
1947 | if (j == 0) | |
1948 | vec_oprnd0 | |
1949 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1950 | else | |
63827fb8 IR |
1951 | { |
1952 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
1953 | vec_oprnd0 | |
1954 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
1955 | } | |
ebfd146a | 1956 | |
9771b263 | 1957 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
1958 | } |
1959 | ||
74bf76ed JJ |
1960 | if (gimple_call_internal_p (stmt) |
1961 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
1962 | { | |
1963 | tree *v = XALLOCAVEC (tree, nunits_out); | |
1964 | int k; | |
1965 | for (k = 0; k < nunits_out; ++k) | |
1966 | v[k] = build_int_cst (unsigned_type_node, j * nunits_out + k); | |
1967 | tree cst = build_vector (vectype_out, v); | |
1968 | tree new_var | |
1969 | = vect_get_new_vect_var (vectype_out, vect_simple_var, "cst_"); | |
1970 | gimple init_stmt = gimple_build_assign (new_var, cst); | |
1971 | new_temp = make_ssa_name (new_var, init_stmt); | |
1972 | gimple_assign_set_lhs (init_stmt, new_temp); | |
1973 | vect_init_vector_1 (stmt, init_stmt, NULL); | |
1974 | new_temp = make_ssa_name (vec_dest, NULL); | |
1975 | new_stmt = gimple_build_assign (new_temp, | |
1976 | gimple_assign_lhs (init_stmt)); | |
1977 | } | |
1978 | else | |
1979 | { | |
1980 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1981 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1982 | gimple_call_set_lhs (new_stmt, new_temp); | |
1983 | } | |
ebfd146a IR |
1984 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1985 | ||
1986 | if (j == 0) | |
1987 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1988 | else | |
1989 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1990 | ||
1991 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1992 | } | |
1993 | ||
1994 | break; | |
1995 | ||
1996 | case NARROW: | |
1997 | for (j = 0; j < ncopies; ++j) | |
1998 | { | |
1999 | /* Build argument list for the vectorized call. */ | |
2000 | if (j == 0) | |
9771b263 | 2001 | vargs.create (nargs * 2); |
ebfd146a | 2002 | else |
9771b263 | 2003 | vargs.truncate (0); |
ebfd146a | 2004 | |
190c2236 JJ |
2005 | if (slp_node) |
2006 | { | |
37b5ec8f | 2007 | vec<vec<tree> > vec_defs; |
9771b263 DN |
2008 | vec_defs.create (nargs); |
2009 | vec<tree> vec_oprnds0; | |
190c2236 JJ |
2010 | |
2011 | for (i = 0; i < nargs; i++) | |
9771b263 | 2012 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2013 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2014 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2015 | |
2016 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2017 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
2018 | { |
2019 | size_t k; | |
9771b263 | 2020 | vargs.truncate (0); |
190c2236 JJ |
2021 | for (k = 0; k < nargs; k++) |
2022 | { | |
37b5ec8f | 2023 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
2024 | vargs.quick_push (vec_oprndsk[i]); |
2025 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 JJ |
2026 | } |
2027 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2028 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2029 | gimple_call_set_lhs (new_stmt, new_temp); | |
2030 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 2031 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2032 | } |
2033 | ||
2034 | for (i = 0; i < nargs; i++) | |
2035 | { | |
37b5ec8f | 2036 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2037 | vec_oprndsi.release (); |
190c2236 | 2038 | } |
9771b263 | 2039 | vec_defs.release (); |
190c2236 JJ |
2040 | continue; |
2041 | } | |
2042 | ||
ebfd146a IR |
2043 | for (i = 0; i < nargs; i++) |
2044 | { | |
2045 | op = gimple_call_arg (stmt, i); | |
2046 | if (j == 0) | |
2047 | { | |
2048 | vec_oprnd0 | |
2049 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
2050 | vec_oprnd1 | |
63827fb8 | 2051 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2052 | } |
2053 | else | |
2054 | { | |
336ecb65 | 2055 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 2056 | vec_oprnd0 |
63827fb8 | 2057 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 2058 | vec_oprnd1 |
63827fb8 | 2059 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2060 | } |
2061 | ||
9771b263 DN |
2062 | vargs.quick_push (vec_oprnd0); |
2063 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
2064 | } |
2065 | ||
2066 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2067 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2068 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
2069 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2070 | ||
2071 | if (j == 0) | |
2072 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2073 | else | |
2074 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2075 | ||
2076 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2077 | } | |
2078 | ||
2079 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
2080 | ||
2081 | break; | |
2082 | ||
2083 | case WIDEN: | |
2084 | /* No current target implements this case. */ | |
2085 | return false; | |
2086 | } | |
2087 | ||
9771b263 | 2088 | vargs.release (); |
ebfd146a IR |
2089 | |
2090 | /* Update the exception handling table with the vector stmt if necessary. */ | |
2091 | if (maybe_clean_or_replace_eh_stmt (stmt, *vec_stmt)) | |
2092 | gimple_purge_dead_eh_edges (gimple_bb (stmt)); | |
2093 | ||
2094 | /* The call in STMT might prevent it from being removed in dce. | |
2095 | We however cannot remove it here, due to the way the ssa name | |
2096 | it defines is mapped to the new definition. So just replace | |
2097 | rhs of the statement with something harmless. */ | |
2098 | ||
dd34c087 JJ |
2099 | if (slp_node) |
2100 | return true; | |
2101 | ||
ebfd146a | 2102 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
2103 | if (is_pattern_stmt_p (stmt_info)) |
2104 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
2105 | else | |
2106 | lhs = gimple_call_lhs (stmt); | |
2107 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
ebfd146a | 2108 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 2109 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
2110 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
2111 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
2112 | |
2113 | return true; | |
2114 | } | |
2115 | ||
2116 | ||
2117 | /* Function vect_gen_widened_results_half | |
2118 | ||
2119 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 2120 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 2121 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
2122 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
2123 | needs to be created (DECL is a function-decl of a target-builtin). | |
2124 | STMT is the original scalar stmt that we are vectorizing. */ | |
2125 | ||
2126 | static gimple | |
2127 | vect_gen_widened_results_half (enum tree_code code, | |
2128 | tree decl, | |
2129 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
2130 | tree vec_dest, gimple_stmt_iterator *gsi, | |
2131 | gimple stmt) | |
b8698a0f | 2132 | { |
ebfd146a | 2133 | gimple new_stmt; |
b8698a0f L |
2134 | tree new_temp; |
2135 | ||
2136 | /* Generate half of the widened result: */ | |
2137 | if (code == CALL_EXPR) | |
2138 | { | |
2139 | /* Target specific support */ | |
ebfd146a IR |
2140 | if (op_type == binary_op) |
2141 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
2142 | else | |
2143 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
2144 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2145 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
2146 | } |
2147 | else | |
ebfd146a | 2148 | { |
b8698a0f L |
2149 | /* Generic support */ |
2150 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
2151 | if (op_type != binary_op) |
2152 | vec_oprnd1 = NULL; | |
2153 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vec_oprnd0, | |
2154 | vec_oprnd1); | |
2155 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2156 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 2157 | } |
ebfd146a IR |
2158 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2159 | ||
ebfd146a IR |
2160 | return new_stmt; |
2161 | } | |
2162 | ||
4a00c761 JJ |
2163 | |
2164 | /* Get vectorized definitions for loop-based vectorization. For the first | |
2165 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
2166 | scalar operand), and for the rest we get a copy with | |
2167 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
2168 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
2169 | The vectors are collected into VEC_OPRNDS. */ | |
2170 | ||
2171 | static void | |
2172 | vect_get_loop_based_defs (tree *oprnd, gimple stmt, enum vect_def_type dt, | |
9771b263 | 2173 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
2174 | { |
2175 | tree vec_oprnd; | |
2176 | ||
2177 | /* Get first vector operand. */ | |
2178 | /* All the vector operands except the very first one (that is scalar oprnd) | |
2179 | are stmt copies. */ | |
2180 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
2181 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt, NULL); | |
2182 | else | |
2183 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
2184 | ||
9771b263 | 2185 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
2186 | |
2187 | /* Get second vector operand. */ | |
2188 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 2189 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
2190 | |
2191 | *oprnd = vec_oprnd; | |
2192 | ||
2193 | /* For conversion in multiple steps, continue to get operands | |
2194 | recursively. */ | |
2195 | if (multi_step_cvt) | |
2196 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
2197 | } | |
2198 | ||
2199 | ||
2200 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
2201 | For multi-step conversions store the resulting vectors and call the function | |
2202 | recursively. */ | |
2203 | ||
2204 | static void | |
9771b263 | 2205 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
4a00c761 | 2206 | int multi_step_cvt, gimple stmt, |
9771b263 | 2207 | vec<tree> vec_dsts, |
4a00c761 JJ |
2208 | gimple_stmt_iterator *gsi, |
2209 | slp_tree slp_node, enum tree_code code, | |
2210 | stmt_vec_info *prev_stmt_info) | |
2211 | { | |
2212 | unsigned int i; | |
2213 | tree vop0, vop1, new_tmp, vec_dest; | |
2214 | gimple new_stmt; | |
2215 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2216 | ||
9771b263 | 2217 | vec_dest = vec_dsts.pop (); |
4a00c761 | 2218 | |
9771b263 | 2219 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
2220 | { |
2221 | /* Create demotion operation. */ | |
9771b263 DN |
2222 | vop0 = (*vec_oprnds)[i]; |
2223 | vop1 = (*vec_oprnds)[i + 1]; | |
4a00c761 JJ |
2224 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); |
2225 | new_tmp = make_ssa_name (vec_dest, new_stmt); | |
2226 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
2227 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2228 | ||
2229 | if (multi_step_cvt) | |
2230 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 2231 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
2232 | else |
2233 | { | |
2234 | /* This is the last step of the conversion sequence. Store the | |
2235 | vectors in SLP_NODE or in vector info of the scalar statement | |
2236 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
2237 | if (slp_node) | |
9771b263 | 2238 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
2239 | else |
2240 | { | |
2241 | if (!*prev_stmt_info) | |
2242 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2243 | else | |
2244 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
2245 | ||
2246 | *prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2247 | } | |
2248 | } | |
2249 | } | |
2250 | ||
2251 | /* For multi-step demotion operations we first generate demotion operations | |
2252 | from the source type to the intermediate types, and then combine the | |
2253 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
2254 | type. */ | |
2255 | if (multi_step_cvt) | |
2256 | { | |
2257 | /* At each level of recursion we have half of the operands we had at the | |
2258 | previous level. */ | |
9771b263 | 2259 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
2260 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
2261 | stmt, vec_dsts, gsi, slp_node, | |
2262 | VEC_PACK_TRUNC_EXPR, | |
2263 | prev_stmt_info); | |
2264 | } | |
2265 | ||
9771b263 | 2266 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
2267 | } |
2268 | ||
2269 | ||
2270 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
2271 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
2272 | the resulting vectors and call the function recursively. */ | |
2273 | ||
2274 | static void | |
9771b263 DN |
2275 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
2276 | vec<tree> *vec_oprnds1, | |
4a00c761 JJ |
2277 | gimple stmt, tree vec_dest, |
2278 | gimple_stmt_iterator *gsi, | |
2279 | enum tree_code code1, | |
2280 | enum tree_code code2, tree decl1, | |
2281 | tree decl2, int op_type) | |
2282 | { | |
2283 | int i; | |
2284 | tree vop0, vop1, new_tmp1, new_tmp2; | |
2285 | gimple new_stmt1, new_stmt2; | |
6e1aa848 | 2286 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 2287 | |
9771b263 DN |
2288 | vec_tmp.create (vec_oprnds0->length () * 2); |
2289 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
2290 | { |
2291 | if (op_type == binary_op) | |
9771b263 | 2292 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
2293 | else |
2294 | vop1 = NULL_TREE; | |
2295 | ||
2296 | /* Generate the two halves of promotion operation. */ | |
2297 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
2298 | op_type, vec_dest, gsi, stmt); | |
2299 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
2300 | op_type, vec_dest, gsi, stmt); | |
2301 | if (is_gimple_call (new_stmt1)) | |
2302 | { | |
2303 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
2304 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
2305 | } | |
2306 | else | |
2307 | { | |
2308 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
2309 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
2310 | } | |
2311 | ||
2312 | /* Store the results for the next step. */ | |
9771b263 DN |
2313 | vec_tmp.quick_push (new_tmp1); |
2314 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
2315 | } |
2316 | ||
689eaba3 | 2317 | vec_oprnds0->release (); |
4a00c761 JJ |
2318 | *vec_oprnds0 = vec_tmp; |
2319 | } | |
2320 | ||
2321 | ||
b8698a0f L |
2322 | /* Check if STMT performs a conversion operation, that can be vectorized. |
2323 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 2324 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
2325 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
2326 | ||
2327 | static bool | |
2328 | vectorizable_conversion (gimple stmt, gimple_stmt_iterator *gsi, | |
2329 | gimple *vec_stmt, slp_tree slp_node) | |
2330 | { | |
2331 | tree vec_dest; | |
2332 | tree scalar_dest; | |
4a00c761 | 2333 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
2334 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
2335 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2336 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2337 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 2338 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
2339 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
2340 | tree new_temp; | |
2341 | tree def; | |
2342 | gimple def_stmt; | |
2343 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
2344 | gimple new_stmt = NULL; | |
2345 | stmt_vec_info prev_stmt_info; | |
2346 | int nunits_in; | |
2347 | int nunits_out; | |
2348 | tree vectype_out, vectype_in; | |
4a00c761 JJ |
2349 | int ncopies, i, j; |
2350 | tree lhs_type, rhs_type; | |
ebfd146a | 2351 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
2352 | vec<tree> vec_oprnds0 = vNULL; |
2353 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 2354 | tree vop0; |
4a00c761 JJ |
2355 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
2356 | int multi_step_cvt = 0; | |
6e1aa848 DN |
2357 | vec<tree> vec_dsts = vNULL; |
2358 | vec<tree> interm_types = vNULL; | |
4a00c761 JJ |
2359 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
2360 | int op_type; | |
2361 | enum machine_mode rhs_mode; | |
2362 | unsigned short fltsz; | |
ebfd146a IR |
2363 | |
2364 | /* Is STMT a vectorizable conversion? */ | |
2365 | ||
4a00c761 | 2366 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2367 | return false; |
2368 | ||
8644a673 | 2369 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2370 | return false; |
2371 | ||
2372 | if (!is_gimple_assign (stmt)) | |
2373 | return false; | |
2374 | ||
2375 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2376 | return false; | |
2377 | ||
2378 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
2379 | if (!CONVERT_EXPR_CODE_P (code) |
2380 | && code != FIX_TRUNC_EXPR | |
2381 | && code != FLOAT_EXPR | |
2382 | && code != WIDEN_MULT_EXPR | |
2383 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
2384 | return false; |
2385 | ||
4a00c761 JJ |
2386 | op_type = TREE_CODE_LENGTH (code); |
2387 | ||
ebfd146a | 2388 | /* Check types of lhs and rhs. */ |
b690cc0f | 2389 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 2390 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
2391 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
2392 | ||
ebfd146a IR |
2393 | op0 = gimple_assign_rhs1 (stmt); |
2394 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
2395 | |
2396 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
2397 | && !((INTEGRAL_TYPE_P (lhs_type) | |
2398 | && INTEGRAL_TYPE_P (rhs_type)) | |
2399 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
2400 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
2401 | return false; | |
2402 | ||
2403 | if ((INTEGRAL_TYPE_P (lhs_type) | |
2404 | && (TYPE_PRECISION (lhs_type) | |
2405 | != GET_MODE_PRECISION (TYPE_MODE (lhs_type)))) | |
2406 | || (INTEGRAL_TYPE_P (rhs_type) | |
2407 | && (TYPE_PRECISION (rhs_type) | |
2408 | != GET_MODE_PRECISION (TYPE_MODE (rhs_type))))) | |
2409 | { | |
73fbfcad | 2410 | if (dump_enabled_p ()) |
78c60e3d | 2411 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
2412 | "type conversion to/from bit-precision unsupported." |
2413 | "\n"); | |
4a00c761 JJ |
2414 | return false; |
2415 | } | |
2416 | ||
b690cc0f | 2417 | /* Check the operands of the operation. */ |
24ee1384 | 2418 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
b690cc0f RG |
2419 | &def_stmt, &def, &dt[0], &vectype_in)) |
2420 | { | |
73fbfcad | 2421 | if (dump_enabled_p ()) |
78c60e3d | 2422 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2423 | "use not simple.\n"); |
b690cc0f RG |
2424 | return false; |
2425 | } | |
4a00c761 JJ |
2426 | if (op_type == binary_op) |
2427 | { | |
2428 | bool ok; | |
2429 | ||
2430 | op1 = gimple_assign_rhs2 (stmt); | |
2431 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
2432 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
2433 | OP1. */ | |
2434 | if (CONSTANT_CLASS_P (op0)) | |
f5709183 | 2435 | ok = vect_is_simple_use_1 (op1, stmt, loop_vinfo, bb_vinfo, |
4a00c761 JJ |
2436 | &def_stmt, &def, &dt[1], &vectype_in); |
2437 | else | |
f5709183 | 2438 | ok = vect_is_simple_use (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
24ee1384 | 2439 | &def, &dt[1]); |
4a00c761 JJ |
2440 | |
2441 | if (!ok) | |
2442 | { | |
73fbfcad | 2443 | if (dump_enabled_p ()) |
78c60e3d | 2444 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2445 | "use not simple.\n"); |
4a00c761 JJ |
2446 | return false; |
2447 | } | |
2448 | } | |
2449 | ||
b690cc0f RG |
2450 | /* If op0 is an external or constant defs use a vector type of |
2451 | the same size as the output vector type. */ | |
ebfd146a | 2452 | if (!vectype_in) |
b690cc0f | 2453 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
2454 | if (vec_stmt) |
2455 | gcc_assert (vectype_in); | |
2456 | if (!vectype_in) | |
2457 | { | |
73fbfcad | 2458 | if (dump_enabled_p ()) |
4a00c761 | 2459 | { |
78c60e3d SS |
2460 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2461 | "no vectype for scalar type "); | |
2462 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 2463 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 2464 | } |
7d8930a0 IR |
2465 | |
2466 | return false; | |
2467 | } | |
ebfd146a | 2468 | |
b690cc0f RG |
2469 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
2470 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4a00c761 | 2471 | if (nunits_in < nunits_out) |
ebfd146a IR |
2472 | modifier = NARROW; |
2473 | else if (nunits_out == nunits_in) | |
2474 | modifier = NONE; | |
ebfd146a | 2475 | else |
4a00c761 | 2476 | modifier = WIDEN; |
ebfd146a | 2477 | |
ff802fa1 IR |
2478 | /* Multiple types in SLP are handled by creating the appropriate number of |
2479 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2480 | case of SLP. */ | |
437f4a00 | 2481 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a | 2482 | ncopies = 1; |
4a00c761 JJ |
2483 | else if (modifier == NARROW) |
2484 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
2485 | else | |
2486 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
b8698a0f | 2487 | |
ebfd146a IR |
2488 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
2489 | needs to be generated. */ | |
2490 | gcc_assert (ncopies >= 1); | |
2491 | ||
ebfd146a | 2492 | /* Supportable by target? */ |
4a00c761 | 2493 | switch (modifier) |
ebfd146a | 2494 | { |
4a00c761 JJ |
2495 | case NONE: |
2496 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
2497 | return false; | |
2498 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
2499 | &decl1, &code1)) | |
2500 | break; | |
2501 | /* FALLTHRU */ | |
2502 | unsupported: | |
73fbfcad | 2503 | if (dump_enabled_p ()) |
78c60e3d | 2504 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2505 | "conversion not supported by target.\n"); |
ebfd146a | 2506 | return false; |
ebfd146a | 2507 | |
4a00c761 JJ |
2508 | case WIDEN: |
2509 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
2510 | &code1, &code2, &multi_step_cvt, |
2511 | &interm_types)) | |
4a00c761 JJ |
2512 | { |
2513 | /* Binary widening operation can only be supported directly by the | |
2514 | architecture. */ | |
2515 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
2516 | break; | |
2517 | } | |
2518 | ||
2519 | if (code != FLOAT_EXPR | |
2520 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
2521 | <= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
2522 | goto unsupported; | |
2523 | ||
2524 | rhs_mode = TYPE_MODE (rhs_type); | |
2525 | fltsz = GET_MODE_SIZE (TYPE_MODE (lhs_type)); | |
2526 | for (rhs_mode = GET_MODE_2XWIDER_MODE (TYPE_MODE (rhs_type)); | |
2527 | rhs_mode != VOIDmode && GET_MODE_SIZE (rhs_mode) <= fltsz; | |
2528 | rhs_mode = GET_MODE_2XWIDER_MODE (rhs_mode)) | |
2529 | { | |
2530 | cvt_type | |
2531 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
2532 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
2533 | if (cvt_type == NULL_TREE) | |
2534 | goto unsupported; | |
2535 | ||
2536 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
2537 | { | |
2538 | if (!supportable_convert_operation (code, vectype_out, | |
2539 | cvt_type, &decl1, &codecvt1)) | |
2540 | goto unsupported; | |
2541 | } | |
2542 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
2543 | cvt_type, &codecvt1, |
2544 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
2545 | &interm_types)) |
2546 | continue; | |
2547 | else | |
2548 | gcc_assert (multi_step_cvt == 0); | |
2549 | ||
2550 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
2551 | vectype_in, &code1, &code2, |
2552 | &multi_step_cvt, &interm_types)) | |
4a00c761 JJ |
2553 | break; |
2554 | } | |
2555 | ||
2556 | if (rhs_mode == VOIDmode || GET_MODE_SIZE (rhs_mode) > fltsz) | |
2557 | goto unsupported; | |
2558 | ||
2559 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
2560 | codecvt2 = ERROR_MARK; | |
2561 | else | |
2562 | { | |
2563 | multi_step_cvt++; | |
9771b263 | 2564 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
2565 | cvt_type = NULL_TREE; |
2566 | } | |
2567 | break; | |
2568 | ||
2569 | case NARROW: | |
2570 | gcc_assert (op_type == unary_op); | |
2571 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
2572 | &code1, &multi_step_cvt, | |
2573 | &interm_types)) | |
2574 | break; | |
2575 | ||
2576 | if (code != FIX_TRUNC_EXPR | |
2577 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
2578 | >= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
2579 | goto unsupported; | |
2580 | ||
2581 | rhs_mode = TYPE_MODE (rhs_type); | |
2582 | cvt_type | |
2583 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
2584 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
2585 | if (cvt_type == NULL_TREE) | |
2586 | goto unsupported; | |
2587 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
2588 | &decl1, &codecvt1)) | |
2589 | goto unsupported; | |
2590 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
2591 | &code1, &multi_step_cvt, | |
2592 | &interm_types)) | |
2593 | break; | |
2594 | goto unsupported; | |
2595 | ||
2596 | default: | |
2597 | gcc_unreachable (); | |
ebfd146a IR |
2598 | } |
2599 | ||
2600 | if (!vec_stmt) /* transformation not required. */ | |
2601 | { | |
73fbfcad | 2602 | if (dump_enabled_p ()) |
78c60e3d | 2603 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 2604 | "=== vectorizable_conversion ===\n"); |
4a00c761 | 2605 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
2606 | { |
2607 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
c3e7ee41 | 2608 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
8bd37302 | 2609 | } |
4a00c761 JJ |
2610 | else if (modifier == NARROW) |
2611 | { | |
2612 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
8bd37302 | 2613 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 JJ |
2614 | } |
2615 | else | |
2616 | { | |
2617 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
8bd37302 | 2618 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 | 2619 | } |
9771b263 | 2620 | interm_types.release (); |
ebfd146a IR |
2621 | return true; |
2622 | } | |
2623 | ||
2624 | /** Transform. **/ | |
73fbfcad | 2625 | if (dump_enabled_p ()) |
78c60e3d | 2626 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 2627 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 2628 | |
4a00c761 JJ |
2629 | if (op_type == binary_op) |
2630 | { | |
2631 | if (CONSTANT_CLASS_P (op0)) | |
2632 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
2633 | else if (CONSTANT_CLASS_P (op1)) | |
2634 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
2635 | } | |
2636 | ||
2637 | /* In case of multi-step conversion, we first generate conversion operations | |
2638 | to the intermediate types, and then from that types to the final one. | |
2639 | We create vector destinations for the intermediate type (TYPES) received | |
2640 | from supportable_*_operation, and store them in the correct order | |
2641 | for future use in vect_create_vectorized_*_stmts (). */ | |
9771b263 | 2642 | vec_dsts.create (multi_step_cvt + 1); |
82294ec1 JJ |
2643 | vec_dest = vect_create_destination_var (scalar_dest, |
2644 | (cvt_type && modifier == WIDEN) | |
2645 | ? cvt_type : vectype_out); | |
9771b263 | 2646 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
2647 | |
2648 | if (multi_step_cvt) | |
2649 | { | |
9771b263 DN |
2650 | for (i = interm_types.length () - 1; |
2651 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
2652 | { |
2653 | vec_dest = vect_create_destination_var (scalar_dest, | |
2654 | intermediate_type); | |
9771b263 | 2655 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
2656 | } |
2657 | } | |
ebfd146a | 2658 | |
4a00c761 | 2659 | if (cvt_type) |
82294ec1 JJ |
2660 | vec_dest = vect_create_destination_var (scalar_dest, |
2661 | modifier == WIDEN | |
2662 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
2663 | |
2664 | if (!slp_node) | |
2665 | { | |
30862efc | 2666 | if (modifier == WIDEN) |
4a00c761 | 2667 | { |
c3284718 | 2668 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 2669 | if (op_type == binary_op) |
9771b263 | 2670 | vec_oprnds1.create (1); |
4a00c761 | 2671 | } |
30862efc | 2672 | else if (modifier == NARROW) |
9771b263 DN |
2673 | vec_oprnds0.create ( |
2674 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
2675 | } |
2676 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 2677 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 2678 | |
4a00c761 | 2679 | last_oprnd = op0; |
ebfd146a IR |
2680 | prev_stmt_info = NULL; |
2681 | switch (modifier) | |
2682 | { | |
2683 | case NONE: | |
2684 | for (j = 0; j < ncopies; j++) | |
2685 | { | |
ebfd146a | 2686 | if (j == 0) |
d092494c IR |
2687 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node, |
2688 | -1); | |
ebfd146a IR |
2689 | else |
2690 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
2691 | ||
9771b263 | 2692 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
2693 | { |
2694 | /* Arguments are ready, create the new vector stmt. */ | |
2695 | if (code1 == CALL_EXPR) | |
2696 | { | |
2697 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
2698 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2699 | gimple_call_set_lhs (new_stmt, new_temp); | |
2700 | } | |
2701 | else | |
2702 | { | |
2703 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
2704 | new_stmt = gimple_build_assign_with_ops (code1, vec_dest, | |
2705 | vop0, NULL); | |
2706 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2707 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2708 | } | |
2709 | ||
2710 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2711 | if (slp_node) | |
9771b263 | 2712 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
2713 | } |
2714 | ||
ebfd146a IR |
2715 | if (j == 0) |
2716 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2717 | else | |
2718 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2719 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2720 | } | |
2721 | break; | |
2722 | ||
2723 | case WIDEN: | |
2724 | /* In case the vectorization factor (VF) is bigger than the number | |
2725 | of elements that we can fit in a vectype (nunits), we have to | |
2726 | generate more than one vector stmt - i.e - we need to "unroll" | |
2727 | the vector stmt by a factor VF/nunits. */ | |
2728 | for (j = 0; j < ncopies; j++) | |
2729 | { | |
4a00c761 | 2730 | /* Handle uses. */ |
ebfd146a | 2731 | if (j == 0) |
4a00c761 JJ |
2732 | { |
2733 | if (slp_node) | |
2734 | { | |
2735 | if (code == WIDEN_LSHIFT_EXPR) | |
2736 | { | |
2737 | unsigned int k; | |
ebfd146a | 2738 | |
4a00c761 JJ |
2739 | vec_oprnd1 = op1; |
2740 | /* Store vec_oprnd1 for every vector stmt to be created | |
2741 | for SLP_NODE. We check during the analysis that all | |
2742 | the shift arguments are the same. */ | |
2743 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 2744 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
2745 | |
2746 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
2747 | slp_node, -1); | |
2748 | } | |
2749 | else | |
2750 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
2751 | &vec_oprnds1, slp_node, -1); | |
2752 | } | |
2753 | else | |
2754 | { | |
2755 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
9771b263 | 2756 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
2757 | if (op_type == binary_op) |
2758 | { | |
2759 | if (code == WIDEN_LSHIFT_EXPR) | |
2760 | vec_oprnd1 = op1; | |
2761 | else | |
2762 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, | |
2763 | NULL); | |
9771b263 | 2764 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
2765 | } |
2766 | } | |
2767 | } | |
ebfd146a | 2768 | else |
4a00c761 JJ |
2769 | { |
2770 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
2771 | vec_oprnds0.truncate (0); |
2772 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
2773 | if (op_type == binary_op) |
2774 | { | |
2775 | if (code == WIDEN_LSHIFT_EXPR) | |
2776 | vec_oprnd1 = op1; | |
2777 | else | |
2778 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
2779 | vec_oprnd1); | |
9771b263 DN |
2780 | vec_oprnds1.truncate (0); |
2781 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
2782 | } |
2783 | } | |
ebfd146a | 2784 | |
4a00c761 JJ |
2785 | /* Arguments are ready. Create the new vector stmts. */ |
2786 | for (i = multi_step_cvt; i >= 0; i--) | |
2787 | { | |
9771b263 | 2788 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
2789 | enum tree_code c1 = code1, c2 = code2; |
2790 | if (i == 0 && codecvt2 != ERROR_MARK) | |
2791 | { | |
2792 | c1 = codecvt1; | |
2793 | c2 = codecvt2; | |
2794 | } | |
2795 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
2796 | &vec_oprnds1, | |
2797 | stmt, this_dest, gsi, | |
2798 | c1, c2, decl1, decl2, | |
2799 | op_type); | |
2800 | } | |
2801 | ||
9771b263 | 2802 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
2803 | { |
2804 | if (cvt_type) | |
2805 | { | |
2806 | if (codecvt1 == CALL_EXPR) | |
2807 | { | |
2808 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
2809 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2810 | gimple_call_set_lhs (new_stmt, new_temp); | |
2811 | } | |
2812 | else | |
2813 | { | |
2814 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
2815 | new_temp = make_ssa_name (vec_dest, NULL); | |
2816 | new_stmt = gimple_build_assign_with_ops (codecvt1, | |
2817 | new_temp, | |
2818 | vop0, NULL); | |
2819 | } | |
2820 | ||
2821 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2822 | } | |
2823 | else | |
2824 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
2825 | ||
2826 | if (slp_node) | |
9771b263 | 2827 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
2828 | else |
2829 | { | |
2830 | if (!prev_stmt_info) | |
2831 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2832 | else | |
2833 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2834 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2835 | } | |
2836 | } | |
ebfd146a | 2837 | } |
4a00c761 JJ |
2838 | |
2839 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
2840 | break; |
2841 | ||
2842 | case NARROW: | |
2843 | /* In case the vectorization factor (VF) is bigger than the number | |
2844 | of elements that we can fit in a vectype (nunits), we have to | |
2845 | generate more than one vector stmt - i.e - we need to "unroll" | |
2846 | the vector stmt by a factor VF/nunits. */ | |
2847 | for (j = 0; j < ncopies; j++) | |
2848 | { | |
2849 | /* Handle uses. */ | |
4a00c761 JJ |
2850 | if (slp_node) |
2851 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
2852 | slp_node, -1); | |
ebfd146a IR |
2853 | else |
2854 | { | |
9771b263 | 2855 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
2856 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
2857 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
2858 | } |
2859 | ||
4a00c761 JJ |
2860 | /* Arguments are ready. Create the new vector stmts. */ |
2861 | if (cvt_type) | |
9771b263 | 2862 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
2863 | { |
2864 | if (codecvt1 == CALL_EXPR) | |
2865 | { | |
2866 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
2867 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2868 | gimple_call_set_lhs (new_stmt, new_temp); | |
2869 | } | |
2870 | else | |
2871 | { | |
2872 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
2873 | new_temp = make_ssa_name (vec_dest, NULL); | |
2874 | new_stmt = gimple_build_assign_with_ops (codecvt1, new_temp, | |
2875 | vop0, NULL); | |
2876 | } | |
ebfd146a | 2877 | |
4a00c761 | 2878 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 2879 | vec_oprnds0[i] = new_temp; |
4a00c761 | 2880 | } |
ebfd146a | 2881 | |
4a00c761 JJ |
2882 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
2883 | stmt, vec_dsts, gsi, | |
2884 | slp_node, code1, | |
2885 | &prev_stmt_info); | |
ebfd146a IR |
2886 | } |
2887 | ||
2888 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 2889 | break; |
ebfd146a IR |
2890 | } |
2891 | ||
9771b263 DN |
2892 | vec_oprnds0.release (); |
2893 | vec_oprnds1.release (); | |
2894 | vec_dsts.release (); | |
2895 | interm_types.release (); | |
ebfd146a IR |
2896 | |
2897 | return true; | |
2898 | } | |
ff802fa1 IR |
2899 | |
2900 | ||
ebfd146a IR |
2901 | /* Function vectorizable_assignment. |
2902 | ||
b8698a0f L |
2903 | Check if STMT performs an assignment (copy) that can be vectorized. |
2904 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
2905 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2906 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2907 | ||
2908 | static bool | |
2909 | vectorizable_assignment (gimple stmt, gimple_stmt_iterator *gsi, | |
2910 | gimple *vec_stmt, slp_tree slp_node) | |
2911 | { | |
2912 | tree vec_dest; | |
2913 | tree scalar_dest; | |
2914 | tree op; | |
2915 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2916 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2917 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2918 | tree new_temp; | |
2919 | tree def; | |
2920 | gimple def_stmt; | |
2921 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
fde9c428 | 2922 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
ebfd146a | 2923 | int ncopies; |
f18b55bd | 2924 | int i, j; |
6e1aa848 | 2925 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 2926 | tree vop; |
a70d6342 | 2927 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
f18b55bd IR |
2928 | gimple new_stmt = NULL; |
2929 | stmt_vec_info prev_stmt_info = NULL; | |
fde9c428 RG |
2930 | enum tree_code code; |
2931 | tree vectype_in; | |
ebfd146a IR |
2932 | |
2933 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2934 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2935 | case of SLP. */ | |
437f4a00 | 2936 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
2937 | ncopies = 1; |
2938 | else | |
2939 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
2940 | ||
2941 | gcc_assert (ncopies >= 1); | |
ebfd146a | 2942 | |
a70d6342 | 2943 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2944 | return false; |
2945 | ||
8644a673 | 2946 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2947 | return false; |
2948 | ||
2949 | /* Is vectorizable assignment? */ | |
2950 | if (!is_gimple_assign (stmt)) | |
2951 | return false; | |
2952 | ||
2953 | scalar_dest = gimple_assign_lhs (stmt); | |
2954 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
2955 | return false; | |
2956 | ||
fde9c428 | 2957 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 2958 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
2959 | || code == PAREN_EXPR |
2960 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
2961 | op = gimple_assign_rhs1 (stmt); |
2962 | else | |
2963 | return false; | |
2964 | ||
7b7ec6c5 RG |
2965 | if (code == VIEW_CONVERT_EXPR) |
2966 | op = TREE_OPERAND (op, 0); | |
2967 | ||
24ee1384 | 2968 | if (!vect_is_simple_use_1 (op, stmt, loop_vinfo, bb_vinfo, |
fde9c428 | 2969 | &def_stmt, &def, &dt[0], &vectype_in)) |
ebfd146a | 2970 | { |
73fbfcad | 2971 | if (dump_enabled_p ()) |
78c60e3d | 2972 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2973 | "use not simple.\n"); |
ebfd146a IR |
2974 | return false; |
2975 | } | |
2976 | ||
fde9c428 RG |
2977 | /* We can handle NOP_EXPR conversions that do not change the number |
2978 | of elements or the vector size. */ | |
7b7ec6c5 RG |
2979 | if ((CONVERT_EXPR_CODE_P (code) |
2980 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 RG |
2981 | && (!vectype_in |
2982 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
2983 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
2984 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
2985 | return false; | |
2986 | ||
7b7b1813 RG |
2987 | /* We do not handle bit-precision changes. */ |
2988 | if ((CONVERT_EXPR_CODE_P (code) | |
2989 | || code == VIEW_CONVERT_EXPR) | |
2990 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
2991 | && ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
2992 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
2993 | || ((TYPE_PRECISION (TREE_TYPE (op)) | |
2994 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))) | |
2995 | /* But a conversion that does not change the bit-pattern is ok. */ | |
2996 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
2997 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2998 | && TYPE_UNSIGNED (TREE_TYPE (op)))) | |
2999 | { | |
73fbfcad | 3000 | if (dump_enabled_p ()) |
78c60e3d SS |
3001 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3002 | "type conversion to/from bit-precision " | |
e645e942 | 3003 | "unsupported.\n"); |
7b7b1813 RG |
3004 | return false; |
3005 | } | |
3006 | ||
ebfd146a IR |
3007 | if (!vec_stmt) /* transformation not required. */ |
3008 | { | |
3009 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 3010 | if (dump_enabled_p ()) |
78c60e3d | 3011 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3012 | "=== vectorizable_assignment ===\n"); |
c3e7ee41 | 3013 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
3014 | return true; |
3015 | } | |
3016 | ||
3017 | /** Transform. **/ | |
73fbfcad | 3018 | if (dump_enabled_p ()) |
e645e942 | 3019 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
3020 | |
3021 | /* Handle def. */ | |
3022 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3023 | ||
3024 | /* Handle use. */ | |
f18b55bd | 3025 | for (j = 0; j < ncopies; j++) |
ebfd146a | 3026 | { |
f18b55bd IR |
3027 | /* Handle uses. */ |
3028 | if (j == 0) | |
d092494c | 3029 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node, -1); |
f18b55bd IR |
3030 | else |
3031 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
3032 | ||
3033 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 3034 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 3035 | { |
7b7ec6c5 RG |
3036 | if (CONVERT_EXPR_CODE_P (code) |
3037 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 3038 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
3039 | new_stmt = gimple_build_assign (vec_dest, vop); |
3040 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3041 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3042 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3043 | if (slp_node) | |
9771b263 | 3044 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 3045 | } |
ebfd146a IR |
3046 | |
3047 | if (slp_node) | |
f18b55bd IR |
3048 | continue; |
3049 | ||
3050 | if (j == 0) | |
3051 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3052 | else | |
3053 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3054 | ||
3055 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3056 | } | |
b8698a0f | 3057 | |
9771b263 | 3058 | vec_oprnds.release (); |
ebfd146a IR |
3059 | return true; |
3060 | } | |
3061 | ||
9dc3f7de | 3062 | |
1107f3ae IR |
3063 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
3064 | either as shift by a scalar or by a vector. */ | |
3065 | ||
3066 | bool | |
3067 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
3068 | { | |
3069 | ||
3070 | enum machine_mode vec_mode; | |
3071 | optab optab; | |
3072 | int icode; | |
3073 | tree vectype; | |
3074 | ||
3075 | vectype = get_vectype_for_scalar_type (scalar_type); | |
3076 | if (!vectype) | |
3077 | return false; | |
3078 | ||
3079 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
3080 | if (!optab | |
3081 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
3082 | { | |
3083 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
3084 | if (!optab | |
3085 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
3086 | == CODE_FOR_nothing)) | |
3087 | return false; | |
3088 | } | |
3089 | ||
3090 | vec_mode = TYPE_MODE (vectype); | |
3091 | icode = (int) optab_handler (optab, vec_mode); | |
3092 | if (icode == CODE_FOR_nothing) | |
3093 | return false; | |
3094 | ||
3095 | return true; | |
3096 | } | |
3097 | ||
3098 | ||
9dc3f7de IR |
3099 | /* Function vectorizable_shift. |
3100 | ||
3101 | Check if STMT performs a shift operation that can be vectorized. | |
3102 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3103 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3104 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3105 | ||
3106 | static bool | |
3107 | vectorizable_shift (gimple stmt, gimple_stmt_iterator *gsi, | |
3108 | gimple *vec_stmt, slp_tree slp_node) | |
3109 | { | |
3110 | tree vec_dest; | |
3111 | tree scalar_dest; | |
3112 | tree op0, op1 = NULL; | |
3113 | tree vec_oprnd1 = NULL_TREE; | |
3114 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3115 | tree vectype; | |
3116 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3117 | enum tree_code code; | |
3118 | enum machine_mode vec_mode; | |
3119 | tree new_temp; | |
3120 | optab optab; | |
3121 | int icode; | |
3122 | enum machine_mode optab_op2_mode; | |
3123 | tree def; | |
3124 | gimple def_stmt; | |
3125 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
3126 | gimple new_stmt = NULL; | |
3127 | stmt_vec_info prev_stmt_info; | |
3128 | int nunits_in; | |
3129 | int nunits_out; | |
3130 | tree vectype_out; | |
cede2577 | 3131 | tree op1_vectype; |
9dc3f7de IR |
3132 | int ncopies; |
3133 | int j, i; | |
6e1aa848 DN |
3134 | vec<tree> vec_oprnds0 = vNULL; |
3135 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
3136 | tree vop0, vop1; |
3137 | unsigned int k; | |
49eab32e | 3138 | bool scalar_shift_arg = true; |
9dc3f7de IR |
3139 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3140 | int vf; | |
3141 | ||
3142 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
3143 | return false; | |
3144 | ||
3145 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
3146 | return false; | |
3147 | ||
3148 | /* Is STMT a vectorizable binary/unary operation? */ | |
3149 | if (!is_gimple_assign (stmt)) | |
3150 | return false; | |
3151 | ||
3152 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3153 | return false; | |
3154 | ||
3155 | code = gimple_assign_rhs_code (stmt); | |
3156 | ||
3157 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
3158 | || code == RROTATE_EXPR)) | |
3159 | return false; | |
3160 | ||
3161 | scalar_dest = gimple_assign_lhs (stmt); | |
3162 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
7b7b1813 RG |
3163 | if (TYPE_PRECISION (TREE_TYPE (scalar_dest)) |
3164 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
3165 | { | |
73fbfcad | 3166 | if (dump_enabled_p ()) |
78c60e3d | 3167 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3168 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
3169 | return false; |
3170 | } | |
9dc3f7de IR |
3171 | |
3172 | op0 = gimple_assign_rhs1 (stmt); | |
24ee1384 | 3173 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
9dc3f7de IR |
3174 | &def_stmt, &def, &dt[0], &vectype)) |
3175 | { | |
73fbfcad | 3176 | if (dump_enabled_p ()) |
78c60e3d | 3177 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3178 | "use not simple.\n"); |
9dc3f7de IR |
3179 | return false; |
3180 | } | |
3181 | /* If op0 is an external or constant def use a vector type with | |
3182 | the same size as the output vector type. */ | |
3183 | if (!vectype) | |
3184 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
3185 | if (vec_stmt) | |
3186 | gcc_assert (vectype); | |
3187 | if (!vectype) | |
3188 | { | |
73fbfcad | 3189 | if (dump_enabled_p ()) |
78c60e3d | 3190 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3191 | "no vectype for scalar type\n"); |
9dc3f7de IR |
3192 | return false; |
3193 | } | |
3194 | ||
3195 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
3196 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
3197 | if (nunits_out != nunits_in) | |
3198 | return false; | |
3199 | ||
3200 | op1 = gimple_assign_rhs2 (stmt); | |
24ee1384 IR |
3201 | if (!vect_is_simple_use_1 (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3202 | &def, &dt[1], &op1_vectype)) | |
9dc3f7de | 3203 | { |
73fbfcad | 3204 | if (dump_enabled_p ()) |
78c60e3d | 3205 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3206 | "use not simple.\n"); |
9dc3f7de IR |
3207 | return false; |
3208 | } | |
3209 | ||
3210 | if (loop_vinfo) | |
3211 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
3212 | else | |
3213 | vf = 1; | |
3214 | ||
3215 | /* Multiple types in SLP are handled by creating the appropriate number of | |
3216 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3217 | case of SLP. */ | |
437f4a00 | 3218 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
9dc3f7de IR |
3219 | ncopies = 1; |
3220 | else | |
3221 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
3222 | ||
3223 | gcc_assert (ncopies >= 1); | |
3224 | ||
3225 | /* Determine whether the shift amount is a vector, or scalar. If the | |
3226 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
3227 | ||
49eab32e JJ |
3228 | if (dt[1] == vect_internal_def && !slp_node) |
3229 | scalar_shift_arg = false; | |
3230 | else if (dt[1] == vect_constant_def | |
3231 | || dt[1] == vect_external_def | |
3232 | || dt[1] == vect_internal_def) | |
3233 | { | |
3234 | /* In SLP, need to check whether the shift count is the same, | |
3235 | in loops if it is a constant or invariant, it is always | |
3236 | a scalar shift. */ | |
3237 | if (slp_node) | |
3238 | { | |
9771b263 | 3239 | vec<gimple> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
49eab32e JJ |
3240 | gimple slpstmt; |
3241 | ||
9771b263 | 3242 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
3243 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
3244 | scalar_shift_arg = false; | |
3245 | } | |
3246 | } | |
3247 | else | |
3248 | { | |
73fbfcad | 3249 | if (dump_enabled_p ()) |
78c60e3d | 3250 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3251 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
3252 | return false; |
3253 | } | |
3254 | ||
9dc3f7de | 3255 | /* Vector shifted by vector. */ |
49eab32e | 3256 | if (!scalar_shift_arg) |
9dc3f7de IR |
3257 | { |
3258 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 3259 | if (dump_enabled_p ()) |
78c60e3d | 3260 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3261 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 3262 | |
aa948027 JJ |
3263 | if (!op1_vectype) |
3264 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
3265 | if (op1_vectype == NULL_TREE | |
3266 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 3267 | { |
73fbfcad | 3268 | if (dump_enabled_p ()) |
78c60e3d SS |
3269 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3270 | "unusable type for last operand in" | |
e645e942 | 3271 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
3272 | return false; |
3273 | } | |
9dc3f7de IR |
3274 | } |
3275 | /* See if the machine has a vector shifted by scalar insn and if not | |
3276 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 3277 | else |
9dc3f7de IR |
3278 | { |
3279 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
3280 | if (optab | |
3281 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
3282 | { | |
73fbfcad | 3283 | if (dump_enabled_p ()) |
78c60e3d | 3284 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3285 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
3286 | } |
3287 | else | |
3288 | { | |
3289 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
3290 | if (optab | |
3291 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
3292 | != CODE_FOR_nothing)) | |
3293 | { | |
49eab32e JJ |
3294 | scalar_shift_arg = false; |
3295 | ||
73fbfcad | 3296 | if (dump_enabled_p ()) |
78c60e3d | 3297 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3298 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
3299 | |
3300 | /* Unlike the other binary operators, shifts/rotates have | |
3301 | the rhs being int, instead of the same type as the lhs, | |
3302 | so make sure the scalar is the right type if we are | |
aa948027 | 3303 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
3304 | if (dt[1] == vect_constant_def) |
3305 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
3306 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
3307 | TREE_TYPE (op1))) | |
3308 | { | |
3309 | if (slp_node | |
3310 | && TYPE_MODE (TREE_TYPE (vectype)) | |
3311 | != TYPE_MODE (TREE_TYPE (op1))) | |
3312 | { | |
73fbfcad | 3313 | if (dump_enabled_p ()) |
78c60e3d SS |
3314 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3315 | "unusable type for last operand in" | |
e645e942 | 3316 | " vector/vector shift/rotate.\n"); |
aa948027 JJ |
3317 | return false; |
3318 | } | |
3319 | if (vec_stmt && !slp_node) | |
3320 | { | |
3321 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
3322 | op1 = vect_init_vector (stmt, op1, | |
3323 | TREE_TYPE (vectype), NULL); | |
3324 | } | |
3325 | } | |
9dc3f7de IR |
3326 | } |
3327 | } | |
3328 | } | |
9dc3f7de IR |
3329 | |
3330 | /* Supportable by target? */ | |
3331 | if (!optab) | |
3332 | { | |
73fbfcad | 3333 | if (dump_enabled_p ()) |
78c60e3d | 3334 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3335 | "no optab.\n"); |
9dc3f7de IR |
3336 | return false; |
3337 | } | |
3338 | vec_mode = TYPE_MODE (vectype); | |
3339 | icode = (int) optab_handler (optab, vec_mode); | |
3340 | if (icode == CODE_FOR_nothing) | |
3341 | { | |
73fbfcad | 3342 | if (dump_enabled_p ()) |
78c60e3d | 3343 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3344 | "op not supported by target.\n"); |
9dc3f7de IR |
3345 | /* Check only during analysis. */ |
3346 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
3347 | || (vf < vect_min_worthwhile_factor (code) | |
3348 | && !vec_stmt)) | |
3349 | return false; | |
73fbfcad | 3350 | if (dump_enabled_p ()) |
e645e942 TJ |
3351 | dump_printf_loc (MSG_NOTE, vect_location, |
3352 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
3353 | } |
3354 | ||
3355 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
3356 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
3357 | && vf < vect_min_worthwhile_factor (code) | |
3358 | && !vec_stmt) | |
3359 | { | |
73fbfcad | 3360 | if (dump_enabled_p ()) |
78c60e3d | 3361 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3362 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
3363 | return false; |
3364 | } | |
3365 | ||
3366 | if (!vec_stmt) /* transformation not required. */ | |
3367 | { | |
3368 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 3369 | if (dump_enabled_p ()) |
e645e942 TJ |
3370 | dump_printf_loc (MSG_NOTE, vect_location, |
3371 | "=== vectorizable_shift ===\n"); | |
c3e7ee41 | 3372 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
9dc3f7de IR |
3373 | return true; |
3374 | } | |
3375 | ||
3376 | /** Transform. **/ | |
3377 | ||
73fbfcad | 3378 | if (dump_enabled_p ()) |
78c60e3d | 3379 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3380 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
3381 | |
3382 | /* Handle def. */ | |
3383 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3384 | ||
9dc3f7de IR |
3385 | prev_stmt_info = NULL; |
3386 | for (j = 0; j < ncopies; j++) | |
3387 | { | |
3388 | /* Handle uses. */ | |
3389 | if (j == 0) | |
3390 | { | |
3391 | if (scalar_shift_arg) | |
3392 | { | |
3393 | /* Vector shl and shr insn patterns can be defined with scalar | |
3394 | operand 2 (shift operand). In this case, use constant or loop | |
3395 | invariant op1 directly, without extending it to vector mode | |
3396 | first. */ | |
3397 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
3398 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
3399 | { | |
73fbfcad | 3400 | if (dump_enabled_p ()) |
78c60e3d | 3401 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3402 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 3403 | vec_oprnd1 = op1; |
8930f723 | 3404 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 3405 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
3406 | if (slp_node) |
3407 | { | |
3408 | /* Store vec_oprnd1 for every vector stmt to be created | |
3409 | for SLP_NODE. We check during the analysis that all | |
3410 | the shift arguments are the same. | |
3411 | TODO: Allow different constants for different vector | |
3412 | stmts generated for an SLP instance. */ | |
3413 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 3414 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
3415 | } |
3416 | } | |
3417 | } | |
3418 | ||
3419 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
3420 | (a special case for certain kind of vector shifts); otherwise, | |
3421 | operand 1 should be of a vector type (the usual case). */ | |
3422 | if (vec_oprnd1) | |
3423 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
d092494c | 3424 | slp_node, -1); |
9dc3f7de IR |
3425 | else |
3426 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
d092494c | 3427 | slp_node, -1); |
9dc3f7de IR |
3428 | } |
3429 | else | |
3430 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
3431 | ||
3432 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3433 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 3434 | { |
9771b263 | 3435 | vop1 = vec_oprnds1[i]; |
9dc3f7de IR |
3436 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); |
3437 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3438 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3439 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3440 | if (slp_node) | |
9771b263 | 3441 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
3442 | } |
3443 | ||
3444 | if (slp_node) | |
3445 | continue; | |
3446 | ||
3447 | if (j == 0) | |
3448 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3449 | else | |
3450 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3451 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3452 | } | |
3453 | ||
9771b263 DN |
3454 | vec_oprnds0.release (); |
3455 | vec_oprnds1.release (); | |
9dc3f7de IR |
3456 | |
3457 | return true; | |
3458 | } | |
3459 | ||
3460 | ||
5deb57cb JJ |
3461 | static tree permute_vec_elements (tree, tree, tree, gimple, |
3462 | gimple_stmt_iterator *); | |
3463 | ||
3464 | ||
ebfd146a IR |
3465 | /* Function vectorizable_operation. |
3466 | ||
16949072 RG |
3467 | Check if STMT performs a binary, unary or ternary operation that can |
3468 | be vectorized. | |
b8698a0f | 3469 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
3470 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3471 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3472 | ||
3473 | static bool | |
3474 | vectorizable_operation (gimple stmt, gimple_stmt_iterator *gsi, | |
3475 | gimple *vec_stmt, slp_tree slp_node) | |
3476 | { | |
00f07b86 | 3477 | tree vec_dest; |
ebfd146a | 3478 | tree scalar_dest; |
16949072 | 3479 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 3480 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 3481 | tree vectype; |
ebfd146a IR |
3482 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
3483 | enum tree_code code; | |
3484 | enum machine_mode vec_mode; | |
3485 | tree new_temp; | |
3486 | int op_type; | |
00f07b86 | 3487 | optab optab; |
ebfd146a | 3488 | int icode; |
ebfd146a IR |
3489 | tree def; |
3490 | gimple def_stmt; | |
16949072 RG |
3491 | enum vect_def_type dt[3] |
3492 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
ebfd146a IR |
3493 | gimple new_stmt = NULL; |
3494 | stmt_vec_info prev_stmt_info; | |
b690cc0f | 3495 | int nunits_in; |
ebfd146a IR |
3496 | int nunits_out; |
3497 | tree vectype_out; | |
3498 | int ncopies; | |
3499 | int j, i; | |
6e1aa848 DN |
3500 | vec<tree> vec_oprnds0 = vNULL; |
3501 | vec<tree> vec_oprnds1 = vNULL; | |
3502 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 3503 | tree vop0, vop1, vop2; |
a70d6342 IR |
3504 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3505 | int vf; | |
3506 | ||
a70d6342 | 3507 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3508 | return false; |
3509 | ||
8644a673 | 3510 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3511 | return false; |
3512 | ||
3513 | /* Is STMT a vectorizable binary/unary operation? */ | |
3514 | if (!is_gimple_assign (stmt)) | |
3515 | return false; | |
3516 | ||
3517 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3518 | return false; | |
3519 | ||
ebfd146a IR |
3520 | code = gimple_assign_rhs_code (stmt); |
3521 | ||
3522 | /* For pointer addition, we should use the normal plus for | |
3523 | the vector addition. */ | |
3524 | if (code == POINTER_PLUS_EXPR) | |
3525 | code = PLUS_EXPR; | |
3526 | ||
3527 | /* Support only unary or binary operations. */ | |
3528 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 3529 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 3530 | { |
73fbfcad | 3531 | if (dump_enabled_p ()) |
78c60e3d | 3532 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3533 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 3534 | op_type); |
ebfd146a IR |
3535 | return false; |
3536 | } | |
3537 | ||
b690cc0f RG |
3538 | scalar_dest = gimple_assign_lhs (stmt); |
3539 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
3540 | ||
7b7b1813 RG |
3541 | /* Most operations cannot handle bit-precision types without extra |
3542 | truncations. */ | |
3543 | if ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
3544 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
3545 | /* Exception are bitwise binary operations. */ | |
3546 | && code != BIT_IOR_EXPR | |
3547 | && code != BIT_XOR_EXPR | |
3548 | && code != BIT_AND_EXPR) | |
3549 | { | |
73fbfcad | 3550 | if (dump_enabled_p ()) |
78c60e3d | 3551 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3552 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
3553 | return false; |
3554 | } | |
3555 | ||
ebfd146a | 3556 | op0 = gimple_assign_rhs1 (stmt); |
24ee1384 | 3557 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
b690cc0f | 3558 | &def_stmt, &def, &dt[0], &vectype)) |
ebfd146a | 3559 | { |
73fbfcad | 3560 | if (dump_enabled_p ()) |
78c60e3d | 3561 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3562 | "use not simple.\n"); |
ebfd146a IR |
3563 | return false; |
3564 | } | |
b690cc0f RG |
3565 | /* If op0 is an external or constant def use a vector type with |
3566 | the same size as the output vector type. */ | |
3567 | if (!vectype) | |
3568 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
3569 | if (vec_stmt) |
3570 | gcc_assert (vectype); | |
3571 | if (!vectype) | |
3572 | { | |
73fbfcad | 3573 | if (dump_enabled_p ()) |
7d8930a0 | 3574 | { |
78c60e3d SS |
3575 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3576 | "no vectype for scalar type "); | |
3577 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
3578 | TREE_TYPE (op0)); | |
e645e942 | 3579 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
3580 | } |
3581 | ||
3582 | return false; | |
3583 | } | |
b690cc0f RG |
3584 | |
3585 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
3586 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
3587 | if (nunits_out != nunits_in) | |
3588 | return false; | |
ebfd146a | 3589 | |
16949072 | 3590 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
3591 | { |
3592 | op1 = gimple_assign_rhs2 (stmt); | |
24ee1384 IR |
3593 | if (!vect_is_simple_use (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3594 | &def, &dt[1])) | |
ebfd146a | 3595 | { |
73fbfcad | 3596 | if (dump_enabled_p ()) |
78c60e3d | 3597 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3598 | "use not simple.\n"); |
ebfd146a IR |
3599 | return false; |
3600 | } | |
3601 | } | |
16949072 RG |
3602 | if (op_type == ternary_op) |
3603 | { | |
3604 | op2 = gimple_assign_rhs3 (stmt); | |
24ee1384 IR |
3605 | if (!vect_is_simple_use (op2, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3606 | &def, &dt[2])) | |
16949072 | 3607 | { |
73fbfcad | 3608 | if (dump_enabled_p ()) |
78c60e3d | 3609 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3610 | "use not simple.\n"); |
16949072 RG |
3611 | return false; |
3612 | } | |
3613 | } | |
ebfd146a | 3614 | |
b690cc0f RG |
3615 | if (loop_vinfo) |
3616 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
3617 | else | |
3618 | vf = 1; | |
3619 | ||
3620 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 3621 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 3622 | case of SLP. */ |
437f4a00 | 3623 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
b690cc0f RG |
3624 | ncopies = 1; |
3625 | else | |
3626 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
3627 | ||
3628 | gcc_assert (ncopies >= 1); | |
3629 | ||
9dc3f7de | 3630 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
3631 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
3632 | || code == RROTATE_EXPR) | |
9dc3f7de | 3633 | return false; |
ebfd146a | 3634 | |
ebfd146a | 3635 | /* Supportable by target? */ |
00f07b86 RH |
3636 | |
3637 | vec_mode = TYPE_MODE (vectype); | |
3638 | if (code == MULT_HIGHPART_EXPR) | |
ebfd146a | 3639 | { |
00f07b86 | 3640 | if (can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype))) |
dee54b6e | 3641 | icode = LAST_INSN_CODE; |
00f07b86 RH |
3642 | else |
3643 | icode = CODE_FOR_nothing; | |
ebfd146a | 3644 | } |
00f07b86 RH |
3645 | else |
3646 | { | |
3647 | optab = optab_for_tree_code (code, vectype, optab_default); | |
3648 | if (!optab) | |
5deb57cb | 3649 | { |
73fbfcad | 3650 | if (dump_enabled_p ()) |
78c60e3d | 3651 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3652 | "no optab.\n"); |
00f07b86 | 3653 | return false; |
5deb57cb | 3654 | } |
00f07b86 | 3655 | icode = (int) optab_handler (optab, vec_mode); |
5deb57cb JJ |
3656 | } |
3657 | ||
ebfd146a IR |
3658 | if (icode == CODE_FOR_nothing) |
3659 | { | |
73fbfcad | 3660 | if (dump_enabled_p ()) |
78c60e3d | 3661 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3662 | "op not supported by target.\n"); |
ebfd146a IR |
3663 | /* Check only during analysis. */ |
3664 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
5deb57cb | 3665 | || (!vec_stmt && vf < vect_min_worthwhile_factor (code))) |
ebfd146a | 3666 | return false; |
73fbfcad | 3667 | if (dump_enabled_p ()) |
e645e942 TJ |
3668 | dump_printf_loc (MSG_NOTE, vect_location, |
3669 | "proceeding using word mode.\n"); | |
383d9c83 IR |
3670 | } |
3671 | ||
4a00c761 | 3672 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
3673 | if (!VECTOR_MODE_P (vec_mode) |
3674 | && !vec_stmt | |
3675 | && vf < vect_min_worthwhile_factor (code)) | |
7d8930a0 | 3676 | { |
73fbfcad | 3677 | if (dump_enabled_p ()) |
78c60e3d | 3678 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3679 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 3680 | return false; |
7d8930a0 | 3681 | } |
ebfd146a | 3682 | |
ebfd146a IR |
3683 | if (!vec_stmt) /* transformation not required. */ |
3684 | { | |
4a00c761 | 3685 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 3686 | if (dump_enabled_p ()) |
78c60e3d | 3687 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3688 | "=== vectorizable_operation ===\n"); |
c3e7ee41 | 3689 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
3690 | return true; |
3691 | } | |
3692 | ||
3693 | /** Transform. **/ | |
3694 | ||
73fbfcad | 3695 | if (dump_enabled_p ()) |
78c60e3d | 3696 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3697 | "transform binary/unary operation.\n"); |
383d9c83 | 3698 | |
ebfd146a | 3699 | /* Handle def. */ |
00f07b86 | 3700 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 3701 | |
ebfd146a IR |
3702 | /* In case the vectorization factor (VF) is bigger than the number |
3703 | of elements that we can fit in a vectype (nunits), we have to generate | |
3704 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
3705 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
3706 | from one copy of the vector stmt to the next, in the field | |
3707 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
3708 | stages to find the correct vector defs to be used when vectorizing | |
3709 | stmts that use the defs of the current stmt. The example below | |
3710 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
3711 | we need to create 4 vectorized stmts): | |
3712 | ||
3713 | before vectorization: | |
3714 | RELATED_STMT VEC_STMT | |
3715 | S1: x = memref - - | |
3716 | S2: z = x + 1 - - | |
3717 | ||
3718 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
3719 | there): | |
3720 | RELATED_STMT VEC_STMT | |
3721 | VS1_0: vx0 = memref0 VS1_1 - | |
3722 | VS1_1: vx1 = memref1 VS1_2 - | |
3723 | VS1_2: vx2 = memref2 VS1_3 - | |
3724 | VS1_3: vx3 = memref3 - - | |
3725 | S1: x = load - VS1_0 | |
3726 | S2: z = x + 1 - - | |
3727 | ||
3728 | step2: vectorize stmt S2 (done here): | |
3729 | To vectorize stmt S2 we first need to find the relevant vector | |
3730 | def for the first operand 'x'. This is, as usual, obtained from | |
3731 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
3732 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
3733 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
3734 | the vector stmt VS2_0, and as usual, record it in the | |
3735 | STMT_VINFO_VEC_STMT of stmt S2. | |
3736 | When creating the second copy (VS2_1), we obtain the relevant vector | |
3737 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
3738 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
3739 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
3740 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
3741 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
3742 | chain of stmts and pointers: | |
3743 | RELATED_STMT VEC_STMT | |
3744 | VS1_0: vx0 = memref0 VS1_1 - | |
3745 | VS1_1: vx1 = memref1 VS1_2 - | |
3746 | VS1_2: vx2 = memref2 VS1_3 - | |
3747 | VS1_3: vx3 = memref3 - - | |
3748 | S1: x = load - VS1_0 | |
3749 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
3750 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
3751 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
3752 | VS2_3: vz3 = vx3 + v1 - - | |
3753 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
3754 | |
3755 | prev_stmt_info = NULL; | |
3756 | for (j = 0; j < ncopies; j++) | |
3757 | { | |
3758 | /* Handle uses. */ | |
3759 | if (j == 0) | |
4a00c761 JJ |
3760 | { |
3761 | if (op_type == binary_op || op_type == ternary_op) | |
3762 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
3763 | slp_node, -1); | |
3764 | else | |
3765 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
3766 | slp_node, -1); | |
3767 | if (op_type == ternary_op) | |
36ba4aae | 3768 | { |
9771b263 DN |
3769 | vec_oprnds2.create (1); |
3770 | vec_oprnds2.quick_push (vect_get_vec_def_for_operand (op2, | |
3771 | stmt, | |
3772 | NULL)); | |
36ba4aae | 3773 | } |
4a00c761 | 3774 | } |
ebfd146a | 3775 | else |
4a00c761 JJ |
3776 | { |
3777 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
3778 | if (op_type == ternary_op) | |
3779 | { | |
9771b263 DN |
3780 | tree vec_oprnd = vec_oprnds2.pop (); |
3781 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
3782 | vec_oprnd)); | |
4a00c761 JJ |
3783 | } |
3784 | } | |
3785 | ||
3786 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3787 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 3788 | { |
4a00c761 | 3789 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 3790 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 3791 | vop2 = ((op_type == ternary_op) |
9771b263 | 3792 | ? vec_oprnds2[i] : NULL_TREE); |
73804b12 RG |
3793 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, |
3794 | vop0, vop1, vop2); | |
4a00c761 JJ |
3795 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3796 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3797 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3798 | if (slp_node) | |
9771b263 | 3799 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
3800 | } |
3801 | ||
4a00c761 JJ |
3802 | if (slp_node) |
3803 | continue; | |
3804 | ||
3805 | if (j == 0) | |
3806 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3807 | else | |
3808 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3809 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
3810 | } |
3811 | ||
9771b263 DN |
3812 | vec_oprnds0.release (); |
3813 | vec_oprnds1.release (); | |
3814 | vec_oprnds2.release (); | |
ebfd146a | 3815 | |
ebfd146a IR |
3816 | return true; |
3817 | } | |
3818 | ||
c716e67f XDL |
3819 | /* A helper function to ensure data reference DR's base alignment |
3820 | for STMT_INFO. */ | |
3821 | ||
3822 | static void | |
3823 | ensure_base_align (stmt_vec_info stmt_info, struct data_reference *dr) | |
3824 | { | |
3825 | if (!dr->aux) | |
3826 | return; | |
3827 | ||
3828 | if (((dataref_aux *)dr->aux)->base_misaligned) | |
3829 | { | |
3830 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3831 | tree base_decl = ((dataref_aux *)dr->aux)->base_decl; | |
3832 | ||
3833 | DECL_ALIGN (base_decl) = TYPE_ALIGN (vectype); | |
3834 | DECL_USER_ALIGN (base_decl) = 1; | |
3835 | ((dataref_aux *)dr->aux)->base_misaligned = false; | |
3836 | } | |
3837 | } | |
3838 | ||
ebfd146a IR |
3839 | |
3840 | /* Function vectorizable_store. | |
3841 | ||
b8698a0f L |
3842 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
3843 | can be vectorized. | |
3844 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
3845 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3846 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3847 | ||
3848 | static bool | |
3849 | vectorizable_store (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
c716e67f | 3850 | slp_tree slp_node) |
ebfd146a IR |
3851 | { |
3852 | tree scalar_dest; | |
3853 | tree data_ref; | |
3854 | tree op; | |
3855 | tree vec_oprnd = NULL_TREE; | |
3856 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3857 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
3858 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
272c6793 | 3859 | tree elem_type; |
ebfd146a | 3860 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 3861 | struct loop *loop = NULL; |
ebfd146a IR |
3862 | enum machine_mode vec_mode; |
3863 | tree dummy; | |
3864 | enum dr_alignment_support alignment_support_scheme; | |
3865 | tree def; | |
3866 | gimple def_stmt; | |
3867 | enum vect_def_type dt; | |
3868 | stmt_vec_info prev_stmt_info = NULL; | |
3869 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 3870 | tree dataref_offset = NULL_TREE; |
fef4d2b3 | 3871 | gimple ptr_incr = NULL; |
ebfd146a IR |
3872 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
3873 | int ncopies; | |
3874 | int j; | |
3875 | gimple next_stmt, first_stmt = NULL; | |
0d0293ac | 3876 | bool grouped_store = false; |
272c6793 | 3877 | bool store_lanes_p = false; |
ebfd146a | 3878 | unsigned int group_size, i; |
6e1aa848 DN |
3879 | vec<tree> dr_chain = vNULL; |
3880 | vec<tree> oprnds = vNULL; | |
3881 | vec<tree> result_chain = vNULL; | |
ebfd146a | 3882 | bool inv_p; |
6e1aa848 | 3883 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 3884 | bool slp = (slp_node != NULL); |
ebfd146a | 3885 | unsigned int vec_num; |
a70d6342 | 3886 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
272c6793 | 3887 | tree aggr_type; |
a70d6342 IR |
3888 | |
3889 | if (loop_vinfo) | |
3890 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
3891 | |
3892 | /* Multiple types in SLP are handled by creating the appropriate number of | |
3893 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3894 | case of SLP. */ | |
437f4a00 | 3895 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
3896 | ncopies = 1; |
3897 | else | |
3898 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3899 | ||
3900 | gcc_assert (ncopies >= 1); | |
3901 | ||
3902 | /* FORNOW. This restriction should be relaxed. */ | |
a70d6342 | 3903 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) |
ebfd146a | 3904 | { |
73fbfcad | 3905 | if (dump_enabled_p ()) |
78c60e3d | 3906 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3907 | "multiple types in nested loop.\n"); |
ebfd146a IR |
3908 | return false; |
3909 | } | |
3910 | ||
a70d6342 | 3911 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3912 | return false; |
3913 | ||
8644a673 | 3914 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3915 | return false; |
3916 | ||
3917 | /* Is vectorizable store? */ | |
3918 | ||
3919 | if (!is_gimple_assign (stmt)) | |
3920 | return false; | |
3921 | ||
3922 | scalar_dest = gimple_assign_lhs (stmt); | |
ab0ef706 JJ |
3923 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
3924 | && is_pattern_stmt_p (stmt_info)) | |
3925 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
ebfd146a | 3926 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
38000232 | 3927 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF |
ebfd146a | 3928 | && TREE_CODE (scalar_dest) != INDIRECT_REF |
e9dbe7bb IR |
3929 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
3930 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
3931 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
3932 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
3933 | return false; |
3934 | ||
3935 | gcc_assert (gimple_assign_single_p (stmt)); | |
3936 | op = gimple_assign_rhs1 (stmt); | |
24ee1384 IR |
3937 | if (!vect_is_simple_use (op, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3938 | &def, &dt)) | |
ebfd146a | 3939 | { |
73fbfcad | 3940 | if (dump_enabled_p ()) |
78c60e3d | 3941 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3942 | "use not simple.\n"); |
ebfd146a IR |
3943 | return false; |
3944 | } | |
3945 | ||
272c6793 | 3946 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 3947 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 3948 | |
ebfd146a IR |
3949 | /* FORNOW. In some cases can vectorize even if data-type not supported |
3950 | (e.g. - array initialization with 0). */ | |
947131ba | 3951 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
3952 | return false; |
3953 | ||
3954 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
3955 | return false; | |
3956 | ||
a7ce6ec3 RG |
3957 | if (tree_int_cst_compare (loop && nested_in_vect_loop_p (loop, stmt) |
3958 | ? STMT_VINFO_DR_STEP (stmt_info) : DR_STEP (dr), | |
3959 | size_zero_node) < 0) | |
a1e53f3f | 3960 | { |
73fbfcad | 3961 | if (dump_enabled_p ()) |
78c60e3d | 3962 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3963 | "negative step for store.\n"); |
a1e53f3f L |
3964 | return false; |
3965 | } | |
3966 | ||
0d0293ac | 3967 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 3968 | { |
0d0293ac | 3969 | grouped_store = true; |
e14c1050 | 3970 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
b602d918 RS |
3971 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
3972 | { | |
e14c1050 | 3973 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
3974 | if (vect_store_lanes_supported (vectype, group_size)) |
3975 | store_lanes_p = true; | |
0d0293ac | 3976 | else if (!vect_grouped_store_supported (vectype, group_size)) |
b602d918 RS |
3977 | return false; |
3978 | } | |
b8698a0f | 3979 | |
ebfd146a IR |
3980 | if (first_stmt == stmt) |
3981 | { | |
3982 | /* STMT is the leader of the group. Check the operands of all the | |
3983 | stmts of the group. */ | |
e14c1050 | 3984 | next_stmt = GROUP_NEXT_ELEMENT (stmt_info); |
ebfd146a IR |
3985 | while (next_stmt) |
3986 | { | |
3987 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
3988 | op = gimple_assign_rhs1 (next_stmt); | |
24ee1384 IR |
3989 | if (!vect_is_simple_use (op, next_stmt, loop_vinfo, bb_vinfo, |
3990 | &def_stmt, &def, &dt)) | |
ebfd146a | 3991 | { |
73fbfcad | 3992 | if (dump_enabled_p ()) |
78c60e3d | 3993 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3994 | "use not simple.\n"); |
ebfd146a IR |
3995 | return false; |
3996 | } | |
e14c1050 | 3997 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
3998 | } |
3999 | } | |
4000 | } | |
4001 | ||
4002 | if (!vec_stmt) /* transformation not required. */ | |
4003 | { | |
4004 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
92345349 BS |
4005 | vect_model_store_cost (stmt_info, ncopies, store_lanes_p, dt, |
4006 | NULL, NULL, NULL); | |
ebfd146a IR |
4007 | return true; |
4008 | } | |
4009 | ||
4010 | /** Transform. **/ | |
4011 | ||
c716e67f XDL |
4012 | ensure_base_align (stmt_info, dr); |
4013 | ||
0d0293ac | 4014 | if (grouped_store) |
ebfd146a IR |
4015 | { |
4016 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 4017 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 4018 | |
e14c1050 | 4019 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
4020 | |
4021 | /* FORNOW */ | |
a70d6342 | 4022 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
4023 | |
4024 | /* We vectorize all the stmts of the interleaving group when we | |
4025 | reach the last stmt in the group. */ | |
e14c1050 IR |
4026 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
4027 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
4028 | && !slp) |
4029 | { | |
4030 | *vec_stmt = NULL; | |
4031 | return true; | |
4032 | } | |
4033 | ||
4034 | if (slp) | |
4b5caab7 | 4035 | { |
0d0293ac | 4036 | grouped_store = false; |
4b5caab7 IR |
4037 | /* VEC_NUM is the number of vect stmts to be created for this |
4038 | group. */ | |
4039 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 4040 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
4b5caab7 | 4041 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
d092494c | 4042 | op = gimple_assign_rhs1 (first_stmt); |
4b5caab7 | 4043 | } |
ebfd146a | 4044 | else |
4b5caab7 IR |
4045 | /* VEC_NUM is the number of vect stmts to be created for this |
4046 | group. */ | |
ebfd146a IR |
4047 | vec_num = group_size; |
4048 | } | |
b8698a0f | 4049 | else |
ebfd146a IR |
4050 | { |
4051 | first_stmt = stmt; | |
4052 | first_dr = dr; | |
4053 | group_size = vec_num = 1; | |
ebfd146a | 4054 | } |
b8698a0f | 4055 | |
73fbfcad | 4056 | if (dump_enabled_p ()) |
78c60e3d | 4057 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4058 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 4059 | |
9771b263 DN |
4060 | dr_chain.create (group_size); |
4061 | oprnds.create (group_size); | |
ebfd146a | 4062 | |
720f5239 | 4063 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 4064 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
4065 | /* Targets with store-lane instructions must not require explicit |
4066 | realignment. */ | |
4067 | gcc_assert (!store_lanes_p | |
4068 | || alignment_support_scheme == dr_aligned | |
4069 | || alignment_support_scheme == dr_unaligned_supported); | |
4070 | ||
4071 | if (store_lanes_p) | |
4072 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
4073 | else | |
4074 | aggr_type = vectype; | |
ebfd146a IR |
4075 | |
4076 | /* In case the vectorization factor (VF) is bigger than the number | |
4077 | of elements that we can fit in a vectype (nunits), we have to generate | |
4078 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 4079 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
4080 | vect_get_vec_def_for_copy_stmt. */ |
4081 | ||
0d0293ac | 4082 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
4083 | |
4084 | S1: &base + 2 = x2 | |
4085 | S2: &base = x0 | |
4086 | S3: &base + 1 = x1 | |
4087 | S4: &base + 3 = x3 | |
4088 | ||
4089 | We create vectorized stores starting from base address (the access of the | |
4090 | first stmt in the chain (S2 in the above example), when the last store stmt | |
4091 | of the chain (S4) is reached: | |
4092 | ||
4093 | VS1: &base = vx2 | |
4094 | VS2: &base + vec_size*1 = vx0 | |
4095 | VS3: &base + vec_size*2 = vx1 | |
4096 | VS4: &base + vec_size*3 = vx3 | |
4097 | ||
4098 | Then permutation statements are generated: | |
4099 | ||
3fcc1b55 JJ |
4100 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
4101 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 4102 | ... |
b8698a0f | 4103 | |
ebfd146a IR |
4104 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
4105 | (the order of the data-refs in the output of vect_permute_store_chain | |
4106 | corresponds to the order of scalar stmts in the interleaving chain - see | |
4107 | the documentation of vect_permute_store_chain()). | |
4108 | ||
4109 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 4110 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 4111 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 4112 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
4113 | */ |
4114 | ||
4115 | prev_stmt_info = NULL; | |
4116 | for (j = 0; j < ncopies; j++) | |
4117 | { | |
4118 | gimple new_stmt; | |
ebfd146a IR |
4119 | |
4120 | if (j == 0) | |
4121 | { | |
4122 | if (slp) | |
4123 | { | |
4124 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c IR |
4125 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
4126 | NULL, slp_node, -1); | |
ebfd146a | 4127 | |
9771b263 | 4128 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
4129 | } |
4130 | else | |
4131 | { | |
b8698a0f L |
4132 | /* For interleaved stores we collect vectorized defs for all the |
4133 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
4134 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
4135 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
4136 | ||
0d0293ac | 4137 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 4138 | OPRNDS are of size 1. */ |
b8698a0f | 4139 | next_stmt = first_stmt; |
ebfd146a IR |
4140 | for (i = 0; i < group_size; i++) |
4141 | { | |
b8698a0f L |
4142 | /* Since gaps are not supported for interleaved stores, |
4143 | GROUP_SIZE is the exact number of stmts in the chain. | |
4144 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
4145 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
4146 | iteration of the loop will be executed. */ |
4147 | gcc_assert (next_stmt | |
4148 | && gimple_assign_single_p (next_stmt)); | |
4149 | op = gimple_assign_rhs1 (next_stmt); | |
4150 | ||
b8698a0f | 4151 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt, |
ebfd146a | 4152 | NULL); |
9771b263 DN |
4153 | dr_chain.quick_push (vec_oprnd); |
4154 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 4155 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
4156 | } |
4157 | } | |
4158 | ||
4159 | /* We should have catched mismatched types earlier. */ | |
4160 | gcc_assert (useless_type_conversion_p (vectype, | |
4161 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
4162 | bool simd_lane_access_p |
4163 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
4164 | if (simd_lane_access_p | |
4165 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
4166 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
4167 | && integer_zerop (DR_OFFSET (first_dr)) | |
4168 | && integer_zerop (DR_INIT (first_dr)) | |
4169 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
4170 | get_alias_set (DR_REF (first_dr)))) | |
4171 | { | |
4172 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
4173 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
4174 | (DR_REF (first_dr)), 0); | |
8928eff3 | 4175 | inv_p = false; |
74bf76ed JJ |
4176 | } |
4177 | else | |
4178 | dataref_ptr | |
4179 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
4180 | simd_lane_access_p ? loop : NULL, | |
4181 | NULL_TREE, &dummy, gsi, &ptr_incr, | |
4182 | simd_lane_access_p, &inv_p); | |
a70d6342 | 4183 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 4184 | } |
b8698a0f | 4185 | else |
ebfd146a | 4186 | { |
b8698a0f L |
4187 | /* For interleaved stores we created vectorized defs for all the |
4188 | defs stored in OPRNDS in the previous iteration (previous copy). | |
4189 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
4190 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
4191 | next copy. | |
0d0293ac | 4192 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
4193 | OPRNDS are of size 1. */ |
4194 | for (i = 0; i < group_size; i++) | |
4195 | { | |
9771b263 | 4196 | op = oprnds[i]; |
24ee1384 IR |
4197 | vect_is_simple_use (op, NULL, loop_vinfo, bb_vinfo, &def_stmt, |
4198 | &def, &dt); | |
b8698a0f | 4199 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
9771b263 DN |
4200 | dr_chain[i] = vec_oprnd; |
4201 | oprnds[i] = vec_oprnd; | |
ebfd146a | 4202 | } |
74bf76ed JJ |
4203 | if (dataref_offset) |
4204 | dataref_offset | |
4205 | = int_const_binop (PLUS_EXPR, dataref_offset, | |
4206 | TYPE_SIZE_UNIT (aggr_type)); | |
4207 | else | |
4208 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
4209 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
4210 | } |
4211 | ||
272c6793 | 4212 | if (store_lanes_p) |
ebfd146a | 4213 | { |
272c6793 | 4214 | tree vec_array; |
267d3070 | 4215 | |
272c6793 RS |
4216 | /* Combine all the vectors into an array. */ |
4217 | vec_array = create_vector_array (vectype, vec_num); | |
4218 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 4219 | { |
9771b263 | 4220 | vec_oprnd = dr_chain[i]; |
272c6793 | 4221 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 4222 | } |
b8698a0f | 4223 | |
272c6793 RS |
4224 | /* Emit: |
4225 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
4226 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
4227 | new_stmt = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array); | |
4228 | gimple_call_set_lhs (new_stmt, data_ref); | |
267d3070 | 4229 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
4230 | } |
4231 | else | |
4232 | { | |
4233 | new_stmt = NULL; | |
0d0293ac | 4234 | if (grouped_store) |
272c6793 | 4235 | { |
b6b9227d JJ |
4236 | if (j == 0) |
4237 | result_chain.create (group_size); | |
272c6793 RS |
4238 | /* Permute. */ |
4239 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
4240 | &result_chain); | |
4241 | } | |
c2d7ab2a | 4242 | |
272c6793 RS |
4243 | next_stmt = first_stmt; |
4244 | for (i = 0; i < vec_num; i++) | |
4245 | { | |
644ffefd | 4246 | unsigned align, misalign; |
272c6793 RS |
4247 | |
4248 | if (i > 0) | |
4249 | /* Bump the vector pointer. */ | |
4250 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
4251 | stmt, NULL_TREE); | |
4252 | ||
4253 | if (slp) | |
9771b263 | 4254 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
4255 | else if (grouped_store) |
4256 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 4257 | vect_permute_store_chain(). */ |
9771b263 | 4258 | vec_oprnd = result_chain[i]; |
272c6793 RS |
4259 | |
4260 | data_ref = build2 (MEM_REF, TREE_TYPE (vec_oprnd), dataref_ptr, | |
74bf76ed JJ |
4261 | dataref_offset |
4262 | ? dataref_offset | |
4263 | : build_int_cst (reference_alias_ptr_type | |
4264 | (DR_REF (first_dr)), 0)); | |
644ffefd | 4265 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 | 4266 | if (aligned_access_p (first_dr)) |
644ffefd | 4267 | misalign = 0; |
272c6793 RS |
4268 | else if (DR_MISALIGNMENT (first_dr) == -1) |
4269 | { | |
4270 | TREE_TYPE (data_ref) | |
4271 | = build_aligned_type (TREE_TYPE (data_ref), | |
4272 | TYPE_ALIGN (elem_type)); | |
644ffefd MJ |
4273 | align = TYPE_ALIGN_UNIT (elem_type); |
4274 | misalign = 0; | |
272c6793 RS |
4275 | } |
4276 | else | |
4277 | { | |
4278 | TREE_TYPE (data_ref) | |
4279 | = build_aligned_type (TREE_TYPE (data_ref), | |
4280 | TYPE_ALIGN (elem_type)); | |
644ffefd | 4281 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 4282 | } |
74bf76ed JJ |
4283 | if (dataref_offset == NULL_TREE) |
4284 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
4285 | misalign); | |
c2d7ab2a | 4286 | |
272c6793 RS |
4287 | /* Arguments are ready. Create the new vector stmt. */ |
4288 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
4289 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
4290 | |
4291 | if (slp) | |
4292 | continue; | |
4293 | ||
e14c1050 | 4294 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
4295 | if (!next_stmt) |
4296 | break; | |
4297 | } | |
ebfd146a | 4298 | } |
1da0876c RS |
4299 | if (!slp) |
4300 | { | |
4301 | if (j == 0) | |
4302 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4303 | else | |
4304 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4305 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4306 | } | |
ebfd146a IR |
4307 | } |
4308 | ||
9771b263 DN |
4309 | dr_chain.release (); |
4310 | oprnds.release (); | |
4311 | result_chain.release (); | |
4312 | vec_oprnds.release (); | |
ebfd146a IR |
4313 | |
4314 | return true; | |
4315 | } | |
4316 | ||
aec7ae7d JJ |
4317 | /* Given a vector type VECTYPE and permutation SEL returns |
4318 | the VECTOR_CST mask that implements the permutation of the | |
4319 | vector elements. If that is impossible to do, returns NULL. */ | |
a1e53f3f | 4320 | |
3fcc1b55 JJ |
4321 | tree |
4322 | vect_gen_perm_mask (tree vectype, unsigned char *sel) | |
a1e53f3f | 4323 | { |
d2a12ae7 | 4324 | tree mask_elt_type, mask_type, mask_vec, *mask_elts; |
2635892a | 4325 | int i, nunits; |
a1e53f3f | 4326 | |
22e4dee7 | 4327 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
22e4dee7 RH |
4328 | |
4329 | if (!can_vec_perm_p (TYPE_MODE (vectype), false, sel)) | |
a1e53f3f L |
4330 | return NULL; |
4331 | ||
96f9265a RG |
4332 | mask_elt_type = lang_hooks.types.type_for_mode |
4333 | (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype))), 1); | |
22e4dee7 | 4334 | mask_type = get_vectype_for_scalar_type (mask_elt_type); |
a1e53f3f | 4335 | |
d2a12ae7 | 4336 | mask_elts = XALLOCAVEC (tree, nunits); |
aec7ae7d | 4337 | for (i = nunits - 1; i >= 0; i--) |
d2a12ae7 RG |
4338 | mask_elts[i] = build_int_cst (mask_elt_type, sel[i]); |
4339 | mask_vec = build_vector (mask_type, mask_elts); | |
a1e53f3f | 4340 | |
2635892a | 4341 | return mask_vec; |
a1e53f3f L |
4342 | } |
4343 | ||
aec7ae7d JJ |
4344 | /* Given a vector type VECTYPE returns the VECTOR_CST mask that implements |
4345 | reversal of the vector elements. If that is impossible to do, | |
4346 | returns NULL. */ | |
4347 | ||
4348 | static tree | |
4349 | perm_mask_for_reverse (tree vectype) | |
4350 | { | |
4351 | int i, nunits; | |
4352 | unsigned char *sel; | |
4353 | ||
4354 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4355 | sel = XALLOCAVEC (unsigned char, nunits); | |
4356 | ||
4357 | for (i = 0; i < nunits; ++i) | |
4358 | sel[i] = nunits - 1 - i; | |
4359 | ||
3fcc1b55 | 4360 | return vect_gen_perm_mask (vectype, sel); |
aec7ae7d JJ |
4361 | } |
4362 | ||
4363 | /* Given a vector variable X and Y, that was generated for the scalar | |
4364 | STMT, generate instructions to permute the vector elements of X and Y | |
4365 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
4366 | permuted vector variable. */ | |
a1e53f3f L |
4367 | |
4368 | static tree | |
aec7ae7d JJ |
4369 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple stmt, |
4370 | gimple_stmt_iterator *gsi) | |
a1e53f3f L |
4371 | { |
4372 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 4373 | tree perm_dest, data_ref; |
a1e53f3f L |
4374 | gimple perm_stmt; |
4375 | ||
a1e53f3f | 4376 | perm_dest = vect_create_destination_var (gimple_assign_lhs (stmt), vectype); |
aec7ae7d | 4377 | data_ref = make_ssa_name (perm_dest, NULL); |
a1e53f3f L |
4378 | |
4379 | /* Generate the permute statement. */ | |
73804b12 RG |
4380 | perm_stmt = gimple_build_assign_with_ops (VEC_PERM_EXPR, data_ref, |
4381 | x, y, mask_vec); | |
a1e53f3f L |
4382 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
4383 | ||
4384 | return data_ref; | |
4385 | } | |
4386 | ||
ebfd146a IR |
4387 | /* vectorizable_load. |
4388 | ||
b8698a0f L |
4389 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
4390 | can be vectorized. | |
4391 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4392 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4393 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4394 | ||
4395 | static bool | |
4396 | vectorizable_load (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
c716e67f | 4397 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a IR |
4398 | { |
4399 | tree scalar_dest; | |
4400 | tree vec_dest = NULL; | |
4401 | tree data_ref = NULL; | |
4402 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 4403 | stmt_vec_info prev_stmt_info; |
ebfd146a | 4404 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 4405 | struct loop *loop = NULL; |
ebfd146a | 4406 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 4407 | bool nested_in_vect_loop = false; |
c716e67f | 4408 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
ebfd146a | 4409 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
272c6793 | 4410 | tree elem_type; |
ebfd146a | 4411 | tree new_temp; |
947131ba | 4412 | enum machine_mode mode; |
ebfd146a IR |
4413 | gimple new_stmt = NULL; |
4414 | tree dummy; | |
4415 | enum dr_alignment_support alignment_support_scheme; | |
4416 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 4417 | tree dataref_offset = NULL_TREE; |
fef4d2b3 | 4418 | gimple ptr_incr = NULL; |
ebfd146a IR |
4419 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
4420 | int ncopies; | |
a64b9c26 | 4421 | int i, j, group_size, group_gap; |
ebfd146a IR |
4422 | tree msq = NULL_TREE, lsq; |
4423 | tree offset = NULL_TREE; | |
4424 | tree realignment_token = NULL_TREE; | |
4425 | gimple phi = NULL; | |
6e1aa848 | 4426 | vec<tree> dr_chain = vNULL; |
0d0293ac | 4427 | bool grouped_load = false; |
272c6793 | 4428 | bool load_lanes_p = false; |
ebfd146a | 4429 | gimple first_stmt; |
ebfd146a | 4430 | bool inv_p; |
319e6439 | 4431 | bool negative = false; |
ebfd146a IR |
4432 | bool compute_in_loop = false; |
4433 | struct loop *at_loop; | |
4434 | int vec_num; | |
4435 | bool slp = (slp_node != NULL); | |
4436 | bool slp_perm = false; | |
4437 | enum tree_code code; | |
a70d6342 IR |
4438 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
4439 | int vf; | |
272c6793 | 4440 | tree aggr_type; |
aec7ae7d JJ |
4441 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; |
4442 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
4443 | int gather_scale = 1; | |
4444 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
a70d6342 IR |
4445 | |
4446 | if (loop_vinfo) | |
4447 | { | |
4448 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
4449 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
4450 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4451 | } | |
4452 | else | |
3533e503 | 4453 | vf = 1; |
ebfd146a IR |
4454 | |
4455 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 4456 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 4457 | case of SLP. */ |
437f4a00 | 4458 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
4459 | ncopies = 1; |
4460 | else | |
4461 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4462 | ||
4463 | gcc_assert (ncopies >= 1); | |
4464 | ||
4465 | /* FORNOW. This restriction should be relaxed. */ | |
4466 | if (nested_in_vect_loop && ncopies > 1) | |
4467 | { | |
73fbfcad | 4468 | if (dump_enabled_p ()) |
78c60e3d | 4469 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4470 | "multiple types in nested loop.\n"); |
ebfd146a IR |
4471 | return false; |
4472 | } | |
4473 | ||
a70d6342 | 4474 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4475 | return false; |
4476 | ||
8644a673 | 4477 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
4478 | return false; |
4479 | ||
4480 | /* Is vectorizable load? */ | |
4481 | if (!is_gimple_assign (stmt)) | |
4482 | return false; | |
4483 | ||
4484 | scalar_dest = gimple_assign_lhs (stmt); | |
4485 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4486 | return false; | |
4487 | ||
4488 | code = gimple_assign_rhs_code (stmt); | |
4489 | if (code != ARRAY_REF | |
38000232 | 4490 | && code != BIT_FIELD_REF |
ebfd146a | 4491 | && code != INDIRECT_REF |
e9dbe7bb IR |
4492 | && code != COMPONENT_REF |
4493 | && code != IMAGPART_EXPR | |
70f34814 | 4494 | && code != REALPART_EXPR |
42373e0b RG |
4495 | && code != MEM_REF |
4496 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
ebfd146a IR |
4497 | return false; |
4498 | ||
4499 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
4500 | return false; | |
4501 | ||
7b7b1813 | 4502 | elem_type = TREE_TYPE (vectype); |
947131ba | 4503 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
4504 | |
4505 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
4506 | (e.g. - data copies). */ | |
947131ba | 4507 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 4508 | { |
73fbfcad | 4509 | if (dump_enabled_p ()) |
78c60e3d | 4510 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4511 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
4512 | return false; |
4513 | } | |
4514 | ||
ebfd146a | 4515 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 4516 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 4517 | { |
0d0293ac | 4518 | grouped_load = true; |
ebfd146a | 4519 | /* FORNOW */ |
aec7ae7d | 4520 | gcc_assert (! nested_in_vect_loop && !STMT_VINFO_GATHER_P (stmt_info)); |
ebfd146a | 4521 | |
e14c1050 | 4522 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
b602d918 RS |
4523 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
4524 | { | |
e14c1050 | 4525 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
4526 | if (vect_load_lanes_supported (vectype, group_size)) |
4527 | load_lanes_p = true; | |
0d0293ac | 4528 | else if (!vect_grouped_load_supported (vectype, group_size)) |
b602d918 RS |
4529 | return false; |
4530 | } | |
ebfd146a IR |
4531 | } |
4532 | ||
a1e53f3f | 4533 | |
aec7ae7d JJ |
4534 | if (STMT_VINFO_GATHER_P (stmt_info)) |
4535 | { | |
4536 | gimple def_stmt; | |
4537 | tree def; | |
4538 | gather_decl = vect_check_gather (stmt, loop_vinfo, &gather_base, | |
4539 | &gather_off, &gather_scale); | |
4540 | gcc_assert (gather_decl); | |
24ee1384 | 4541 | if (!vect_is_simple_use_1 (gather_off, NULL, loop_vinfo, bb_vinfo, |
aec7ae7d JJ |
4542 | &def_stmt, &def, &gather_dt, |
4543 | &gather_off_vectype)) | |
4544 | { | |
73fbfcad | 4545 | if (dump_enabled_p ()) |
78c60e3d | 4546 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4547 | "gather index use not simple.\n"); |
aec7ae7d JJ |
4548 | return false; |
4549 | } | |
4550 | } | |
7d75abc8 | 4551 | else if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
14ac6aa2 | 4552 | ; |
319e6439 RG |
4553 | else |
4554 | { | |
4555 | negative = tree_int_cst_compare (nested_in_vect_loop | |
4556 | ? STMT_VINFO_DR_STEP (stmt_info) | |
4557 | : DR_STEP (dr), | |
4558 | size_zero_node) < 0; | |
4559 | if (negative && ncopies > 1) | |
4560 | { | |
73fbfcad | 4561 | if (dump_enabled_p ()) |
78c60e3d | 4562 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4563 | "multiple types with negative step.\n"); |
319e6439 RG |
4564 | return false; |
4565 | } | |
4566 | ||
4567 | if (negative) | |
4568 | { | |
08940f33 RB |
4569 | if (grouped_load) |
4570 | { | |
4571 | if (dump_enabled_p ()) | |
4572 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 TJ |
4573 | "negative step for group load not supported" |
4574 | "\n"); | |
08940f33 RB |
4575 | return false; |
4576 | } | |
319e6439 RG |
4577 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
4578 | if (alignment_support_scheme != dr_aligned | |
4579 | && alignment_support_scheme != dr_unaligned_supported) | |
4580 | { | |
73fbfcad | 4581 | if (dump_enabled_p ()) |
78c60e3d | 4582 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4583 | "negative step but alignment required.\n"); |
319e6439 RG |
4584 | return false; |
4585 | } | |
4586 | if (!perm_mask_for_reverse (vectype)) | |
4587 | { | |
73fbfcad | 4588 | if (dump_enabled_p ()) |
78c60e3d | 4589 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
4590 | "negative step and reversing not supported." |
4591 | "\n"); | |
319e6439 RG |
4592 | return false; |
4593 | } | |
4594 | } | |
7d75abc8 | 4595 | } |
aec7ae7d | 4596 | |
ebfd146a IR |
4597 | if (!vec_stmt) /* transformation not required. */ |
4598 | { | |
4599 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
92345349 | 4600 | vect_model_load_cost (stmt_info, ncopies, load_lanes_p, NULL, NULL, NULL); |
ebfd146a IR |
4601 | return true; |
4602 | } | |
4603 | ||
73fbfcad | 4604 | if (dump_enabled_p ()) |
78c60e3d | 4605 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4606 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a IR |
4607 | |
4608 | /** Transform. **/ | |
4609 | ||
c716e67f XDL |
4610 | ensure_base_align (stmt_info, dr); |
4611 | ||
aec7ae7d JJ |
4612 | if (STMT_VINFO_GATHER_P (stmt_info)) |
4613 | { | |
4614 | tree vec_oprnd0 = NULL_TREE, op; | |
4615 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
4616 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
4617 | tree ptr, mask, var, scale, perm_mask = NULL_TREE, prev_res = NULL_TREE; | |
4618 | edge pe = loop_preheader_edge (loop); | |
4619 | gimple_seq seq; | |
4620 | basic_block new_bb; | |
4621 | enum { NARROW, NONE, WIDEN } modifier; | |
4622 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
4623 | ||
4624 | if (nunits == gather_off_nunits) | |
4625 | modifier = NONE; | |
4626 | else if (nunits == gather_off_nunits / 2) | |
4627 | { | |
4628 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
4629 | modifier = WIDEN; | |
4630 | ||
4631 | for (i = 0; i < gather_off_nunits; ++i) | |
4632 | sel[i] = i | nunits; | |
4633 | ||
3fcc1b55 | 4634 | perm_mask = vect_gen_perm_mask (gather_off_vectype, sel); |
aec7ae7d JJ |
4635 | gcc_assert (perm_mask != NULL_TREE); |
4636 | } | |
4637 | else if (nunits == gather_off_nunits * 2) | |
4638 | { | |
4639 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
4640 | modifier = NARROW; | |
4641 | ||
4642 | for (i = 0; i < nunits; ++i) | |
4643 | sel[i] = i < gather_off_nunits | |
4644 | ? i : i + nunits - gather_off_nunits; | |
4645 | ||
3fcc1b55 | 4646 | perm_mask = vect_gen_perm_mask (vectype, sel); |
aec7ae7d JJ |
4647 | gcc_assert (perm_mask != NULL_TREE); |
4648 | ncopies *= 2; | |
4649 | } | |
4650 | else | |
4651 | gcc_unreachable (); | |
4652 | ||
4653 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); | |
4654 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4655 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4656 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4657 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4658 | scaletype = TREE_VALUE (arglist); | |
4659 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
4660 | && types_compatible_p (srctype, masktype)); | |
4661 | ||
4662 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4663 | ||
4664 | ptr = fold_convert (ptrtype, gather_base); | |
4665 | if (!is_gimple_min_invariant (ptr)) | |
4666 | { | |
4667 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
4668 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
4669 | gcc_assert (!new_bb); | |
4670 | } | |
4671 | ||
4672 | /* Currently we support only unconditional gather loads, | |
4673 | so mask should be all ones. */ | |
4674 | if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
4675 | mask = build_int_cst (TREE_TYPE (masktype), -1); | |
4676 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) | |
4677 | { | |
4678 | REAL_VALUE_TYPE r; | |
4679 | long tmp[6]; | |
4680 | for (j = 0; j < 6; ++j) | |
4681 | tmp[j] = -1; | |
4682 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
4683 | mask = build_real (TREE_TYPE (masktype), r); | |
4684 | } | |
4685 | else | |
4686 | gcc_unreachable (); | |
4687 | mask = build_vector_from_val (masktype, mask); | |
4688 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
4689 | ||
4690 | scale = build_int_cst (scaletype, gather_scale); | |
4691 | ||
4692 | prev_stmt_info = NULL; | |
4693 | for (j = 0; j < ncopies; ++j) | |
4694 | { | |
4695 | if (modifier == WIDEN && (j & 1)) | |
4696 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
4697 | perm_mask, stmt, gsi); | |
4698 | else if (j == 0) | |
4699 | op = vec_oprnd0 | |
4700 | = vect_get_vec_def_for_operand (gather_off, stmt, NULL); | |
4701 | else | |
4702 | op = vec_oprnd0 | |
4703 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
4704 | ||
4705 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
4706 | { | |
4707 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
4708 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
4709 | var = vect_get_new_vect_var (idxtype, vect_simple_var, NULL); | |
aec7ae7d JJ |
4710 | var = make_ssa_name (var, NULL); |
4711 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); | |
4712 | new_stmt | |
4713 | = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var, | |
4714 | op, NULL_TREE); | |
4715 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4716 | op = var; | |
4717 | } | |
4718 | ||
4719 | new_stmt | |
4720 | = gimple_build_call (gather_decl, 5, mask, ptr, op, mask, scale); | |
4721 | ||
4722 | if (!useless_type_conversion_p (vectype, rettype)) | |
4723 | { | |
4724 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
4725 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
4726 | var = vect_get_new_vect_var (rettype, vect_simple_var, NULL); | |
aec7ae7d JJ |
4727 | op = make_ssa_name (var, new_stmt); |
4728 | gimple_call_set_lhs (new_stmt, op); | |
4729 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4730 | var = make_ssa_name (vec_dest, NULL); | |
4731 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); | |
4732 | new_stmt | |
4733 | = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var, op, | |
4734 | NULL_TREE); | |
4735 | } | |
4736 | else | |
4737 | { | |
4738 | var = make_ssa_name (vec_dest, new_stmt); | |
4739 | gimple_call_set_lhs (new_stmt, var); | |
4740 | } | |
4741 | ||
4742 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4743 | ||
4744 | if (modifier == NARROW) | |
4745 | { | |
4746 | if ((j & 1) == 0) | |
4747 | { | |
4748 | prev_res = var; | |
4749 | continue; | |
4750 | } | |
4751 | var = permute_vec_elements (prev_res, var, | |
4752 | perm_mask, stmt, gsi); | |
4753 | new_stmt = SSA_NAME_DEF_STMT (var); | |
4754 | } | |
4755 | ||
4756 | if (prev_stmt_info == NULL) | |
4757 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4758 | else | |
4759 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4760 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4761 | } | |
4762 | return true; | |
4763 | } | |
7d75abc8 MM |
4764 | else if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
4765 | { | |
4766 | gimple_stmt_iterator incr_gsi; | |
4767 | bool insert_after; | |
4768 | gimple incr; | |
4769 | tree offvar; | |
7d75abc8 MM |
4770 | tree ivstep; |
4771 | tree running_off; | |
9771b263 | 4772 | vec<constructor_elt, va_gc> *v = NULL; |
7d75abc8 | 4773 | gimple_seq stmts = NULL; |
14ac6aa2 RB |
4774 | tree stride_base, stride_step, alias_off; |
4775 | ||
4776 | gcc_assert (!nested_in_vect_loop); | |
7d75abc8 | 4777 | |
14ac6aa2 RB |
4778 | stride_base |
4779 | = fold_build_pointer_plus | |
4780 | (unshare_expr (DR_BASE_ADDRESS (dr)), | |
4781 | size_binop (PLUS_EXPR, | |
4782 | convert_to_ptrofftype (unshare_expr (DR_OFFSET (dr))), | |
c3284718 | 4783 | convert_to_ptrofftype (DR_INIT (dr)))); |
14ac6aa2 | 4784 | stride_step = fold_convert (sizetype, unshare_expr (DR_STEP (dr))); |
7d75abc8 MM |
4785 | |
4786 | /* For a load with loop-invariant (but other than power-of-2) | |
4787 | stride (i.e. not a grouped access) like so: | |
4788 | ||
4789 | for (i = 0; i < n; i += stride) | |
4790 | ... = array[i]; | |
4791 | ||
4792 | we generate a new induction variable and new accesses to | |
4793 | form a new vector (or vectors, depending on ncopies): | |
4794 | ||
4795 | for (j = 0; ; j += VF*stride) | |
4796 | tmp1 = array[j]; | |
4797 | tmp2 = array[j + stride]; | |
4798 | ... | |
4799 | vectemp = {tmp1, tmp2, ...} | |
4800 | */ | |
4801 | ||
4802 | ivstep = stride_step; | |
4803 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
4804 | build_int_cst (TREE_TYPE (ivstep), vf)); | |
4805 | ||
4806 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
4807 | ||
4808 | create_iv (stride_base, ivstep, NULL, | |
4809 | loop, &incr_gsi, insert_after, | |
4810 | &offvar, NULL); | |
4811 | incr = gsi_stmt (incr_gsi); | |
4812 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo, NULL)); | |
4813 | ||
4814 | stride_step = force_gimple_operand (stride_step, &stmts, true, NULL_TREE); | |
4815 | if (stmts) | |
4816 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
4817 | ||
4818 | prev_stmt_info = NULL; | |
4819 | running_off = offvar; | |
14ac6aa2 | 4820 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (dr)), 0); |
7d75abc8 MM |
4821 | for (j = 0; j < ncopies; j++) |
4822 | { | |
4823 | tree vec_inv; | |
4824 | ||
9771b263 | 4825 | vec_alloc (v, nunits); |
7d75abc8 MM |
4826 | for (i = 0; i < nunits; i++) |
4827 | { | |
4828 | tree newref, newoff; | |
4829 | gimple incr; | |
14ac6aa2 RB |
4830 | newref = build2 (MEM_REF, TREE_TYPE (vectype), |
4831 | running_off, alias_off); | |
7d75abc8 MM |
4832 | |
4833 | newref = force_gimple_operand_gsi (gsi, newref, true, | |
4834 | NULL_TREE, true, | |
4835 | GSI_SAME_STMT); | |
4836 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, newref); | |
070ecdfd | 4837 | newoff = copy_ssa_name (running_off, NULL); |
14ac6aa2 RB |
4838 | incr = gimple_build_assign_with_ops (POINTER_PLUS_EXPR, newoff, |
4839 | running_off, stride_step); | |
7d75abc8 MM |
4840 | vect_finish_stmt_generation (stmt, incr, gsi); |
4841 | ||
4842 | running_off = newoff; | |
4843 | } | |
4844 | ||
4845 | vec_inv = build_constructor (vectype, v); | |
4846 | new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi); | |
4847 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
7d75abc8 MM |
4848 | |
4849 | if (j == 0) | |
4850 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4851 | else | |
4852 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4853 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4854 | } | |
4855 | return true; | |
4856 | } | |
aec7ae7d | 4857 | |
0d0293ac | 4858 | if (grouped_load) |
ebfd146a | 4859 | { |
e14c1050 | 4860 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
6aa904c4 | 4861 | if (slp |
01d8bf07 | 4862 | && !SLP_TREE_LOAD_PERMUTATION (slp_node).exists () |
9771b263 DN |
4863 | && first_stmt != SLP_TREE_SCALAR_STMTS (slp_node)[0]) |
4864 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 4865 | |
ebfd146a | 4866 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
4867 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
4868 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
4869 | ??? But we can only do so if there is exactly one | |
4870 | as we have no way to get at the rest. Leave the CSE | |
4871 | opportunity alone. | |
4872 | ??? With the group load eventually participating | |
4873 | in multiple different permutations (having multiple | |
4874 | slp nodes which refer to the same group) the CSE | |
4875 | is even wrong code. See PR56270. */ | |
4876 | && !slp) | |
ebfd146a IR |
4877 | { |
4878 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4879 | return true; | |
4880 | } | |
4881 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 4882 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a IR |
4883 | |
4884 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
4885 | if (slp) | |
4886 | { | |
0d0293ac | 4887 | grouped_load = false; |
ebfd146a | 4888 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
01d8bf07 | 4889 | if (SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
a70d6342 | 4890 | slp_perm = true; |
a64b9c26 | 4891 | group_gap = GROUP_GAP (vinfo_for_stmt (first_stmt)); |
a70d6342 | 4892 | } |
ebfd146a | 4893 | else |
a64b9c26 RB |
4894 | { |
4895 | vec_num = group_size; | |
4896 | group_gap = 0; | |
4897 | } | |
ebfd146a IR |
4898 | } |
4899 | else | |
4900 | { | |
4901 | first_stmt = stmt; | |
4902 | first_dr = dr; | |
4903 | group_size = vec_num = 1; | |
a64b9c26 | 4904 | group_gap = 0; |
ebfd146a IR |
4905 | } |
4906 | ||
720f5239 | 4907 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 4908 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
4909 | /* Targets with load-lane instructions must not require explicit |
4910 | realignment. */ | |
4911 | gcc_assert (!load_lanes_p | |
4912 | || alignment_support_scheme == dr_aligned | |
4913 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
4914 | |
4915 | /* In case the vectorization factor (VF) is bigger than the number | |
4916 | of elements that we can fit in a vectype (nunits), we have to generate | |
4917 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 4918 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 4919 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 4920 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 4921 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
4922 | stmts that use the defs of the current stmt. The example below |
4923 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
4924 | need to create 4 vectorized stmts): | |
ebfd146a IR |
4925 | |
4926 | before vectorization: | |
4927 | RELATED_STMT VEC_STMT | |
4928 | S1: x = memref - - | |
4929 | S2: z = x + 1 - - | |
4930 | ||
4931 | step 1: vectorize stmt S1: | |
4932 | We first create the vector stmt VS1_0, and, as usual, record a | |
4933 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
4934 | Next, we create the vector stmt VS1_1, and record a pointer to | |
4935 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 4936 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
4937 | stmts and pointers: |
4938 | RELATED_STMT VEC_STMT | |
4939 | VS1_0: vx0 = memref0 VS1_1 - | |
4940 | VS1_1: vx1 = memref1 VS1_2 - | |
4941 | VS1_2: vx2 = memref2 VS1_3 - | |
4942 | VS1_3: vx3 = memref3 - - | |
4943 | S1: x = load - VS1_0 | |
4944 | S2: z = x + 1 - - | |
4945 | ||
b8698a0f L |
4946 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
4947 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
4948 | stmt S2. */ |
4949 | ||
0d0293ac | 4950 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
4951 | |
4952 | S1: x2 = &base + 2 | |
4953 | S2: x0 = &base | |
4954 | S3: x1 = &base + 1 | |
4955 | S4: x3 = &base + 3 | |
4956 | ||
b8698a0f | 4957 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
4958 | starting from the access of the first stmt of the chain: |
4959 | ||
4960 | VS1: vx0 = &base | |
4961 | VS2: vx1 = &base + vec_size*1 | |
4962 | VS3: vx3 = &base + vec_size*2 | |
4963 | VS4: vx4 = &base + vec_size*3 | |
4964 | ||
4965 | Then permutation statements are generated: | |
4966 | ||
e2c83630 RH |
4967 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
4968 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
4969 | ... |
4970 | ||
4971 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
4972 | (the order of the data-refs in the output of vect_permute_load_chain | |
4973 | corresponds to the order of scalar stmts in the interleaving chain - see | |
4974 | the documentation of vect_permute_load_chain()). | |
4975 | The generation of permutation stmts and recording them in | |
0d0293ac | 4976 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 4977 | |
b8698a0f | 4978 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
4979 | permutation stmts above are created for every copy. The result vector |
4980 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
4981 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
4982 | |
4983 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
4984 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
4985 | we generate the following code: | |
4986 | p = initial_addr; | |
4987 | indx = 0; | |
4988 | loop { | |
4989 | p = p + indx * vectype_size; | |
4990 | vec_dest = *(p); | |
4991 | indx = indx + 1; | |
4992 | } | |
4993 | ||
4994 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 4995 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
4996 | then generate the following code, in which the data in each iteration is |
4997 | obtained by two vector loads, one from the previous iteration, and one | |
4998 | from the current iteration: | |
4999 | p1 = initial_addr; | |
5000 | msq_init = *(floor(p1)) | |
5001 | p2 = initial_addr + VS - 1; | |
5002 | realignment_token = call target_builtin; | |
5003 | indx = 0; | |
5004 | loop { | |
5005 | p2 = p2 + indx * vectype_size | |
5006 | lsq = *(floor(p2)) | |
5007 | vec_dest = realign_load (msq, lsq, realignment_token) | |
5008 | indx = indx + 1; | |
5009 | msq = lsq; | |
5010 | } */ | |
5011 | ||
5012 | /* If the misalignment remains the same throughout the execution of the | |
5013 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 5014 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
5015 | This can only occur when vectorizing memory accesses in the inner-loop |
5016 | nested within an outer-loop that is being vectorized. */ | |
5017 | ||
d1e4b493 | 5018 | if (nested_in_vect_loop |
211bea38 | 5019 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
ebfd146a IR |
5020 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) |
5021 | { | |
5022 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
5023 | compute_in_loop = true; | |
5024 | } | |
5025 | ||
5026 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
5027 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 5028 | && !compute_in_loop) |
ebfd146a IR |
5029 | { |
5030 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
5031 | alignment_support_scheme, NULL_TREE, | |
5032 | &at_loop); | |
5033 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
5034 | { | |
5035 | phi = SSA_NAME_DEF_STMT (msq); | |
5036 | offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
5037 | } | |
5038 | } | |
5039 | else | |
5040 | at_loop = loop; | |
5041 | ||
a1e53f3f L |
5042 | if (negative) |
5043 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
5044 | ||
272c6793 RS |
5045 | if (load_lanes_p) |
5046 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
5047 | else | |
5048 | aggr_type = vectype; | |
5049 | ||
ebfd146a IR |
5050 | prev_stmt_info = NULL; |
5051 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 5052 | { |
272c6793 | 5053 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 5054 | if (j == 0) |
74bf76ed JJ |
5055 | { |
5056 | bool simd_lane_access_p | |
5057 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
5058 | if (simd_lane_access_p | |
5059 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
5060 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
5061 | && integer_zerop (DR_OFFSET (first_dr)) | |
5062 | && integer_zerop (DR_INIT (first_dr)) | |
5063 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
5064 | get_alias_set (DR_REF (first_dr))) | |
5065 | && (alignment_support_scheme == dr_aligned | |
5066 | || alignment_support_scheme == dr_unaligned_supported)) | |
5067 | { | |
5068 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
5069 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
5070 | (DR_REF (first_dr)), 0); | |
8928eff3 | 5071 | inv_p = false; |
74bf76ed JJ |
5072 | } |
5073 | else | |
5074 | dataref_ptr | |
5075 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
5076 | offset, &dummy, gsi, &ptr_incr, | |
5077 | simd_lane_access_p, &inv_p); | |
5078 | } | |
5079 | else if (dataref_offset) | |
5080 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
5081 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 5082 | else |
272c6793 RS |
5083 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
5084 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 5085 | |
0d0293ac | 5086 | if (grouped_load || slp_perm) |
9771b263 | 5087 | dr_chain.create (vec_num); |
5ce1ee7f | 5088 | |
272c6793 | 5089 | if (load_lanes_p) |
ebfd146a | 5090 | { |
272c6793 RS |
5091 | tree vec_array; |
5092 | ||
5093 | vec_array = create_vector_array (vectype, vec_num); | |
5094 | ||
5095 | /* Emit: | |
5096 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
5097 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
5098 | new_stmt = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
5099 | gimple_call_set_lhs (new_stmt, vec_array); | |
5100 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a | 5101 | |
272c6793 RS |
5102 | /* Extract each vector into an SSA_NAME. */ |
5103 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 5104 | { |
272c6793 RS |
5105 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
5106 | vec_array, i); | |
9771b263 | 5107 | dr_chain.quick_push (new_temp); |
272c6793 RS |
5108 | } |
5109 | ||
5110 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 5111 | vect_record_grouped_load_vectors (stmt, dr_chain); |
272c6793 RS |
5112 | } |
5113 | else | |
5114 | { | |
5115 | for (i = 0; i < vec_num; i++) | |
5116 | { | |
5117 | if (i > 0) | |
5118 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
5119 | stmt, NULL_TREE); | |
5120 | ||
5121 | /* 2. Create the vector-load in the loop. */ | |
5122 | switch (alignment_support_scheme) | |
5123 | { | |
5124 | case dr_aligned: | |
5125 | case dr_unaligned_supported: | |
be1ac4ec | 5126 | { |
644ffefd MJ |
5127 | unsigned int align, misalign; |
5128 | ||
272c6793 RS |
5129 | data_ref |
5130 | = build2 (MEM_REF, vectype, dataref_ptr, | |
74bf76ed JJ |
5131 | dataref_offset |
5132 | ? dataref_offset | |
5133 | : build_int_cst (reference_alias_ptr_type | |
5134 | (DR_REF (first_dr)), 0)); | |
644ffefd | 5135 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 RS |
5136 | if (alignment_support_scheme == dr_aligned) |
5137 | { | |
5138 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 5139 | misalign = 0; |
272c6793 RS |
5140 | } |
5141 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
5142 | { | |
5143 | TREE_TYPE (data_ref) | |
5144 | = build_aligned_type (TREE_TYPE (data_ref), | |
5145 | TYPE_ALIGN (elem_type)); | |
644ffefd MJ |
5146 | align = TYPE_ALIGN_UNIT (elem_type); |
5147 | misalign = 0; | |
272c6793 RS |
5148 | } |
5149 | else | |
5150 | { | |
5151 | TREE_TYPE (data_ref) | |
5152 | = build_aligned_type (TREE_TYPE (data_ref), | |
5153 | TYPE_ALIGN (elem_type)); | |
644ffefd | 5154 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 5155 | } |
74bf76ed JJ |
5156 | if (dataref_offset == NULL_TREE) |
5157 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), | |
5158 | align, misalign); | |
272c6793 | 5159 | break; |
be1ac4ec | 5160 | } |
272c6793 | 5161 | case dr_explicit_realign: |
267d3070 | 5162 | { |
272c6793 RS |
5163 | tree ptr, bump; |
5164 | tree vs_minus_1; | |
5165 | ||
5166 | vs_minus_1 = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
5167 | ||
5168 | if (compute_in_loop) | |
5169 | msq = vect_setup_realignment (first_stmt, gsi, | |
5170 | &realignment_token, | |
5171 | dr_explicit_realign, | |
5172 | dataref_ptr, NULL); | |
5173 | ||
070ecdfd | 5174 | ptr = copy_ssa_name (dataref_ptr, NULL); |
272c6793 | 5175 | new_stmt = gimple_build_assign_with_ops |
070ecdfd | 5176 | (BIT_AND_EXPR, ptr, dataref_ptr, |
272c6793 RS |
5177 | build_int_cst |
5178 | (TREE_TYPE (dataref_ptr), | |
5179 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
5180 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
5181 | data_ref | |
5182 | = build2 (MEM_REF, vectype, ptr, | |
5183 | build_int_cst (reference_alias_ptr_type | |
5184 | (DR_REF (first_dr)), 0)); | |
5185 | vec_dest = vect_create_destination_var (scalar_dest, | |
5186 | vectype); | |
5187 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
5188 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5189 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5190 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
5191 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
5192 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5193 | msq = new_temp; | |
5194 | ||
5195 | bump = size_binop (MULT_EXPR, vs_minus_1, | |
7b7b1813 | 5196 | TYPE_SIZE_UNIT (elem_type)); |
272c6793 RS |
5197 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
5198 | new_stmt = gimple_build_assign_with_ops | |
5199 | (BIT_AND_EXPR, NULL_TREE, ptr, | |
5200 | build_int_cst | |
5201 | (TREE_TYPE (ptr), | |
5202 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
070ecdfd | 5203 | ptr = copy_ssa_name (dataref_ptr, new_stmt); |
272c6793 RS |
5204 | gimple_assign_set_lhs (new_stmt, ptr); |
5205 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5206 | data_ref | |
5207 | = build2 (MEM_REF, vectype, ptr, | |
5208 | build_int_cst (reference_alias_ptr_type | |
5209 | (DR_REF (first_dr)), 0)); | |
5210 | break; | |
267d3070 | 5211 | } |
272c6793 | 5212 | case dr_explicit_realign_optimized: |
070ecdfd | 5213 | new_temp = copy_ssa_name (dataref_ptr, NULL); |
272c6793 | 5214 | new_stmt = gimple_build_assign_with_ops |
070ecdfd | 5215 | (BIT_AND_EXPR, new_temp, dataref_ptr, |
272c6793 RS |
5216 | build_int_cst |
5217 | (TREE_TYPE (dataref_ptr), | |
5218 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
5219 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
5220 | data_ref | |
5221 | = build2 (MEM_REF, vectype, new_temp, | |
5222 | build_int_cst (reference_alias_ptr_type | |
5223 | (DR_REF (first_dr)), 0)); | |
5224 | break; | |
5225 | default: | |
5226 | gcc_unreachable (); | |
5227 | } | |
ebfd146a | 5228 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 5229 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
5230 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5231 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5232 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5233 | ||
272c6793 RS |
5234 | /* 3. Handle explicit realignment if necessary/supported. |
5235 | Create in loop: | |
5236 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
5237 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
5238 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 5239 | { |
272c6793 RS |
5240 | lsq = gimple_assign_lhs (new_stmt); |
5241 | if (!realignment_token) | |
5242 | realignment_token = dataref_ptr; | |
5243 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5244 | new_stmt | |
73804b12 RG |
5245 | = gimple_build_assign_with_ops (REALIGN_LOAD_EXPR, |
5246 | vec_dest, msq, lsq, | |
5247 | realignment_token); | |
272c6793 RS |
5248 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5249 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5250 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5251 | ||
5252 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
5253 | { | |
5254 | gcc_assert (phi); | |
5255 | if (i == vec_num - 1 && j == ncopies - 1) | |
5256 | add_phi_arg (phi, lsq, | |
5257 | loop_latch_edge (containing_loop), | |
9e227d60 | 5258 | UNKNOWN_LOCATION); |
272c6793 RS |
5259 | msq = lsq; |
5260 | } | |
ebfd146a | 5261 | } |
ebfd146a | 5262 | |
59fd17e3 RB |
5263 | /* 4. Handle invariant-load. */ |
5264 | if (inv_p && !bb_vinfo) | |
5265 | { | |
5266 | gimple_stmt_iterator gsi2 = *gsi; | |
5267 | gcc_assert (!grouped_load); | |
5268 | gsi_next (&gsi2); | |
5269 | new_temp = vect_init_vector (stmt, scalar_dest, | |
5270 | vectype, &gsi2); | |
5271 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
5272 | } | |
5273 | ||
272c6793 RS |
5274 | if (negative) |
5275 | { | |
aec7ae7d JJ |
5276 | tree perm_mask = perm_mask_for_reverse (vectype); |
5277 | new_temp = permute_vec_elements (new_temp, new_temp, | |
5278 | perm_mask, stmt, gsi); | |
ebfd146a IR |
5279 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
5280 | } | |
267d3070 | 5281 | |
272c6793 | 5282 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
5283 | vect_transform_grouped_load (). */ |
5284 | if (grouped_load || slp_perm) | |
9771b263 | 5285 | dr_chain.quick_push (new_temp); |
267d3070 | 5286 | |
272c6793 RS |
5287 | /* Store vector loads in the corresponding SLP_NODE. */ |
5288 | if (slp && !slp_perm) | |
9771b263 | 5289 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
272c6793 | 5290 | } |
a64b9c26 RB |
5291 | /* Bump the vector pointer to account for a gap. */ |
5292 | if (slp && group_gap != 0) | |
5293 | { | |
5294 | tree bump = size_binop (MULT_EXPR, | |
5295 | TYPE_SIZE_UNIT (elem_type), | |
5296 | size_int (group_gap)); | |
5297 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
5298 | stmt, bump); | |
5299 | } | |
ebfd146a IR |
5300 | } |
5301 | ||
5302 | if (slp && !slp_perm) | |
5303 | continue; | |
5304 | ||
5305 | if (slp_perm) | |
5306 | { | |
01d8bf07 | 5307 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
ebfd146a IR |
5308 | slp_node_instance, false)) |
5309 | { | |
9771b263 | 5310 | dr_chain.release (); |
ebfd146a IR |
5311 | return false; |
5312 | } | |
5313 | } | |
5314 | else | |
5315 | { | |
0d0293ac | 5316 | if (grouped_load) |
ebfd146a | 5317 | { |
272c6793 | 5318 | if (!load_lanes_p) |
0d0293ac | 5319 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 5320 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
5321 | } |
5322 | else | |
5323 | { | |
5324 | if (j == 0) | |
5325 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5326 | else | |
5327 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5328 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5329 | } | |
5330 | } | |
9771b263 | 5331 | dr_chain.release (); |
ebfd146a IR |
5332 | } |
5333 | ||
ebfd146a IR |
5334 | return true; |
5335 | } | |
5336 | ||
5337 | /* Function vect_is_simple_cond. | |
b8698a0f | 5338 | |
ebfd146a IR |
5339 | Input: |
5340 | LOOP - the loop that is being vectorized. | |
5341 | COND - Condition that is checked for simple use. | |
5342 | ||
e9e1d143 RG |
5343 | Output: |
5344 | *COMP_VECTYPE - the vector type for the comparison. | |
5345 | ||
ebfd146a IR |
5346 | Returns whether a COND can be vectorized. Checks whether |
5347 | condition operands are supportable using vec_is_simple_use. */ | |
5348 | ||
87aab9b2 | 5349 | static bool |
24ee1384 IR |
5350 | vect_is_simple_cond (tree cond, gimple stmt, loop_vec_info loop_vinfo, |
5351 | bb_vec_info bb_vinfo, tree *comp_vectype) | |
ebfd146a IR |
5352 | { |
5353 | tree lhs, rhs; | |
5354 | tree def; | |
5355 | enum vect_def_type dt; | |
e9e1d143 | 5356 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a IR |
5357 | |
5358 | if (!COMPARISON_CLASS_P (cond)) | |
5359 | return false; | |
5360 | ||
5361 | lhs = TREE_OPERAND (cond, 0); | |
5362 | rhs = TREE_OPERAND (cond, 1); | |
5363 | ||
5364 | if (TREE_CODE (lhs) == SSA_NAME) | |
5365 | { | |
5366 | gimple lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); | |
24ee1384 IR |
5367 | if (!vect_is_simple_use_1 (lhs, stmt, loop_vinfo, bb_vinfo, |
5368 | &lhs_def_stmt, &def, &dt, &vectype1)) | |
ebfd146a IR |
5369 | return false; |
5370 | } | |
5371 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
5372 | && TREE_CODE (lhs) != FIXED_CST) | |
5373 | return false; | |
5374 | ||
5375 | if (TREE_CODE (rhs) == SSA_NAME) | |
5376 | { | |
5377 | gimple rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); | |
24ee1384 IR |
5378 | if (!vect_is_simple_use_1 (rhs, stmt, loop_vinfo, bb_vinfo, |
5379 | &rhs_def_stmt, &def, &dt, &vectype2)) | |
ebfd146a IR |
5380 | return false; |
5381 | } | |
f7e531cf | 5382 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST |
ebfd146a IR |
5383 | && TREE_CODE (rhs) != FIXED_CST) |
5384 | return false; | |
5385 | ||
e9e1d143 | 5386 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
ebfd146a IR |
5387 | return true; |
5388 | } | |
5389 | ||
5390 | /* vectorizable_condition. | |
5391 | ||
b8698a0f L |
5392 | Check if STMT is conditional modify expression that can be vectorized. |
5393 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
5394 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
5395 | at GSI. |
5396 | ||
5397 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
5398 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
5399 | else caluse if it is 2). | |
ebfd146a IR |
5400 | |
5401 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5402 | ||
4bbe8262 | 5403 | bool |
ebfd146a | 5404 | vectorizable_condition (gimple stmt, gimple_stmt_iterator *gsi, |
f7e531cf IR |
5405 | gimple *vec_stmt, tree reduc_def, int reduc_index, |
5406 | slp_tree slp_node) | |
ebfd146a IR |
5407 | { |
5408 | tree scalar_dest = NULL_TREE; | |
5409 | tree vec_dest = NULL_TREE; | |
ebfd146a IR |
5410 | tree cond_expr, then_clause, else_clause; |
5411 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5412 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
df11cc78 | 5413 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
5414 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
5415 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
ebfd146a IR |
5416 | tree vec_compare, vec_cond_expr; |
5417 | tree new_temp; | |
5418 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
ebfd146a | 5419 | tree def; |
a855b1b1 | 5420 | enum vect_def_type dt, dts[4]; |
ebfd146a | 5421 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
f7e531cf | 5422 | int ncopies; |
ebfd146a | 5423 | enum tree_code code; |
a855b1b1 | 5424 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
5425 | int i, j; |
5426 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
5427 | vec<tree> vec_oprnds0 = vNULL; |
5428 | vec<tree> vec_oprnds1 = vNULL; | |
5429 | vec<tree> vec_oprnds2 = vNULL; | |
5430 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 5431 | tree vec_cmp_type; |
b8698a0f | 5432 | |
f7e531cf IR |
5433 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
5434 | ncopies = 1; | |
5435 | else | |
5436 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
437f4a00 | 5437 | |
ebfd146a | 5438 | gcc_assert (ncopies >= 1); |
a855b1b1 | 5439 | if (reduc_index && ncopies > 1) |
ebfd146a IR |
5440 | return false; /* FORNOW */ |
5441 | ||
f7e531cf IR |
5442 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
5443 | return false; | |
5444 | ||
5445 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
ebfd146a IR |
5446 | return false; |
5447 | ||
4bbe8262 IR |
5448 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5449 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
5450 | && reduc_def)) | |
ebfd146a IR |
5451 | return false; |
5452 | ||
ebfd146a | 5453 | /* FORNOW: not yet supported. */ |
b8698a0f | 5454 | if (STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 5455 | { |
73fbfcad | 5456 | if (dump_enabled_p ()) |
78c60e3d | 5457 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5458 | "value used after loop.\n"); |
ebfd146a IR |
5459 | return false; |
5460 | } | |
5461 | ||
5462 | /* Is vectorizable conditional operation? */ | |
5463 | if (!is_gimple_assign (stmt)) | |
5464 | return false; | |
5465 | ||
5466 | code = gimple_assign_rhs_code (stmt); | |
5467 | ||
5468 | if (code != COND_EXPR) | |
5469 | return false; | |
5470 | ||
4e71066d RG |
5471 | cond_expr = gimple_assign_rhs1 (stmt); |
5472 | then_clause = gimple_assign_rhs2 (stmt); | |
5473 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 5474 | |
24ee1384 IR |
5475 | if (!vect_is_simple_cond (cond_expr, stmt, loop_vinfo, bb_vinfo, |
5476 | &comp_vectype) | |
e9e1d143 | 5477 | || !comp_vectype) |
ebfd146a IR |
5478 | return false; |
5479 | ||
5480 | if (TREE_CODE (then_clause) == SSA_NAME) | |
5481 | { | |
5482 | gimple then_def_stmt = SSA_NAME_DEF_STMT (then_clause); | |
24ee1384 | 5483 | if (!vect_is_simple_use (then_clause, stmt, loop_vinfo, bb_vinfo, |
ebfd146a IR |
5484 | &then_def_stmt, &def, &dt)) |
5485 | return false; | |
5486 | } | |
b8698a0f | 5487 | else if (TREE_CODE (then_clause) != INTEGER_CST |
ebfd146a IR |
5488 | && TREE_CODE (then_clause) != REAL_CST |
5489 | && TREE_CODE (then_clause) != FIXED_CST) | |
5490 | return false; | |
5491 | ||
5492 | if (TREE_CODE (else_clause) == SSA_NAME) | |
5493 | { | |
5494 | gimple else_def_stmt = SSA_NAME_DEF_STMT (else_clause); | |
24ee1384 | 5495 | if (!vect_is_simple_use (else_clause, stmt, loop_vinfo, bb_vinfo, |
ebfd146a IR |
5496 | &else_def_stmt, &def, &dt)) |
5497 | return false; | |
5498 | } | |
b8698a0f | 5499 | else if (TREE_CODE (else_clause) != INTEGER_CST |
ebfd146a IR |
5500 | && TREE_CODE (else_clause) != REAL_CST |
5501 | && TREE_CODE (else_clause) != FIXED_CST) | |
5502 | return false; | |
5503 | ||
74946978 MP |
5504 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (vectype))); |
5505 | /* The result of a vector comparison should be signed type. */ | |
5506 | tree cmp_type = build_nonstandard_integer_type (prec, 0); | |
5507 | vec_cmp_type = get_same_sized_vectype (cmp_type, vectype); | |
5508 | if (vec_cmp_type == NULL_TREE) | |
5509 | return false; | |
784fb9b3 | 5510 | |
b8698a0f | 5511 | if (!vec_stmt) |
ebfd146a IR |
5512 | { |
5513 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
e9e1d143 | 5514 | return expand_vec_cond_expr_p (vectype, comp_vectype); |
ebfd146a IR |
5515 | } |
5516 | ||
f7e531cf IR |
5517 | /* Transform. */ |
5518 | ||
5519 | if (!slp_node) | |
5520 | { | |
9771b263 DN |
5521 | vec_oprnds0.create (1); |
5522 | vec_oprnds1.create (1); | |
5523 | vec_oprnds2.create (1); | |
5524 | vec_oprnds3.create (1); | |
f7e531cf | 5525 | } |
ebfd146a IR |
5526 | |
5527 | /* Handle def. */ | |
5528 | scalar_dest = gimple_assign_lhs (stmt); | |
5529 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5530 | ||
5531 | /* Handle cond expr. */ | |
a855b1b1 MM |
5532 | for (j = 0; j < ncopies; j++) |
5533 | { | |
f7e531cf | 5534 | gimple new_stmt = NULL; |
a855b1b1 MM |
5535 | if (j == 0) |
5536 | { | |
f7e531cf IR |
5537 | if (slp_node) |
5538 | { | |
07687835 TS |
5539 | stack_vec<tree, 4> ops; |
5540 | stack_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 5541 | |
9771b263 DN |
5542 | ops.safe_push (TREE_OPERAND (cond_expr, 0)); |
5543 | ops.safe_push (TREE_OPERAND (cond_expr, 1)); | |
5544 | ops.safe_push (then_clause); | |
5545 | ops.safe_push (else_clause); | |
f7e531cf | 5546 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); |
37b5ec8f JJ |
5547 | vec_oprnds3 = vec_defs.pop (); |
5548 | vec_oprnds2 = vec_defs.pop (); | |
5549 | vec_oprnds1 = vec_defs.pop (); | |
5550 | vec_oprnds0 = vec_defs.pop (); | |
f7e531cf | 5551 | |
9771b263 DN |
5552 | ops.release (); |
5553 | vec_defs.release (); | |
f7e531cf IR |
5554 | } |
5555 | else | |
5556 | { | |
5557 | gimple gtemp; | |
5558 | vec_cond_lhs = | |
a855b1b1 MM |
5559 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), |
5560 | stmt, NULL); | |
24ee1384 IR |
5561 | vect_is_simple_use (TREE_OPERAND (cond_expr, 0), stmt, |
5562 | loop_vinfo, NULL, >emp, &def, &dts[0]); | |
f7e531cf IR |
5563 | |
5564 | vec_cond_rhs = | |
5565 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), | |
5566 | stmt, NULL); | |
24ee1384 IR |
5567 | vect_is_simple_use (TREE_OPERAND (cond_expr, 1), stmt, |
5568 | loop_vinfo, NULL, >emp, &def, &dts[1]); | |
f7e531cf IR |
5569 | if (reduc_index == 1) |
5570 | vec_then_clause = reduc_def; | |
5571 | else | |
5572 | { | |
5573 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
5574 | stmt, NULL); | |
24ee1384 | 5575 | vect_is_simple_use (then_clause, stmt, loop_vinfo, |
f7e531cf IR |
5576 | NULL, >emp, &def, &dts[2]); |
5577 | } | |
5578 | if (reduc_index == 2) | |
5579 | vec_else_clause = reduc_def; | |
5580 | else | |
5581 | { | |
5582 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
a855b1b1 | 5583 | stmt, NULL); |
24ee1384 | 5584 | vect_is_simple_use (else_clause, stmt, loop_vinfo, |
a855b1b1 | 5585 | NULL, >emp, &def, &dts[3]); |
f7e531cf | 5586 | } |
a855b1b1 MM |
5587 | } |
5588 | } | |
5589 | else | |
5590 | { | |
f7e531cf | 5591 | vec_cond_lhs = vect_get_vec_def_for_stmt_copy (dts[0], |
9771b263 | 5592 | vec_oprnds0.pop ()); |
f7e531cf | 5593 | vec_cond_rhs = vect_get_vec_def_for_stmt_copy (dts[1], |
9771b263 | 5594 | vec_oprnds1.pop ()); |
a855b1b1 | 5595 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 5596 | vec_oprnds2.pop ()); |
a855b1b1 | 5597 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 5598 | vec_oprnds3.pop ()); |
f7e531cf IR |
5599 | } |
5600 | ||
5601 | if (!slp_node) | |
5602 | { | |
9771b263 DN |
5603 | vec_oprnds0.quick_push (vec_cond_lhs); |
5604 | vec_oprnds1.quick_push (vec_cond_rhs); | |
5605 | vec_oprnds2.quick_push (vec_then_clause); | |
5606 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
5607 | } |
5608 | ||
9dc3f7de | 5609 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 5610 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 5611 | { |
9771b263 DN |
5612 | vec_cond_rhs = vec_oprnds1[i]; |
5613 | vec_then_clause = vec_oprnds2[i]; | |
5614 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 5615 | |
784fb9b3 JJ |
5616 | vec_compare = build2 (TREE_CODE (cond_expr), vec_cmp_type, |
5617 | vec_cond_lhs, vec_cond_rhs); | |
f7e531cf IR |
5618 | vec_cond_expr = build3 (VEC_COND_EXPR, vectype, |
5619 | vec_compare, vec_then_clause, vec_else_clause); | |
a855b1b1 | 5620 | |
f7e531cf IR |
5621 | new_stmt = gimple_build_assign (vec_dest, vec_cond_expr); |
5622 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5623 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5624 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5625 | if (slp_node) | |
9771b263 | 5626 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
5627 | } |
5628 | ||
5629 | if (slp_node) | |
5630 | continue; | |
5631 | ||
5632 | if (j == 0) | |
5633 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5634 | else | |
5635 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5636 | ||
5637 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 5638 | } |
b8698a0f | 5639 | |
9771b263 DN |
5640 | vec_oprnds0.release (); |
5641 | vec_oprnds1.release (); | |
5642 | vec_oprnds2.release (); | |
5643 | vec_oprnds3.release (); | |
f7e531cf | 5644 | |
ebfd146a IR |
5645 | return true; |
5646 | } | |
5647 | ||
5648 | ||
8644a673 | 5649 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
5650 | |
5651 | bool | |
a70d6342 | 5652 | vect_analyze_stmt (gimple stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 5653 | { |
8644a673 | 5654 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 5655 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 5656 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 5657 | bool ok; |
a70d6342 | 5658 | tree scalar_type, vectype; |
363477c0 JJ |
5659 | gimple pattern_stmt; |
5660 | gimple_seq pattern_def_seq; | |
ebfd146a | 5661 | |
73fbfcad | 5662 | if (dump_enabled_p ()) |
ebfd146a | 5663 | { |
78c60e3d SS |
5664 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
5665 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 5666 | dump_printf (MSG_NOTE, "\n"); |
8644a673 | 5667 | } |
ebfd146a | 5668 | |
1825a1f3 | 5669 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 5670 | { |
73fbfcad | 5671 | if (dump_enabled_p ()) |
78c60e3d | 5672 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5673 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
5674 | |
5675 | return false; | |
5676 | } | |
b8698a0f L |
5677 | |
5678 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
5679 | to include: |
5680 | - the COND_EXPR which is the loop exit condition | |
5681 | - any LABEL_EXPRs in the loop | |
b8698a0f | 5682 | - computations that are used only for array indexing or loop control. |
8644a673 | 5683 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 5684 | instance, therefore, all the statements are relevant. |
ebfd146a | 5685 | |
d092494c | 5686 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 5687 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
5688 | statements. In basic blocks we are called from some SLP instance |
5689 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
5690 | already will be part of SLP instance. */ | |
83197f37 IR |
5691 | |
5692 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 5693 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 5694 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 5695 | { |
9d5e7640 | 5696 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 5697 | && pattern_stmt |
9d5e7640 IR |
5698 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
5699 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
5700 | { | |
83197f37 | 5701 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
5702 | stmt = pattern_stmt; |
5703 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 5704 | if (dump_enabled_p ()) |
9d5e7640 | 5705 | { |
78c60e3d SS |
5706 | dump_printf_loc (MSG_NOTE, vect_location, |
5707 | "==> examining pattern statement: "); | |
5708 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 5709 | dump_printf (MSG_NOTE, "\n"); |
9d5e7640 IR |
5710 | } |
5711 | } | |
5712 | else | |
5713 | { | |
73fbfcad | 5714 | if (dump_enabled_p ()) |
e645e942 | 5715 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 5716 | |
9d5e7640 IR |
5717 | return true; |
5718 | } | |
8644a673 | 5719 | } |
83197f37 | 5720 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 5721 | && node == NULL |
83197f37 IR |
5722 | && pattern_stmt |
5723 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
5724 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
5725 | { | |
5726 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 5727 | if (dump_enabled_p ()) |
83197f37 | 5728 | { |
78c60e3d SS |
5729 | dump_printf_loc (MSG_NOTE, vect_location, |
5730 | "==> examining pattern statement: "); | |
5731 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
e645e942 | 5732 | dump_printf (MSG_NOTE, "\n"); |
83197f37 IR |
5733 | } |
5734 | ||
5735 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node)) | |
5736 | return false; | |
5737 | } | |
ebfd146a | 5738 | |
1107f3ae | 5739 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 5740 | && node == NULL |
363477c0 | 5741 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 5742 | { |
363477c0 | 5743 | gimple_stmt_iterator si; |
1107f3ae | 5744 | |
363477c0 JJ |
5745 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
5746 | { | |
5747 | gimple pattern_def_stmt = gsi_stmt (si); | |
5748 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) | |
5749 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
5750 | { | |
5751 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 5752 | if (dump_enabled_p ()) |
363477c0 | 5753 | { |
78c60e3d SS |
5754 | dump_printf_loc (MSG_NOTE, vect_location, |
5755 | "==> examining pattern def statement: "); | |
5756 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
e645e942 | 5757 | dump_printf (MSG_NOTE, "\n"); |
363477c0 | 5758 | } |
1107f3ae | 5759 | |
363477c0 JJ |
5760 | if (!vect_analyze_stmt (pattern_def_stmt, |
5761 | need_to_vectorize, node)) | |
5762 | return false; | |
5763 | } | |
5764 | } | |
5765 | } | |
1107f3ae | 5766 | |
8644a673 IR |
5767 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
5768 | { | |
5769 | case vect_internal_def: | |
5770 | break; | |
ebfd146a | 5771 | |
8644a673 | 5772 | case vect_reduction_def: |
7c5222ff | 5773 | case vect_nested_cycle: |
a70d6342 | 5774 | gcc_assert (!bb_vinfo && (relevance == vect_used_in_outer |
8644a673 | 5775 | || relevance == vect_used_in_outer_by_reduction |
a70d6342 | 5776 | || relevance == vect_unused_in_scope)); |
8644a673 IR |
5777 | break; |
5778 | ||
5779 | case vect_induction_def: | |
5780 | case vect_constant_def: | |
5781 | case vect_external_def: | |
5782 | case vect_unknown_def_type: | |
5783 | default: | |
5784 | gcc_unreachable (); | |
5785 | } | |
ebfd146a | 5786 | |
a70d6342 IR |
5787 | if (bb_vinfo) |
5788 | { | |
5789 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
5790 | ||
b690cc0f | 5791 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
73fbfcad | 5792 | if (dump_enabled_p ()) |
a70d6342 | 5793 | { |
78c60e3d SS |
5794 | dump_printf_loc (MSG_NOTE, vect_location, |
5795 | "get vectype for scalar type: "); | |
5796 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
e645e942 | 5797 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
5798 | } |
5799 | ||
5800 | vectype = get_vectype_for_scalar_type (scalar_type); | |
5801 | if (!vectype) | |
5802 | { | |
73fbfcad | 5803 | if (dump_enabled_p ()) |
a70d6342 | 5804 | { |
78c60e3d SS |
5805 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5806 | "not SLPed: unsupported data-type "); | |
5807 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
5808 | scalar_type); | |
e645e942 | 5809 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
a70d6342 IR |
5810 | } |
5811 | return false; | |
5812 | } | |
5813 | ||
73fbfcad | 5814 | if (dump_enabled_p ()) |
a70d6342 | 5815 | { |
78c60e3d SS |
5816 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); |
5817 | dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); | |
e645e942 | 5818 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
5819 | } |
5820 | ||
5821 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
5822 | } | |
5823 | ||
8644a673 | 5824 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 5825 | { |
8644a673 IR |
5826 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
5827 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info)); | |
5828 | *need_to_vectorize = true; | |
ebfd146a IR |
5829 | } |
5830 | ||
8644a673 | 5831 | ok = true; |
b8698a0f | 5832 | if (!bb_vinfo |
a70d6342 IR |
5833 | && (STMT_VINFO_RELEVANT_P (stmt_info) |
5834 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
4a00c761 | 5835 | ok = (vectorizable_conversion (stmt, NULL, NULL, NULL) |
9dc3f7de | 5836 | || vectorizable_shift (stmt, NULL, NULL, NULL) |
8644a673 IR |
5837 | || vectorizable_operation (stmt, NULL, NULL, NULL) |
5838 | || vectorizable_assignment (stmt, NULL, NULL, NULL) | |
5839 | || vectorizable_load (stmt, NULL, NULL, NULL, NULL) | |
190c2236 | 5840 | || vectorizable_call (stmt, NULL, NULL, NULL) |
8644a673 | 5841 | || vectorizable_store (stmt, NULL, NULL, NULL) |
b5aeb3bb | 5842 | || vectorizable_reduction (stmt, NULL, NULL, NULL) |
f7e531cf | 5843 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, NULL)); |
a70d6342 IR |
5844 | else |
5845 | { | |
5846 | if (bb_vinfo) | |
4a00c761 JJ |
5847 | ok = (vectorizable_conversion (stmt, NULL, NULL, node) |
5848 | || vectorizable_shift (stmt, NULL, NULL, node) | |
9dc3f7de | 5849 | || vectorizable_operation (stmt, NULL, NULL, node) |
a70d6342 IR |
5850 | || vectorizable_assignment (stmt, NULL, NULL, node) |
5851 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
190c2236 | 5852 | || vectorizable_call (stmt, NULL, NULL, node) |
f7e531cf IR |
5853 | || vectorizable_store (stmt, NULL, NULL, node) |
5854 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node)); | |
b8698a0f | 5855 | } |
8644a673 IR |
5856 | |
5857 | if (!ok) | |
ebfd146a | 5858 | { |
73fbfcad | 5859 | if (dump_enabled_p ()) |
8644a673 | 5860 | { |
78c60e3d SS |
5861 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5862 | "not vectorized: relevant stmt not "); | |
5863 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
5864 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 5865 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
8644a673 | 5866 | } |
b8698a0f | 5867 | |
ebfd146a IR |
5868 | return false; |
5869 | } | |
5870 | ||
a70d6342 IR |
5871 | if (bb_vinfo) |
5872 | return true; | |
5873 | ||
8644a673 IR |
5874 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
5875 | need extra handling, except for vectorizable reductions. */ | |
5876 | if (STMT_VINFO_LIVE_P (stmt_info) | |
5877 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
5878 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 5879 | |
8644a673 | 5880 | if (!ok) |
ebfd146a | 5881 | { |
73fbfcad | 5882 | if (dump_enabled_p ()) |
8644a673 | 5883 | { |
78c60e3d SS |
5884 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5885 | "not vectorized: live stmt not "); | |
5886 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
5887 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
e645e942 | 5888 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
8644a673 | 5889 | } |
b8698a0f | 5890 | |
8644a673 | 5891 | return false; |
ebfd146a IR |
5892 | } |
5893 | ||
ebfd146a IR |
5894 | return true; |
5895 | } | |
5896 | ||
5897 | ||
5898 | /* Function vect_transform_stmt. | |
5899 | ||
5900 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
5901 | ||
5902 | bool | |
5903 | vect_transform_stmt (gimple stmt, gimple_stmt_iterator *gsi, | |
0d0293ac | 5904 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
5905 | slp_instance slp_node_instance) |
5906 | { | |
5907 | bool is_store = false; | |
5908 | gimple vec_stmt = NULL; | |
5909 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a | 5910 | bool done; |
ebfd146a IR |
5911 | |
5912 | switch (STMT_VINFO_TYPE (stmt_info)) | |
5913 | { | |
5914 | case type_demotion_vec_info_type: | |
ebfd146a | 5915 | case type_promotion_vec_info_type: |
ebfd146a IR |
5916 | case type_conversion_vec_info_type: |
5917 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
5918 | gcc_assert (done); | |
5919 | break; | |
5920 | ||
5921 | case induc_vec_info_type: | |
5922 | gcc_assert (!slp_node); | |
5923 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
5924 | gcc_assert (done); | |
5925 | break; | |
5926 | ||
9dc3f7de IR |
5927 | case shift_vec_info_type: |
5928 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
5929 | gcc_assert (done); | |
5930 | break; | |
5931 | ||
ebfd146a IR |
5932 | case op_vec_info_type: |
5933 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
5934 | gcc_assert (done); | |
5935 | break; | |
5936 | ||
5937 | case assignment_vec_info_type: | |
5938 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
5939 | gcc_assert (done); | |
5940 | break; | |
5941 | ||
5942 | case load_vec_info_type: | |
b8698a0f | 5943 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
5944 | slp_node_instance); |
5945 | gcc_assert (done); | |
5946 | break; | |
5947 | ||
5948 | case store_vec_info_type: | |
5949 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
5950 | gcc_assert (done); | |
0d0293ac | 5951 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
5952 | { |
5953 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 5954 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
5955 | one are skipped, and there vec_stmt_info shouldn't be freed |
5956 | meanwhile. */ | |
0d0293ac | 5957 | *grouped_store = true; |
ebfd146a IR |
5958 | if (STMT_VINFO_VEC_STMT (stmt_info)) |
5959 | is_store = true; | |
5960 | } | |
5961 | else | |
5962 | is_store = true; | |
5963 | break; | |
5964 | ||
5965 | case condition_vec_info_type: | |
f7e531cf | 5966 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
5967 | gcc_assert (done); |
5968 | break; | |
5969 | ||
5970 | case call_vec_info_type: | |
190c2236 | 5971 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 5972 | stmt = gsi_stmt (*gsi); |
ebfd146a IR |
5973 | break; |
5974 | ||
5975 | case reduc_vec_info_type: | |
b5aeb3bb | 5976 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
5977 | gcc_assert (done); |
5978 | break; | |
5979 | ||
5980 | default: | |
5981 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
5982 | { | |
73fbfcad | 5983 | if (dump_enabled_p ()) |
78c60e3d | 5984 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5985 | "stmt not supported.\n"); |
ebfd146a IR |
5986 | gcc_unreachable (); |
5987 | } | |
5988 | } | |
5989 | ||
5990 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that | |
5991 | is being vectorized, but outside the immediately enclosing loop. */ | |
5992 | if (vec_stmt | |
a70d6342 IR |
5993 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
5994 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
5995 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
5996 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
5997 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 5998 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 5999 | vect_used_in_outer_by_reduction)) |
ebfd146a | 6000 | { |
a70d6342 IR |
6001 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
6002 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
6003 | imm_use_iterator imm_iter; |
6004 | use_operand_p use_p; | |
6005 | tree scalar_dest; | |
6006 | gimple exit_phi; | |
6007 | ||
73fbfcad | 6008 | if (dump_enabled_p ()) |
78c60e3d | 6009 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6010 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
6011 | |
6012 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
6013 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
6014 | STMT). */ | |
6015 | if (gimple_code (stmt) == GIMPLE_PHI) | |
6016 | scalar_dest = PHI_RESULT (stmt); | |
6017 | else | |
6018 | scalar_dest = gimple_assign_lhs (stmt); | |
6019 | ||
6020 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
6021 | { | |
6022 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
6023 | { | |
6024 | exit_phi = USE_STMT (use_p); | |
6025 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
6026 | } | |
6027 | } | |
6028 | } | |
6029 | ||
6030 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
6031 | being vectorized. */ | |
6032 | if (STMT_VINFO_LIVE_P (stmt_info) | |
6033 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
6034 | { | |
6035 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
6036 | gcc_assert (done); | |
6037 | } | |
6038 | ||
6039 | if (vec_stmt) | |
83197f37 | 6040 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 6041 | |
b8698a0f | 6042 | return is_store; |
ebfd146a IR |
6043 | } |
6044 | ||
6045 | ||
b8698a0f | 6046 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
6047 | stmt_vec_info. */ |
6048 | ||
6049 | void | |
6050 | vect_remove_stores (gimple first_stmt) | |
6051 | { | |
6052 | gimple next = first_stmt; | |
6053 | gimple tmp; | |
6054 | gimple_stmt_iterator next_si; | |
6055 | ||
6056 | while (next) | |
6057 | { | |
78048b1c JJ |
6058 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
6059 | ||
6060 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
6061 | if (is_pattern_stmt_p (stmt_info)) | |
6062 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
6063 | /* Free the attached stmt_vec_info and remove the stmt. */ |
6064 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 6065 | unlink_stmt_vdef (next); |
ebfd146a | 6066 | gsi_remove (&next_si, true); |
3d3f2249 | 6067 | release_defs (next); |
ebfd146a IR |
6068 | free_stmt_vec_info (next); |
6069 | next = tmp; | |
6070 | } | |
6071 | } | |
6072 | ||
6073 | ||
6074 | /* Function new_stmt_vec_info. | |
6075 | ||
6076 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
6077 | ||
6078 | stmt_vec_info | |
b8698a0f | 6079 | new_stmt_vec_info (gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 6080 | bb_vec_info bb_vinfo) |
ebfd146a IR |
6081 | { |
6082 | stmt_vec_info res; | |
6083 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
6084 | ||
6085 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
6086 | STMT_VINFO_STMT (res) = stmt; | |
6087 | STMT_VINFO_LOOP_VINFO (res) = loop_vinfo; | |
a70d6342 | 6088 | STMT_VINFO_BB_VINFO (res) = bb_vinfo; |
8644a673 | 6089 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
6090 | STMT_VINFO_LIVE_P (res) = false; |
6091 | STMT_VINFO_VECTYPE (res) = NULL; | |
6092 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 6093 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
6094 | STMT_VINFO_IN_PATTERN_P (res) = false; |
6095 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 6096 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a IR |
6097 | STMT_VINFO_DATA_REF (res) = NULL; |
6098 | ||
6099 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
6100 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
6101 | STMT_VINFO_DR_INIT (res) = NULL; | |
6102 | STMT_VINFO_DR_STEP (res) = NULL; | |
6103 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
6104 | ||
6105 | if (gimple_code (stmt) == GIMPLE_PHI | |
6106 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
6107 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
6108 | else | |
8644a673 IR |
6109 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
6110 | ||
9771b263 | 6111 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 6112 | STMT_SLP_TYPE (res) = loop_vect; |
e14c1050 IR |
6113 | GROUP_FIRST_ELEMENT (res) = NULL; |
6114 | GROUP_NEXT_ELEMENT (res) = NULL; | |
6115 | GROUP_SIZE (res) = 0; | |
6116 | GROUP_STORE_COUNT (res) = 0; | |
6117 | GROUP_GAP (res) = 0; | |
6118 | GROUP_SAME_DR_STMT (res) = NULL; | |
ebfd146a IR |
6119 | |
6120 | return res; | |
6121 | } | |
6122 | ||
6123 | ||
6124 | /* Create a hash table for stmt_vec_info. */ | |
6125 | ||
6126 | void | |
6127 | init_stmt_vec_info_vec (void) | |
6128 | { | |
9771b263 DN |
6129 | gcc_assert (!stmt_vec_info_vec.exists ()); |
6130 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
6131 | } |
6132 | ||
6133 | ||
6134 | /* Free hash table for stmt_vec_info. */ | |
6135 | ||
6136 | void | |
6137 | free_stmt_vec_info_vec (void) | |
6138 | { | |
93675444 JJ |
6139 | unsigned int i; |
6140 | vec_void_p info; | |
6141 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) | |
6142 | if (info != NULL) | |
6143 | free_stmt_vec_info (STMT_VINFO_STMT ((stmt_vec_info) info)); | |
9771b263 DN |
6144 | gcc_assert (stmt_vec_info_vec.exists ()); |
6145 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
6146 | } |
6147 | ||
6148 | ||
6149 | /* Free stmt vectorization related info. */ | |
6150 | ||
6151 | void | |
6152 | free_stmt_vec_info (gimple stmt) | |
6153 | { | |
6154 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
6155 | ||
6156 | if (!stmt_info) | |
6157 | return; | |
6158 | ||
78048b1c JJ |
6159 | /* Check if this statement has a related "pattern stmt" |
6160 | (introduced by the vectorizer during the pattern recognition | |
6161 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
6162 | too. */ | |
6163 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
6164 | { | |
6165 | stmt_vec_info patt_info | |
6166 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
6167 | if (patt_info) | |
6168 | { | |
363477c0 JJ |
6169 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
6170 | if (seq) | |
6171 | { | |
6172 | gimple_stmt_iterator si; | |
6173 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
6174 | free_stmt_vec_info (gsi_stmt (si)); | |
6175 | } | |
78048b1c JJ |
6176 | free_stmt_vec_info (STMT_VINFO_RELATED_STMT (stmt_info)); |
6177 | } | |
6178 | } | |
6179 | ||
9771b263 | 6180 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
ebfd146a IR |
6181 | set_vinfo_for_stmt (stmt, NULL); |
6182 | free (stmt_info); | |
6183 | } | |
6184 | ||
6185 | ||
bb67d9c7 | 6186 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 6187 | |
bb67d9c7 | 6188 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
6189 | by the target. */ |
6190 | ||
bb67d9c7 RG |
6191 | static tree |
6192 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a IR |
6193 | { |
6194 | enum machine_mode inner_mode = TYPE_MODE (scalar_type); | |
cc4b5170 | 6195 | enum machine_mode simd_mode; |
2f816591 | 6196 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
6197 | int nunits; |
6198 | tree vectype; | |
6199 | ||
cc4b5170 | 6200 | if (nbytes == 0) |
ebfd146a IR |
6201 | return NULL_TREE; |
6202 | ||
48f2e373 RB |
6203 | if (GET_MODE_CLASS (inner_mode) != MODE_INT |
6204 | && GET_MODE_CLASS (inner_mode) != MODE_FLOAT) | |
6205 | return NULL_TREE; | |
6206 | ||
7b7b1813 RG |
6207 | /* For vector types of elements whose mode precision doesn't |
6208 | match their types precision we use a element type of mode | |
6209 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
6210 | they support the proper result truncation/extension. |
6211 | We also make sure to build vector types with INTEGER_TYPE | |
6212 | component type only. */ | |
6d7971b8 | 6213 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
6214 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
6215 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
6216 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
6217 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 6218 | |
ccbf5bb4 RG |
6219 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
6220 | When the component mode passes the above test simply use a type | |
6221 | corresponding to that mode. The theory is that any use that | |
6222 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 6223 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 6224 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
6225 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
6226 | ||
6227 | /* We can't build a vector type of elements with alignment bigger than | |
6228 | their size. */ | |
dfc2e2ac | 6229 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
6230 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
6231 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 6232 | |
dfc2e2ac RB |
6233 | /* If we felt back to using the mode fail if there was |
6234 | no scalar type for it. */ | |
6235 | if (scalar_type == NULL_TREE) | |
6236 | return NULL_TREE; | |
6237 | ||
bb67d9c7 RG |
6238 | /* If no size was supplied use the mode the target prefers. Otherwise |
6239 | lookup a vector mode of the specified size. */ | |
6240 | if (size == 0) | |
6241 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
6242 | else | |
6243 | simd_mode = mode_for_vector (inner_mode, size / nbytes); | |
cc4b5170 RG |
6244 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
6245 | if (nunits <= 1) | |
6246 | return NULL_TREE; | |
ebfd146a IR |
6247 | |
6248 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
6249 | |
6250 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
6251 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 6252 | return NULL_TREE; |
ebfd146a IR |
6253 | |
6254 | return vectype; | |
6255 | } | |
6256 | ||
bb67d9c7 RG |
6257 | unsigned int current_vector_size; |
6258 | ||
6259 | /* Function get_vectype_for_scalar_type. | |
6260 | ||
6261 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
6262 | by the target. */ | |
6263 | ||
6264 | tree | |
6265 | get_vectype_for_scalar_type (tree scalar_type) | |
6266 | { | |
6267 | tree vectype; | |
6268 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
6269 | current_vector_size); | |
6270 | if (vectype | |
6271 | && current_vector_size == 0) | |
6272 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
6273 | return vectype; | |
6274 | } | |
6275 | ||
b690cc0f RG |
6276 | /* Function get_same_sized_vectype |
6277 | ||
6278 | Returns a vector type corresponding to SCALAR_TYPE of size | |
6279 | VECTOR_TYPE if supported by the target. */ | |
6280 | ||
6281 | tree | |
bb67d9c7 | 6282 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 6283 | { |
bb67d9c7 RG |
6284 | return get_vectype_for_scalar_type_and_size |
6285 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
6286 | } |
6287 | ||
ebfd146a IR |
6288 | /* Function vect_is_simple_use. |
6289 | ||
6290 | Input: | |
a70d6342 IR |
6291 | LOOP_VINFO - the vect info of the loop that is being vectorized. |
6292 | BB_VINFO - the vect info of the basic block that is being vectorized. | |
24ee1384 | 6293 | OPERAND - operand of STMT in the loop or bb. |
ebfd146a IR |
6294 | DEF - the defining stmt in case OPERAND is an SSA_NAME. |
6295 | ||
6296 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 6297 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 6298 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 6299 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
6300 | is the case in reduction/induction computations). |
6301 | For basic blocks, supportable operands are constants and bb invariants. | |
6302 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
6303 | |
6304 | bool | |
24ee1384 | 6305 | vect_is_simple_use (tree operand, gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 6306 | bb_vec_info bb_vinfo, gimple *def_stmt, |
ebfd146a | 6307 | tree *def, enum vect_def_type *dt) |
b8698a0f | 6308 | { |
ebfd146a IR |
6309 | basic_block bb; |
6310 | stmt_vec_info stmt_vinfo; | |
a70d6342 | 6311 | struct loop *loop = NULL; |
b8698a0f | 6312 | |
a70d6342 IR |
6313 | if (loop_vinfo) |
6314 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
6315 | |
6316 | *def_stmt = NULL; | |
6317 | *def = NULL_TREE; | |
b8698a0f | 6318 | |
73fbfcad | 6319 | if (dump_enabled_p ()) |
ebfd146a | 6320 | { |
78c60e3d SS |
6321 | dump_printf_loc (MSG_NOTE, vect_location, |
6322 | "vect_is_simple_use: operand "); | |
6323 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
e645e942 | 6324 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 6325 | } |
b8698a0f | 6326 | |
b758f602 | 6327 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
6328 | { |
6329 | *dt = vect_constant_def; | |
6330 | return true; | |
6331 | } | |
b8698a0f | 6332 | |
ebfd146a IR |
6333 | if (is_gimple_min_invariant (operand)) |
6334 | { | |
6335 | *def = operand; | |
8644a673 | 6336 | *dt = vect_external_def; |
ebfd146a IR |
6337 | return true; |
6338 | } | |
6339 | ||
6340 | if (TREE_CODE (operand) == PAREN_EXPR) | |
6341 | { | |
73fbfcad | 6342 | if (dump_enabled_p ()) |
e645e942 | 6343 | dump_printf_loc (MSG_NOTE, vect_location, "non-associatable copy.\n"); |
ebfd146a IR |
6344 | operand = TREE_OPERAND (operand, 0); |
6345 | } | |
b8698a0f | 6346 | |
ebfd146a IR |
6347 | if (TREE_CODE (operand) != SSA_NAME) |
6348 | { | |
73fbfcad | 6349 | if (dump_enabled_p ()) |
78c60e3d | 6350 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6351 | "not ssa-name.\n"); |
ebfd146a IR |
6352 | return false; |
6353 | } | |
b8698a0f | 6354 | |
ebfd146a IR |
6355 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
6356 | if (*def_stmt == NULL) | |
6357 | { | |
73fbfcad | 6358 | if (dump_enabled_p ()) |
78c60e3d | 6359 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6360 | "no def_stmt.\n"); |
ebfd146a IR |
6361 | return false; |
6362 | } | |
6363 | ||
73fbfcad | 6364 | if (dump_enabled_p ()) |
ebfd146a | 6365 | { |
78c60e3d SS |
6366 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
6367 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
e645e942 | 6368 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
6369 | } |
6370 | ||
8644a673 | 6371 | /* Empty stmt is expected only in case of a function argument. |
ebfd146a IR |
6372 | (Otherwise - we expect a phi_node or a GIMPLE_ASSIGN). */ |
6373 | if (gimple_nop_p (*def_stmt)) | |
6374 | { | |
6375 | *def = operand; | |
8644a673 | 6376 | *dt = vect_external_def; |
ebfd146a IR |
6377 | return true; |
6378 | } | |
6379 | ||
6380 | bb = gimple_bb (*def_stmt); | |
a70d6342 IR |
6381 | |
6382 | if ((loop && !flow_bb_inside_loop_p (loop, bb)) | |
6383 | || (!loop && bb != BB_VINFO_BB (bb_vinfo)) | |
b8698a0f | 6384 | || (!loop && gimple_code (*def_stmt) == GIMPLE_PHI)) |
8644a673 | 6385 | *dt = vect_external_def; |
ebfd146a IR |
6386 | else |
6387 | { | |
6388 | stmt_vinfo = vinfo_for_stmt (*def_stmt); | |
6389 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
6390 | } | |
6391 | ||
24ee1384 IR |
6392 | if (*dt == vect_unknown_def_type |
6393 | || (stmt | |
6394 | && *dt == vect_double_reduction_def | |
6395 | && gimple_code (stmt) != GIMPLE_PHI)) | |
ebfd146a | 6396 | { |
73fbfcad | 6397 | if (dump_enabled_p ()) |
78c60e3d | 6398 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6399 | "Unsupported pattern.\n"); |
ebfd146a IR |
6400 | return false; |
6401 | } | |
6402 | ||
73fbfcad | 6403 | if (dump_enabled_p ()) |
e645e942 | 6404 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: %d.\n", *dt); |
ebfd146a IR |
6405 | |
6406 | switch (gimple_code (*def_stmt)) | |
6407 | { | |
6408 | case GIMPLE_PHI: | |
6409 | *def = gimple_phi_result (*def_stmt); | |
6410 | break; | |
6411 | ||
6412 | case GIMPLE_ASSIGN: | |
6413 | *def = gimple_assign_lhs (*def_stmt); | |
6414 | break; | |
6415 | ||
6416 | case GIMPLE_CALL: | |
6417 | *def = gimple_call_lhs (*def_stmt); | |
6418 | if (*def != NULL) | |
6419 | break; | |
6420 | /* FALLTHRU */ | |
6421 | default: | |
73fbfcad | 6422 | if (dump_enabled_p ()) |
78c60e3d | 6423 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6424 | "unsupported defining stmt:\n"); |
ebfd146a IR |
6425 | return false; |
6426 | } | |
6427 | ||
6428 | return true; | |
6429 | } | |
6430 | ||
b690cc0f RG |
6431 | /* Function vect_is_simple_use_1. |
6432 | ||
6433 | Same as vect_is_simple_use_1 but also determines the vector operand | |
6434 | type of OPERAND and stores it to *VECTYPE. If the definition of | |
6435 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
6436 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
6437 | is responsible to compute the best suited vector type for the | |
6438 | scalar operand. */ | |
6439 | ||
6440 | bool | |
24ee1384 | 6441 | vect_is_simple_use_1 (tree operand, gimple stmt, loop_vec_info loop_vinfo, |
b690cc0f RG |
6442 | bb_vec_info bb_vinfo, gimple *def_stmt, |
6443 | tree *def, enum vect_def_type *dt, tree *vectype) | |
6444 | { | |
24ee1384 IR |
6445 | if (!vect_is_simple_use (operand, stmt, loop_vinfo, bb_vinfo, def_stmt, |
6446 | def, dt)) | |
b690cc0f RG |
6447 | return false; |
6448 | ||
6449 | /* Now get a vector type if the def is internal, otherwise supply | |
6450 | NULL_TREE and leave it up to the caller to figure out a proper | |
6451 | type for the use stmt. */ | |
6452 | if (*dt == vect_internal_def | |
6453 | || *dt == vect_induction_def | |
6454 | || *dt == vect_reduction_def | |
6455 | || *dt == vect_double_reduction_def | |
6456 | || *dt == vect_nested_cycle) | |
6457 | { | |
6458 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
6459 | |
6460 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
6461 | && !STMT_VINFO_RELEVANT (stmt_info) | |
6462 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 6463 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 6464 | |
b690cc0f RG |
6465 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
6466 | gcc_assert (*vectype != NULL_TREE); | |
6467 | } | |
6468 | else if (*dt == vect_uninitialized_def | |
6469 | || *dt == vect_constant_def | |
6470 | || *dt == vect_external_def) | |
6471 | *vectype = NULL_TREE; | |
6472 | else | |
6473 | gcc_unreachable (); | |
6474 | ||
6475 | return true; | |
6476 | } | |
6477 | ||
ebfd146a IR |
6478 | |
6479 | /* Function supportable_widening_operation | |
6480 | ||
b8698a0f L |
6481 | Check whether an operation represented by the code CODE is a |
6482 | widening operation that is supported by the target platform in | |
b690cc0f RG |
6483 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
6484 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 6485 | |
ebfd146a IR |
6486 | Widening operations we currently support are NOP (CONVERT), FLOAT |
6487 | and WIDEN_MULT. This function checks if these operations are supported | |
6488 | by the target platform either directly (via vector tree-codes), or via | |
6489 | target builtins. | |
6490 | ||
6491 | Output: | |
b8698a0f L |
6492 | - CODE1 and CODE2 are codes of vector operations to be used when |
6493 | vectorizing the operation, if available. | |
ebfd146a IR |
6494 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
6495 | case of multi-step conversion (like char->short->int - in that case | |
6496 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
6497 | - INTERM_TYPES contains the intermediate type required to perform the |
6498 | widening operation (short in the above example). */ | |
ebfd146a IR |
6499 | |
6500 | bool | |
b690cc0f RG |
6501 | supportable_widening_operation (enum tree_code code, gimple stmt, |
6502 | tree vectype_out, tree vectype_in, | |
ebfd146a IR |
6503 | enum tree_code *code1, enum tree_code *code2, |
6504 | int *multi_step_cvt, | |
9771b263 | 6505 | vec<tree> *interm_types) |
ebfd146a IR |
6506 | { |
6507 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
6508 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 6509 | struct loop *vect_loop = NULL; |
ebfd146a | 6510 | enum machine_mode vec_mode; |
81f40b79 | 6511 | enum insn_code icode1, icode2; |
ebfd146a | 6512 | optab optab1, optab2; |
b690cc0f RG |
6513 | tree vectype = vectype_in; |
6514 | tree wide_vectype = vectype_out; | |
ebfd146a | 6515 | enum tree_code c1, c2; |
4a00c761 JJ |
6516 | int i; |
6517 | tree prev_type, intermediate_type; | |
6518 | enum machine_mode intermediate_mode, prev_mode; | |
6519 | optab optab3, optab4; | |
ebfd146a | 6520 | |
4a00c761 | 6521 | *multi_step_cvt = 0; |
4ef69dfc IR |
6522 | if (loop_info) |
6523 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
6524 | ||
ebfd146a IR |
6525 | switch (code) |
6526 | { | |
6527 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
6528 | /* The result of a vectorized widening operation usually requires |
6529 | two vectors (because the widened results do not fit into one vector). | |
6530 | The generated vector results would normally be expected to be | |
6531 | generated in the same order as in the original scalar computation, | |
6532 | i.e. if 8 results are generated in each vector iteration, they are | |
6533 | to be organized as follows: | |
6534 | vect1: [res1,res2,res3,res4], | |
6535 | vect2: [res5,res6,res7,res8]. | |
6536 | ||
6537 | However, in the special case that the result of the widening | |
6538 | operation is used in a reduction computation only, the order doesn't | |
6539 | matter (because when vectorizing a reduction we change the order of | |
6540 | the computation). Some targets can take advantage of this and | |
6541 | generate more efficient code. For example, targets like Altivec, | |
6542 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
6543 | generate the following vectors: | |
6544 | vect1: [res1,res3,res5,res7], | |
6545 | vect2: [res2,res4,res6,res8]. | |
6546 | ||
6547 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
6548 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
6549 | iterations in parallel). We therefore don't allow to change the | |
6550 | order of the computation in the inner-loop during outer-loop | |
6551 | vectorization. */ | |
6552 | /* TODO: Another case in which order doesn't *really* matter is when we | |
6553 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
6554 | Normally, pack_trunc performs an even/odd permute, whereas the | |
6555 | repack from an even/odd expansion would be an interleave, which | |
6556 | would be significantly simpler for e.g. AVX2. */ | |
6557 | /* In any case, in order to avoid duplicating the code below, recurse | |
6558 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
6559 | are properly set up for the caller. If we fail, we'll continue with | |
6560 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
6561 | if (vect_loop | |
6562 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
6563 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
6564 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
6565 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
6566 | code1, code2, multi_step_cvt, |
6567 | interm_types)) | |
6ae6116f | 6568 | return true; |
4a00c761 JJ |
6569 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
6570 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
6571 | break; |
6572 | ||
6ae6116f RH |
6573 | case VEC_WIDEN_MULT_EVEN_EXPR: |
6574 | /* Support the recursion induced just above. */ | |
6575 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
6576 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
6577 | break; | |
6578 | ||
36ba4aae | 6579 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
6580 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
6581 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
6582 | break; |
6583 | ||
ebfd146a | 6584 | CASE_CONVERT: |
4a00c761 JJ |
6585 | c1 = VEC_UNPACK_LO_EXPR; |
6586 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
6587 | break; |
6588 | ||
6589 | case FLOAT_EXPR: | |
4a00c761 JJ |
6590 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
6591 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
6592 | break; |
6593 | ||
6594 | case FIX_TRUNC_EXPR: | |
6595 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
6596 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
6597 | computing the operation. */ | |
6598 | return false; | |
6599 | ||
6600 | default: | |
6601 | gcc_unreachable (); | |
6602 | } | |
6603 | ||
6ae6116f | 6604 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
4a00c761 JJ |
6605 | { |
6606 | enum tree_code ctmp = c1; | |
6607 | c1 = c2; | |
6608 | c2 = ctmp; | |
6609 | } | |
6610 | ||
ebfd146a IR |
6611 | if (code == FIX_TRUNC_EXPR) |
6612 | { | |
6613 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
6614 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
6615 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
6616 | } |
6617 | else | |
6618 | { | |
6619 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
6620 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
6621 | } | |
6622 | ||
6623 | if (!optab1 || !optab2) | |
6624 | return false; | |
6625 | ||
6626 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
6627 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
6628 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
6629 | return false; |
6630 | ||
4a00c761 JJ |
6631 | *code1 = c1; |
6632 | *code2 = c2; | |
6633 | ||
6634 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
6635 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
6636 | return true; | |
6637 | ||
b8698a0f | 6638 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 6639 | types. */ |
ebfd146a | 6640 | |
4a00c761 JJ |
6641 | prev_type = vectype; |
6642 | prev_mode = vec_mode; | |
b8698a0f | 6643 | |
4a00c761 JJ |
6644 | if (!CONVERT_EXPR_CODE_P (code)) |
6645 | return false; | |
b8698a0f | 6646 | |
4a00c761 JJ |
6647 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
6648 | intermediate steps in promotion sequence. We try | |
6649 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
6650 | not. */ | |
9771b263 | 6651 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
6652 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
6653 | { | |
6654 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
6655 | intermediate_type | |
6656 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
6657 | TYPE_UNSIGNED (prev_type)); | |
6658 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); | |
6659 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
6660 | ||
6661 | if (!optab3 || !optab4 | |
6662 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
6663 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
6664 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
6665 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
6666 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
6667 | == CODE_FOR_nothing) | |
6668 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
6669 | == CODE_FOR_nothing)) | |
6670 | break; | |
ebfd146a | 6671 | |
9771b263 | 6672 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
6673 | (*multi_step_cvt)++; |
6674 | ||
6675 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
6676 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
6677 | return true; | |
6678 | ||
6679 | prev_type = intermediate_type; | |
6680 | prev_mode = intermediate_mode; | |
ebfd146a IR |
6681 | } |
6682 | ||
9771b263 | 6683 | interm_types->release (); |
4a00c761 | 6684 | return false; |
ebfd146a IR |
6685 | } |
6686 | ||
6687 | ||
6688 | /* Function supportable_narrowing_operation | |
6689 | ||
b8698a0f L |
6690 | Check whether an operation represented by the code CODE is a |
6691 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
6692 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
6693 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 6694 | |
ebfd146a | 6695 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 6696 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
6697 | the target platform directly via vector tree-codes. |
6698 | ||
6699 | Output: | |
b8698a0f L |
6700 | - CODE1 is the code of a vector operation to be used when |
6701 | vectorizing the operation, if available. | |
ebfd146a IR |
6702 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
6703 | case of multi-step conversion (like int->short->char - in that case | |
6704 | MULTI_STEP_CVT will be 1). | |
6705 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 6706 | narrowing operation (short in the above example). */ |
ebfd146a IR |
6707 | |
6708 | bool | |
6709 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 6710 | tree vectype_out, tree vectype_in, |
ebfd146a | 6711 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 6712 | vec<tree> *interm_types) |
ebfd146a IR |
6713 | { |
6714 | enum machine_mode vec_mode; | |
6715 | enum insn_code icode1; | |
6716 | optab optab1, interm_optab; | |
b690cc0f RG |
6717 | tree vectype = vectype_in; |
6718 | tree narrow_vectype = vectype_out; | |
ebfd146a | 6719 | enum tree_code c1; |
4a00c761 JJ |
6720 | tree intermediate_type; |
6721 | enum machine_mode intermediate_mode, prev_mode; | |
ebfd146a | 6722 | int i; |
4a00c761 | 6723 | bool uns; |
ebfd146a | 6724 | |
4a00c761 | 6725 | *multi_step_cvt = 0; |
ebfd146a IR |
6726 | switch (code) |
6727 | { | |
6728 | CASE_CONVERT: | |
6729 | c1 = VEC_PACK_TRUNC_EXPR; | |
6730 | break; | |
6731 | ||
6732 | case FIX_TRUNC_EXPR: | |
6733 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
6734 | break; | |
6735 | ||
6736 | case FLOAT_EXPR: | |
6737 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
6738 | tree code and optabs used for computing the operation. */ | |
6739 | return false; | |
6740 | ||
6741 | default: | |
6742 | gcc_unreachable (); | |
6743 | } | |
6744 | ||
6745 | if (code == FIX_TRUNC_EXPR) | |
6746 | /* The signedness is determined from output operand. */ | |
b690cc0f | 6747 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
6748 | else |
6749 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
6750 | ||
6751 | if (!optab1) | |
6752 | return false; | |
6753 | ||
6754 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 6755 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
6756 | return false; |
6757 | ||
4a00c761 JJ |
6758 | *code1 = c1; |
6759 | ||
6760 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
6761 | return true; | |
6762 | ||
ebfd146a IR |
6763 | /* Check if it's a multi-step conversion that can be done using intermediate |
6764 | types. */ | |
4a00c761 JJ |
6765 | prev_mode = vec_mode; |
6766 | if (code == FIX_TRUNC_EXPR) | |
6767 | uns = TYPE_UNSIGNED (vectype_out); | |
6768 | else | |
6769 | uns = TYPE_UNSIGNED (vectype); | |
6770 | ||
6771 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
6772 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
6773 | costly than signed. */ | |
6774 | if (code == FIX_TRUNC_EXPR && uns) | |
6775 | { | |
6776 | enum insn_code icode2; | |
6777 | ||
6778 | intermediate_type | |
6779 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
6780 | interm_optab | |
6781 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 6782 | if (interm_optab != unknown_optab |
4a00c761 JJ |
6783 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
6784 | && insn_data[icode1].operand[0].mode | |
6785 | == insn_data[icode2].operand[0].mode) | |
6786 | { | |
6787 | uns = false; | |
6788 | optab1 = interm_optab; | |
6789 | icode1 = icode2; | |
6790 | } | |
6791 | } | |
ebfd146a | 6792 | |
4a00c761 JJ |
6793 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
6794 | intermediate steps in promotion sequence. We try | |
6795 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 6796 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
6797 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
6798 | { | |
6799 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
6800 | intermediate_type | |
6801 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
6802 | interm_optab | |
6803 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
6804 | optab_default); | |
6805 | if (!interm_optab | |
6806 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
6807 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
6808 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
6809 | == CODE_FOR_nothing)) | |
6810 | break; | |
6811 | ||
9771b263 | 6812 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
6813 | (*multi_step_cvt)++; |
6814 | ||
6815 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
6816 | return true; | |
6817 | ||
6818 | prev_mode = intermediate_mode; | |
6819 | optab1 = interm_optab; | |
ebfd146a IR |
6820 | } |
6821 | ||
9771b263 | 6822 | interm_types->release (); |
4a00c761 | 6823 | return false; |
ebfd146a | 6824 | } |