]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
818ab71a | 2 | Copyright (C) 2003-2016 Free Software Foundation, Inc. |
b8698a0f | 3 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
4 | and Ira Rosen <irar@il.ibm.com> |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 3, or (at your option) any later | |
11 | version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
c7131fb2 | 25 | #include "backend.h" |
957060b5 AM |
26 | #include "target.h" |
27 | #include "rtl.h" | |
ebfd146a | 28 | #include "tree.h" |
c7131fb2 | 29 | #include "gimple.h" |
c7131fb2 | 30 | #include "ssa.h" |
957060b5 AM |
31 | #include "optabs-tree.h" |
32 | #include "insn-config.h" | |
33 | #include "recog.h" /* FIXME: for insn_data */ | |
34 | #include "cgraph.h" | |
957060b5 | 35 | #include "dumpfile.h" |
c7131fb2 | 36 | #include "alias.h" |
40e23961 | 37 | #include "fold-const.h" |
d8a2d370 | 38 | #include "stor-layout.h" |
2fb9a547 | 39 | #include "tree-eh.h" |
45b0be94 | 40 | #include "gimplify.h" |
5be5c238 | 41 | #include "gimple-iterator.h" |
18f429e2 | 42 | #include "gimplify-me.h" |
442b4905 | 43 | #include "tree-cfg.h" |
e28030cf | 44 | #include "tree-ssa-loop-manip.h" |
ebfd146a | 45 | #include "cfgloop.h" |
0136f8f0 AH |
46 | #include "tree-ssa-loop.h" |
47 | #include "tree-scalar-evolution.h" | |
ebfd146a | 48 | #include "tree-vectorizer.h" |
9b2b7279 | 49 | #include "builtins.h" |
70439f0d | 50 | #include "internal-fn.h" |
ebfd146a | 51 | |
7ee2468b SB |
52 | /* For lang_hooks.types.type_for_mode. */ |
53 | #include "langhooks.h" | |
ebfd146a | 54 | |
c3e7ee41 BS |
55 | /* Return the vectorized type for the given statement. */ |
56 | ||
57 | tree | |
58 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
59 | { | |
60 | return STMT_VINFO_VECTYPE (stmt_info); | |
61 | } | |
62 | ||
63 | /* Return TRUE iff the given statement is in an inner loop relative to | |
64 | the loop being vectorized. */ | |
65 | bool | |
66 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
67 | { | |
355fe088 | 68 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
c3e7ee41 BS |
69 | basic_block bb = gimple_bb (stmt); |
70 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
71 | struct loop* loop; | |
72 | ||
73 | if (!loop_vinfo) | |
74 | return false; | |
75 | ||
76 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
77 | ||
78 | return (bb->loop_father == loop->inner); | |
79 | } | |
80 | ||
81 | /* Record the cost of a statement, either by directly informing the | |
82 | target model or by saving it in a vector for later processing. | |
83 | Return a preliminary estimate of the statement's cost. */ | |
84 | ||
85 | unsigned | |
92345349 | 86 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 87 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 88 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 89 | { |
92345349 | 90 | if (body_cost_vec) |
c3e7ee41 | 91 | { |
92345349 | 92 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
ddf56386 RB |
93 | stmt_info_for_cost si = { count, kind, |
94 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
95 | misalign }; | |
96 | body_cost_vec->safe_push (si); | |
c3e7ee41 | 97 | return (unsigned) |
92345349 | 98 | (builtin_vectorization_cost (kind, vectype, misalign) * count); |
c3e7ee41 BS |
99 | } |
100 | else | |
310213d4 RB |
101 | return add_stmt_cost (stmt_info->vinfo->target_cost_data, |
102 | count, kind, stmt_info, misalign, where); | |
c3e7ee41 BS |
103 | } |
104 | ||
272c6793 RS |
105 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
106 | ||
107 | static tree | |
108 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
109 | { | |
110 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
111 | "vect_array"); | |
112 | } | |
113 | ||
114 | /* ARRAY is an array of vectors created by create_vector_array. | |
115 | Return an SSA_NAME for the vector in index N. The reference | |
116 | is part of the vectorization of STMT and the vector is associated | |
117 | with scalar destination SCALAR_DEST. */ | |
118 | ||
119 | static tree | |
355fe088 | 120 | read_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree scalar_dest, |
272c6793 RS |
121 | tree array, unsigned HOST_WIDE_INT n) |
122 | { | |
123 | tree vect_type, vect, vect_name, array_ref; | |
355fe088 | 124 | gimple *new_stmt; |
272c6793 RS |
125 | |
126 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
127 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
128 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
129 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
130 | build_int_cst (size_type_node, n), | |
131 | NULL_TREE, NULL_TREE); | |
132 | ||
133 | new_stmt = gimple_build_assign (vect, array_ref); | |
134 | vect_name = make_ssa_name (vect, new_stmt); | |
135 | gimple_assign_set_lhs (new_stmt, vect_name); | |
136 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
137 | |
138 | return vect_name; | |
139 | } | |
140 | ||
141 | /* ARRAY is an array of vectors created by create_vector_array. | |
142 | Emit code to store SSA_NAME VECT in index N of the array. | |
143 | The store is part of the vectorization of STMT. */ | |
144 | ||
145 | static void | |
355fe088 | 146 | write_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree vect, |
272c6793 RS |
147 | tree array, unsigned HOST_WIDE_INT n) |
148 | { | |
149 | tree array_ref; | |
355fe088 | 150 | gimple *new_stmt; |
272c6793 RS |
151 | |
152 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
153 | build_int_cst (size_type_node, n), | |
154 | NULL_TREE, NULL_TREE); | |
155 | ||
156 | new_stmt = gimple_build_assign (array_ref, vect); | |
157 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
158 | } |
159 | ||
160 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
161 | of *PTR. The memory reference replaces those in FIRST_DR | |
162 | (and its group). */ | |
163 | ||
164 | static tree | |
165 | create_array_ref (tree type, tree ptr, struct data_reference *first_dr) | |
166 | { | |
272c6793 RS |
167 | tree mem_ref, alias_ptr_type; |
168 | ||
169 | alias_ptr_type = reference_alias_ptr_type (DR_REF (first_dr)); | |
170 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); | |
171 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 172 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
173 | return mem_ref; |
174 | } | |
175 | ||
ebfd146a IR |
176 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
177 | ||
178 | /* Function vect_mark_relevant. | |
179 | ||
180 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
181 | ||
182 | static void | |
355fe088 | 183 | vect_mark_relevant (vec<gimple *> *worklist, gimple *stmt, |
83197f37 IR |
184 | enum vect_relevant relevant, bool live_p, |
185 | bool used_in_pattern) | |
ebfd146a IR |
186 | { |
187 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
188 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
189 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
355fe088 | 190 | gimple *pattern_stmt; |
ebfd146a | 191 | |
73fbfcad | 192 | if (dump_enabled_p ()) |
66c16fd9 RB |
193 | { |
194 | dump_printf_loc (MSG_NOTE, vect_location, | |
195 | "mark relevant %d, live %d: ", relevant, live_p); | |
196 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
197 | } | |
ebfd146a | 198 | |
83197f37 IR |
199 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
200 | related pattern stmt instead of the original stmt. However, such stmts | |
201 | may have their own uses that are not in any pattern, in such cases the | |
202 | stmt itself should be marked. */ | |
ebfd146a IR |
203 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
204 | { | |
83197f37 IR |
205 | bool found = false; |
206 | if (!used_in_pattern) | |
207 | { | |
208 | imm_use_iterator imm_iter; | |
209 | use_operand_p use_p; | |
355fe088 | 210 | gimple *use_stmt; |
83197f37 | 211 | tree lhs; |
13c931c9 JJ |
212 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
213 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a | 214 | |
83197f37 IR |
215 | if (is_gimple_assign (stmt)) |
216 | lhs = gimple_assign_lhs (stmt); | |
217 | else | |
218 | lhs = gimple_call_lhs (stmt); | |
ebfd146a | 219 | |
83197f37 IR |
220 | /* This use is out of pattern use, if LHS has other uses that are |
221 | pattern uses, we should mark the stmt itself, and not the pattern | |
222 | stmt. */ | |
5ce9450f | 223 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
ab0ef706 JJ |
224 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs) |
225 | { | |
226 | if (is_gimple_debug (USE_STMT (use_p))) | |
227 | continue; | |
228 | use_stmt = USE_STMT (use_p); | |
229 | ||
13c931c9 JJ |
230 | if (!flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) |
231 | continue; | |
232 | ||
ab0ef706 JJ |
233 | if (vinfo_for_stmt (use_stmt) |
234 | && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (use_stmt))) | |
235 | { | |
236 | found = true; | |
237 | break; | |
238 | } | |
239 | } | |
83197f37 IR |
240 | } |
241 | ||
242 | if (!found) | |
243 | { | |
244 | /* This is the last stmt in a sequence that was detected as a | |
245 | pattern that can potentially be vectorized. Don't mark the stmt | |
246 | as relevant/live because it's not going to be vectorized. | |
247 | Instead mark the pattern-stmt that replaces it. */ | |
248 | ||
249 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
250 | ||
73fbfcad | 251 | if (dump_enabled_p ()) |
78c60e3d SS |
252 | dump_printf_loc (MSG_NOTE, vect_location, |
253 | "last stmt in pattern. don't mark" | |
e645e942 | 254 | " relevant/live.\n"); |
83197f37 IR |
255 | stmt_info = vinfo_for_stmt (pattern_stmt); |
256 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
257 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
258 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
259 | stmt = pattern_stmt; | |
260 | } | |
ebfd146a IR |
261 | } |
262 | ||
263 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
264 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
265 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
266 | ||
267 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
268 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
269 | { | |
73fbfcad | 270 | if (dump_enabled_p ()) |
78c60e3d | 271 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 272 | "already marked relevant/live.\n"); |
ebfd146a IR |
273 | return; |
274 | } | |
275 | ||
9771b263 | 276 | worklist->safe_push (stmt); |
ebfd146a IR |
277 | } |
278 | ||
279 | ||
280 | /* Function vect_stmt_relevant_p. | |
281 | ||
282 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
283 | "relevant for vectorization". | |
284 | ||
285 | A stmt is considered "relevant for vectorization" if: | |
286 | - it has uses outside the loop. | |
287 | - it has vdefs (it alters memory). | |
288 | - control stmts in the loop (except for the exit condition). | |
289 | ||
290 | CHECKME: what other side effects would the vectorizer allow? */ | |
291 | ||
292 | static bool | |
355fe088 | 293 | vect_stmt_relevant_p (gimple *stmt, loop_vec_info loop_vinfo, |
ebfd146a IR |
294 | enum vect_relevant *relevant, bool *live_p) |
295 | { | |
296 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
297 | ssa_op_iter op_iter; | |
298 | imm_use_iterator imm_iter; | |
299 | use_operand_p use_p; | |
300 | def_operand_p def_p; | |
301 | ||
8644a673 | 302 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
303 | *live_p = false; |
304 | ||
305 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
306 | if (is_ctrl_stmt (stmt) |
307 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
308 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 309 | *relevant = vect_used_in_scope; |
ebfd146a IR |
310 | |
311 | /* changing memory. */ | |
312 | if (gimple_code (stmt) != GIMPLE_PHI) | |
ac6aeab4 RB |
313 | if (gimple_vdef (stmt) |
314 | && !gimple_clobber_p (stmt)) | |
ebfd146a | 315 | { |
73fbfcad | 316 | if (dump_enabled_p ()) |
78c60e3d | 317 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 318 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 319 | *relevant = vect_used_in_scope; |
ebfd146a IR |
320 | } |
321 | ||
322 | /* uses outside the loop. */ | |
323 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
324 | { | |
325 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
326 | { | |
327 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
328 | if (!flow_bb_inside_loop_p (loop, bb)) | |
329 | { | |
73fbfcad | 330 | if (dump_enabled_p ()) |
78c60e3d | 331 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 332 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 333 | |
3157b0c2 AO |
334 | if (is_gimple_debug (USE_STMT (use_p))) |
335 | continue; | |
336 | ||
ebfd146a IR |
337 | /* We expect all such uses to be in the loop exit phis |
338 | (because of loop closed form) */ | |
339 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
340 | gcc_assert (bb == single_exit (loop)->dest); | |
341 | ||
342 | *live_p = true; | |
343 | } | |
344 | } | |
345 | } | |
346 | ||
347 | return (*live_p || *relevant); | |
348 | } | |
349 | ||
350 | ||
b8698a0f | 351 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 352 | |
ff802fa1 | 353 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
354 | used in STMT for anything other than indexing an array. */ |
355 | ||
356 | static bool | |
355fe088 | 357 | exist_non_indexing_operands_for_use_p (tree use, gimple *stmt) |
ebfd146a IR |
358 | { |
359 | tree operand; | |
360 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 361 | |
ff802fa1 | 362 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
363 | reference in STMT, then any operand that corresponds to USE |
364 | is not indexing an array. */ | |
365 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
366 | return true; | |
59a05b0c | 367 | |
ebfd146a IR |
368 | /* STMT has a data_ref. FORNOW this means that its of one of |
369 | the following forms: | |
370 | -1- ARRAY_REF = var | |
371 | -2- var = ARRAY_REF | |
372 | (This should have been verified in analyze_data_refs). | |
373 | ||
374 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 375 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
376 | for array indexing. |
377 | ||
378 | Therefore, all we need to check is if STMT falls into the | |
379 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
380 | |
381 | if (!gimple_assign_copy_p (stmt)) | |
5ce9450f JJ |
382 | { |
383 | if (is_gimple_call (stmt) | |
384 | && gimple_call_internal_p (stmt)) | |
385 | switch (gimple_call_internal_fn (stmt)) | |
386 | { | |
387 | case IFN_MASK_STORE: | |
388 | operand = gimple_call_arg (stmt, 3); | |
389 | if (operand == use) | |
390 | return true; | |
391 | /* FALLTHRU */ | |
392 | case IFN_MASK_LOAD: | |
393 | operand = gimple_call_arg (stmt, 2); | |
394 | if (operand == use) | |
395 | return true; | |
396 | break; | |
397 | default: | |
398 | break; | |
399 | } | |
400 | return false; | |
401 | } | |
402 | ||
59a05b0c EB |
403 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
404 | return false; | |
ebfd146a | 405 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
406 | if (TREE_CODE (operand) != SSA_NAME) |
407 | return false; | |
408 | ||
409 | if (operand == use) | |
410 | return true; | |
411 | ||
412 | return false; | |
413 | } | |
414 | ||
415 | ||
b8698a0f | 416 | /* |
ebfd146a IR |
417 | Function process_use. |
418 | ||
419 | Inputs: | |
420 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b8698a0f | 421 | - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt |
ff802fa1 | 422 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 423 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
424 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
425 | be performed. | |
ebfd146a IR |
426 | |
427 | Outputs: | |
428 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
429 | relevance info of the DEF_STMT of this USE: | |
430 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
431 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
432 | Exceptions: | |
433 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 434 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 435 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
436 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
437 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
438 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
439 | be modified accordingly. | |
440 | ||
441 | Return true if everything is as expected. Return false otherwise. */ | |
442 | ||
443 | static bool | |
355fe088 TS |
444 | process_use (gimple *stmt, tree use, loop_vec_info loop_vinfo, bool live_p, |
445 | enum vect_relevant relevant, vec<gimple *> *worklist, | |
aec7ae7d | 446 | bool force) |
ebfd146a IR |
447 | { |
448 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
449 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
450 | stmt_vec_info dstmt_vinfo; | |
451 | basic_block bb, def_bb; | |
355fe088 | 452 | gimple *def_stmt; |
ebfd146a IR |
453 | enum vect_def_type dt; |
454 | ||
b8698a0f | 455 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 456 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 457 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
458 | return true; |
459 | ||
81c40241 | 460 | if (!vect_is_simple_use (use, loop_vinfo, &def_stmt, &dt)) |
b8698a0f | 461 | { |
73fbfcad | 462 | if (dump_enabled_p ()) |
78c60e3d | 463 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 464 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
465 | return false; |
466 | } | |
467 | ||
468 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
469 | return true; | |
470 | ||
471 | def_bb = gimple_bb (def_stmt); | |
472 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
473 | { | |
73fbfcad | 474 | if (dump_enabled_p ()) |
e645e942 | 475 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n"); |
ebfd146a IR |
476 | return true; |
477 | } | |
478 | ||
b8698a0f L |
479 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
480 | DEF_STMT must have already been processed, because this should be the | |
481 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
482 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
483 | check that everything is as expected, and we are done. */ |
484 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
485 | bb = gimple_bb (stmt); | |
486 | if (gimple_code (stmt) == GIMPLE_PHI | |
487 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
488 | && gimple_code (def_stmt) != GIMPLE_PHI | |
489 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
490 | && bb->loop_father == def_bb->loop_father) | |
491 | { | |
73fbfcad | 492 | if (dump_enabled_p ()) |
78c60e3d | 493 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 494 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a IR |
495 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
496 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
497 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 498 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 499 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
500 | return true; |
501 | } | |
502 | ||
503 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
504 | outer-loop-header-bb: | |
505 | d = def_stmt | |
506 | inner-loop: | |
507 | stmt # use (d) | |
508 | outer-loop-tail-bb: | |
509 | ... */ | |
510 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
511 | { | |
73fbfcad | 512 | if (dump_enabled_p ()) |
78c60e3d | 513 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 514 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 515 | |
ebfd146a IR |
516 | switch (relevant) |
517 | { | |
8644a673 | 518 | case vect_unused_in_scope: |
7c5222ff IR |
519 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
520 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 521 | break; |
7c5222ff | 522 | |
ebfd146a | 523 | case vect_used_in_outer_by_reduction: |
7c5222ff | 524 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
525 | relevant = vect_used_by_reduction; |
526 | break; | |
7c5222ff | 527 | |
ebfd146a | 528 | case vect_used_in_outer: |
7c5222ff | 529 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 530 | relevant = vect_used_in_scope; |
ebfd146a | 531 | break; |
7c5222ff | 532 | |
8644a673 | 533 | case vect_used_in_scope: |
ebfd146a IR |
534 | break; |
535 | ||
536 | default: | |
537 | gcc_unreachable (); | |
b8698a0f | 538 | } |
ebfd146a IR |
539 | } |
540 | ||
541 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
542 | outer-loop-header-bb: | |
543 | ... | |
544 | inner-loop: | |
545 | d = def_stmt | |
06066f92 | 546 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
547 | stmt # use (d) */ |
548 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
549 | { | |
73fbfcad | 550 | if (dump_enabled_p ()) |
78c60e3d | 551 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 552 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 553 | |
ebfd146a IR |
554 | switch (relevant) |
555 | { | |
8644a673 | 556 | case vect_unused_in_scope: |
b8698a0f | 557 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 558 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 559 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
560 | break; |
561 | ||
ebfd146a IR |
562 | case vect_used_by_reduction: |
563 | relevant = vect_used_in_outer_by_reduction; | |
564 | break; | |
565 | ||
8644a673 | 566 | case vect_used_in_scope: |
ebfd146a IR |
567 | relevant = vect_used_in_outer; |
568 | break; | |
569 | ||
570 | default: | |
571 | gcc_unreachable (); | |
572 | } | |
573 | } | |
574 | ||
83197f37 IR |
575 | vect_mark_relevant (worklist, def_stmt, relevant, live_p, |
576 | is_pattern_stmt_p (stmt_vinfo)); | |
ebfd146a IR |
577 | return true; |
578 | } | |
579 | ||
580 | ||
581 | /* Function vect_mark_stmts_to_be_vectorized. | |
582 | ||
583 | Not all stmts in the loop need to be vectorized. For example: | |
584 | ||
585 | for i... | |
586 | for j... | |
587 | 1. T0 = i + j | |
588 | 2. T1 = a[T0] | |
589 | ||
590 | 3. j = j + 1 | |
591 | ||
592 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
593 | addressing of vectorized data-refs are handled differently. | |
594 | ||
595 | This pass detects such stmts. */ | |
596 | ||
597 | bool | |
598 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
599 | { | |
ebfd146a IR |
600 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
601 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
602 | unsigned int nbbs = loop->num_nodes; | |
603 | gimple_stmt_iterator si; | |
355fe088 | 604 | gimple *stmt; |
ebfd146a IR |
605 | unsigned int i; |
606 | stmt_vec_info stmt_vinfo; | |
607 | basic_block bb; | |
355fe088 | 608 | gimple *phi; |
ebfd146a | 609 | bool live_p; |
06066f92 IR |
610 | enum vect_relevant relevant, tmp_relevant; |
611 | enum vect_def_type def_type; | |
ebfd146a | 612 | |
73fbfcad | 613 | if (dump_enabled_p ()) |
78c60e3d | 614 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 615 | "=== vect_mark_stmts_to_be_vectorized ===\n"); |
ebfd146a | 616 | |
355fe088 | 617 | auto_vec<gimple *, 64> worklist; |
ebfd146a IR |
618 | |
619 | /* 1. Init worklist. */ | |
620 | for (i = 0; i < nbbs; i++) | |
621 | { | |
622 | bb = bbs[i]; | |
623 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 624 | { |
ebfd146a | 625 | phi = gsi_stmt (si); |
73fbfcad | 626 | if (dump_enabled_p ()) |
ebfd146a | 627 | { |
78c60e3d SS |
628 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
629 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
630 | } |
631 | ||
632 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 633 | vect_mark_relevant (&worklist, phi, relevant, live_p, false); |
ebfd146a IR |
634 | } |
635 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
636 | { | |
637 | stmt = gsi_stmt (si); | |
73fbfcad | 638 | if (dump_enabled_p ()) |
ebfd146a | 639 | { |
78c60e3d SS |
640 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
641 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
b8698a0f | 642 | } |
ebfd146a IR |
643 | |
644 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 645 | vect_mark_relevant (&worklist, stmt, relevant, live_p, false); |
ebfd146a IR |
646 | } |
647 | } | |
648 | ||
649 | /* 2. Process_worklist */ | |
9771b263 | 650 | while (worklist.length () > 0) |
ebfd146a IR |
651 | { |
652 | use_operand_p use_p; | |
653 | ssa_op_iter iter; | |
654 | ||
9771b263 | 655 | stmt = worklist.pop (); |
73fbfcad | 656 | if (dump_enabled_p ()) |
ebfd146a | 657 | { |
78c60e3d SS |
658 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
659 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
ebfd146a IR |
660 | } |
661 | ||
b8698a0f L |
662 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
663 | (DEF_STMT) as relevant/irrelevant and live/dead according to the | |
ebfd146a IR |
664 | liveness and relevance properties of STMT. */ |
665 | stmt_vinfo = vinfo_for_stmt (stmt); | |
666 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
667 | live_p = STMT_VINFO_LIVE_P (stmt_vinfo); | |
668 | ||
669 | /* Generally, the liveness and relevance properties of STMT are | |
670 | propagated as is to the DEF_STMTs of its USEs: | |
671 | live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO) | |
672 | relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO) | |
673 | ||
674 | One exception is when STMT has been identified as defining a reduction | |
675 | variable; in this case we set the liveness/relevance as follows: | |
676 | live_p = false | |
677 | relevant = vect_used_by_reduction | |
678 | This is because we distinguish between two kinds of relevant stmts - | |
b8698a0f | 679 | those that are used by a reduction computation, and those that are |
ff802fa1 | 680 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 681 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 682 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 683 | |
06066f92 IR |
684 | def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
685 | tmp_relevant = relevant; | |
686 | switch (def_type) | |
ebfd146a | 687 | { |
06066f92 IR |
688 | case vect_reduction_def: |
689 | switch (tmp_relevant) | |
690 | { | |
691 | case vect_unused_in_scope: | |
692 | relevant = vect_used_by_reduction; | |
693 | break; | |
694 | ||
695 | case vect_used_by_reduction: | |
696 | if (gimple_code (stmt) == GIMPLE_PHI) | |
697 | break; | |
698 | /* fall through */ | |
699 | ||
700 | default: | |
73fbfcad | 701 | if (dump_enabled_p ()) |
78c60e3d | 702 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 703 | "unsupported use of reduction.\n"); |
06066f92 IR |
704 | return false; |
705 | } | |
706 | ||
b8698a0f | 707 | live_p = false; |
06066f92 | 708 | break; |
b8698a0f | 709 | |
06066f92 IR |
710 | case vect_nested_cycle: |
711 | if (tmp_relevant != vect_unused_in_scope | |
712 | && tmp_relevant != vect_used_in_outer_by_reduction | |
713 | && tmp_relevant != vect_used_in_outer) | |
714 | { | |
73fbfcad | 715 | if (dump_enabled_p ()) |
78c60e3d | 716 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 717 | "unsupported use of nested cycle.\n"); |
7c5222ff | 718 | |
06066f92 IR |
719 | return false; |
720 | } | |
7c5222ff | 721 | |
b8698a0f L |
722 | live_p = false; |
723 | break; | |
724 | ||
06066f92 IR |
725 | case vect_double_reduction_def: |
726 | if (tmp_relevant != vect_unused_in_scope | |
727 | && tmp_relevant != vect_used_by_reduction) | |
728 | { | |
73fbfcad | 729 | if (dump_enabled_p ()) |
78c60e3d | 730 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 731 | "unsupported use of double reduction.\n"); |
7c5222ff | 732 | |
7c5222ff | 733 | return false; |
06066f92 IR |
734 | } |
735 | ||
736 | live_p = false; | |
b8698a0f | 737 | break; |
7c5222ff | 738 | |
06066f92 IR |
739 | default: |
740 | break; | |
7c5222ff | 741 | } |
b8698a0f | 742 | |
aec7ae7d | 743 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
744 | { |
745 | /* Pattern statements are not inserted into the code, so | |
746 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
747 | have to scan the RHS or function arguments instead. */ | |
748 | if (is_gimple_assign (stmt)) | |
749 | { | |
69d2aade JJ |
750 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
751 | tree op = gimple_assign_rhs1 (stmt); | |
752 | ||
753 | i = 1; | |
754 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
755 | { | |
756 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
aec7ae7d | 757 | live_p, relevant, &worklist, false) |
69d2aade | 758 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
aec7ae7d | 759 | live_p, relevant, &worklist, false)) |
566d377a | 760 | return false; |
69d2aade JJ |
761 | i = 2; |
762 | } | |
763 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 764 | { |
69d2aade | 765 | op = gimple_op (stmt, i); |
afbe6325 RB |
766 | if (TREE_CODE (op) == SSA_NAME |
767 | && !process_use (stmt, op, loop_vinfo, live_p, relevant, | |
768 | &worklist, false)) | |
07687835 | 769 | return false; |
9d5e7640 IR |
770 | } |
771 | } | |
772 | else if (is_gimple_call (stmt)) | |
773 | { | |
774 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
775 | { | |
776 | tree arg = gimple_call_arg (stmt, i); | |
777 | if (!process_use (stmt, arg, loop_vinfo, live_p, relevant, | |
aec7ae7d | 778 | &worklist, false)) |
07687835 | 779 | return false; |
9d5e7640 IR |
780 | } |
781 | } | |
782 | } | |
783 | else | |
784 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
785 | { | |
786 | tree op = USE_FROM_PTR (use_p); | |
787 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, | |
aec7ae7d | 788 | &worklist, false)) |
07687835 | 789 | return false; |
9d5e7640 | 790 | } |
aec7ae7d | 791 | |
3bab6342 | 792 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo)) |
aec7ae7d JJ |
793 | { |
794 | tree off; | |
3bab6342 | 795 | tree decl = vect_check_gather_scatter (stmt, loop_vinfo, NULL, &off, NULL); |
aec7ae7d JJ |
796 | gcc_assert (decl); |
797 | if (!process_use (stmt, off, loop_vinfo, live_p, relevant, | |
798 | &worklist, true)) | |
566d377a | 799 | return false; |
aec7ae7d | 800 | } |
ebfd146a IR |
801 | } /* while worklist */ |
802 | ||
ebfd146a IR |
803 | return true; |
804 | } | |
805 | ||
806 | ||
b8698a0f | 807 | /* Function vect_model_simple_cost. |
ebfd146a | 808 | |
b8698a0f | 809 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
810 | single op. Right now, this does not account for multiple insns that could |
811 | be generated for the single vector op. We will handle that shortly. */ | |
812 | ||
813 | void | |
b8698a0f | 814 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 BS |
815 | enum vect_def_type *dt, |
816 | stmt_vector_for_cost *prologue_cost_vec, | |
817 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
818 | { |
819 | int i; | |
92345349 | 820 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
821 | |
822 | /* The SLP costs were already calculated during SLP tree build. */ | |
823 | if (PURE_SLP_STMT (stmt_info)) | |
824 | return; | |
825 | ||
ebfd146a IR |
826 | /* FORNOW: Assuming maximum 2 args per stmts. */ |
827 | for (i = 0; i < 2; i++) | |
92345349 BS |
828 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
829 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, vector_stmt, | |
830 | stmt_info, 0, vect_prologue); | |
c3e7ee41 BS |
831 | |
832 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
92345349 BS |
833 | inside_cost = record_stmt_cost (body_cost_vec, ncopies, vector_stmt, |
834 | stmt_info, 0, vect_body); | |
c3e7ee41 | 835 | |
73fbfcad | 836 | if (dump_enabled_p ()) |
78c60e3d SS |
837 | dump_printf_loc (MSG_NOTE, vect_location, |
838 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 839 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
840 | } |
841 | ||
842 | ||
8bd37302 BS |
843 | /* Model cost for type demotion and promotion operations. PWR is normally |
844 | zero for single-step promotions and demotions. It will be one if | |
845 | two-step promotion/demotion is required, and so on. Each additional | |
846 | step doubles the number of instructions required. */ | |
847 | ||
848 | static void | |
849 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
850 | enum vect_def_type *dt, int pwr) | |
851 | { | |
852 | int i, tmp; | |
92345349 | 853 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 BS |
854 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
855 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
856 | void *target_cost_data; | |
8bd37302 BS |
857 | |
858 | /* The SLP costs were already calculated during SLP tree build. */ | |
859 | if (PURE_SLP_STMT (stmt_info)) | |
860 | return; | |
861 | ||
c3e7ee41 BS |
862 | if (loop_vinfo) |
863 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
864 | else | |
865 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
866 | ||
8bd37302 BS |
867 | for (i = 0; i < pwr + 1; i++) |
868 | { | |
869 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
870 | (i + 1) : i; | |
c3e7ee41 | 871 | inside_cost += add_stmt_cost (target_cost_data, vect_pow2 (tmp), |
92345349 BS |
872 | vec_promote_demote, stmt_info, 0, |
873 | vect_body); | |
8bd37302 BS |
874 | } |
875 | ||
876 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
877 | for (i = 0; i < 2; i++) | |
92345349 BS |
878 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
879 | prologue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt, | |
880 | stmt_info, 0, vect_prologue); | |
8bd37302 | 881 | |
73fbfcad | 882 | if (dump_enabled_p ()) |
78c60e3d SS |
883 | dump_printf_loc (MSG_NOTE, vect_location, |
884 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 885 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
886 | } |
887 | ||
0d0293ac | 888 | /* Function vect_cost_group_size |
b8698a0f | 889 | |
0d0293ac | 890 | For grouped load or store, return the group_size only if it is the first |
ebfd146a IR |
891 | load or store of a group, else return 1. This ensures that group size is |
892 | only returned once per group. */ | |
893 | ||
894 | static int | |
0d0293ac | 895 | vect_cost_group_size (stmt_vec_info stmt_info) |
ebfd146a | 896 | { |
355fe088 | 897 | gimple *first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a IR |
898 | |
899 | if (first_stmt == STMT_VINFO_STMT (stmt_info)) | |
e14c1050 | 900 | return GROUP_SIZE (stmt_info); |
ebfd146a IR |
901 | |
902 | return 1; | |
903 | } | |
904 | ||
905 | ||
906 | /* Function vect_model_store_cost | |
907 | ||
0d0293ac MM |
908 | Models cost for stores. In the case of grouped accesses, one access |
909 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a IR |
910 | |
911 | void | |
b8698a0f | 912 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
272c6793 | 913 | bool store_lanes_p, enum vect_def_type dt, |
92345349 BS |
914 | slp_tree slp_node, |
915 | stmt_vector_for_cost *prologue_cost_vec, | |
916 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
917 | { |
918 | int group_size; | |
92345349 | 919 | unsigned int inside_cost = 0, prologue_cost = 0; |
720f5239 | 920 | struct data_reference *first_dr; |
355fe088 | 921 | gimple *first_stmt; |
ebfd146a | 922 | |
8644a673 | 923 | if (dt == vect_constant_def || dt == vect_external_def) |
92345349 BS |
924 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
925 | stmt_info, 0, vect_prologue); | |
ebfd146a | 926 | |
0d0293ac MM |
927 | /* Grouped access? */ |
928 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
720f5239 IR |
929 | { |
930 | if (slp_node) | |
931 | { | |
9771b263 | 932 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
720f5239 IR |
933 | group_size = 1; |
934 | } | |
935 | else | |
936 | { | |
e14c1050 | 937 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 938 | group_size = vect_cost_group_size (stmt_info); |
720f5239 IR |
939 | } |
940 | ||
941 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
942 | } | |
0d0293ac | 943 | /* Not a grouped access. */ |
ebfd146a | 944 | else |
720f5239 IR |
945 | { |
946 | group_size = 1; | |
947 | first_dr = STMT_VINFO_DATA_REF (stmt_info); | |
948 | } | |
ebfd146a | 949 | |
272c6793 | 950 | /* We assume that the cost of a single store-lanes instruction is |
0d0293ac | 951 | equivalent to the cost of GROUP_SIZE separate stores. If a grouped |
272c6793 RS |
952 | access is instead being provided by a permute-and-store operation, |
953 | include the cost of the permutes. */ | |
cee62fee MM |
954 | if (!store_lanes_p && group_size > 1 |
955 | && !STMT_VINFO_STRIDED_P (stmt_info)) | |
ebfd146a | 956 | { |
e1377713 ES |
957 | /* Uses a high and low interleave or shuffle operations for each |
958 | needed permute. */ | |
959 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; | |
92345349 BS |
960 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
961 | stmt_info, 0, vect_body); | |
ebfd146a | 962 | |
73fbfcad | 963 | if (dump_enabled_p ()) |
78c60e3d | 964 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 965 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 966 | group_size); |
ebfd146a IR |
967 | } |
968 | ||
cee62fee | 969 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a | 970 | /* Costs of the stores. */ |
cee62fee MM |
971 | if (STMT_VINFO_STRIDED_P (stmt_info) |
972 | && !STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
f2e2a985 MM |
973 | { |
974 | /* N scalar stores plus extracting the elements. */ | |
f2e2a985 MM |
975 | inside_cost += record_stmt_cost (body_cost_vec, |
976 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), | |
977 | scalar_store, stmt_info, 0, vect_body); | |
f2e2a985 MM |
978 | } |
979 | else | |
980 | vect_get_store_cost (first_dr, ncopies, &inside_cost, body_cost_vec); | |
ebfd146a | 981 | |
cee62fee MM |
982 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
983 | inside_cost += record_stmt_cost (body_cost_vec, | |
984 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), | |
985 | vec_to_scalar, stmt_info, 0, vect_body); | |
986 | ||
73fbfcad | 987 | if (dump_enabled_p ()) |
78c60e3d SS |
988 | dump_printf_loc (MSG_NOTE, vect_location, |
989 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 990 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
991 | } |
992 | ||
993 | ||
720f5239 IR |
994 | /* Calculate cost of DR's memory access. */ |
995 | void | |
996 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 997 | unsigned int *inside_cost, |
92345349 | 998 | stmt_vector_for_cost *body_cost_vec) |
720f5239 IR |
999 | { |
1000 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 1001 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 1002 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
1003 | |
1004 | switch (alignment_support_scheme) | |
1005 | { | |
1006 | case dr_aligned: | |
1007 | { | |
92345349 BS |
1008 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
1009 | vector_store, stmt_info, 0, | |
1010 | vect_body); | |
720f5239 | 1011 | |
73fbfcad | 1012 | if (dump_enabled_p ()) |
78c60e3d | 1013 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1014 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
1015 | break; |
1016 | } | |
1017 | ||
1018 | case dr_unaligned_supported: | |
1019 | { | |
720f5239 | 1020 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1021 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1022 | unaligned_store, stmt_info, |
92345349 | 1023 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1024 | if (dump_enabled_p ()) |
78c60e3d SS |
1025 | dump_printf_loc (MSG_NOTE, vect_location, |
1026 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1027 | "hardware.\n"); |
720f5239 IR |
1028 | break; |
1029 | } | |
1030 | ||
38eec4c6 UW |
1031 | case dr_unaligned_unsupported: |
1032 | { | |
1033 | *inside_cost = VECT_MAX_COST; | |
1034 | ||
73fbfcad | 1035 | if (dump_enabled_p ()) |
78c60e3d | 1036 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1037 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1038 | break; |
1039 | } | |
1040 | ||
720f5239 IR |
1041 | default: |
1042 | gcc_unreachable (); | |
1043 | } | |
1044 | } | |
1045 | ||
1046 | ||
ebfd146a IR |
1047 | /* Function vect_model_load_cost |
1048 | ||
0d0293ac MM |
1049 | Models cost for loads. In the case of grouped accesses, the last access |
1050 | has the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1051 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1052 | access scheme chosen. */ |
1053 | ||
1054 | void | |
92345349 BS |
1055 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, |
1056 | bool load_lanes_p, slp_tree slp_node, | |
1057 | stmt_vector_for_cost *prologue_cost_vec, | |
1058 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
1059 | { |
1060 | int group_size; | |
355fe088 | 1061 | gimple *first_stmt; |
ebfd146a | 1062 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; |
92345349 | 1063 | unsigned int inside_cost = 0, prologue_cost = 0; |
ebfd146a | 1064 | |
0d0293ac | 1065 | /* Grouped accesses? */ |
e14c1050 | 1066 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 1067 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && first_stmt && !slp_node) |
ebfd146a | 1068 | { |
0d0293ac | 1069 | group_size = vect_cost_group_size (stmt_info); |
ebfd146a IR |
1070 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
1071 | } | |
0d0293ac | 1072 | /* Not a grouped access. */ |
ebfd146a IR |
1073 | else |
1074 | { | |
1075 | group_size = 1; | |
1076 | first_dr = dr; | |
1077 | } | |
1078 | ||
272c6793 | 1079 | /* We assume that the cost of a single load-lanes instruction is |
0d0293ac | 1080 | equivalent to the cost of GROUP_SIZE separate loads. If a grouped |
272c6793 RS |
1081 | access is instead being provided by a load-and-permute operation, |
1082 | include the cost of the permutes. */ | |
7b5fc413 | 1083 | if (!load_lanes_p && group_size > 1 |
f2e2a985 | 1084 | && !STMT_VINFO_STRIDED_P (stmt_info)) |
ebfd146a | 1085 | { |
2c23db6d ES |
1086 | /* Uses an even and odd extract operations or shuffle operations |
1087 | for each needed permute. */ | |
1088 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; | |
1089 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, | |
1090 | stmt_info, 0, vect_body); | |
ebfd146a | 1091 | |
73fbfcad | 1092 | if (dump_enabled_p ()) |
e645e942 TJ |
1093 | dump_printf_loc (MSG_NOTE, vect_location, |
1094 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1095 | group_size); |
ebfd146a IR |
1096 | } |
1097 | ||
1098 | /* The loads themselves. */ | |
f2e2a985 | 1099 | if (STMT_VINFO_STRIDED_P (stmt_info) |
7b5fc413 | 1100 | && !STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
a82960aa | 1101 | { |
a21892ad BS |
1102 | /* N scalar loads plus gathering them into a vector. */ |
1103 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
92345349 | 1104 | inside_cost += record_stmt_cost (body_cost_vec, |
c3e7ee41 | 1105 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), |
92345349 | 1106 | scalar_load, stmt_info, 0, vect_body); |
a82960aa RG |
1107 | } |
1108 | else | |
1109 | vect_get_load_cost (first_dr, ncopies, | |
1110 | ((!STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1111 | || group_size > 1 || slp_node), | |
92345349 BS |
1112 | &inside_cost, &prologue_cost, |
1113 | prologue_cost_vec, body_cost_vec, true); | |
f2e2a985 | 1114 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
7b5fc413 RB |
1115 | inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct, |
1116 | stmt_info, 0, vect_body); | |
720f5239 | 1117 | |
73fbfcad | 1118 | if (dump_enabled_p ()) |
78c60e3d SS |
1119 | dump_printf_loc (MSG_NOTE, vect_location, |
1120 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1121 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1122 | } |
1123 | ||
1124 | ||
1125 | /* Calculate cost of DR's memory access. */ | |
1126 | void | |
1127 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1128 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1129 | unsigned int *prologue_cost, |
1130 | stmt_vector_for_cost *prologue_cost_vec, | |
1131 | stmt_vector_for_cost *body_cost_vec, | |
1132 | bool record_prologue_costs) | |
720f5239 IR |
1133 | { |
1134 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 1135 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 1136 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
1137 | |
1138 | switch (alignment_support_scheme) | |
ebfd146a IR |
1139 | { |
1140 | case dr_aligned: | |
1141 | { | |
92345349 BS |
1142 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1143 | stmt_info, 0, vect_body); | |
ebfd146a | 1144 | |
73fbfcad | 1145 | if (dump_enabled_p ()) |
78c60e3d | 1146 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1147 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1148 | |
1149 | break; | |
1150 | } | |
1151 | case dr_unaligned_supported: | |
1152 | { | |
720f5239 | 1153 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1154 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1155 | unaligned_load, stmt_info, |
92345349 | 1156 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1157 | |
73fbfcad | 1158 | if (dump_enabled_p ()) |
78c60e3d SS |
1159 | dump_printf_loc (MSG_NOTE, vect_location, |
1160 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1161 | "hardware.\n"); |
ebfd146a IR |
1162 | |
1163 | break; | |
1164 | } | |
1165 | case dr_explicit_realign: | |
1166 | { | |
92345349 BS |
1167 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1168 | vector_load, stmt_info, 0, vect_body); | |
1169 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1170 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1171 | |
1172 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1173 | the containing loop, the following cost should be added to the | |
92345349 | 1174 | prologue costs. */ |
ebfd146a | 1175 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1176 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1177 | stmt_info, 0, vect_body); | |
ebfd146a | 1178 | |
73fbfcad | 1179 | if (dump_enabled_p ()) |
e645e942 TJ |
1180 | dump_printf_loc (MSG_NOTE, vect_location, |
1181 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1182 | |
ebfd146a IR |
1183 | break; |
1184 | } | |
1185 | case dr_explicit_realign_optimized: | |
1186 | { | |
73fbfcad | 1187 | if (dump_enabled_p ()) |
e645e942 | 1188 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1189 | "vect_model_load_cost: unaligned software " |
e645e942 | 1190 | "pipelined.\n"); |
ebfd146a IR |
1191 | |
1192 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1193 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1194 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1195 | access, then the above cost should only be considered for one |
ff802fa1 | 1196 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1197 | and a realignment op. */ |
1198 | ||
92345349 | 1199 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1200 | { |
92345349 BS |
1201 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1202 | vector_stmt, stmt_info, | |
1203 | 0, vect_prologue); | |
ebfd146a | 1204 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1205 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1206 | vector_stmt, stmt_info, | |
1207 | 0, vect_prologue); | |
ebfd146a IR |
1208 | } |
1209 | ||
92345349 BS |
1210 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1211 | stmt_info, 0, vect_body); | |
1212 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1213 | stmt_info, 0, vect_body); | |
8bd37302 | 1214 | |
73fbfcad | 1215 | if (dump_enabled_p ()) |
78c60e3d | 1216 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1217 | "vect_model_load_cost: explicit realign optimized" |
1218 | "\n"); | |
8bd37302 | 1219 | |
ebfd146a IR |
1220 | break; |
1221 | } | |
1222 | ||
38eec4c6 UW |
1223 | case dr_unaligned_unsupported: |
1224 | { | |
1225 | *inside_cost = VECT_MAX_COST; | |
1226 | ||
73fbfcad | 1227 | if (dump_enabled_p ()) |
78c60e3d | 1228 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1229 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1230 | break; |
1231 | } | |
1232 | ||
ebfd146a IR |
1233 | default: |
1234 | gcc_unreachable (); | |
1235 | } | |
ebfd146a IR |
1236 | } |
1237 | ||
418b7df3 RG |
1238 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1239 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1240 | |
418b7df3 | 1241 | static void |
355fe088 | 1242 | vect_init_vector_1 (gimple *stmt, gimple *new_stmt, gimple_stmt_iterator *gsi) |
ebfd146a | 1243 | { |
ebfd146a | 1244 | if (gsi) |
418b7df3 | 1245 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1246 | else |
1247 | { | |
418b7df3 | 1248 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1249 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1250 | |
a70d6342 IR |
1251 | if (loop_vinfo) |
1252 | { | |
1253 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1254 | basic_block new_bb; |
1255 | edge pe; | |
a70d6342 IR |
1256 | |
1257 | if (nested_in_vect_loop_p (loop, stmt)) | |
1258 | loop = loop->inner; | |
b8698a0f | 1259 | |
a70d6342 | 1260 | pe = loop_preheader_edge (loop); |
418b7df3 | 1261 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1262 | gcc_assert (!new_bb); |
1263 | } | |
1264 | else | |
1265 | { | |
1266 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1267 | basic_block bb; | |
1268 | gimple_stmt_iterator gsi_bb_start; | |
1269 | ||
1270 | gcc_assert (bb_vinfo); | |
1271 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1272 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1273 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1274 | } |
ebfd146a IR |
1275 | } |
1276 | ||
73fbfcad | 1277 | if (dump_enabled_p ()) |
ebfd146a | 1278 | { |
78c60e3d SS |
1279 | dump_printf_loc (MSG_NOTE, vect_location, |
1280 | "created new init_stmt: "); | |
1281 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
ebfd146a | 1282 | } |
418b7df3 RG |
1283 | } |
1284 | ||
1285 | /* Function vect_init_vector. | |
ebfd146a | 1286 | |
5467ee52 RG |
1287 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1288 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1289 | vector type a vector with all elements equal to VAL is created first. | |
1290 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1291 | initialization at the loop preheader. | |
418b7df3 RG |
1292 | Return the DEF of INIT_STMT. |
1293 | It will be used in the vectorization of STMT. */ | |
1294 | ||
1295 | tree | |
355fe088 | 1296 | vect_init_vector (gimple *stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 | 1297 | { |
355fe088 | 1298 | gimple *init_stmt; |
418b7df3 RG |
1299 | tree new_temp; |
1300 | ||
5467ee52 RG |
1301 | if (TREE_CODE (type) == VECTOR_TYPE |
1302 | && TREE_CODE (TREE_TYPE (val)) != VECTOR_TYPE) | |
418b7df3 | 1303 | { |
5467ee52 | 1304 | if (!types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) |
418b7df3 | 1305 | { |
5a308cf1 IE |
1306 | /* Scalar boolean value should be transformed into |
1307 | all zeros or all ones value before building a vector. */ | |
1308 | if (VECTOR_BOOLEAN_TYPE_P (type)) | |
1309 | { | |
b3d51f23 IE |
1310 | tree true_val = build_all_ones_cst (TREE_TYPE (type)); |
1311 | tree false_val = build_zero_cst (TREE_TYPE (type)); | |
5a308cf1 IE |
1312 | |
1313 | if (CONSTANT_CLASS_P (val)) | |
1314 | val = integer_zerop (val) ? false_val : true_val; | |
1315 | else | |
1316 | { | |
1317 | new_temp = make_ssa_name (TREE_TYPE (type)); | |
1318 | init_stmt = gimple_build_assign (new_temp, COND_EXPR, | |
1319 | val, true_val, false_val); | |
1320 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
1321 | val = new_temp; | |
1322 | } | |
1323 | } | |
1324 | else if (CONSTANT_CLASS_P (val)) | |
42fd8198 | 1325 | val = fold_convert (TREE_TYPE (type), val); |
418b7df3 RG |
1326 | else |
1327 | { | |
b731b390 | 1328 | new_temp = make_ssa_name (TREE_TYPE (type)); |
0d0e4a03 | 1329 | init_stmt = gimple_build_assign (new_temp, NOP_EXPR, val); |
418b7df3 | 1330 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1331 | val = new_temp; |
418b7df3 RG |
1332 | } |
1333 | } | |
5467ee52 | 1334 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1335 | } |
1336 | ||
0e22bb5a RB |
1337 | new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_"); |
1338 | init_stmt = gimple_build_assign (new_temp, val); | |
418b7df3 | 1339 | vect_init_vector_1 (stmt, init_stmt, gsi); |
0e22bb5a | 1340 | return new_temp; |
ebfd146a IR |
1341 | } |
1342 | ||
a70d6342 | 1343 | |
ebfd146a IR |
1344 | /* Function vect_get_vec_def_for_operand. |
1345 | ||
ff802fa1 | 1346 | OP is an operand in STMT. This function returns a (vector) def that will be |
ebfd146a IR |
1347 | used in the vectorized stmt for STMT. |
1348 | ||
1349 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1350 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1351 | ||
1352 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
42fd8198 IE |
1353 | needs to be introduced. VECTYPE may be used to specify a required type for |
1354 | vector invariant. */ | |
ebfd146a IR |
1355 | |
1356 | tree | |
42fd8198 | 1357 | vect_get_vec_def_for_operand (tree op, gimple *stmt, tree vectype) |
ebfd146a IR |
1358 | { |
1359 | tree vec_oprnd; | |
355fe088 TS |
1360 | gimple *vec_stmt; |
1361 | gimple *def_stmt; | |
ebfd146a IR |
1362 | stmt_vec_info def_stmt_info = NULL; |
1363 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
42fd8198 | 1364 | tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo); |
ebfd146a | 1365 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
ebfd146a IR |
1366 | enum vect_def_type dt; |
1367 | bool is_simple_use; | |
1368 | tree vector_type; | |
1369 | ||
73fbfcad | 1370 | if (dump_enabled_p ()) |
ebfd146a | 1371 | { |
78c60e3d SS |
1372 | dump_printf_loc (MSG_NOTE, vect_location, |
1373 | "vect_get_vec_def_for_operand: "); | |
1374 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
e645e942 | 1375 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1376 | } |
1377 | ||
81c40241 | 1378 | is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt); |
ebfd146a | 1379 | gcc_assert (is_simple_use); |
73fbfcad | 1380 | if (dump_enabled_p ()) |
ebfd146a | 1381 | { |
78c60e3d | 1382 | int loc_printed = 0; |
ebfd146a IR |
1383 | if (def_stmt) |
1384 | { | |
78c60e3d SS |
1385 | if (loc_printed) |
1386 | dump_printf (MSG_NOTE, " def_stmt = "); | |
1387 | else | |
1388 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1389 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
ebfd146a IR |
1390 | } |
1391 | } | |
1392 | ||
1393 | switch (dt) | |
1394 | { | |
81c40241 | 1395 | /* operand is a constant or a loop invariant. */ |
ebfd146a | 1396 | case vect_constant_def: |
81c40241 | 1397 | case vect_external_def: |
ebfd146a | 1398 | { |
42fd8198 IE |
1399 | if (vectype) |
1400 | vector_type = vectype; | |
1401 | else if (TREE_CODE (TREE_TYPE (op)) == BOOLEAN_TYPE | |
1402 | && VECTOR_BOOLEAN_TYPE_P (stmt_vectype)) | |
1403 | vector_type = build_same_sized_truth_vector_type (stmt_vectype); | |
1404 | else | |
1405 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
1406 | ||
7569a6cc | 1407 | gcc_assert (vector_type); |
418b7df3 | 1408 | return vect_init_vector (stmt, op, vector_type, NULL); |
ebfd146a IR |
1409 | } |
1410 | ||
81c40241 | 1411 | /* operand is defined inside the loop. */ |
8644a673 | 1412 | case vect_internal_def: |
ebfd146a | 1413 | { |
ebfd146a IR |
1414 | /* Get the def from the vectorized stmt. */ |
1415 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1416 | |
ebfd146a | 1417 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1418 | /* Get vectorized pattern statement. */ |
1419 | if (!vec_stmt | |
1420 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1421 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1422 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1423 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1424 | gcc_assert (vec_stmt); |
1425 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1426 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1427 | else if (is_gimple_call (vec_stmt)) | |
1428 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1429 | else | |
1430 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1431 | return vec_oprnd; | |
1432 | } | |
1433 | ||
81c40241 | 1434 | /* operand is defined by a loop header phi - reduction */ |
ebfd146a | 1435 | case vect_reduction_def: |
06066f92 | 1436 | case vect_double_reduction_def: |
7c5222ff | 1437 | case vect_nested_cycle: |
81c40241 RB |
1438 | /* Code should use get_initial_def_for_reduction. */ |
1439 | gcc_unreachable (); | |
ebfd146a | 1440 | |
81c40241 | 1441 | /* operand is defined by loop-header phi - induction. */ |
ebfd146a IR |
1442 | case vect_induction_def: |
1443 | { | |
1444 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1445 | ||
1446 | /* Get the def from the vectorized stmt. */ | |
1447 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1448 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1449 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1450 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1451 | else | |
1452 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1453 | return vec_oprnd; |
1454 | } | |
1455 | ||
1456 | default: | |
1457 | gcc_unreachable (); | |
1458 | } | |
1459 | } | |
1460 | ||
1461 | ||
1462 | /* Function vect_get_vec_def_for_stmt_copy | |
1463 | ||
ff802fa1 | 1464 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1465 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1466 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1467 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1468 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1469 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1470 | DT is the type of the vector def VEC_OPRND. |
1471 | ||
1472 | Context: | |
1473 | In case the vectorization factor (VF) is bigger than the number | |
1474 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1475 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1476 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1477 | smallest data-type determines the VF, and as a result, when vectorizing |
1478 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1479 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1480 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1481 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1482 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1483 | ||
1484 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1485 | |
ebfd146a IR |
1486 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1487 | VS1.1: vx.1 = memref1 VS1.2 | |
1488 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1489 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1490 | |
1491 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1492 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1493 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1494 | VSnew.3: vz3 = vx.3 + ... | |
1495 | ||
1496 | The vectorization of S1 is explained in vectorizable_load. | |
1497 | The vectorization of S2: | |
b8698a0f L |
1498 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1499 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1500 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1501 | returns the vector-def 'vx.0'. |
1502 | ||
b8698a0f L |
1503 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1504 | function is called to get the relevant vector-def for each operand. It is | |
1505 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1506 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1507 | ||
b8698a0f L |
1508 | For example, to obtain the vector-def 'vx.1' in order to create the |
1509 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1510 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1511 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1512 | and return its def ('vx.1'). | |
1513 | Overall, to create the above sequence this function will be called 3 times: | |
1514 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1515 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1516 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1517 | ||
1518 | tree | |
1519 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1520 | { | |
355fe088 | 1521 | gimple *vec_stmt_for_operand; |
ebfd146a IR |
1522 | stmt_vec_info def_stmt_info; |
1523 | ||
1524 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1525 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1526 | return vec_oprnd; |
1527 | ||
1528 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1529 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1530 | gcc_assert (def_stmt_info); | |
1531 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1532 | gcc_assert (vec_stmt_for_operand); | |
ebfd146a IR |
1533 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) |
1534 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1535 | else | |
1536 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1537 | return vec_oprnd; | |
1538 | } | |
1539 | ||
1540 | ||
1541 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1542 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a IR |
1543 | |
1544 | static void | |
b8698a0f | 1545 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1546 | vec<tree> *vec_oprnds0, |
1547 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1548 | { |
9771b263 | 1549 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1550 | |
1551 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1552 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1553 | |
9771b263 | 1554 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1555 | { |
9771b263 | 1556 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1557 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1558 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1559 | } |
1560 | } | |
1561 | ||
1562 | ||
d092494c IR |
1563 | /* Get vectorized definitions for OP0 and OP1. |
1564 | REDUC_INDEX is the index of reduction operand in case of reduction, | |
1565 | and -1 otherwise. */ | |
ebfd146a | 1566 | |
d092494c | 1567 | void |
355fe088 | 1568 | vect_get_vec_defs (tree op0, tree op1, gimple *stmt, |
9771b263 DN |
1569 | vec<tree> *vec_oprnds0, |
1570 | vec<tree> *vec_oprnds1, | |
d092494c | 1571 | slp_tree slp_node, int reduc_index) |
ebfd146a IR |
1572 | { |
1573 | if (slp_node) | |
d092494c IR |
1574 | { |
1575 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1576 | auto_vec<tree> ops (nops); |
1577 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1578 | |
9771b263 | 1579 | ops.quick_push (op0); |
d092494c | 1580 | if (op1) |
9771b263 | 1581 | ops.quick_push (op1); |
d092494c IR |
1582 | |
1583 | vect_get_slp_defs (ops, slp_node, &vec_defs, reduc_index); | |
1584 | ||
37b5ec8f | 1585 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1586 | if (op1) |
37b5ec8f | 1587 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1588 | } |
ebfd146a IR |
1589 | else |
1590 | { | |
1591 | tree vec_oprnd; | |
1592 | ||
9771b263 | 1593 | vec_oprnds0->create (1); |
81c40241 | 1594 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 1595 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1596 | |
1597 | if (op1) | |
1598 | { | |
9771b263 | 1599 | vec_oprnds1->create (1); |
81c40241 | 1600 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 1601 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1602 | } |
1603 | } | |
1604 | } | |
1605 | ||
1606 | ||
1607 | /* Function vect_finish_stmt_generation. | |
1608 | ||
1609 | Insert a new stmt. */ | |
1610 | ||
1611 | void | |
355fe088 | 1612 | vect_finish_stmt_generation (gimple *stmt, gimple *vec_stmt, |
ebfd146a IR |
1613 | gimple_stmt_iterator *gsi) |
1614 | { | |
1615 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
310213d4 | 1616 | vec_info *vinfo = stmt_info->vinfo; |
ebfd146a IR |
1617 | |
1618 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1619 | ||
54e8e2c3 RG |
1620 | if (!gsi_end_p (*gsi) |
1621 | && gimple_has_mem_ops (vec_stmt)) | |
1622 | { | |
355fe088 | 1623 | gimple *at_stmt = gsi_stmt (*gsi); |
54e8e2c3 RG |
1624 | tree vuse = gimple_vuse (at_stmt); |
1625 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1626 | { | |
1627 | tree vdef = gimple_vdef (at_stmt); | |
1628 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1629 | /* If we have an SSA vuse and insert a store, update virtual | |
1630 | SSA form to avoid triggering the renamer. Do so only | |
1631 | if we can easily see all uses - which is what almost always | |
1632 | happens with the way vectorized stmts are inserted. */ | |
1633 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1634 | && ((is_gimple_assign (vec_stmt) | |
1635 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1636 | || (is_gimple_call (vec_stmt) | |
1637 | && !(gimple_call_flags (vec_stmt) | |
1638 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1639 | { | |
1640 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1641 | gimple_set_vdef (vec_stmt, new_vdef); | |
1642 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1643 | } | |
1644 | } | |
1645 | } | |
ebfd146a IR |
1646 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1647 | ||
310213d4 | 1648 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, vinfo)); |
ebfd146a | 1649 | |
73fbfcad | 1650 | if (dump_enabled_p ()) |
ebfd146a | 1651 | { |
78c60e3d SS |
1652 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); |
1653 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
ebfd146a IR |
1654 | } |
1655 | ||
ad885386 | 1656 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
8e91d222 JJ |
1657 | |
1658 | /* While EH edges will generally prevent vectorization, stmt might | |
1659 | e.g. be in a must-not-throw region. Ensure newly created stmts | |
1660 | that could throw are part of the same region. */ | |
1661 | int lp_nr = lookup_stmt_eh_lp (stmt); | |
1662 | if (lp_nr != 0 && stmt_could_throw_p (vec_stmt)) | |
1663 | add_stmt_to_eh_lp (vec_stmt, lp_nr); | |
ebfd146a IR |
1664 | } |
1665 | ||
70439f0d RS |
1666 | /* We want to vectorize a call to combined function CFN with function |
1667 | decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN | |
1668 | as the types of all inputs. Check whether this is possible using | |
1669 | an internal function, returning its code if so or IFN_LAST if not. */ | |
ebfd146a | 1670 | |
70439f0d RS |
1671 | static internal_fn |
1672 | vectorizable_internal_function (combined_fn cfn, tree fndecl, | |
1673 | tree vectype_out, tree vectype_in) | |
ebfd146a | 1674 | { |
70439f0d RS |
1675 | internal_fn ifn; |
1676 | if (internal_fn_p (cfn)) | |
1677 | ifn = as_internal_fn (cfn); | |
1678 | else | |
1679 | ifn = associated_internal_fn (fndecl); | |
1680 | if (ifn != IFN_LAST && direct_internal_fn_p (ifn)) | |
1681 | { | |
1682 | const direct_internal_fn_info &info = direct_internal_fn (ifn); | |
1683 | if (info.vectorizable) | |
1684 | { | |
1685 | tree type0 = (info.type0 < 0 ? vectype_out : vectype_in); | |
1686 | tree type1 = (info.type1 < 0 ? vectype_out : vectype_in); | |
d95ab70a RS |
1687 | if (direct_internal_fn_supported_p (ifn, tree_pair (type0, type1), |
1688 | OPTIMIZE_FOR_SPEED)) | |
70439f0d RS |
1689 | return ifn; |
1690 | } | |
1691 | } | |
1692 | return IFN_LAST; | |
ebfd146a IR |
1693 | } |
1694 | ||
5ce9450f | 1695 | |
355fe088 | 1696 | static tree permute_vec_elements (tree, tree, tree, gimple *, |
5ce9450f JJ |
1697 | gimple_stmt_iterator *); |
1698 | ||
1699 | ||
1700 | /* Function vectorizable_mask_load_store. | |
1701 | ||
1702 | Check if STMT performs a conditional load or store that can be vectorized. | |
1703 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
1704 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. | |
1705 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1706 | ||
1707 | static bool | |
355fe088 TS |
1708 | vectorizable_mask_load_store (gimple *stmt, gimple_stmt_iterator *gsi, |
1709 | gimple **vec_stmt, slp_tree slp_node) | |
5ce9450f JJ |
1710 | { |
1711 | tree vec_dest = NULL; | |
1712 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1713 | stmt_vec_info prev_stmt_info; | |
1714 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1715 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1716 | bool nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
1717 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
1718 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
57e2f6ad | 1719 | tree rhs_vectype = NULL_TREE; |
045c1278 | 1720 | tree mask_vectype; |
5ce9450f | 1721 | tree elem_type; |
355fe088 | 1722 | gimple *new_stmt; |
5ce9450f JJ |
1723 | tree dummy; |
1724 | tree dataref_ptr = NULL_TREE; | |
355fe088 | 1725 | gimple *ptr_incr; |
5ce9450f JJ |
1726 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
1727 | int ncopies; | |
1728 | int i, j; | |
1729 | bool inv_p; | |
1730 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; | |
1731 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
1732 | int gather_scale = 1; | |
1733 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
1734 | bool is_store; | |
1735 | tree mask; | |
355fe088 | 1736 | gimple *def_stmt; |
5ce9450f JJ |
1737 | enum vect_def_type dt; |
1738 | ||
1739 | if (slp_node != NULL) | |
1740 | return false; | |
1741 | ||
1742 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
1743 | gcc_assert (ncopies >= 1); | |
1744 | ||
1745 | is_store = gimple_call_internal_fn (stmt) == IFN_MASK_STORE; | |
1746 | mask = gimple_call_arg (stmt, 2); | |
045c1278 IE |
1747 | |
1748 | if (TREE_CODE (TREE_TYPE (mask)) != BOOLEAN_TYPE) | |
5ce9450f JJ |
1749 | return false; |
1750 | ||
1751 | /* FORNOW. This restriction should be relaxed. */ | |
1752 | if (nested_in_vect_loop && ncopies > 1) | |
1753 | { | |
1754 | if (dump_enabled_p ()) | |
1755 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1756 | "multiple types in nested loop."); | |
1757 | return false; | |
1758 | } | |
1759 | ||
1760 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) | |
1761 | return false; | |
1762 | ||
66c16fd9 RB |
1763 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
1764 | && ! vec_stmt) | |
5ce9450f JJ |
1765 | return false; |
1766 | ||
1767 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
1768 | return false; | |
1769 | ||
1770 | elem_type = TREE_TYPE (vectype); | |
1771 | ||
1772 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1773 | return false; | |
1774 | ||
f2e2a985 | 1775 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
5ce9450f JJ |
1776 | return false; |
1777 | ||
045c1278 IE |
1778 | if (TREE_CODE (mask) != SSA_NAME) |
1779 | return false; | |
1780 | ||
1781 | if (!vect_is_simple_use (mask, loop_vinfo, &def_stmt, &dt, &mask_vectype)) | |
1782 | return false; | |
1783 | ||
1784 | if (!mask_vectype) | |
1785 | mask_vectype = get_mask_type_for_scalar_type (TREE_TYPE (vectype)); | |
1786 | ||
56e39820 | 1787 | if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype)) |
045c1278 IE |
1788 | return false; |
1789 | ||
57e2f6ad IE |
1790 | if (is_store) |
1791 | { | |
1792 | tree rhs = gimple_call_arg (stmt, 3); | |
1793 | if (!vect_is_simple_use (rhs, loop_vinfo, &def_stmt, &dt, &rhs_vectype)) | |
1794 | return false; | |
1795 | } | |
1796 | ||
3bab6342 | 1797 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5ce9450f | 1798 | { |
355fe088 | 1799 | gimple *def_stmt; |
3bab6342 | 1800 | gather_decl = vect_check_gather_scatter (stmt, loop_vinfo, &gather_base, |
5ce9450f JJ |
1801 | &gather_off, &gather_scale); |
1802 | gcc_assert (gather_decl); | |
81c40241 RB |
1803 | if (!vect_is_simple_use (gather_off, loop_vinfo, &def_stmt, &gather_dt, |
1804 | &gather_off_vectype)) | |
5ce9450f JJ |
1805 | { |
1806 | if (dump_enabled_p ()) | |
1807 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1808 | "gather index use not simple."); | |
1809 | return false; | |
1810 | } | |
03b9e8e4 JJ |
1811 | |
1812 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
1813 | tree masktype | |
1814 | = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (TREE_CHAIN (arglist)))); | |
1815 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
1816 | { | |
1817 | if (dump_enabled_p ()) | |
1818 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1819 | "masked gather with integer mask not supported."); | |
1820 | return false; | |
1821 | } | |
5ce9450f JJ |
1822 | } |
1823 | else if (tree_int_cst_compare (nested_in_vect_loop | |
1824 | ? STMT_VINFO_DR_STEP (stmt_info) | |
1825 | : DR_STEP (dr), size_zero_node) <= 0) | |
1826 | return false; | |
1827 | else if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
045c1278 IE |
1828 | || !can_vec_mask_load_store_p (TYPE_MODE (vectype), |
1829 | TYPE_MODE (mask_vectype), | |
57e2f6ad IE |
1830 | !is_store) |
1831 | || (rhs_vectype | |
1832 | && !useless_type_conversion_p (vectype, rhs_vectype))) | |
5ce9450f JJ |
1833 | return false; |
1834 | ||
5ce9450f JJ |
1835 | if (!vec_stmt) /* transformation not required. */ |
1836 | { | |
1837 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
1838 | if (is_store) | |
1839 | vect_model_store_cost (stmt_info, ncopies, false, dt, | |
1840 | NULL, NULL, NULL); | |
1841 | else | |
1842 | vect_model_load_cost (stmt_info, ncopies, false, NULL, NULL, NULL); | |
1843 | return true; | |
1844 | } | |
1845 | ||
1846 | /** Transform. **/ | |
1847 | ||
3bab6342 | 1848 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5ce9450f JJ |
1849 | { |
1850 | tree vec_oprnd0 = NULL_TREE, op; | |
1851 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
1852 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
acdcd61b | 1853 | tree ptr, vec_mask = NULL_TREE, mask_op = NULL_TREE, var, scale; |
5ce9450f | 1854 | tree perm_mask = NULL_TREE, prev_res = NULL_TREE; |
acdcd61b | 1855 | tree mask_perm_mask = NULL_TREE; |
5ce9450f JJ |
1856 | edge pe = loop_preheader_edge (loop); |
1857 | gimple_seq seq; | |
1858 | basic_block new_bb; | |
1859 | enum { NARROW, NONE, WIDEN } modifier; | |
1860 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
1861 | ||
acdcd61b JJ |
1862 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); |
1863 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1864 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1865 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1866 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1867 | scaletype = TREE_VALUE (arglist); | |
1868 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
1869 | && types_compatible_p (srctype, masktype)); | |
1870 | ||
5ce9450f JJ |
1871 | if (nunits == gather_off_nunits) |
1872 | modifier = NONE; | |
1873 | else if (nunits == gather_off_nunits / 2) | |
1874 | { | |
1875 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
1876 | modifier = WIDEN; | |
1877 | ||
1878 | for (i = 0; i < gather_off_nunits; ++i) | |
1879 | sel[i] = i | nunits; | |
1880 | ||
557be5a8 | 1881 | perm_mask = vect_gen_perm_mask_checked (gather_off_vectype, sel); |
5ce9450f JJ |
1882 | } |
1883 | else if (nunits == gather_off_nunits * 2) | |
1884 | { | |
1885 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
1886 | modifier = NARROW; | |
1887 | ||
1888 | for (i = 0; i < nunits; ++i) | |
1889 | sel[i] = i < gather_off_nunits | |
1890 | ? i : i + nunits - gather_off_nunits; | |
1891 | ||
557be5a8 | 1892 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); |
5ce9450f | 1893 | ncopies *= 2; |
acdcd61b JJ |
1894 | for (i = 0; i < nunits; ++i) |
1895 | sel[i] = i | gather_off_nunits; | |
557be5a8 | 1896 | mask_perm_mask = vect_gen_perm_mask_checked (masktype, sel); |
5ce9450f JJ |
1897 | } |
1898 | else | |
1899 | gcc_unreachable (); | |
1900 | ||
5ce9450f JJ |
1901 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); |
1902 | ||
1903 | ptr = fold_convert (ptrtype, gather_base); | |
1904 | if (!is_gimple_min_invariant (ptr)) | |
1905 | { | |
1906 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
1907 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
1908 | gcc_assert (!new_bb); | |
1909 | } | |
1910 | ||
1911 | scale = build_int_cst (scaletype, gather_scale); | |
1912 | ||
1913 | prev_stmt_info = NULL; | |
1914 | for (j = 0; j < ncopies; ++j) | |
1915 | { | |
1916 | if (modifier == WIDEN && (j & 1)) | |
1917 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
1918 | perm_mask, stmt, gsi); | |
1919 | else if (j == 0) | |
1920 | op = vec_oprnd0 | |
81c40241 | 1921 | = vect_get_vec_def_for_operand (gather_off, stmt); |
5ce9450f JJ |
1922 | else |
1923 | op = vec_oprnd0 | |
1924 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
1925 | ||
1926 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
1927 | { | |
1928 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
1929 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 1930 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
5ce9450f JJ |
1931 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
1932 | new_stmt | |
0d0e4a03 | 1933 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
1934 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1935 | op = var; | |
1936 | } | |
1937 | ||
acdcd61b JJ |
1938 | if (mask_perm_mask && (j & 1)) |
1939 | mask_op = permute_vec_elements (mask_op, mask_op, | |
1940 | mask_perm_mask, stmt, gsi); | |
5ce9450f JJ |
1941 | else |
1942 | { | |
acdcd61b | 1943 | if (j == 0) |
81c40241 | 1944 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); |
acdcd61b JJ |
1945 | else |
1946 | { | |
81c40241 | 1947 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
acdcd61b JJ |
1948 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
1949 | } | |
5ce9450f | 1950 | |
acdcd61b JJ |
1951 | mask_op = vec_mask; |
1952 | if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask))) | |
1953 | { | |
1954 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)) | |
1955 | == TYPE_VECTOR_SUBPARTS (masktype)); | |
0e22bb5a | 1956 | var = vect_get_new_ssa_name (masktype, vect_simple_var); |
acdcd61b JJ |
1957 | mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op); |
1958 | new_stmt | |
0d0e4a03 | 1959 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, mask_op); |
acdcd61b JJ |
1960 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1961 | mask_op = var; | |
1962 | } | |
5ce9450f JJ |
1963 | } |
1964 | ||
1965 | new_stmt | |
1966 | = gimple_build_call (gather_decl, 5, mask_op, ptr, op, mask_op, | |
1967 | scale); | |
1968 | ||
1969 | if (!useless_type_conversion_p (vectype, rettype)) | |
1970 | { | |
1971 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
1972 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 1973 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
5ce9450f JJ |
1974 | gimple_call_set_lhs (new_stmt, op); |
1975 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 1976 | var = make_ssa_name (vec_dest); |
5ce9450f | 1977 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
0d0e4a03 | 1978 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
1979 | } |
1980 | else | |
1981 | { | |
1982 | var = make_ssa_name (vec_dest, new_stmt); | |
1983 | gimple_call_set_lhs (new_stmt, var); | |
1984 | } | |
1985 | ||
1986 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1987 | ||
1988 | if (modifier == NARROW) | |
1989 | { | |
1990 | if ((j & 1) == 0) | |
1991 | { | |
1992 | prev_res = var; | |
1993 | continue; | |
1994 | } | |
1995 | var = permute_vec_elements (prev_res, var, | |
1996 | perm_mask, stmt, gsi); | |
1997 | new_stmt = SSA_NAME_DEF_STMT (var); | |
1998 | } | |
1999 | ||
2000 | if (prev_stmt_info == NULL) | |
2001 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2002 | else | |
2003 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2004 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2005 | } | |
3efe2e2c JJ |
2006 | |
2007 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
2008 | from the IL. */ | |
e6f5c25d IE |
2009 | if (STMT_VINFO_RELATED_STMT (stmt_info)) |
2010 | { | |
2011 | stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
2012 | stmt_info = vinfo_for_stmt (stmt); | |
2013 | } | |
3efe2e2c JJ |
2014 | tree lhs = gimple_call_lhs (stmt); |
2015 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
2016 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
2017 | set_vinfo_for_stmt (stmt, NULL); | |
2018 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2019 | gsi_replace (gsi, new_stmt, true); | |
5ce9450f JJ |
2020 | return true; |
2021 | } | |
2022 | else if (is_store) | |
2023 | { | |
2024 | tree vec_rhs = NULL_TREE, vec_mask = NULL_TREE; | |
2025 | prev_stmt_info = NULL; | |
2d4dc223 | 2026 | LOOP_VINFO_HAS_MASK_STORE (loop_vinfo) = true; |
5ce9450f JJ |
2027 | for (i = 0; i < ncopies; i++) |
2028 | { | |
2029 | unsigned align, misalign; | |
2030 | ||
2031 | if (i == 0) | |
2032 | { | |
2033 | tree rhs = gimple_call_arg (stmt, 3); | |
81c40241 RB |
2034 | vec_rhs = vect_get_vec_def_for_operand (rhs, stmt); |
2035 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); | |
5ce9450f JJ |
2036 | /* We should have catched mismatched types earlier. */ |
2037 | gcc_assert (useless_type_conversion_p (vectype, | |
2038 | TREE_TYPE (vec_rhs))); | |
2039 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, | |
2040 | NULL_TREE, &dummy, gsi, | |
2041 | &ptr_incr, false, &inv_p); | |
2042 | gcc_assert (!inv_p); | |
2043 | } | |
2044 | else | |
2045 | { | |
81c40241 | 2046 | vect_is_simple_use (vec_rhs, loop_vinfo, &def_stmt, &dt); |
5ce9450f | 2047 | vec_rhs = vect_get_vec_def_for_stmt_copy (dt, vec_rhs); |
81c40241 | 2048 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2049 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2050 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2051 | TYPE_SIZE_UNIT (vectype)); | |
2052 | } | |
2053 | ||
2054 | align = TYPE_ALIGN_UNIT (vectype); | |
2055 | if (aligned_access_p (dr)) | |
2056 | misalign = 0; | |
2057 | else if (DR_MISALIGNMENT (dr) == -1) | |
2058 | { | |
2059 | align = TYPE_ALIGN_UNIT (elem_type); | |
2060 | misalign = 0; | |
2061 | } | |
2062 | else | |
2063 | misalign = DR_MISALIGNMENT (dr); | |
2064 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2065 | misalign); | |
08554c26 JJ |
2066 | tree ptr = build_int_cst (TREE_TYPE (gimple_call_arg (stmt, 1)), |
2067 | misalign ? misalign & -misalign : align); | |
5ce9450f JJ |
2068 | new_stmt |
2069 | = gimple_build_call_internal (IFN_MASK_STORE, 4, dataref_ptr, | |
08554c26 | 2070 | ptr, vec_mask, vec_rhs); |
5ce9450f JJ |
2071 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2072 | if (i == 0) | |
2073 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2074 | else | |
2075 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2076 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2077 | } | |
2078 | } | |
2079 | else | |
2080 | { | |
2081 | tree vec_mask = NULL_TREE; | |
2082 | prev_stmt_info = NULL; | |
2083 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); | |
2084 | for (i = 0; i < ncopies; i++) | |
2085 | { | |
2086 | unsigned align, misalign; | |
2087 | ||
2088 | if (i == 0) | |
2089 | { | |
81c40241 | 2090 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); |
5ce9450f JJ |
2091 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, |
2092 | NULL_TREE, &dummy, gsi, | |
2093 | &ptr_incr, false, &inv_p); | |
2094 | gcc_assert (!inv_p); | |
2095 | } | |
2096 | else | |
2097 | { | |
81c40241 | 2098 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2099 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2100 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2101 | TYPE_SIZE_UNIT (vectype)); | |
2102 | } | |
2103 | ||
2104 | align = TYPE_ALIGN_UNIT (vectype); | |
2105 | if (aligned_access_p (dr)) | |
2106 | misalign = 0; | |
2107 | else if (DR_MISALIGNMENT (dr) == -1) | |
2108 | { | |
2109 | align = TYPE_ALIGN_UNIT (elem_type); | |
2110 | misalign = 0; | |
2111 | } | |
2112 | else | |
2113 | misalign = DR_MISALIGNMENT (dr); | |
2114 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2115 | misalign); | |
08554c26 JJ |
2116 | tree ptr = build_int_cst (TREE_TYPE (gimple_call_arg (stmt, 1)), |
2117 | misalign ? misalign & -misalign : align); | |
5ce9450f JJ |
2118 | new_stmt |
2119 | = gimple_build_call_internal (IFN_MASK_LOAD, 3, dataref_ptr, | |
08554c26 | 2120 | ptr, vec_mask); |
b731b390 | 2121 | gimple_call_set_lhs (new_stmt, make_ssa_name (vec_dest)); |
5ce9450f JJ |
2122 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2123 | if (i == 0) | |
2124 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2125 | else | |
2126 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2127 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2128 | } | |
2129 | } | |
2130 | ||
3efe2e2c JJ |
2131 | if (!is_store) |
2132 | { | |
2133 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
2134 | from the IL. */ | |
e6f5c25d IE |
2135 | if (STMT_VINFO_RELATED_STMT (stmt_info)) |
2136 | { | |
2137 | stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
2138 | stmt_info = vinfo_for_stmt (stmt); | |
2139 | } | |
3efe2e2c JJ |
2140 | tree lhs = gimple_call_lhs (stmt); |
2141 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
2142 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
2143 | set_vinfo_for_stmt (stmt, NULL); | |
2144 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2145 | gsi_replace (gsi, new_stmt, true); | |
2146 | } | |
2147 | ||
5ce9450f JJ |
2148 | return true; |
2149 | } | |
2150 | ||
b1b6836e RS |
2151 | /* Return true if vector types VECTYPE_IN and VECTYPE_OUT have |
2152 | integer elements and if we can narrow VECTYPE_IN to VECTYPE_OUT | |
2153 | in a single step. On success, store the binary pack code in | |
2154 | *CONVERT_CODE. */ | |
2155 | ||
2156 | static bool | |
2157 | simple_integer_narrowing (tree vectype_out, tree vectype_in, | |
2158 | tree_code *convert_code) | |
2159 | { | |
2160 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_out)) | |
2161 | || !INTEGRAL_TYPE_P (TREE_TYPE (vectype_in))) | |
2162 | return false; | |
2163 | ||
2164 | tree_code code; | |
2165 | int multi_step_cvt = 0; | |
2166 | auto_vec <tree, 8> interm_types; | |
2167 | if (!supportable_narrowing_operation (NOP_EXPR, vectype_out, vectype_in, | |
2168 | &code, &multi_step_cvt, | |
2169 | &interm_types) | |
2170 | || multi_step_cvt) | |
2171 | return false; | |
2172 | ||
2173 | *convert_code = code; | |
2174 | return true; | |
2175 | } | |
5ce9450f | 2176 | |
ebfd146a IR |
2177 | /* Function vectorizable_call. |
2178 | ||
538dd0b7 | 2179 | Check if GS performs a function call that can be vectorized. |
b8698a0f | 2180 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
2181 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2182 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2183 | ||
2184 | static bool | |
355fe088 | 2185 | vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
190c2236 | 2186 | slp_tree slp_node) |
ebfd146a | 2187 | { |
538dd0b7 | 2188 | gcall *stmt; |
ebfd146a IR |
2189 | tree vec_dest; |
2190 | tree scalar_dest; | |
2191 | tree op, type; | |
2192 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
538dd0b7 | 2193 | stmt_vec_info stmt_info = vinfo_for_stmt (gs), prev_stmt_info; |
ebfd146a IR |
2194 | tree vectype_out, vectype_in; |
2195 | int nunits_in; | |
2196 | int nunits_out; | |
2197 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
190c2236 | 2198 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 2199 | vec_info *vinfo = stmt_info->vinfo; |
81c40241 | 2200 | tree fndecl, new_temp, rhs_type; |
355fe088 | 2201 | gimple *def_stmt; |
0502fb85 UB |
2202 | enum vect_def_type dt[3] |
2203 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
355fe088 | 2204 | gimple *new_stmt = NULL; |
ebfd146a | 2205 | int ncopies, j; |
6e1aa848 | 2206 | vec<tree> vargs = vNULL; |
ebfd146a IR |
2207 | enum { NARROW, NONE, WIDEN } modifier; |
2208 | size_t i, nargs; | |
9d5e7640 | 2209 | tree lhs; |
ebfd146a | 2210 | |
190c2236 | 2211 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2212 | return false; |
2213 | ||
66c16fd9 RB |
2214 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
2215 | && ! vec_stmt) | |
ebfd146a IR |
2216 | return false; |
2217 | ||
538dd0b7 DM |
2218 | /* Is GS a vectorizable call? */ |
2219 | stmt = dyn_cast <gcall *> (gs); | |
2220 | if (!stmt) | |
ebfd146a IR |
2221 | return false; |
2222 | ||
5ce9450f JJ |
2223 | if (gimple_call_internal_p (stmt) |
2224 | && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD | |
2225 | || gimple_call_internal_fn (stmt) == IFN_MASK_STORE)) | |
2226 | return vectorizable_mask_load_store (stmt, gsi, vec_stmt, | |
2227 | slp_node); | |
2228 | ||
0136f8f0 AH |
2229 | if (gimple_call_lhs (stmt) == NULL_TREE |
2230 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
ebfd146a IR |
2231 | return false; |
2232 | ||
0136f8f0 | 2233 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); |
5a2c1986 | 2234 | |
b690cc0f RG |
2235 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
2236 | ||
ebfd146a IR |
2237 | /* Process function arguments. */ |
2238 | rhs_type = NULL_TREE; | |
b690cc0f | 2239 | vectype_in = NULL_TREE; |
ebfd146a IR |
2240 | nargs = gimple_call_num_args (stmt); |
2241 | ||
1b1562a5 MM |
2242 | /* Bail out if the function has more than three arguments, we do not have |
2243 | interesting builtin functions to vectorize with more than two arguments | |
2244 | except for fma. No arguments is also not good. */ | |
2245 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
2246 | return false; |
2247 | ||
74bf76ed JJ |
2248 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
2249 | if (gimple_call_internal_p (stmt) | |
2250 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2251 | { | |
2252 | nargs = 0; | |
2253 | rhs_type = unsigned_type_node; | |
2254 | } | |
2255 | ||
ebfd146a IR |
2256 | for (i = 0; i < nargs; i++) |
2257 | { | |
b690cc0f RG |
2258 | tree opvectype; |
2259 | ||
ebfd146a IR |
2260 | op = gimple_call_arg (stmt, i); |
2261 | ||
2262 | /* We can only handle calls with arguments of the same type. */ | |
2263 | if (rhs_type | |
8533c9d8 | 2264 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 2265 | { |
73fbfcad | 2266 | if (dump_enabled_p ()) |
78c60e3d | 2267 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2268 | "argument types differ.\n"); |
ebfd146a IR |
2269 | return false; |
2270 | } | |
b690cc0f RG |
2271 | if (!rhs_type) |
2272 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 2273 | |
81c40241 | 2274 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[i], &opvectype)) |
ebfd146a | 2275 | { |
73fbfcad | 2276 | if (dump_enabled_p ()) |
78c60e3d | 2277 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2278 | "use not simple.\n"); |
ebfd146a IR |
2279 | return false; |
2280 | } | |
ebfd146a | 2281 | |
b690cc0f RG |
2282 | if (!vectype_in) |
2283 | vectype_in = opvectype; | |
2284 | else if (opvectype | |
2285 | && opvectype != vectype_in) | |
2286 | { | |
73fbfcad | 2287 | if (dump_enabled_p ()) |
78c60e3d | 2288 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2289 | "argument vector types differ.\n"); |
b690cc0f RG |
2290 | return false; |
2291 | } | |
2292 | } | |
2293 | /* If all arguments are external or constant defs use a vector type with | |
2294 | the same size as the output vector type. */ | |
ebfd146a | 2295 | if (!vectype_in) |
b690cc0f | 2296 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
2297 | if (vec_stmt) |
2298 | gcc_assert (vectype_in); | |
2299 | if (!vectype_in) | |
2300 | { | |
73fbfcad | 2301 | if (dump_enabled_p ()) |
7d8930a0 | 2302 | { |
78c60e3d SS |
2303 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2304 | "no vectype for scalar type "); | |
2305 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 2306 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
2307 | } |
2308 | ||
2309 | return false; | |
2310 | } | |
ebfd146a IR |
2311 | |
2312 | /* FORNOW */ | |
b690cc0f RG |
2313 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
2314 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
2315 | if (nunits_in == nunits_out / 2) |
2316 | modifier = NARROW; | |
2317 | else if (nunits_out == nunits_in) | |
2318 | modifier = NONE; | |
2319 | else if (nunits_out == nunits_in / 2) | |
2320 | modifier = WIDEN; | |
2321 | else | |
2322 | return false; | |
2323 | ||
70439f0d RS |
2324 | /* We only handle functions that do not read or clobber memory. */ |
2325 | if (gimple_vuse (stmt)) | |
2326 | { | |
2327 | if (dump_enabled_p ()) | |
2328 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2329 | "function reads from or writes to memory.\n"); | |
2330 | return false; | |
2331 | } | |
2332 | ||
ebfd146a IR |
2333 | /* For now, we only vectorize functions if a target specific builtin |
2334 | is available. TODO -- in some cases, it might be profitable to | |
2335 | insert the calls for pieces of the vector, in order to be able | |
2336 | to vectorize other operations in the loop. */ | |
70439f0d RS |
2337 | fndecl = NULL_TREE; |
2338 | internal_fn ifn = IFN_LAST; | |
2339 | combined_fn cfn = gimple_call_combined_fn (stmt); | |
2340 | tree callee = gimple_call_fndecl (stmt); | |
2341 | ||
2342 | /* First try using an internal function. */ | |
b1b6836e RS |
2343 | tree_code convert_code = ERROR_MARK; |
2344 | if (cfn != CFN_LAST | |
2345 | && (modifier == NONE | |
2346 | || (modifier == NARROW | |
2347 | && simple_integer_narrowing (vectype_out, vectype_in, | |
2348 | &convert_code)))) | |
70439f0d RS |
2349 | ifn = vectorizable_internal_function (cfn, callee, vectype_out, |
2350 | vectype_in); | |
2351 | ||
2352 | /* If that fails, try asking for a target-specific built-in function. */ | |
2353 | if (ifn == IFN_LAST) | |
2354 | { | |
2355 | if (cfn != CFN_LAST) | |
2356 | fndecl = targetm.vectorize.builtin_vectorized_function | |
2357 | (cfn, vectype_out, vectype_in); | |
2358 | else | |
2359 | fndecl = targetm.vectorize.builtin_md_vectorized_function | |
2360 | (callee, vectype_out, vectype_in); | |
2361 | } | |
2362 | ||
2363 | if (ifn == IFN_LAST && !fndecl) | |
ebfd146a | 2364 | { |
70439f0d | 2365 | if (cfn == CFN_GOMP_SIMD_LANE |
74bf76ed JJ |
2366 | && !slp_node |
2367 | && loop_vinfo | |
2368 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2369 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
2370 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2371 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
2372 | { | |
2373 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
2374 | { 0, 1, 2, ... vf - 1 } vector. */ | |
2375 | gcc_assert (nargs == 0); | |
2376 | } | |
2377 | else | |
2378 | { | |
2379 | if (dump_enabled_p ()) | |
2380 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 2381 | "function is not vectorizable.\n"); |
74bf76ed JJ |
2382 | return false; |
2383 | } | |
ebfd146a IR |
2384 | } |
2385 | ||
190c2236 JJ |
2386 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
2387 | ncopies = 1; | |
b1b6836e | 2388 | else if (modifier == NARROW && ifn == IFN_LAST) |
ebfd146a IR |
2389 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; |
2390 | else | |
2391 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2392 | ||
2393 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
2394 | needs to be generated. */ | |
2395 | gcc_assert (ncopies >= 1); | |
2396 | ||
2397 | if (!vec_stmt) /* transformation not required. */ | |
2398 | { | |
2399 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 2400 | if (dump_enabled_p ()) |
e645e942 TJ |
2401 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ===" |
2402 | "\n"); | |
c3e7ee41 | 2403 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
b1b6836e RS |
2404 | if (ifn != IFN_LAST && modifier == NARROW && !slp_node) |
2405 | add_stmt_cost (stmt_info->vinfo->target_cost_data, ncopies / 2, | |
2406 | vec_promote_demote, stmt_info, 0, vect_body); | |
2407 | ||
ebfd146a IR |
2408 | return true; |
2409 | } | |
2410 | ||
2411 | /** Transform. **/ | |
2412 | ||
73fbfcad | 2413 | if (dump_enabled_p ()) |
e645e942 | 2414 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
2415 | |
2416 | /* Handle def. */ | |
2417 | scalar_dest = gimple_call_lhs (stmt); | |
2418 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
2419 | ||
2420 | prev_stmt_info = NULL; | |
b1b6836e | 2421 | if (modifier == NONE || ifn != IFN_LAST) |
ebfd146a | 2422 | { |
b1b6836e | 2423 | tree prev_res = NULL_TREE; |
ebfd146a IR |
2424 | for (j = 0; j < ncopies; ++j) |
2425 | { | |
2426 | /* Build argument list for the vectorized call. */ | |
2427 | if (j == 0) | |
9771b263 | 2428 | vargs.create (nargs); |
ebfd146a | 2429 | else |
9771b263 | 2430 | vargs.truncate (0); |
ebfd146a | 2431 | |
190c2236 JJ |
2432 | if (slp_node) |
2433 | { | |
ef062b13 | 2434 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2435 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2436 | |
2437 | for (i = 0; i < nargs; i++) | |
9771b263 | 2438 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2439 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2440 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2441 | |
2442 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2443 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
2444 | { |
2445 | size_t k; | |
2446 | for (k = 0; k < nargs; k++) | |
2447 | { | |
37b5ec8f | 2448 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 2449 | vargs[k] = vec_oprndsk[i]; |
190c2236 | 2450 | } |
b1b6836e RS |
2451 | if (modifier == NARROW) |
2452 | { | |
2453 | tree half_res = make_ssa_name (vectype_in); | |
2454 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2455 | gimple_call_set_lhs (new_stmt, half_res); | |
2456 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2457 | if ((i & 1) == 0) | |
2458 | { | |
2459 | prev_res = half_res; | |
2460 | continue; | |
2461 | } | |
2462 | new_temp = make_ssa_name (vec_dest); | |
2463 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
2464 | prev_res, half_res); | |
2465 | } | |
70439f0d | 2466 | else |
b1b6836e RS |
2467 | { |
2468 | if (ifn != IFN_LAST) | |
2469 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2470 | else | |
2471 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2472 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2473 | gimple_call_set_lhs (new_stmt, new_temp); | |
2474 | } | |
190c2236 | 2475 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 2476 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2477 | } |
2478 | ||
2479 | for (i = 0; i < nargs; i++) | |
2480 | { | |
37b5ec8f | 2481 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2482 | vec_oprndsi.release (); |
190c2236 | 2483 | } |
190c2236 JJ |
2484 | continue; |
2485 | } | |
2486 | ||
ebfd146a IR |
2487 | for (i = 0; i < nargs; i++) |
2488 | { | |
2489 | op = gimple_call_arg (stmt, i); | |
2490 | if (j == 0) | |
2491 | vec_oprnd0 | |
81c40241 | 2492 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 2493 | else |
63827fb8 IR |
2494 | { |
2495 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
2496 | vec_oprnd0 | |
2497 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
2498 | } | |
ebfd146a | 2499 | |
9771b263 | 2500 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
2501 | } |
2502 | ||
74bf76ed JJ |
2503 | if (gimple_call_internal_p (stmt) |
2504 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2505 | { | |
2506 | tree *v = XALLOCAVEC (tree, nunits_out); | |
2507 | int k; | |
2508 | for (k = 0; k < nunits_out; ++k) | |
2509 | v[k] = build_int_cst (unsigned_type_node, j * nunits_out + k); | |
2510 | tree cst = build_vector (vectype_out, v); | |
2511 | tree new_var | |
0e22bb5a | 2512 | = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_"); |
355fe088 | 2513 | gimple *init_stmt = gimple_build_assign (new_var, cst); |
74bf76ed | 2514 | vect_init_vector_1 (stmt, init_stmt, NULL); |
b731b390 | 2515 | new_temp = make_ssa_name (vec_dest); |
0e22bb5a | 2516 | new_stmt = gimple_build_assign (new_temp, new_var); |
74bf76ed | 2517 | } |
b1b6836e RS |
2518 | else if (modifier == NARROW) |
2519 | { | |
2520 | tree half_res = make_ssa_name (vectype_in); | |
2521 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2522 | gimple_call_set_lhs (new_stmt, half_res); | |
2523 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2524 | if ((j & 1) == 0) | |
2525 | { | |
2526 | prev_res = half_res; | |
2527 | continue; | |
2528 | } | |
2529 | new_temp = make_ssa_name (vec_dest); | |
2530 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
2531 | prev_res, half_res); | |
2532 | } | |
74bf76ed JJ |
2533 | else |
2534 | { | |
70439f0d RS |
2535 | if (ifn != IFN_LAST) |
2536 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2537 | else | |
2538 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
74bf76ed JJ |
2539 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2540 | gimple_call_set_lhs (new_stmt, new_temp); | |
2541 | } | |
ebfd146a IR |
2542 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2543 | ||
b1b6836e | 2544 | if (j == (modifier == NARROW ? 1 : 0)) |
ebfd146a IR |
2545 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; |
2546 | else | |
2547 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2548 | ||
2549 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2550 | } | |
b1b6836e RS |
2551 | } |
2552 | else if (modifier == NARROW) | |
2553 | { | |
ebfd146a IR |
2554 | for (j = 0; j < ncopies; ++j) |
2555 | { | |
2556 | /* Build argument list for the vectorized call. */ | |
2557 | if (j == 0) | |
9771b263 | 2558 | vargs.create (nargs * 2); |
ebfd146a | 2559 | else |
9771b263 | 2560 | vargs.truncate (0); |
ebfd146a | 2561 | |
190c2236 JJ |
2562 | if (slp_node) |
2563 | { | |
ef062b13 | 2564 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2565 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2566 | |
2567 | for (i = 0; i < nargs; i++) | |
9771b263 | 2568 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2569 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2570 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2571 | |
2572 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2573 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
2574 | { |
2575 | size_t k; | |
9771b263 | 2576 | vargs.truncate (0); |
190c2236 JJ |
2577 | for (k = 0; k < nargs; k++) |
2578 | { | |
37b5ec8f | 2579 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
2580 | vargs.quick_push (vec_oprndsk[i]); |
2581 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 | 2582 | } |
70439f0d RS |
2583 | if (ifn != IFN_LAST) |
2584 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2585 | else | |
2586 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
190c2236 JJ |
2587 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2588 | gimple_call_set_lhs (new_stmt, new_temp); | |
2589 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 2590 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2591 | } |
2592 | ||
2593 | for (i = 0; i < nargs; i++) | |
2594 | { | |
37b5ec8f | 2595 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2596 | vec_oprndsi.release (); |
190c2236 | 2597 | } |
190c2236 JJ |
2598 | continue; |
2599 | } | |
2600 | ||
ebfd146a IR |
2601 | for (i = 0; i < nargs; i++) |
2602 | { | |
2603 | op = gimple_call_arg (stmt, i); | |
2604 | if (j == 0) | |
2605 | { | |
2606 | vec_oprnd0 | |
81c40241 | 2607 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 2608 | vec_oprnd1 |
63827fb8 | 2609 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2610 | } |
2611 | else | |
2612 | { | |
336ecb65 | 2613 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 2614 | vec_oprnd0 |
63827fb8 | 2615 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 2616 | vec_oprnd1 |
63827fb8 | 2617 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2618 | } |
2619 | ||
9771b263 DN |
2620 | vargs.quick_push (vec_oprnd0); |
2621 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
2622 | } |
2623 | ||
b1b6836e | 2624 | new_stmt = gimple_build_call_vec (fndecl, vargs); |
ebfd146a IR |
2625 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2626 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
2627 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2628 | ||
2629 | if (j == 0) | |
2630 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2631 | else | |
2632 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2633 | ||
2634 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2635 | } | |
2636 | ||
2637 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a | 2638 | } |
b1b6836e RS |
2639 | else |
2640 | /* No current target implements this case. */ | |
2641 | return false; | |
ebfd146a | 2642 | |
9771b263 | 2643 | vargs.release (); |
ebfd146a | 2644 | |
ebfd146a IR |
2645 | /* The call in STMT might prevent it from being removed in dce. |
2646 | We however cannot remove it here, due to the way the ssa name | |
2647 | it defines is mapped to the new definition. So just replace | |
2648 | rhs of the statement with something harmless. */ | |
2649 | ||
dd34c087 JJ |
2650 | if (slp_node) |
2651 | return true; | |
2652 | ||
ebfd146a | 2653 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
2654 | if (is_pattern_stmt_p (stmt_info)) |
2655 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
2656 | else | |
2657 | lhs = gimple_call_lhs (stmt); | |
3cc2fa2a JJ |
2658 | |
2659 | if (gimple_call_internal_p (stmt) | |
2660 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2661 | { | |
2662 | /* Replace uses of the lhs of GOMP_SIMD_LANE call outside the loop | |
2663 | with vf - 1 rather than 0, that is the last iteration of the | |
2664 | vectorized loop. */ | |
2665 | imm_use_iterator iter; | |
2666 | use_operand_p use_p; | |
355fe088 | 2667 | gimple *use_stmt; |
3cc2fa2a JJ |
2668 | FOR_EACH_IMM_USE_STMT (use_stmt, iter, lhs) |
2669 | { | |
2670 | basic_block use_bb = gimple_bb (use_stmt); | |
2671 | if (use_bb | |
2672 | && !flow_bb_inside_loop_p (LOOP_VINFO_LOOP (loop_vinfo), use_bb)) | |
2673 | { | |
2674 | FOR_EACH_IMM_USE_ON_STMT (use_p, iter) | |
2675 | SET_USE (use_p, build_int_cst (TREE_TYPE (lhs), | |
2676 | ncopies * nunits_out - 1)); | |
2677 | update_stmt (use_stmt); | |
2678 | } | |
2679 | } | |
2680 | } | |
2681 | ||
9d5e7640 | 2682 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); |
ebfd146a | 2683 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 2684 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
2685 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
2686 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
2687 | |
2688 | return true; | |
2689 | } | |
2690 | ||
2691 | ||
0136f8f0 AH |
2692 | struct simd_call_arg_info |
2693 | { | |
2694 | tree vectype; | |
2695 | tree op; | |
2696 | enum vect_def_type dt; | |
2697 | HOST_WIDE_INT linear_step; | |
2698 | unsigned int align; | |
17b658af | 2699 | bool simd_lane_linear; |
0136f8f0 AH |
2700 | }; |
2701 | ||
17b658af JJ |
2702 | /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME, |
2703 | is linear within simd lane (but not within whole loop), note it in | |
2704 | *ARGINFO. */ | |
2705 | ||
2706 | static void | |
2707 | vect_simd_lane_linear (tree op, struct loop *loop, | |
2708 | struct simd_call_arg_info *arginfo) | |
2709 | { | |
355fe088 | 2710 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
17b658af JJ |
2711 | |
2712 | if (!is_gimple_assign (def_stmt) | |
2713 | || gimple_assign_rhs_code (def_stmt) != POINTER_PLUS_EXPR | |
2714 | || !is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))) | |
2715 | return; | |
2716 | ||
2717 | tree base = gimple_assign_rhs1 (def_stmt); | |
2718 | HOST_WIDE_INT linear_step = 0; | |
2719 | tree v = gimple_assign_rhs2 (def_stmt); | |
2720 | while (TREE_CODE (v) == SSA_NAME) | |
2721 | { | |
2722 | tree t; | |
2723 | def_stmt = SSA_NAME_DEF_STMT (v); | |
2724 | if (is_gimple_assign (def_stmt)) | |
2725 | switch (gimple_assign_rhs_code (def_stmt)) | |
2726 | { | |
2727 | case PLUS_EXPR: | |
2728 | t = gimple_assign_rhs2 (def_stmt); | |
2729 | if (linear_step || TREE_CODE (t) != INTEGER_CST) | |
2730 | return; | |
2731 | base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t); | |
2732 | v = gimple_assign_rhs1 (def_stmt); | |
2733 | continue; | |
2734 | case MULT_EXPR: | |
2735 | t = gimple_assign_rhs2 (def_stmt); | |
2736 | if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t)) | |
2737 | return; | |
2738 | linear_step = tree_to_shwi (t); | |
2739 | v = gimple_assign_rhs1 (def_stmt); | |
2740 | continue; | |
2741 | CASE_CONVERT: | |
2742 | t = gimple_assign_rhs1 (def_stmt); | |
2743 | if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE | |
2744 | || (TYPE_PRECISION (TREE_TYPE (v)) | |
2745 | < TYPE_PRECISION (TREE_TYPE (t)))) | |
2746 | return; | |
2747 | if (!linear_step) | |
2748 | linear_step = 1; | |
2749 | v = t; | |
2750 | continue; | |
2751 | default: | |
2752 | return; | |
2753 | } | |
2754 | else if (is_gimple_call (def_stmt) | |
2755 | && gimple_call_internal_p (def_stmt) | |
2756 | && gimple_call_internal_fn (def_stmt) == IFN_GOMP_SIMD_LANE | |
2757 | && loop->simduid | |
2758 | && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME | |
2759 | && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0)) | |
2760 | == loop->simduid)) | |
2761 | { | |
2762 | if (!linear_step) | |
2763 | linear_step = 1; | |
2764 | arginfo->linear_step = linear_step; | |
2765 | arginfo->op = base; | |
2766 | arginfo->simd_lane_linear = true; | |
2767 | return; | |
2768 | } | |
2769 | } | |
2770 | } | |
2771 | ||
0136f8f0 AH |
2772 | /* Function vectorizable_simd_clone_call. |
2773 | ||
2774 | Check if STMT performs a function call that can be vectorized | |
2775 | by calling a simd clone of the function. | |
2776 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2777 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2778 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2779 | ||
2780 | static bool | |
355fe088 TS |
2781 | vectorizable_simd_clone_call (gimple *stmt, gimple_stmt_iterator *gsi, |
2782 | gimple **vec_stmt, slp_tree slp_node) | |
0136f8f0 AH |
2783 | { |
2784 | tree vec_dest; | |
2785 | tree scalar_dest; | |
2786 | tree op, type; | |
2787 | tree vec_oprnd0 = NULL_TREE; | |
2788 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
2789 | tree vectype; | |
2790 | unsigned int nunits; | |
2791 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2792 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
310213d4 | 2793 | vec_info *vinfo = stmt_info->vinfo; |
0136f8f0 | 2794 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
81c40241 | 2795 | tree fndecl, new_temp; |
355fe088 TS |
2796 | gimple *def_stmt; |
2797 | gimple *new_stmt = NULL; | |
0136f8f0 AH |
2798 | int ncopies, j; |
2799 | vec<simd_call_arg_info> arginfo = vNULL; | |
2800 | vec<tree> vargs = vNULL; | |
2801 | size_t i, nargs; | |
2802 | tree lhs, rtype, ratype; | |
2803 | vec<constructor_elt, va_gc> *ret_ctor_elts; | |
2804 | ||
2805 | /* Is STMT a vectorizable call? */ | |
2806 | if (!is_gimple_call (stmt)) | |
2807 | return false; | |
2808 | ||
2809 | fndecl = gimple_call_fndecl (stmt); | |
2810 | if (fndecl == NULL_TREE) | |
2811 | return false; | |
2812 | ||
d52f5295 | 2813 | struct cgraph_node *node = cgraph_node::get (fndecl); |
0136f8f0 AH |
2814 | if (node == NULL || node->simd_clones == NULL) |
2815 | return false; | |
2816 | ||
2817 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
2818 | return false; | |
2819 | ||
66c16fd9 RB |
2820 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
2821 | && ! vec_stmt) | |
0136f8f0 AH |
2822 | return false; |
2823 | ||
2824 | if (gimple_call_lhs (stmt) | |
2825 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
2826 | return false; | |
2827 | ||
2828 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); | |
2829 | ||
2830 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2831 | ||
2832 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt)) | |
2833 | return false; | |
2834 | ||
2835 | /* FORNOW */ | |
2836 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
2837 | return false; | |
2838 | ||
2839 | /* Process function arguments. */ | |
2840 | nargs = gimple_call_num_args (stmt); | |
2841 | ||
2842 | /* Bail out if the function has zero arguments. */ | |
2843 | if (nargs == 0) | |
2844 | return false; | |
2845 | ||
2846 | arginfo.create (nargs); | |
2847 | ||
2848 | for (i = 0; i < nargs; i++) | |
2849 | { | |
2850 | simd_call_arg_info thisarginfo; | |
2851 | affine_iv iv; | |
2852 | ||
2853 | thisarginfo.linear_step = 0; | |
2854 | thisarginfo.align = 0; | |
2855 | thisarginfo.op = NULL_TREE; | |
17b658af | 2856 | thisarginfo.simd_lane_linear = false; |
0136f8f0 AH |
2857 | |
2858 | op = gimple_call_arg (stmt, i); | |
81c40241 RB |
2859 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &thisarginfo.dt, |
2860 | &thisarginfo.vectype) | |
0136f8f0 AH |
2861 | || thisarginfo.dt == vect_uninitialized_def) |
2862 | { | |
2863 | if (dump_enabled_p ()) | |
2864 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2865 | "use not simple.\n"); | |
2866 | arginfo.release (); | |
2867 | return false; | |
2868 | } | |
2869 | ||
2870 | if (thisarginfo.dt == vect_constant_def | |
2871 | || thisarginfo.dt == vect_external_def) | |
2872 | gcc_assert (thisarginfo.vectype == NULL_TREE); | |
2873 | else | |
2874 | gcc_assert (thisarginfo.vectype != NULL_TREE); | |
2875 | ||
6c9e85fb JJ |
2876 | /* For linear arguments, the analyze phase should have saved |
2877 | the base and step in STMT_VINFO_SIMD_CLONE_INFO. */ | |
17b658af JJ |
2878 | if (i * 3 + 4 <= STMT_VINFO_SIMD_CLONE_INFO (stmt_info).length () |
2879 | && STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]) | |
6c9e85fb JJ |
2880 | { |
2881 | gcc_assert (vec_stmt); | |
2882 | thisarginfo.linear_step | |
17b658af | 2883 | = tree_to_shwi (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]); |
6c9e85fb | 2884 | thisarginfo.op |
17b658af JJ |
2885 | = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 1]; |
2886 | thisarginfo.simd_lane_linear | |
2887 | = (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 3] | |
2888 | == boolean_true_node); | |
6c9e85fb JJ |
2889 | /* If loop has been peeled for alignment, we need to adjust it. */ |
2890 | tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo); | |
2891 | tree n2 = LOOP_VINFO_NITERS (loop_vinfo); | |
17b658af | 2892 | if (n1 != n2 && !thisarginfo.simd_lane_linear) |
6c9e85fb JJ |
2893 | { |
2894 | tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2); | |
17b658af | 2895 | tree step = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]; |
6c9e85fb JJ |
2896 | tree opt = TREE_TYPE (thisarginfo.op); |
2897 | bias = fold_convert (TREE_TYPE (step), bias); | |
2898 | bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step); | |
2899 | thisarginfo.op | |
2900 | = fold_build2 (POINTER_TYPE_P (opt) | |
2901 | ? POINTER_PLUS_EXPR : PLUS_EXPR, opt, | |
2902 | thisarginfo.op, bias); | |
2903 | } | |
2904 | } | |
2905 | else if (!vec_stmt | |
2906 | && thisarginfo.dt != vect_constant_def | |
2907 | && thisarginfo.dt != vect_external_def | |
2908 | && loop_vinfo | |
2909 | && TREE_CODE (op) == SSA_NAME | |
2910 | && simple_iv (loop, loop_containing_stmt (stmt), op, | |
2911 | &iv, false) | |
2912 | && tree_fits_shwi_p (iv.step)) | |
0136f8f0 AH |
2913 | { |
2914 | thisarginfo.linear_step = tree_to_shwi (iv.step); | |
2915 | thisarginfo.op = iv.base; | |
2916 | } | |
2917 | else if ((thisarginfo.dt == vect_constant_def | |
2918 | || thisarginfo.dt == vect_external_def) | |
2919 | && POINTER_TYPE_P (TREE_TYPE (op))) | |
2920 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; | |
17b658af JJ |
2921 | /* Addresses of array elements indexed by GOMP_SIMD_LANE are |
2922 | linear too. */ | |
2923 | if (POINTER_TYPE_P (TREE_TYPE (op)) | |
2924 | && !thisarginfo.linear_step | |
2925 | && !vec_stmt | |
2926 | && thisarginfo.dt != vect_constant_def | |
2927 | && thisarginfo.dt != vect_external_def | |
2928 | && loop_vinfo | |
2929 | && !slp_node | |
2930 | && TREE_CODE (op) == SSA_NAME) | |
2931 | vect_simd_lane_linear (op, loop, &thisarginfo); | |
0136f8f0 AH |
2932 | |
2933 | arginfo.quick_push (thisarginfo); | |
2934 | } | |
2935 | ||
2936 | unsigned int badness = 0; | |
2937 | struct cgraph_node *bestn = NULL; | |
6c9e85fb JJ |
2938 | if (STMT_VINFO_SIMD_CLONE_INFO (stmt_info).exists ()) |
2939 | bestn = cgraph_node::get (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[0]); | |
0136f8f0 AH |
2940 | else |
2941 | for (struct cgraph_node *n = node->simd_clones; n != NULL; | |
2942 | n = n->simdclone->next_clone) | |
2943 | { | |
2944 | unsigned int this_badness = 0; | |
2945 | if (n->simdclone->simdlen | |
2946 | > (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
2947 | || n->simdclone->nargs != nargs) | |
2948 | continue; | |
2949 | if (n->simdclone->simdlen | |
2950 | < (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
2951 | this_badness += (exact_log2 (LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
2952 | - exact_log2 (n->simdclone->simdlen)) * 1024; | |
2953 | if (n->simdclone->inbranch) | |
2954 | this_badness += 2048; | |
2955 | int target_badness = targetm.simd_clone.usable (n); | |
2956 | if (target_badness < 0) | |
2957 | continue; | |
2958 | this_badness += target_badness * 512; | |
2959 | /* FORNOW: Have to add code to add the mask argument. */ | |
2960 | if (n->simdclone->inbranch) | |
2961 | continue; | |
2962 | for (i = 0; i < nargs; i++) | |
2963 | { | |
2964 | switch (n->simdclone->args[i].arg_type) | |
2965 | { | |
2966 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
2967 | if (!useless_type_conversion_p | |
2968 | (n->simdclone->args[i].orig_type, | |
2969 | TREE_TYPE (gimple_call_arg (stmt, i)))) | |
2970 | i = -1; | |
2971 | else if (arginfo[i].dt == vect_constant_def | |
2972 | || arginfo[i].dt == vect_external_def | |
2973 | || arginfo[i].linear_step) | |
2974 | this_badness += 64; | |
2975 | break; | |
2976 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
2977 | if (arginfo[i].dt != vect_constant_def | |
2978 | && arginfo[i].dt != vect_external_def) | |
2979 | i = -1; | |
2980 | break; | |
2981 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
d9a6bd32 | 2982 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
2983 | if (arginfo[i].dt == vect_constant_def |
2984 | || arginfo[i].dt == vect_external_def | |
2985 | || (arginfo[i].linear_step | |
2986 | != n->simdclone->args[i].linear_step)) | |
2987 | i = -1; | |
2988 | break; | |
2989 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
d9a6bd32 JJ |
2990 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
2991 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
e01d41e5 JJ |
2992 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
2993 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
2994 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
2995 | /* FORNOW */ |
2996 | i = -1; | |
2997 | break; | |
2998 | case SIMD_CLONE_ARG_TYPE_MASK: | |
2999 | gcc_unreachable (); | |
3000 | } | |
3001 | if (i == (size_t) -1) | |
3002 | break; | |
3003 | if (n->simdclone->args[i].alignment > arginfo[i].align) | |
3004 | { | |
3005 | i = -1; | |
3006 | break; | |
3007 | } | |
3008 | if (arginfo[i].align) | |
3009 | this_badness += (exact_log2 (arginfo[i].align) | |
3010 | - exact_log2 (n->simdclone->args[i].alignment)); | |
3011 | } | |
3012 | if (i == (size_t) -1) | |
3013 | continue; | |
3014 | if (bestn == NULL || this_badness < badness) | |
3015 | { | |
3016 | bestn = n; | |
3017 | badness = this_badness; | |
3018 | } | |
3019 | } | |
3020 | ||
3021 | if (bestn == NULL) | |
3022 | { | |
3023 | arginfo.release (); | |
3024 | return false; | |
3025 | } | |
3026 | ||
3027 | for (i = 0; i < nargs; i++) | |
3028 | if ((arginfo[i].dt == vect_constant_def | |
3029 | || arginfo[i].dt == vect_external_def) | |
3030 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
3031 | { | |
3032 | arginfo[i].vectype | |
3033 | = get_vectype_for_scalar_type (TREE_TYPE (gimple_call_arg (stmt, | |
3034 | i))); | |
3035 | if (arginfo[i].vectype == NULL | |
3036 | || (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
3037 | > bestn->simdclone->simdlen)) | |
3038 | { | |
3039 | arginfo.release (); | |
3040 | return false; | |
3041 | } | |
3042 | } | |
3043 | ||
3044 | fndecl = bestn->decl; | |
3045 | nunits = bestn->simdclone->simdlen; | |
3046 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3047 | ||
3048 | /* If the function isn't const, only allow it in simd loops where user | |
3049 | has asserted that at least nunits consecutive iterations can be | |
3050 | performed using SIMD instructions. */ | |
3051 | if ((loop == NULL || (unsigned) loop->safelen < nunits) | |
3052 | && gimple_vuse (stmt)) | |
3053 | { | |
3054 | arginfo.release (); | |
3055 | return false; | |
3056 | } | |
3057 | ||
3058 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
3059 | needs to be generated. */ | |
3060 | gcc_assert (ncopies >= 1); | |
3061 | ||
3062 | if (!vec_stmt) /* transformation not required. */ | |
3063 | { | |
6c9e85fb JJ |
3064 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (bestn->decl); |
3065 | for (i = 0; i < nargs; i++) | |
3066 | if (bestn->simdclone->args[i].arg_type | |
3067 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
3068 | { | |
17b658af | 3069 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_grow_cleared (i * 3 |
6c9e85fb JJ |
3070 | + 1); |
3071 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (arginfo[i].op); | |
3072 | tree lst = POINTER_TYPE_P (TREE_TYPE (arginfo[i].op)) | |
3073 | ? size_type_node : TREE_TYPE (arginfo[i].op); | |
3074 | tree ls = build_int_cst (lst, arginfo[i].linear_step); | |
3075 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (ls); | |
17b658af JJ |
3076 | tree sll = arginfo[i].simd_lane_linear |
3077 | ? boolean_true_node : boolean_false_node; | |
3078 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (sll); | |
6c9e85fb | 3079 | } |
0136f8f0 AH |
3080 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; |
3081 | if (dump_enabled_p ()) | |
3082 | dump_printf_loc (MSG_NOTE, vect_location, | |
3083 | "=== vectorizable_simd_clone_call ===\n"); | |
3084 | /* vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); */ | |
3085 | arginfo.release (); | |
3086 | return true; | |
3087 | } | |
3088 | ||
3089 | /** Transform. **/ | |
3090 | ||
3091 | if (dump_enabled_p ()) | |
3092 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); | |
3093 | ||
3094 | /* Handle def. */ | |
3095 | scalar_dest = gimple_call_lhs (stmt); | |
3096 | vec_dest = NULL_TREE; | |
3097 | rtype = NULL_TREE; | |
3098 | ratype = NULL_TREE; | |
3099 | if (scalar_dest) | |
3100 | { | |
3101 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3102 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); | |
3103 | if (TREE_CODE (rtype) == ARRAY_TYPE) | |
3104 | { | |
3105 | ratype = rtype; | |
3106 | rtype = TREE_TYPE (ratype); | |
3107 | } | |
3108 | } | |
3109 | ||
3110 | prev_stmt_info = NULL; | |
3111 | for (j = 0; j < ncopies; ++j) | |
3112 | { | |
3113 | /* Build argument list for the vectorized call. */ | |
3114 | if (j == 0) | |
3115 | vargs.create (nargs); | |
3116 | else | |
3117 | vargs.truncate (0); | |
3118 | ||
3119 | for (i = 0; i < nargs; i++) | |
3120 | { | |
3121 | unsigned int k, l, m, o; | |
3122 | tree atype; | |
3123 | op = gimple_call_arg (stmt, i); | |
3124 | switch (bestn->simdclone->args[i].arg_type) | |
3125 | { | |
3126 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3127 | atype = bestn->simdclone->args[i].vector_type; | |
3128 | o = nunits / TYPE_VECTOR_SUBPARTS (atype); | |
3129 | for (m = j * o; m < (j + 1) * o; m++) | |
3130 | { | |
3131 | if (TYPE_VECTOR_SUBPARTS (atype) | |
3132 | < TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)) | |
3133 | { | |
3134 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); | |
3135 | k = (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
3136 | / TYPE_VECTOR_SUBPARTS (atype)); | |
3137 | gcc_assert ((k & (k - 1)) == 0); | |
3138 | if (m == 0) | |
3139 | vec_oprnd0 | |
81c40241 | 3140 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3141 | else |
3142 | { | |
3143 | vec_oprnd0 = arginfo[i].op; | |
3144 | if ((m & (k - 1)) == 0) | |
3145 | vec_oprnd0 | |
3146 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3147 | vec_oprnd0); | |
3148 | } | |
3149 | arginfo[i].op = vec_oprnd0; | |
3150 | vec_oprnd0 | |
3151 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, | |
3152 | size_int (prec), | |
3153 | bitsize_int ((m & (k - 1)) * prec)); | |
3154 | new_stmt | |
b731b390 | 3155 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3156 | vec_oprnd0); |
3157 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3158 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3159 | } | |
3160 | else | |
3161 | { | |
3162 | k = (TYPE_VECTOR_SUBPARTS (atype) | |
3163 | / TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)); | |
3164 | gcc_assert ((k & (k - 1)) == 0); | |
3165 | vec<constructor_elt, va_gc> *ctor_elts; | |
3166 | if (k != 1) | |
3167 | vec_alloc (ctor_elts, k); | |
3168 | else | |
3169 | ctor_elts = NULL; | |
3170 | for (l = 0; l < k; l++) | |
3171 | { | |
3172 | if (m == 0 && l == 0) | |
3173 | vec_oprnd0 | |
81c40241 | 3174 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3175 | else |
3176 | vec_oprnd0 | |
3177 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3178 | arginfo[i].op); | |
3179 | arginfo[i].op = vec_oprnd0; | |
3180 | if (k == 1) | |
3181 | break; | |
3182 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, | |
3183 | vec_oprnd0); | |
3184 | } | |
3185 | if (k == 1) | |
3186 | vargs.safe_push (vec_oprnd0); | |
3187 | else | |
3188 | { | |
3189 | vec_oprnd0 = build_constructor (atype, ctor_elts); | |
3190 | new_stmt | |
b731b390 | 3191 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3192 | vec_oprnd0); |
3193 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3194 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3195 | } | |
3196 | } | |
3197 | } | |
3198 | break; | |
3199 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3200 | vargs.safe_push (op); | |
3201 | break; | |
3202 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
3203 | if (j == 0) | |
3204 | { | |
3205 | gimple_seq stmts; | |
3206 | arginfo[i].op | |
3207 | = force_gimple_operand (arginfo[i].op, &stmts, true, | |
3208 | NULL_TREE); | |
3209 | if (stmts != NULL) | |
3210 | { | |
3211 | basic_block new_bb; | |
3212 | edge pe = loop_preheader_edge (loop); | |
3213 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
3214 | gcc_assert (!new_bb); | |
3215 | } | |
17b658af JJ |
3216 | if (arginfo[i].simd_lane_linear) |
3217 | { | |
3218 | vargs.safe_push (arginfo[i].op); | |
3219 | break; | |
3220 | } | |
b731b390 | 3221 | tree phi_res = copy_ssa_name (op); |
538dd0b7 | 3222 | gphi *new_phi = create_phi_node (phi_res, loop->header); |
0136f8f0 | 3223 | set_vinfo_for_stmt (new_phi, |
310213d4 | 3224 | new_stmt_vec_info (new_phi, loop_vinfo)); |
0136f8f0 AH |
3225 | add_phi_arg (new_phi, arginfo[i].op, |
3226 | loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
3227 | enum tree_code code | |
3228 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3229 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3230 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3231 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3232 | widest_int cst |
3233 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3234 | ncopies * nunits); | |
3235 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3236 | tree phi_arg = copy_ssa_name (op); |
0d0e4a03 JJ |
3237 | new_stmt |
3238 | = gimple_build_assign (phi_arg, code, phi_res, tcst); | |
0136f8f0 AH |
3239 | gimple_stmt_iterator si = gsi_after_labels (loop->header); |
3240 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); | |
3241 | set_vinfo_for_stmt (new_stmt, | |
310213d4 | 3242 | new_stmt_vec_info (new_stmt, loop_vinfo)); |
0136f8f0 AH |
3243 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), |
3244 | UNKNOWN_LOCATION); | |
3245 | arginfo[i].op = phi_res; | |
3246 | vargs.safe_push (phi_res); | |
3247 | } | |
3248 | else | |
3249 | { | |
3250 | enum tree_code code | |
3251 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3252 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3253 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3254 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3255 | widest_int cst |
3256 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3257 | j * nunits); | |
3258 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3259 | new_temp = make_ssa_name (TREE_TYPE (op)); |
0d0e4a03 JJ |
3260 | new_stmt = gimple_build_assign (new_temp, code, |
3261 | arginfo[i].op, tcst); | |
0136f8f0 AH |
3262 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3263 | vargs.safe_push (new_temp); | |
3264 | } | |
3265 | break; | |
3266 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
e01d41e5 JJ |
3267 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3268 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3269 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3270 | default: |
3271 | gcc_unreachable (); | |
3272 | } | |
3273 | } | |
3274 | ||
3275 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
3276 | if (vec_dest) | |
3277 | { | |
3278 | gcc_assert (ratype || TYPE_VECTOR_SUBPARTS (rtype) == nunits); | |
3279 | if (ratype) | |
b731b390 | 3280 | new_temp = create_tmp_var (ratype); |
0136f8f0 AH |
3281 | else if (TYPE_VECTOR_SUBPARTS (vectype) |
3282 | == TYPE_VECTOR_SUBPARTS (rtype)) | |
3283 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3284 | else | |
3285 | new_temp = make_ssa_name (rtype, new_stmt); | |
3286 | gimple_call_set_lhs (new_stmt, new_temp); | |
3287 | } | |
3288 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3289 | ||
3290 | if (vec_dest) | |
3291 | { | |
3292 | if (TYPE_VECTOR_SUBPARTS (vectype) < nunits) | |
3293 | { | |
3294 | unsigned int k, l; | |
3295 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); | |
3296 | k = nunits / TYPE_VECTOR_SUBPARTS (vectype); | |
3297 | gcc_assert ((k & (k - 1)) == 0); | |
3298 | for (l = 0; l < k; l++) | |
3299 | { | |
3300 | tree t; | |
3301 | if (ratype) | |
3302 | { | |
3303 | t = build_fold_addr_expr (new_temp); | |
3304 | t = build2 (MEM_REF, vectype, t, | |
3305 | build_int_cst (TREE_TYPE (t), | |
3306 | l * prec / BITS_PER_UNIT)); | |
3307 | } | |
3308 | else | |
3309 | t = build3 (BIT_FIELD_REF, vectype, new_temp, | |
3310 | size_int (prec), bitsize_int (l * prec)); | |
3311 | new_stmt | |
b731b390 | 3312 | = gimple_build_assign (make_ssa_name (vectype), t); |
0136f8f0 AH |
3313 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3314 | if (j == 0 && l == 0) | |
3315 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3316 | else | |
3317 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3318 | ||
3319 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3320 | } | |
3321 | ||
3322 | if (ratype) | |
3323 | { | |
3324 | tree clobber = build_constructor (ratype, NULL); | |
3325 | TREE_THIS_VOLATILE (clobber) = 1; | |
3326 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3327 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3328 | } | |
3329 | continue; | |
3330 | } | |
3331 | else if (TYPE_VECTOR_SUBPARTS (vectype) > nunits) | |
3332 | { | |
3333 | unsigned int k = (TYPE_VECTOR_SUBPARTS (vectype) | |
3334 | / TYPE_VECTOR_SUBPARTS (rtype)); | |
3335 | gcc_assert ((k & (k - 1)) == 0); | |
3336 | if ((j & (k - 1)) == 0) | |
3337 | vec_alloc (ret_ctor_elts, k); | |
3338 | if (ratype) | |
3339 | { | |
3340 | unsigned int m, o = nunits / TYPE_VECTOR_SUBPARTS (rtype); | |
3341 | for (m = 0; m < o; m++) | |
3342 | { | |
3343 | tree tem = build4 (ARRAY_REF, rtype, new_temp, | |
3344 | size_int (m), NULL_TREE, NULL_TREE); | |
3345 | new_stmt | |
b731b390 | 3346 | = gimple_build_assign (make_ssa_name (rtype), tem); |
0136f8f0 AH |
3347 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3348 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, | |
3349 | gimple_assign_lhs (new_stmt)); | |
3350 | } | |
3351 | tree clobber = build_constructor (ratype, NULL); | |
3352 | TREE_THIS_VOLATILE (clobber) = 1; | |
3353 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3354 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3355 | } | |
3356 | else | |
3357 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); | |
3358 | if ((j & (k - 1)) != k - 1) | |
3359 | continue; | |
3360 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); | |
3361 | new_stmt | |
b731b390 | 3362 | = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0); |
0136f8f0 AH |
3363 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3364 | ||
3365 | if ((unsigned) j == k - 1) | |
3366 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3367 | else | |
3368 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3369 | ||
3370 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3371 | continue; | |
3372 | } | |
3373 | else if (ratype) | |
3374 | { | |
3375 | tree t = build_fold_addr_expr (new_temp); | |
3376 | t = build2 (MEM_REF, vectype, t, | |
3377 | build_int_cst (TREE_TYPE (t), 0)); | |
3378 | new_stmt | |
b731b390 | 3379 | = gimple_build_assign (make_ssa_name (vec_dest), t); |
0136f8f0 AH |
3380 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3381 | tree clobber = build_constructor (ratype, NULL); | |
3382 | TREE_THIS_VOLATILE (clobber) = 1; | |
3383 | vect_finish_stmt_generation (stmt, | |
3384 | gimple_build_assign (new_temp, | |
3385 | clobber), gsi); | |
3386 | } | |
3387 | } | |
3388 | ||
3389 | if (j == 0) | |
3390 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3391 | else | |
3392 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3393 | ||
3394 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3395 | } | |
3396 | ||
3397 | vargs.release (); | |
3398 | ||
3399 | /* The call in STMT might prevent it from being removed in dce. | |
3400 | We however cannot remove it here, due to the way the ssa name | |
3401 | it defines is mapped to the new definition. So just replace | |
3402 | rhs of the statement with something harmless. */ | |
3403 | ||
3404 | if (slp_node) | |
3405 | return true; | |
3406 | ||
3407 | if (scalar_dest) | |
3408 | { | |
3409 | type = TREE_TYPE (scalar_dest); | |
3410 | if (is_pattern_stmt_p (stmt_info)) | |
3411 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
3412 | else | |
3413 | lhs = gimple_call_lhs (stmt); | |
3414 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
3415 | } | |
3416 | else | |
3417 | new_stmt = gimple_build_nop (); | |
3418 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
3419 | set_vinfo_for_stmt (stmt, NULL); | |
3420 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2865f32a | 3421 | gsi_replace (gsi, new_stmt, true); |
0136f8f0 AH |
3422 | unlink_stmt_vdef (stmt); |
3423 | ||
3424 | return true; | |
3425 | } | |
3426 | ||
3427 | ||
ebfd146a IR |
3428 | /* Function vect_gen_widened_results_half |
3429 | ||
3430 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 3431 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 3432 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
3433 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
3434 | needs to be created (DECL is a function-decl of a target-builtin). | |
3435 | STMT is the original scalar stmt that we are vectorizing. */ | |
3436 | ||
355fe088 | 3437 | static gimple * |
ebfd146a IR |
3438 | vect_gen_widened_results_half (enum tree_code code, |
3439 | tree decl, | |
3440 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
3441 | tree vec_dest, gimple_stmt_iterator *gsi, | |
355fe088 | 3442 | gimple *stmt) |
b8698a0f | 3443 | { |
355fe088 | 3444 | gimple *new_stmt; |
b8698a0f L |
3445 | tree new_temp; |
3446 | ||
3447 | /* Generate half of the widened result: */ | |
3448 | if (code == CALL_EXPR) | |
3449 | { | |
3450 | /* Target specific support */ | |
ebfd146a IR |
3451 | if (op_type == binary_op) |
3452 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
3453 | else | |
3454 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
3455 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3456 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
3457 | } |
3458 | else | |
ebfd146a | 3459 | { |
b8698a0f L |
3460 | /* Generic support */ |
3461 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
3462 | if (op_type != binary_op) |
3463 | vec_oprnd1 = NULL; | |
0d0e4a03 | 3464 | new_stmt = gimple_build_assign (vec_dest, code, vec_oprnd0, vec_oprnd1); |
ebfd146a IR |
3465 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3466 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 3467 | } |
ebfd146a IR |
3468 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3469 | ||
ebfd146a IR |
3470 | return new_stmt; |
3471 | } | |
3472 | ||
4a00c761 JJ |
3473 | |
3474 | /* Get vectorized definitions for loop-based vectorization. For the first | |
3475 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
3476 | scalar operand), and for the rest we get a copy with | |
3477 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
3478 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
3479 | The vectors are collected into VEC_OPRNDS. */ | |
3480 | ||
3481 | static void | |
355fe088 | 3482 | vect_get_loop_based_defs (tree *oprnd, gimple *stmt, enum vect_def_type dt, |
9771b263 | 3483 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
3484 | { |
3485 | tree vec_oprnd; | |
3486 | ||
3487 | /* Get first vector operand. */ | |
3488 | /* All the vector operands except the very first one (that is scalar oprnd) | |
3489 | are stmt copies. */ | |
3490 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
81c40241 | 3491 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt); |
4a00c761 JJ |
3492 | else |
3493 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
3494 | ||
9771b263 | 3495 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3496 | |
3497 | /* Get second vector operand. */ | |
3498 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 3499 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3500 | |
3501 | *oprnd = vec_oprnd; | |
3502 | ||
3503 | /* For conversion in multiple steps, continue to get operands | |
3504 | recursively. */ | |
3505 | if (multi_step_cvt) | |
3506 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
3507 | } | |
3508 | ||
3509 | ||
3510 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
3511 | For multi-step conversions store the resulting vectors and call the function | |
3512 | recursively. */ | |
3513 | ||
3514 | static void | |
9771b263 | 3515 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
355fe088 | 3516 | int multi_step_cvt, gimple *stmt, |
9771b263 | 3517 | vec<tree> vec_dsts, |
4a00c761 JJ |
3518 | gimple_stmt_iterator *gsi, |
3519 | slp_tree slp_node, enum tree_code code, | |
3520 | stmt_vec_info *prev_stmt_info) | |
3521 | { | |
3522 | unsigned int i; | |
3523 | tree vop0, vop1, new_tmp, vec_dest; | |
355fe088 | 3524 | gimple *new_stmt; |
4a00c761 JJ |
3525 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
3526 | ||
9771b263 | 3527 | vec_dest = vec_dsts.pop (); |
4a00c761 | 3528 | |
9771b263 | 3529 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
3530 | { |
3531 | /* Create demotion operation. */ | |
9771b263 DN |
3532 | vop0 = (*vec_oprnds)[i]; |
3533 | vop1 = (*vec_oprnds)[i + 1]; | |
0d0e4a03 | 3534 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
4a00c761 JJ |
3535 | new_tmp = make_ssa_name (vec_dest, new_stmt); |
3536 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
3537 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3538 | ||
3539 | if (multi_step_cvt) | |
3540 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 3541 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
3542 | else |
3543 | { | |
3544 | /* This is the last step of the conversion sequence. Store the | |
3545 | vectors in SLP_NODE or in vector info of the scalar statement | |
3546 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
3547 | if (slp_node) | |
9771b263 | 3548 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 3549 | else |
c689ce1e RB |
3550 | { |
3551 | if (!*prev_stmt_info) | |
3552 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3553 | else | |
3554 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
4a00c761 | 3555 | |
c689ce1e RB |
3556 | *prev_stmt_info = vinfo_for_stmt (new_stmt); |
3557 | } | |
4a00c761 JJ |
3558 | } |
3559 | } | |
3560 | ||
3561 | /* For multi-step demotion operations we first generate demotion operations | |
3562 | from the source type to the intermediate types, and then combine the | |
3563 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
3564 | type. */ | |
3565 | if (multi_step_cvt) | |
3566 | { | |
3567 | /* At each level of recursion we have half of the operands we had at the | |
3568 | previous level. */ | |
9771b263 | 3569 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
3570 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
3571 | stmt, vec_dsts, gsi, slp_node, | |
3572 | VEC_PACK_TRUNC_EXPR, | |
3573 | prev_stmt_info); | |
3574 | } | |
3575 | ||
9771b263 | 3576 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3577 | } |
3578 | ||
3579 | ||
3580 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
3581 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
3582 | the resulting vectors and call the function recursively. */ | |
3583 | ||
3584 | static void | |
9771b263 DN |
3585 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
3586 | vec<tree> *vec_oprnds1, | |
355fe088 | 3587 | gimple *stmt, tree vec_dest, |
4a00c761 JJ |
3588 | gimple_stmt_iterator *gsi, |
3589 | enum tree_code code1, | |
3590 | enum tree_code code2, tree decl1, | |
3591 | tree decl2, int op_type) | |
3592 | { | |
3593 | int i; | |
3594 | tree vop0, vop1, new_tmp1, new_tmp2; | |
355fe088 | 3595 | gimple *new_stmt1, *new_stmt2; |
6e1aa848 | 3596 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 3597 | |
9771b263 DN |
3598 | vec_tmp.create (vec_oprnds0->length () * 2); |
3599 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
3600 | { |
3601 | if (op_type == binary_op) | |
9771b263 | 3602 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
3603 | else |
3604 | vop1 = NULL_TREE; | |
3605 | ||
3606 | /* Generate the two halves of promotion operation. */ | |
3607 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
3608 | op_type, vec_dest, gsi, stmt); | |
3609 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
3610 | op_type, vec_dest, gsi, stmt); | |
3611 | if (is_gimple_call (new_stmt1)) | |
3612 | { | |
3613 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
3614 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
3615 | } | |
3616 | else | |
3617 | { | |
3618 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
3619 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
3620 | } | |
3621 | ||
3622 | /* Store the results for the next step. */ | |
9771b263 DN |
3623 | vec_tmp.quick_push (new_tmp1); |
3624 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
3625 | } |
3626 | ||
689eaba3 | 3627 | vec_oprnds0->release (); |
4a00c761 JJ |
3628 | *vec_oprnds0 = vec_tmp; |
3629 | } | |
3630 | ||
3631 | ||
b8698a0f L |
3632 | /* Check if STMT performs a conversion operation, that can be vectorized. |
3633 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 3634 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
3635 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
3636 | ||
3637 | static bool | |
355fe088 TS |
3638 | vectorizable_conversion (gimple *stmt, gimple_stmt_iterator *gsi, |
3639 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
3640 | { |
3641 | tree vec_dest; | |
3642 | tree scalar_dest; | |
4a00c761 | 3643 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
3644 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
3645 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3646 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3647 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 3648 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
3649 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
3650 | tree new_temp; | |
355fe088 | 3651 | gimple *def_stmt; |
ebfd146a | 3652 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
355fe088 | 3653 | gimple *new_stmt = NULL; |
ebfd146a IR |
3654 | stmt_vec_info prev_stmt_info; |
3655 | int nunits_in; | |
3656 | int nunits_out; | |
3657 | tree vectype_out, vectype_in; | |
4a00c761 JJ |
3658 | int ncopies, i, j; |
3659 | tree lhs_type, rhs_type; | |
ebfd146a | 3660 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
3661 | vec<tree> vec_oprnds0 = vNULL; |
3662 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 3663 | tree vop0; |
4a00c761 | 3664 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 3665 | vec_info *vinfo = stmt_info->vinfo; |
4a00c761 | 3666 | int multi_step_cvt = 0; |
6e1aa848 DN |
3667 | vec<tree> vec_dsts = vNULL; |
3668 | vec<tree> interm_types = vNULL; | |
4a00c761 JJ |
3669 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
3670 | int op_type; | |
ef4bddc2 | 3671 | machine_mode rhs_mode; |
4a00c761 | 3672 | unsigned short fltsz; |
ebfd146a IR |
3673 | |
3674 | /* Is STMT a vectorizable conversion? */ | |
3675 | ||
4a00c761 | 3676 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3677 | return false; |
3678 | ||
66c16fd9 RB |
3679 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3680 | && ! vec_stmt) | |
ebfd146a IR |
3681 | return false; |
3682 | ||
3683 | if (!is_gimple_assign (stmt)) | |
3684 | return false; | |
3685 | ||
3686 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3687 | return false; | |
3688 | ||
3689 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
3690 | if (!CONVERT_EXPR_CODE_P (code) |
3691 | && code != FIX_TRUNC_EXPR | |
3692 | && code != FLOAT_EXPR | |
3693 | && code != WIDEN_MULT_EXPR | |
3694 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
3695 | return false; |
3696 | ||
4a00c761 JJ |
3697 | op_type = TREE_CODE_LENGTH (code); |
3698 | ||
ebfd146a | 3699 | /* Check types of lhs and rhs. */ |
b690cc0f | 3700 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 3701 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
3702 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
3703 | ||
ebfd146a IR |
3704 | op0 = gimple_assign_rhs1 (stmt); |
3705 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
3706 | |
3707 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
3708 | && !((INTEGRAL_TYPE_P (lhs_type) | |
3709 | && INTEGRAL_TYPE_P (rhs_type)) | |
3710 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
3711 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
3712 | return false; | |
3713 | ||
e6f5c25d IE |
3714 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
3715 | && ((INTEGRAL_TYPE_P (lhs_type) | |
3716 | && (TYPE_PRECISION (lhs_type) | |
3717 | != GET_MODE_PRECISION (TYPE_MODE (lhs_type)))) | |
3718 | || (INTEGRAL_TYPE_P (rhs_type) | |
3719 | && (TYPE_PRECISION (rhs_type) | |
3720 | != GET_MODE_PRECISION (TYPE_MODE (rhs_type)))))) | |
4a00c761 | 3721 | { |
73fbfcad | 3722 | if (dump_enabled_p ()) |
78c60e3d | 3723 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
3724 | "type conversion to/from bit-precision unsupported." |
3725 | "\n"); | |
4a00c761 JJ |
3726 | return false; |
3727 | } | |
3728 | ||
b690cc0f | 3729 | /* Check the operands of the operation. */ |
81c40241 | 3730 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype_in)) |
b690cc0f | 3731 | { |
73fbfcad | 3732 | if (dump_enabled_p ()) |
78c60e3d | 3733 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3734 | "use not simple.\n"); |
b690cc0f RG |
3735 | return false; |
3736 | } | |
4a00c761 JJ |
3737 | if (op_type == binary_op) |
3738 | { | |
3739 | bool ok; | |
3740 | ||
3741 | op1 = gimple_assign_rhs2 (stmt); | |
3742 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
3743 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
3744 | OP1. */ | |
3745 | if (CONSTANT_CLASS_P (op0)) | |
81c40241 | 3746 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &vectype_in); |
4a00c761 | 3747 | else |
81c40241 | 3748 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1]); |
4a00c761 JJ |
3749 | |
3750 | if (!ok) | |
3751 | { | |
73fbfcad | 3752 | if (dump_enabled_p ()) |
78c60e3d | 3753 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3754 | "use not simple.\n"); |
4a00c761 JJ |
3755 | return false; |
3756 | } | |
3757 | } | |
3758 | ||
b690cc0f RG |
3759 | /* If op0 is an external or constant defs use a vector type of |
3760 | the same size as the output vector type. */ | |
ebfd146a | 3761 | if (!vectype_in) |
b690cc0f | 3762 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
3763 | if (vec_stmt) |
3764 | gcc_assert (vectype_in); | |
3765 | if (!vectype_in) | |
3766 | { | |
73fbfcad | 3767 | if (dump_enabled_p ()) |
4a00c761 | 3768 | { |
78c60e3d SS |
3769 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3770 | "no vectype for scalar type "); | |
3771 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 3772 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 3773 | } |
7d8930a0 IR |
3774 | |
3775 | return false; | |
3776 | } | |
ebfd146a | 3777 | |
e6f5c25d IE |
3778 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
3779 | && !VECTOR_BOOLEAN_TYPE_P (vectype_in)) | |
3780 | { | |
3781 | if (dump_enabled_p ()) | |
3782 | { | |
3783 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3784 | "can't convert between boolean and non " | |
3785 | "boolean vectors"); | |
3786 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
3787 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
3788 | } | |
3789 | ||
3790 | return false; | |
3791 | } | |
3792 | ||
b690cc0f RG |
3793 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
3794 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4a00c761 | 3795 | if (nunits_in < nunits_out) |
ebfd146a IR |
3796 | modifier = NARROW; |
3797 | else if (nunits_out == nunits_in) | |
3798 | modifier = NONE; | |
ebfd146a | 3799 | else |
4a00c761 | 3800 | modifier = WIDEN; |
ebfd146a | 3801 | |
ff802fa1 IR |
3802 | /* Multiple types in SLP are handled by creating the appropriate number of |
3803 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3804 | case of SLP. */ | |
437f4a00 | 3805 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a | 3806 | ncopies = 1; |
4a00c761 JJ |
3807 | else if (modifier == NARROW) |
3808 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
3809 | else | |
3810 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
b8698a0f | 3811 | |
ebfd146a IR |
3812 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
3813 | needs to be generated. */ | |
3814 | gcc_assert (ncopies >= 1); | |
3815 | ||
ebfd146a | 3816 | /* Supportable by target? */ |
4a00c761 | 3817 | switch (modifier) |
ebfd146a | 3818 | { |
4a00c761 JJ |
3819 | case NONE: |
3820 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
3821 | return false; | |
3822 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
3823 | &decl1, &code1)) | |
3824 | break; | |
3825 | /* FALLTHRU */ | |
3826 | unsupported: | |
73fbfcad | 3827 | if (dump_enabled_p ()) |
78c60e3d | 3828 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3829 | "conversion not supported by target.\n"); |
ebfd146a | 3830 | return false; |
ebfd146a | 3831 | |
4a00c761 JJ |
3832 | case WIDEN: |
3833 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
3834 | &code1, &code2, &multi_step_cvt, |
3835 | &interm_types)) | |
4a00c761 JJ |
3836 | { |
3837 | /* Binary widening operation can only be supported directly by the | |
3838 | architecture. */ | |
3839 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
3840 | break; | |
3841 | } | |
3842 | ||
3843 | if (code != FLOAT_EXPR | |
3844 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
3845 | <= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
3846 | goto unsupported; | |
3847 | ||
3848 | rhs_mode = TYPE_MODE (rhs_type); | |
3849 | fltsz = GET_MODE_SIZE (TYPE_MODE (lhs_type)); | |
3850 | for (rhs_mode = GET_MODE_2XWIDER_MODE (TYPE_MODE (rhs_type)); | |
3851 | rhs_mode != VOIDmode && GET_MODE_SIZE (rhs_mode) <= fltsz; | |
3852 | rhs_mode = GET_MODE_2XWIDER_MODE (rhs_mode)) | |
3853 | { | |
3854 | cvt_type | |
3855 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
3856 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
3857 | if (cvt_type == NULL_TREE) | |
3858 | goto unsupported; | |
3859 | ||
3860 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
3861 | { | |
3862 | if (!supportable_convert_operation (code, vectype_out, | |
3863 | cvt_type, &decl1, &codecvt1)) | |
3864 | goto unsupported; | |
3865 | } | |
3866 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
3867 | cvt_type, &codecvt1, |
3868 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
3869 | &interm_types)) |
3870 | continue; | |
3871 | else | |
3872 | gcc_assert (multi_step_cvt == 0); | |
3873 | ||
3874 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
3875 | vectype_in, &code1, &code2, |
3876 | &multi_step_cvt, &interm_types)) | |
4a00c761 JJ |
3877 | break; |
3878 | } | |
3879 | ||
3880 | if (rhs_mode == VOIDmode || GET_MODE_SIZE (rhs_mode) > fltsz) | |
3881 | goto unsupported; | |
3882 | ||
3883 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
3884 | codecvt2 = ERROR_MARK; | |
3885 | else | |
3886 | { | |
3887 | multi_step_cvt++; | |
9771b263 | 3888 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
3889 | cvt_type = NULL_TREE; |
3890 | } | |
3891 | break; | |
3892 | ||
3893 | case NARROW: | |
3894 | gcc_assert (op_type == unary_op); | |
3895 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
3896 | &code1, &multi_step_cvt, | |
3897 | &interm_types)) | |
3898 | break; | |
3899 | ||
3900 | if (code != FIX_TRUNC_EXPR | |
3901 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
3902 | >= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
3903 | goto unsupported; | |
3904 | ||
3905 | rhs_mode = TYPE_MODE (rhs_type); | |
3906 | cvt_type | |
3907 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
3908 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
3909 | if (cvt_type == NULL_TREE) | |
3910 | goto unsupported; | |
3911 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
3912 | &decl1, &codecvt1)) | |
3913 | goto unsupported; | |
3914 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
3915 | &code1, &multi_step_cvt, | |
3916 | &interm_types)) | |
3917 | break; | |
3918 | goto unsupported; | |
3919 | ||
3920 | default: | |
3921 | gcc_unreachable (); | |
ebfd146a IR |
3922 | } |
3923 | ||
3924 | if (!vec_stmt) /* transformation not required. */ | |
3925 | { | |
73fbfcad | 3926 | if (dump_enabled_p ()) |
78c60e3d | 3927 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3928 | "=== vectorizable_conversion ===\n"); |
4a00c761 | 3929 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
3930 | { |
3931 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
c3e7ee41 | 3932 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
8bd37302 | 3933 | } |
4a00c761 JJ |
3934 | else if (modifier == NARROW) |
3935 | { | |
3936 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
8bd37302 | 3937 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 JJ |
3938 | } |
3939 | else | |
3940 | { | |
3941 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
8bd37302 | 3942 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 | 3943 | } |
9771b263 | 3944 | interm_types.release (); |
ebfd146a IR |
3945 | return true; |
3946 | } | |
3947 | ||
3948 | /** Transform. **/ | |
73fbfcad | 3949 | if (dump_enabled_p ()) |
78c60e3d | 3950 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3951 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 3952 | |
4a00c761 JJ |
3953 | if (op_type == binary_op) |
3954 | { | |
3955 | if (CONSTANT_CLASS_P (op0)) | |
3956 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
3957 | else if (CONSTANT_CLASS_P (op1)) | |
3958 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
3959 | } | |
3960 | ||
3961 | /* In case of multi-step conversion, we first generate conversion operations | |
3962 | to the intermediate types, and then from that types to the final one. | |
3963 | We create vector destinations for the intermediate type (TYPES) received | |
3964 | from supportable_*_operation, and store them in the correct order | |
3965 | for future use in vect_create_vectorized_*_stmts (). */ | |
9771b263 | 3966 | vec_dsts.create (multi_step_cvt + 1); |
82294ec1 JJ |
3967 | vec_dest = vect_create_destination_var (scalar_dest, |
3968 | (cvt_type && modifier == WIDEN) | |
3969 | ? cvt_type : vectype_out); | |
9771b263 | 3970 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3971 | |
3972 | if (multi_step_cvt) | |
3973 | { | |
9771b263 DN |
3974 | for (i = interm_types.length () - 1; |
3975 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
3976 | { |
3977 | vec_dest = vect_create_destination_var (scalar_dest, | |
3978 | intermediate_type); | |
9771b263 | 3979 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3980 | } |
3981 | } | |
ebfd146a | 3982 | |
4a00c761 | 3983 | if (cvt_type) |
82294ec1 JJ |
3984 | vec_dest = vect_create_destination_var (scalar_dest, |
3985 | modifier == WIDEN | |
3986 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
3987 | |
3988 | if (!slp_node) | |
3989 | { | |
30862efc | 3990 | if (modifier == WIDEN) |
4a00c761 | 3991 | { |
c3284718 | 3992 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 3993 | if (op_type == binary_op) |
9771b263 | 3994 | vec_oprnds1.create (1); |
4a00c761 | 3995 | } |
30862efc | 3996 | else if (modifier == NARROW) |
9771b263 DN |
3997 | vec_oprnds0.create ( |
3998 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
3999 | } |
4000 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 4001 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 4002 | |
4a00c761 | 4003 | last_oprnd = op0; |
ebfd146a IR |
4004 | prev_stmt_info = NULL; |
4005 | switch (modifier) | |
4006 | { | |
4007 | case NONE: | |
4008 | for (j = 0; j < ncopies; j++) | |
4009 | { | |
ebfd146a | 4010 | if (j == 0) |
d092494c IR |
4011 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node, |
4012 | -1); | |
ebfd146a IR |
4013 | else |
4014 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
4015 | ||
9771b263 | 4016 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4017 | { |
4018 | /* Arguments are ready, create the new vector stmt. */ | |
4019 | if (code1 == CALL_EXPR) | |
4020 | { | |
4021 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4022 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4023 | gimple_call_set_lhs (new_stmt, new_temp); | |
4024 | } | |
4025 | else | |
4026 | { | |
4027 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
0d0e4a03 | 4028 | new_stmt = gimple_build_assign (vec_dest, code1, vop0); |
4a00c761 JJ |
4029 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4030 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4031 | } | |
4032 | ||
4033 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4034 | if (slp_node) | |
9771b263 | 4035 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
225ce44b RB |
4036 | else |
4037 | { | |
4038 | if (!prev_stmt_info) | |
4039 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4040 | else | |
4041 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4042 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4043 | } | |
4a00c761 | 4044 | } |
ebfd146a IR |
4045 | } |
4046 | break; | |
4047 | ||
4048 | case WIDEN: | |
4049 | /* In case the vectorization factor (VF) is bigger than the number | |
4050 | of elements that we can fit in a vectype (nunits), we have to | |
4051 | generate more than one vector stmt - i.e - we need to "unroll" | |
4052 | the vector stmt by a factor VF/nunits. */ | |
4053 | for (j = 0; j < ncopies; j++) | |
4054 | { | |
4a00c761 | 4055 | /* Handle uses. */ |
ebfd146a | 4056 | if (j == 0) |
4a00c761 JJ |
4057 | { |
4058 | if (slp_node) | |
4059 | { | |
4060 | if (code == WIDEN_LSHIFT_EXPR) | |
4061 | { | |
4062 | unsigned int k; | |
ebfd146a | 4063 | |
4a00c761 JJ |
4064 | vec_oprnd1 = op1; |
4065 | /* Store vec_oprnd1 for every vector stmt to be created | |
4066 | for SLP_NODE. We check during the analysis that all | |
4067 | the shift arguments are the same. */ | |
4068 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4069 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4070 | |
4071 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
4072 | slp_node, -1); | |
4073 | } | |
4074 | else | |
4075 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
4076 | &vec_oprnds1, slp_node, -1); | |
4077 | } | |
4078 | else | |
4079 | { | |
81c40241 | 4080 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 4081 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
4082 | if (op_type == binary_op) |
4083 | { | |
4084 | if (code == WIDEN_LSHIFT_EXPR) | |
4085 | vec_oprnd1 = op1; | |
4086 | else | |
81c40241 | 4087 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 4088 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4089 | } |
4090 | } | |
4091 | } | |
ebfd146a | 4092 | else |
4a00c761 JJ |
4093 | { |
4094 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
4095 | vec_oprnds0.truncate (0); |
4096 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
4097 | if (op_type == binary_op) |
4098 | { | |
4099 | if (code == WIDEN_LSHIFT_EXPR) | |
4100 | vec_oprnd1 = op1; | |
4101 | else | |
4102 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
4103 | vec_oprnd1); | |
9771b263 DN |
4104 | vec_oprnds1.truncate (0); |
4105 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
4106 | } |
4107 | } | |
ebfd146a | 4108 | |
4a00c761 JJ |
4109 | /* Arguments are ready. Create the new vector stmts. */ |
4110 | for (i = multi_step_cvt; i >= 0; i--) | |
4111 | { | |
9771b263 | 4112 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
4113 | enum tree_code c1 = code1, c2 = code2; |
4114 | if (i == 0 && codecvt2 != ERROR_MARK) | |
4115 | { | |
4116 | c1 = codecvt1; | |
4117 | c2 = codecvt2; | |
4118 | } | |
4119 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
4120 | &vec_oprnds1, | |
4121 | stmt, this_dest, gsi, | |
4122 | c1, c2, decl1, decl2, | |
4123 | op_type); | |
4124 | } | |
4125 | ||
9771b263 | 4126 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4127 | { |
4128 | if (cvt_type) | |
4129 | { | |
4130 | if (codecvt1 == CALL_EXPR) | |
4131 | { | |
4132 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4133 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4134 | gimple_call_set_lhs (new_stmt, new_temp); | |
4135 | } | |
4136 | else | |
4137 | { | |
4138 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4139 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4140 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4141 | vop0); | |
4a00c761 JJ |
4142 | } |
4143 | ||
4144 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4145 | } | |
4146 | else | |
4147 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
4148 | ||
4149 | if (slp_node) | |
9771b263 | 4150 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 4151 | else |
c689ce1e RB |
4152 | { |
4153 | if (!prev_stmt_info) | |
4154 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
4155 | else | |
4156 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4157 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4158 | } | |
4a00c761 | 4159 | } |
ebfd146a | 4160 | } |
4a00c761 JJ |
4161 | |
4162 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
4163 | break; |
4164 | ||
4165 | case NARROW: | |
4166 | /* In case the vectorization factor (VF) is bigger than the number | |
4167 | of elements that we can fit in a vectype (nunits), we have to | |
4168 | generate more than one vector stmt - i.e - we need to "unroll" | |
4169 | the vector stmt by a factor VF/nunits. */ | |
4170 | for (j = 0; j < ncopies; j++) | |
4171 | { | |
4172 | /* Handle uses. */ | |
4a00c761 JJ |
4173 | if (slp_node) |
4174 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
4175 | slp_node, -1); | |
ebfd146a IR |
4176 | else |
4177 | { | |
9771b263 | 4178 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
4179 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
4180 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
4181 | } |
4182 | ||
4a00c761 JJ |
4183 | /* Arguments are ready. Create the new vector stmts. */ |
4184 | if (cvt_type) | |
9771b263 | 4185 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4186 | { |
4187 | if (codecvt1 == CALL_EXPR) | |
4188 | { | |
4189 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4190 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4191 | gimple_call_set_lhs (new_stmt, new_temp); | |
4192 | } | |
4193 | else | |
4194 | { | |
4195 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4196 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4197 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4198 | vop0); | |
4a00c761 | 4199 | } |
ebfd146a | 4200 | |
4a00c761 | 4201 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 4202 | vec_oprnds0[i] = new_temp; |
4a00c761 | 4203 | } |
ebfd146a | 4204 | |
4a00c761 JJ |
4205 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
4206 | stmt, vec_dsts, gsi, | |
4207 | slp_node, code1, | |
4208 | &prev_stmt_info); | |
ebfd146a IR |
4209 | } |
4210 | ||
4211 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 4212 | break; |
ebfd146a IR |
4213 | } |
4214 | ||
9771b263 DN |
4215 | vec_oprnds0.release (); |
4216 | vec_oprnds1.release (); | |
4217 | vec_dsts.release (); | |
4218 | interm_types.release (); | |
ebfd146a IR |
4219 | |
4220 | return true; | |
4221 | } | |
ff802fa1 IR |
4222 | |
4223 | ||
ebfd146a IR |
4224 | /* Function vectorizable_assignment. |
4225 | ||
b8698a0f L |
4226 | Check if STMT performs an assignment (copy) that can be vectorized. |
4227 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4228 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4229 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4230 | ||
4231 | static bool | |
355fe088 TS |
4232 | vectorizable_assignment (gimple *stmt, gimple_stmt_iterator *gsi, |
4233 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
4234 | { |
4235 | tree vec_dest; | |
4236 | tree scalar_dest; | |
4237 | tree op; | |
4238 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a IR |
4239 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4240 | tree new_temp; | |
355fe088 | 4241 | gimple *def_stmt; |
ebfd146a | 4242 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
ebfd146a | 4243 | int ncopies; |
f18b55bd | 4244 | int i, j; |
6e1aa848 | 4245 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 4246 | tree vop; |
a70d6342 | 4247 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4248 | vec_info *vinfo = stmt_info->vinfo; |
355fe088 | 4249 | gimple *new_stmt = NULL; |
f18b55bd | 4250 | stmt_vec_info prev_stmt_info = NULL; |
fde9c428 RG |
4251 | enum tree_code code; |
4252 | tree vectype_in; | |
ebfd146a | 4253 | |
a70d6342 | 4254 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4255 | return false; |
4256 | ||
66c16fd9 RB |
4257 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4258 | && ! vec_stmt) | |
ebfd146a IR |
4259 | return false; |
4260 | ||
4261 | /* Is vectorizable assignment? */ | |
4262 | if (!is_gimple_assign (stmt)) | |
4263 | return false; | |
4264 | ||
4265 | scalar_dest = gimple_assign_lhs (stmt); | |
4266 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4267 | return false; | |
4268 | ||
fde9c428 | 4269 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 4270 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
4271 | || code == PAREN_EXPR |
4272 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
4273 | op = gimple_assign_rhs1 (stmt); |
4274 | else | |
4275 | return false; | |
4276 | ||
7b7ec6c5 RG |
4277 | if (code == VIEW_CONVERT_EXPR) |
4278 | op = TREE_OPERAND (op, 0); | |
4279 | ||
465c8c19 JJ |
4280 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
4281 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4282 | ||
4283 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4284 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4285 | case of SLP. */ | |
4286 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
4287 | ncopies = 1; | |
4288 | else | |
4289 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4290 | ||
4291 | gcc_assert (ncopies >= 1); | |
4292 | ||
81c40241 | 4293 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[0], &vectype_in)) |
ebfd146a | 4294 | { |
73fbfcad | 4295 | if (dump_enabled_p ()) |
78c60e3d | 4296 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4297 | "use not simple.\n"); |
ebfd146a IR |
4298 | return false; |
4299 | } | |
4300 | ||
fde9c428 RG |
4301 | /* We can handle NOP_EXPR conversions that do not change the number |
4302 | of elements or the vector size. */ | |
7b7ec6c5 RG |
4303 | if ((CONVERT_EXPR_CODE_P (code) |
4304 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 RG |
4305 | && (!vectype_in |
4306 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
4307 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
4308 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
4309 | return false; | |
4310 | ||
7b7b1813 RG |
4311 | /* We do not handle bit-precision changes. */ |
4312 | if ((CONVERT_EXPR_CODE_P (code) | |
4313 | || code == VIEW_CONVERT_EXPR) | |
4314 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
4315 | && ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4316 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4317 | || ((TYPE_PRECISION (TREE_TYPE (op)) | |
4318 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))) | |
4319 | /* But a conversion that does not change the bit-pattern is ok. */ | |
4320 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4321 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2dab46d5 IE |
4322 | && TYPE_UNSIGNED (TREE_TYPE (op))) |
4323 | /* Conversion between boolean types of different sizes is | |
4324 | a simple assignment in case their vectypes are same | |
4325 | boolean vectors. */ | |
4326 | && (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
4327 | || !VECTOR_BOOLEAN_TYPE_P (vectype_in))) | |
7b7b1813 | 4328 | { |
73fbfcad | 4329 | if (dump_enabled_p ()) |
78c60e3d SS |
4330 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4331 | "type conversion to/from bit-precision " | |
e645e942 | 4332 | "unsupported.\n"); |
7b7b1813 RG |
4333 | return false; |
4334 | } | |
4335 | ||
ebfd146a IR |
4336 | if (!vec_stmt) /* transformation not required. */ |
4337 | { | |
4338 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 4339 | if (dump_enabled_p ()) |
78c60e3d | 4340 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4341 | "=== vectorizable_assignment ===\n"); |
c3e7ee41 | 4342 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
4343 | return true; |
4344 | } | |
4345 | ||
4346 | /** Transform. **/ | |
73fbfcad | 4347 | if (dump_enabled_p ()) |
e645e942 | 4348 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
4349 | |
4350 | /* Handle def. */ | |
4351 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4352 | ||
4353 | /* Handle use. */ | |
f18b55bd | 4354 | for (j = 0; j < ncopies; j++) |
ebfd146a | 4355 | { |
f18b55bd IR |
4356 | /* Handle uses. */ |
4357 | if (j == 0) | |
d092494c | 4358 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node, -1); |
f18b55bd IR |
4359 | else |
4360 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
4361 | ||
4362 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 4363 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 4364 | { |
7b7ec6c5 RG |
4365 | if (CONVERT_EXPR_CODE_P (code) |
4366 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 4367 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
4368 | new_stmt = gimple_build_assign (vec_dest, vop); |
4369 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4370 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4371 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4372 | if (slp_node) | |
9771b263 | 4373 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 4374 | } |
ebfd146a IR |
4375 | |
4376 | if (slp_node) | |
f18b55bd IR |
4377 | continue; |
4378 | ||
4379 | if (j == 0) | |
4380 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4381 | else | |
4382 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4383 | ||
4384 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4385 | } | |
b8698a0f | 4386 | |
9771b263 | 4387 | vec_oprnds.release (); |
ebfd146a IR |
4388 | return true; |
4389 | } | |
4390 | ||
9dc3f7de | 4391 | |
1107f3ae IR |
4392 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
4393 | either as shift by a scalar or by a vector. */ | |
4394 | ||
4395 | bool | |
4396 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
4397 | { | |
4398 | ||
ef4bddc2 | 4399 | machine_mode vec_mode; |
1107f3ae IR |
4400 | optab optab; |
4401 | int icode; | |
4402 | tree vectype; | |
4403 | ||
4404 | vectype = get_vectype_for_scalar_type (scalar_type); | |
4405 | if (!vectype) | |
4406 | return false; | |
4407 | ||
4408 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4409 | if (!optab | |
4410 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
4411 | { | |
4412 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4413 | if (!optab | |
4414 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
4415 | == CODE_FOR_nothing)) | |
4416 | return false; | |
4417 | } | |
4418 | ||
4419 | vec_mode = TYPE_MODE (vectype); | |
4420 | icode = (int) optab_handler (optab, vec_mode); | |
4421 | if (icode == CODE_FOR_nothing) | |
4422 | return false; | |
4423 | ||
4424 | return true; | |
4425 | } | |
4426 | ||
4427 | ||
9dc3f7de IR |
4428 | /* Function vectorizable_shift. |
4429 | ||
4430 | Check if STMT performs a shift operation that can be vectorized. | |
4431 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4432 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
4433 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4434 | ||
4435 | static bool | |
355fe088 TS |
4436 | vectorizable_shift (gimple *stmt, gimple_stmt_iterator *gsi, |
4437 | gimple **vec_stmt, slp_tree slp_node) | |
9dc3f7de IR |
4438 | { |
4439 | tree vec_dest; | |
4440 | tree scalar_dest; | |
4441 | tree op0, op1 = NULL; | |
4442 | tree vec_oprnd1 = NULL_TREE; | |
4443 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4444 | tree vectype; | |
4445 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4446 | enum tree_code code; | |
ef4bddc2 | 4447 | machine_mode vec_mode; |
9dc3f7de IR |
4448 | tree new_temp; |
4449 | optab optab; | |
4450 | int icode; | |
ef4bddc2 | 4451 | machine_mode optab_op2_mode; |
355fe088 | 4452 | gimple *def_stmt; |
9dc3f7de | 4453 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
355fe088 | 4454 | gimple *new_stmt = NULL; |
9dc3f7de IR |
4455 | stmt_vec_info prev_stmt_info; |
4456 | int nunits_in; | |
4457 | int nunits_out; | |
4458 | tree vectype_out; | |
cede2577 | 4459 | tree op1_vectype; |
9dc3f7de IR |
4460 | int ncopies; |
4461 | int j, i; | |
6e1aa848 DN |
4462 | vec<tree> vec_oprnds0 = vNULL; |
4463 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
4464 | tree vop0, vop1; |
4465 | unsigned int k; | |
49eab32e | 4466 | bool scalar_shift_arg = true; |
9dc3f7de | 4467 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4468 | vec_info *vinfo = stmt_info->vinfo; |
9dc3f7de IR |
4469 | int vf; |
4470 | ||
4471 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
4472 | return false; | |
4473 | ||
66c16fd9 RB |
4474 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4475 | && ! vec_stmt) | |
9dc3f7de IR |
4476 | return false; |
4477 | ||
4478 | /* Is STMT a vectorizable binary/unary operation? */ | |
4479 | if (!is_gimple_assign (stmt)) | |
4480 | return false; | |
4481 | ||
4482 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4483 | return false; | |
4484 | ||
4485 | code = gimple_assign_rhs_code (stmt); | |
4486 | ||
4487 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
4488 | || code == RROTATE_EXPR)) | |
4489 | return false; | |
4490 | ||
4491 | scalar_dest = gimple_assign_lhs (stmt); | |
4492 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
7b7b1813 RG |
4493 | if (TYPE_PRECISION (TREE_TYPE (scalar_dest)) |
4494 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4495 | { | |
73fbfcad | 4496 | if (dump_enabled_p ()) |
78c60e3d | 4497 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4498 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
4499 | return false; |
4500 | } | |
9dc3f7de IR |
4501 | |
4502 | op0 = gimple_assign_rhs1 (stmt); | |
81c40241 | 4503 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
9dc3f7de | 4504 | { |
73fbfcad | 4505 | if (dump_enabled_p ()) |
78c60e3d | 4506 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4507 | "use not simple.\n"); |
9dc3f7de IR |
4508 | return false; |
4509 | } | |
4510 | /* If op0 is an external or constant def use a vector type with | |
4511 | the same size as the output vector type. */ | |
4512 | if (!vectype) | |
4513 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
4514 | if (vec_stmt) | |
4515 | gcc_assert (vectype); | |
4516 | if (!vectype) | |
4517 | { | |
73fbfcad | 4518 | if (dump_enabled_p ()) |
78c60e3d | 4519 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4520 | "no vectype for scalar type\n"); |
9dc3f7de IR |
4521 | return false; |
4522 | } | |
4523 | ||
4524 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4525 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4526 | if (nunits_out != nunits_in) | |
4527 | return false; | |
4528 | ||
4529 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 4530 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &op1_vectype)) |
9dc3f7de | 4531 | { |
73fbfcad | 4532 | if (dump_enabled_p ()) |
78c60e3d | 4533 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4534 | "use not simple.\n"); |
9dc3f7de IR |
4535 | return false; |
4536 | } | |
4537 | ||
4538 | if (loop_vinfo) | |
4539 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4540 | else | |
4541 | vf = 1; | |
4542 | ||
4543 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4544 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4545 | case of SLP. */ | |
437f4a00 | 4546 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
9dc3f7de IR |
4547 | ncopies = 1; |
4548 | else | |
4549 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
4550 | ||
4551 | gcc_assert (ncopies >= 1); | |
4552 | ||
4553 | /* Determine whether the shift amount is a vector, or scalar. If the | |
4554 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
4555 | ||
dbfa87aa YR |
4556 | if ((dt[1] == vect_internal_def |
4557 | || dt[1] == vect_induction_def) | |
4558 | && !slp_node) | |
49eab32e JJ |
4559 | scalar_shift_arg = false; |
4560 | else if (dt[1] == vect_constant_def | |
4561 | || dt[1] == vect_external_def | |
4562 | || dt[1] == vect_internal_def) | |
4563 | { | |
4564 | /* In SLP, need to check whether the shift count is the same, | |
4565 | in loops if it is a constant or invariant, it is always | |
4566 | a scalar shift. */ | |
4567 | if (slp_node) | |
4568 | { | |
355fe088 TS |
4569 | vec<gimple *> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
4570 | gimple *slpstmt; | |
49eab32e | 4571 | |
9771b263 | 4572 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
4573 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
4574 | scalar_shift_arg = false; | |
4575 | } | |
4576 | } | |
4577 | else | |
4578 | { | |
73fbfcad | 4579 | if (dump_enabled_p ()) |
78c60e3d | 4580 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4581 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
4582 | return false; |
4583 | } | |
4584 | ||
9dc3f7de | 4585 | /* Vector shifted by vector. */ |
49eab32e | 4586 | if (!scalar_shift_arg) |
9dc3f7de IR |
4587 | { |
4588 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 4589 | if (dump_enabled_p ()) |
78c60e3d | 4590 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4591 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 4592 | |
aa948027 JJ |
4593 | if (!op1_vectype) |
4594 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
4595 | if (op1_vectype == NULL_TREE | |
4596 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 4597 | { |
73fbfcad | 4598 | if (dump_enabled_p ()) |
78c60e3d SS |
4599 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4600 | "unusable type for last operand in" | |
e645e942 | 4601 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
4602 | return false; |
4603 | } | |
9dc3f7de IR |
4604 | } |
4605 | /* See if the machine has a vector shifted by scalar insn and if not | |
4606 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 4607 | else |
9dc3f7de IR |
4608 | { |
4609 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4610 | if (optab | |
4611 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
4612 | { | |
73fbfcad | 4613 | if (dump_enabled_p ()) |
78c60e3d | 4614 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4615 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
4616 | } |
4617 | else | |
4618 | { | |
4619 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4620 | if (optab | |
4621 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
4622 | != CODE_FOR_nothing)) | |
4623 | { | |
49eab32e JJ |
4624 | scalar_shift_arg = false; |
4625 | ||
73fbfcad | 4626 | if (dump_enabled_p ()) |
78c60e3d | 4627 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4628 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
4629 | |
4630 | /* Unlike the other binary operators, shifts/rotates have | |
4631 | the rhs being int, instead of the same type as the lhs, | |
4632 | so make sure the scalar is the right type if we are | |
aa948027 | 4633 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
4634 | if (dt[1] == vect_constant_def) |
4635 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
4636 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
4637 | TREE_TYPE (op1))) | |
4638 | { | |
4639 | if (slp_node | |
4640 | && TYPE_MODE (TREE_TYPE (vectype)) | |
4641 | != TYPE_MODE (TREE_TYPE (op1))) | |
4642 | { | |
73fbfcad | 4643 | if (dump_enabled_p ()) |
78c60e3d SS |
4644 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4645 | "unusable type for last operand in" | |
e645e942 | 4646 | " vector/vector shift/rotate.\n"); |
21c0a521 | 4647 | return false; |
aa948027 JJ |
4648 | } |
4649 | if (vec_stmt && !slp_node) | |
4650 | { | |
4651 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
4652 | op1 = vect_init_vector (stmt, op1, | |
4653 | TREE_TYPE (vectype), NULL); | |
4654 | } | |
4655 | } | |
9dc3f7de IR |
4656 | } |
4657 | } | |
4658 | } | |
9dc3f7de IR |
4659 | |
4660 | /* Supportable by target? */ | |
4661 | if (!optab) | |
4662 | { | |
73fbfcad | 4663 | if (dump_enabled_p ()) |
78c60e3d | 4664 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4665 | "no optab.\n"); |
9dc3f7de IR |
4666 | return false; |
4667 | } | |
4668 | vec_mode = TYPE_MODE (vectype); | |
4669 | icode = (int) optab_handler (optab, vec_mode); | |
4670 | if (icode == CODE_FOR_nothing) | |
4671 | { | |
73fbfcad | 4672 | if (dump_enabled_p ()) |
78c60e3d | 4673 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4674 | "op not supported by target.\n"); |
9dc3f7de IR |
4675 | /* Check only during analysis. */ |
4676 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
4677 | || (vf < vect_min_worthwhile_factor (code) | |
4678 | && !vec_stmt)) | |
4679 | return false; | |
73fbfcad | 4680 | if (dump_enabled_p ()) |
e645e942 TJ |
4681 | dump_printf_loc (MSG_NOTE, vect_location, |
4682 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
4683 | } |
4684 | ||
4685 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
4686 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
4687 | && vf < vect_min_worthwhile_factor (code) | |
4688 | && !vec_stmt) | |
4689 | { | |
73fbfcad | 4690 | if (dump_enabled_p ()) |
78c60e3d | 4691 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4692 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
4693 | return false; |
4694 | } | |
4695 | ||
4696 | if (!vec_stmt) /* transformation not required. */ | |
4697 | { | |
4698 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 4699 | if (dump_enabled_p ()) |
e645e942 TJ |
4700 | dump_printf_loc (MSG_NOTE, vect_location, |
4701 | "=== vectorizable_shift ===\n"); | |
c3e7ee41 | 4702 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
9dc3f7de IR |
4703 | return true; |
4704 | } | |
4705 | ||
4706 | /** Transform. **/ | |
4707 | ||
73fbfcad | 4708 | if (dump_enabled_p ()) |
78c60e3d | 4709 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4710 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
4711 | |
4712 | /* Handle def. */ | |
4713 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4714 | ||
9dc3f7de IR |
4715 | prev_stmt_info = NULL; |
4716 | for (j = 0; j < ncopies; j++) | |
4717 | { | |
4718 | /* Handle uses. */ | |
4719 | if (j == 0) | |
4720 | { | |
4721 | if (scalar_shift_arg) | |
4722 | { | |
4723 | /* Vector shl and shr insn patterns can be defined with scalar | |
4724 | operand 2 (shift operand). In this case, use constant or loop | |
4725 | invariant op1 directly, without extending it to vector mode | |
4726 | first. */ | |
4727 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
4728 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
4729 | { | |
73fbfcad | 4730 | if (dump_enabled_p ()) |
78c60e3d | 4731 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4732 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 4733 | vec_oprnd1 = op1; |
8930f723 | 4734 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 4735 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
4736 | if (slp_node) |
4737 | { | |
4738 | /* Store vec_oprnd1 for every vector stmt to be created | |
4739 | for SLP_NODE. We check during the analysis that all | |
4740 | the shift arguments are the same. | |
4741 | TODO: Allow different constants for different vector | |
4742 | stmts generated for an SLP instance. */ | |
4743 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4744 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
4745 | } |
4746 | } | |
4747 | } | |
4748 | ||
4749 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
4750 | (a special case for certain kind of vector shifts); otherwise, | |
4751 | operand 1 should be of a vector type (the usual case). */ | |
4752 | if (vec_oprnd1) | |
4753 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
d092494c | 4754 | slp_node, -1); |
9dc3f7de IR |
4755 | else |
4756 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
d092494c | 4757 | slp_node, -1); |
9dc3f7de IR |
4758 | } |
4759 | else | |
4760 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
4761 | ||
4762 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 4763 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 4764 | { |
9771b263 | 4765 | vop1 = vec_oprnds1[i]; |
0d0e4a03 | 4766 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
9dc3f7de IR |
4767 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4768 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4769 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4770 | if (slp_node) | |
9771b263 | 4771 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
4772 | } |
4773 | ||
4774 | if (slp_node) | |
4775 | continue; | |
4776 | ||
4777 | if (j == 0) | |
4778 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4779 | else | |
4780 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4781 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4782 | } | |
4783 | ||
9771b263 DN |
4784 | vec_oprnds0.release (); |
4785 | vec_oprnds1.release (); | |
9dc3f7de IR |
4786 | |
4787 | return true; | |
4788 | } | |
4789 | ||
4790 | ||
ebfd146a IR |
4791 | /* Function vectorizable_operation. |
4792 | ||
16949072 RG |
4793 | Check if STMT performs a binary, unary or ternary operation that can |
4794 | be vectorized. | |
b8698a0f | 4795 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
4796 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4797 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4798 | ||
4799 | static bool | |
355fe088 TS |
4800 | vectorizable_operation (gimple *stmt, gimple_stmt_iterator *gsi, |
4801 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a | 4802 | { |
00f07b86 | 4803 | tree vec_dest; |
ebfd146a | 4804 | tree scalar_dest; |
16949072 | 4805 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 4806 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 4807 | tree vectype; |
ebfd146a IR |
4808 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4809 | enum tree_code code; | |
ef4bddc2 | 4810 | machine_mode vec_mode; |
ebfd146a IR |
4811 | tree new_temp; |
4812 | int op_type; | |
00f07b86 | 4813 | optab optab; |
523ba738 | 4814 | bool target_support_p; |
355fe088 | 4815 | gimple *def_stmt; |
16949072 RG |
4816 | enum vect_def_type dt[3] |
4817 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
355fe088 | 4818 | gimple *new_stmt = NULL; |
ebfd146a | 4819 | stmt_vec_info prev_stmt_info; |
b690cc0f | 4820 | int nunits_in; |
ebfd146a IR |
4821 | int nunits_out; |
4822 | tree vectype_out; | |
4823 | int ncopies; | |
4824 | int j, i; | |
6e1aa848 DN |
4825 | vec<tree> vec_oprnds0 = vNULL; |
4826 | vec<tree> vec_oprnds1 = vNULL; | |
4827 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 4828 | tree vop0, vop1, vop2; |
a70d6342 | 4829 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4830 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 IR |
4831 | int vf; |
4832 | ||
a70d6342 | 4833 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4834 | return false; |
4835 | ||
66c16fd9 RB |
4836 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4837 | && ! vec_stmt) | |
ebfd146a IR |
4838 | return false; |
4839 | ||
4840 | /* Is STMT a vectorizable binary/unary operation? */ | |
4841 | if (!is_gimple_assign (stmt)) | |
4842 | return false; | |
4843 | ||
4844 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4845 | return false; | |
4846 | ||
ebfd146a IR |
4847 | code = gimple_assign_rhs_code (stmt); |
4848 | ||
4849 | /* For pointer addition, we should use the normal plus for | |
4850 | the vector addition. */ | |
4851 | if (code == POINTER_PLUS_EXPR) | |
4852 | code = PLUS_EXPR; | |
4853 | ||
4854 | /* Support only unary or binary operations. */ | |
4855 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 4856 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 4857 | { |
73fbfcad | 4858 | if (dump_enabled_p ()) |
78c60e3d | 4859 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4860 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 4861 | op_type); |
ebfd146a IR |
4862 | return false; |
4863 | } | |
4864 | ||
b690cc0f RG |
4865 | scalar_dest = gimple_assign_lhs (stmt); |
4866 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
4867 | ||
7b7b1813 RG |
4868 | /* Most operations cannot handle bit-precision types without extra |
4869 | truncations. */ | |
045c1278 IE |
4870 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4871 | && (TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4872 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
7b7b1813 RG |
4873 | /* Exception are bitwise binary operations. */ |
4874 | && code != BIT_IOR_EXPR | |
4875 | && code != BIT_XOR_EXPR | |
4876 | && code != BIT_AND_EXPR) | |
4877 | { | |
73fbfcad | 4878 | if (dump_enabled_p ()) |
78c60e3d | 4879 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4880 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
4881 | return false; |
4882 | } | |
4883 | ||
ebfd146a | 4884 | op0 = gimple_assign_rhs1 (stmt); |
81c40241 | 4885 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
ebfd146a | 4886 | { |
73fbfcad | 4887 | if (dump_enabled_p ()) |
78c60e3d | 4888 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4889 | "use not simple.\n"); |
ebfd146a IR |
4890 | return false; |
4891 | } | |
b690cc0f RG |
4892 | /* If op0 is an external or constant def use a vector type with |
4893 | the same size as the output vector type. */ | |
4894 | if (!vectype) | |
b036c6c5 IE |
4895 | { |
4896 | /* For boolean type we cannot determine vectype by | |
4897 | invariant value (don't know whether it is a vector | |
4898 | of booleans or vector of integers). We use output | |
4899 | vectype because operations on boolean don't change | |
4900 | type. */ | |
4901 | if (TREE_CODE (TREE_TYPE (op0)) == BOOLEAN_TYPE) | |
4902 | { | |
4903 | if (TREE_CODE (TREE_TYPE (scalar_dest)) != BOOLEAN_TYPE) | |
4904 | { | |
4905 | if (dump_enabled_p ()) | |
4906 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
4907 | "not supported operation on bool value.\n"); | |
4908 | return false; | |
4909 | } | |
4910 | vectype = vectype_out; | |
4911 | } | |
4912 | else | |
4913 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
4914 | } | |
7d8930a0 IR |
4915 | if (vec_stmt) |
4916 | gcc_assert (vectype); | |
4917 | if (!vectype) | |
4918 | { | |
73fbfcad | 4919 | if (dump_enabled_p ()) |
7d8930a0 | 4920 | { |
78c60e3d SS |
4921 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4922 | "no vectype for scalar type "); | |
4923 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
4924 | TREE_TYPE (op0)); | |
e645e942 | 4925 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
4926 | } |
4927 | ||
4928 | return false; | |
4929 | } | |
b690cc0f RG |
4930 | |
4931 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4932 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4933 | if (nunits_out != nunits_in) | |
4934 | return false; | |
ebfd146a | 4935 | |
16949072 | 4936 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
4937 | { |
4938 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 4939 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1])) |
ebfd146a | 4940 | { |
73fbfcad | 4941 | if (dump_enabled_p ()) |
78c60e3d | 4942 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4943 | "use not simple.\n"); |
ebfd146a IR |
4944 | return false; |
4945 | } | |
4946 | } | |
16949072 RG |
4947 | if (op_type == ternary_op) |
4948 | { | |
4949 | op2 = gimple_assign_rhs3 (stmt); | |
81c40241 | 4950 | if (!vect_is_simple_use (op2, vinfo, &def_stmt, &dt[2])) |
16949072 | 4951 | { |
73fbfcad | 4952 | if (dump_enabled_p ()) |
78c60e3d | 4953 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4954 | "use not simple.\n"); |
16949072 RG |
4955 | return false; |
4956 | } | |
4957 | } | |
ebfd146a | 4958 | |
b690cc0f RG |
4959 | if (loop_vinfo) |
4960 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4961 | else | |
4962 | vf = 1; | |
4963 | ||
4964 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 4965 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 4966 | case of SLP. */ |
437f4a00 | 4967 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
b690cc0f RG |
4968 | ncopies = 1; |
4969 | else | |
4970 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
4971 | ||
4972 | gcc_assert (ncopies >= 1); | |
4973 | ||
9dc3f7de | 4974 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
4975 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
4976 | || code == RROTATE_EXPR) | |
9dc3f7de | 4977 | return false; |
ebfd146a | 4978 | |
ebfd146a | 4979 | /* Supportable by target? */ |
00f07b86 RH |
4980 | |
4981 | vec_mode = TYPE_MODE (vectype); | |
4982 | if (code == MULT_HIGHPART_EXPR) | |
523ba738 | 4983 | target_support_p = can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype)); |
00f07b86 RH |
4984 | else |
4985 | { | |
4986 | optab = optab_for_tree_code (code, vectype, optab_default); | |
4987 | if (!optab) | |
5deb57cb | 4988 | { |
73fbfcad | 4989 | if (dump_enabled_p ()) |
78c60e3d | 4990 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4991 | "no optab.\n"); |
00f07b86 | 4992 | return false; |
5deb57cb | 4993 | } |
523ba738 RS |
4994 | target_support_p = (optab_handler (optab, vec_mode) |
4995 | != CODE_FOR_nothing); | |
5deb57cb JJ |
4996 | } |
4997 | ||
523ba738 | 4998 | if (!target_support_p) |
ebfd146a | 4999 | { |
73fbfcad | 5000 | if (dump_enabled_p ()) |
78c60e3d | 5001 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5002 | "op not supported by target.\n"); |
ebfd146a IR |
5003 | /* Check only during analysis. */ |
5004 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
5deb57cb | 5005 | || (!vec_stmt && vf < vect_min_worthwhile_factor (code))) |
ebfd146a | 5006 | return false; |
73fbfcad | 5007 | if (dump_enabled_p ()) |
e645e942 TJ |
5008 | dump_printf_loc (MSG_NOTE, vect_location, |
5009 | "proceeding using word mode.\n"); | |
383d9c83 IR |
5010 | } |
5011 | ||
4a00c761 | 5012 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
5013 | if (!VECTOR_MODE_P (vec_mode) |
5014 | && !vec_stmt | |
5015 | && vf < vect_min_worthwhile_factor (code)) | |
7d8930a0 | 5016 | { |
73fbfcad | 5017 | if (dump_enabled_p ()) |
78c60e3d | 5018 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5019 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 5020 | return false; |
7d8930a0 | 5021 | } |
ebfd146a | 5022 | |
ebfd146a IR |
5023 | if (!vec_stmt) /* transformation not required. */ |
5024 | { | |
4a00c761 | 5025 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 5026 | if (dump_enabled_p ()) |
78c60e3d | 5027 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5028 | "=== vectorizable_operation ===\n"); |
c3e7ee41 | 5029 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
5030 | return true; |
5031 | } | |
5032 | ||
5033 | /** Transform. **/ | |
5034 | ||
73fbfcad | 5035 | if (dump_enabled_p ()) |
78c60e3d | 5036 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5037 | "transform binary/unary operation.\n"); |
383d9c83 | 5038 | |
ebfd146a | 5039 | /* Handle def. */ |
00f07b86 | 5040 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 5041 | |
ebfd146a IR |
5042 | /* In case the vectorization factor (VF) is bigger than the number |
5043 | of elements that we can fit in a vectype (nunits), we have to generate | |
5044 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
5045 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
5046 | from one copy of the vector stmt to the next, in the field | |
5047 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
5048 | stages to find the correct vector defs to be used when vectorizing | |
5049 | stmts that use the defs of the current stmt. The example below | |
5050 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
5051 | we need to create 4 vectorized stmts): | |
5052 | ||
5053 | before vectorization: | |
5054 | RELATED_STMT VEC_STMT | |
5055 | S1: x = memref - - | |
5056 | S2: z = x + 1 - - | |
5057 | ||
5058 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
5059 | there): | |
5060 | RELATED_STMT VEC_STMT | |
5061 | VS1_0: vx0 = memref0 VS1_1 - | |
5062 | VS1_1: vx1 = memref1 VS1_2 - | |
5063 | VS1_2: vx2 = memref2 VS1_3 - | |
5064 | VS1_3: vx3 = memref3 - - | |
5065 | S1: x = load - VS1_0 | |
5066 | S2: z = x + 1 - - | |
5067 | ||
5068 | step2: vectorize stmt S2 (done here): | |
5069 | To vectorize stmt S2 we first need to find the relevant vector | |
5070 | def for the first operand 'x'. This is, as usual, obtained from | |
5071 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
5072 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
5073 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
5074 | the vector stmt VS2_0, and as usual, record it in the | |
5075 | STMT_VINFO_VEC_STMT of stmt S2. | |
5076 | When creating the second copy (VS2_1), we obtain the relevant vector | |
5077 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
5078 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
5079 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
5080 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
5081 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
5082 | chain of stmts and pointers: | |
5083 | RELATED_STMT VEC_STMT | |
5084 | VS1_0: vx0 = memref0 VS1_1 - | |
5085 | VS1_1: vx1 = memref1 VS1_2 - | |
5086 | VS1_2: vx2 = memref2 VS1_3 - | |
5087 | VS1_3: vx3 = memref3 - - | |
5088 | S1: x = load - VS1_0 | |
5089 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
5090 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
5091 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
5092 | VS2_3: vz3 = vx3 + v1 - - | |
5093 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
5094 | |
5095 | prev_stmt_info = NULL; | |
5096 | for (j = 0; j < ncopies; j++) | |
5097 | { | |
5098 | /* Handle uses. */ | |
5099 | if (j == 0) | |
4a00c761 JJ |
5100 | { |
5101 | if (op_type == binary_op || op_type == ternary_op) | |
5102 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
5103 | slp_node, -1); | |
5104 | else | |
5105 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
5106 | slp_node, -1); | |
5107 | if (op_type == ternary_op) | |
36ba4aae | 5108 | { |
9771b263 DN |
5109 | vec_oprnds2.create (1); |
5110 | vec_oprnds2.quick_push (vect_get_vec_def_for_operand (op2, | |
81c40241 | 5111 | stmt)); |
36ba4aae | 5112 | } |
4a00c761 | 5113 | } |
ebfd146a | 5114 | else |
4a00c761 JJ |
5115 | { |
5116 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5117 | if (op_type == ternary_op) | |
5118 | { | |
9771b263 DN |
5119 | tree vec_oprnd = vec_oprnds2.pop (); |
5120 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
5121 | vec_oprnd)); | |
4a00c761 JJ |
5122 | } |
5123 | } | |
5124 | ||
5125 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 5126 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 5127 | { |
4a00c761 | 5128 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 5129 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 5130 | vop2 = ((op_type == ternary_op) |
9771b263 | 5131 | ? vec_oprnds2[i] : NULL_TREE); |
0d0e4a03 | 5132 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1, vop2); |
4a00c761 JJ |
5133 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5134 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5135 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5136 | if (slp_node) | |
9771b263 | 5137 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
5138 | } |
5139 | ||
4a00c761 JJ |
5140 | if (slp_node) |
5141 | continue; | |
5142 | ||
5143 | if (j == 0) | |
5144 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5145 | else | |
5146 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5147 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
5148 | } |
5149 | ||
9771b263 DN |
5150 | vec_oprnds0.release (); |
5151 | vec_oprnds1.release (); | |
5152 | vec_oprnds2.release (); | |
ebfd146a | 5153 | |
ebfd146a IR |
5154 | return true; |
5155 | } | |
5156 | ||
c716e67f XDL |
5157 | /* A helper function to ensure data reference DR's base alignment |
5158 | for STMT_INFO. */ | |
5159 | ||
5160 | static void | |
5161 | ensure_base_align (stmt_vec_info stmt_info, struct data_reference *dr) | |
5162 | { | |
5163 | if (!dr->aux) | |
5164 | return; | |
5165 | ||
52639a61 | 5166 | if (DR_VECT_AUX (dr)->base_misaligned) |
c716e67f XDL |
5167 | { |
5168 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
52639a61 | 5169 | tree base_decl = DR_VECT_AUX (dr)->base_decl; |
c716e67f | 5170 | |
428f0c67 JH |
5171 | if (decl_in_symtab_p (base_decl)) |
5172 | symtab_node::get (base_decl)->increase_alignment (TYPE_ALIGN (vectype)); | |
5173 | else | |
5174 | { | |
5175 | DECL_ALIGN (base_decl) = TYPE_ALIGN (vectype); | |
5176 | DECL_USER_ALIGN (base_decl) = 1; | |
5177 | } | |
52639a61 | 5178 | DR_VECT_AUX (dr)->base_misaligned = false; |
c716e67f XDL |
5179 | } |
5180 | } | |
5181 | ||
ebfd146a | 5182 | |
09dfa495 BM |
5183 | /* Given a vector type VECTYPE returns the VECTOR_CST mask that implements |
5184 | reversal of the vector elements. If that is impossible to do, | |
5185 | returns NULL. */ | |
5186 | ||
5187 | static tree | |
5188 | perm_mask_for_reverse (tree vectype) | |
5189 | { | |
5190 | int i, nunits; | |
5191 | unsigned char *sel; | |
5192 | ||
5193 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
5194 | sel = XALLOCAVEC (unsigned char, nunits); | |
5195 | ||
5196 | for (i = 0; i < nunits; ++i) | |
5197 | sel[i] = nunits - 1 - i; | |
5198 | ||
557be5a8 AL |
5199 | if (!can_vec_perm_p (TYPE_MODE (vectype), false, sel)) |
5200 | return NULL_TREE; | |
5201 | return vect_gen_perm_mask_checked (vectype, sel); | |
09dfa495 BM |
5202 | } |
5203 | ||
ebfd146a IR |
5204 | /* Function vectorizable_store. |
5205 | ||
b8698a0f L |
5206 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
5207 | can be vectorized. | |
5208 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
5209 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5210 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5211 | ||
5212 | static bool | |
355fe088 | 5213 | vectorizable_store (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 5214 | slp_tree slp_node) |
ebfd146a IR |
5215 | { |
5216 | tree scalar_dest; | |
5217 | tree data_ref; | |
5218 | tree op; | |
5219 | tree vec_oprnd = NULL_TREE; | |
5220 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5221 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
272c6793 | 5222 | tree elem_type; |
ebfd146a | 5223 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 5224 | struct loop *loop = NULL; |
ef4bddc2 | 5225 | machine_mode vec_mode; |
ebfd146a IR |
5226 | tree dummy; |
5227 | enum dr_alignment_support alignment_support_scheme; | |
355fe088 | 5228 | gimple *def_stmt; |
ebfd146a IR |
5229 | enum vect_def_type dt; |
5230 | stmt_vec_info prev_stmt_info = NULL; | |
5231 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 5232 | tree dataref_offset = NULL_TREE; |
355fe088 | 5233 | gimple *ptr_incr = NULL; |
ebfd146a IR |
5234 | int ncopies; |
5235 | int j; | |
355fe088 | 5236 | gimple *next_stmt, *first_stmt = NULL; |
0d0293ac | 5237 | bool grouped_store = false; |
272c6793 | 5238 | bool store_lanes_p = false; |
ebfd146a | 5239 | unsigned int group_size, i; |
6e1aa848 DN |
5240 | vec<tree> dr_chain = vNULL; |
5241 | vec<tree> oprnds = vNULL; | |
5242 | vec<tree> result_chain = vNULL; | |
ebfd146a | 5243 | bool inv_p; |
09dfa495 BM |
5244 | bool negative = false; |
5245 | tree offset = NULL_TREE; | |
6e1aa848 | 5246 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 5247 | bool slp = (slp_node != NULL); |
ebfd146a | 5248 | unsigned int vec_num; |
a70d6342 | 5249 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5250 | vec_info *vinfo = stmt_info->vinfo; |
272c6793 | 5251 | tree aggr_type; |
3bab6342 AT |
5252 | tree scatter_base = NULL_TREE, scatter_off = NULL_TREE; |
5253 | tree scatter_off_vectype = NULL_TREE, scatter_decl = NULL_TREE; | |
5254 | int scatter_scale = 1; | |
5255 | enum vect_def_type scatter_idx_dt = vect_unknown_def_type; | |
5256 | enum vect_def_type scatter_src_dt = vect_unknown_def_type; | |
355fe088 | 5257 | gimple *new_stmt; |
a70d6342 | 5258 | |
a70d6342 | 5259 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5260 | return false; |
5261 | ||
66c16fd9 RB |
5262 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5263 | && ! vec_stmt) | |
ebfd146a IR |
5264 | return false; |
5265 | ||
5266 | /* Is vectorizable store? */ | |
5267 | ||
5268 | if (!is_gimple_assign (stmt)) | |
5269 | return false; | |
5270 | ||
5271 | scalar_dest = gimple_assign_lhs (stmt); | |
ab0ef706 JJ |
5272 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
5273 | && is_pattern_stmt_p (stmt_info)) | |
5274 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
ebfd146a | 5275 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
38000232 | 5276 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF |
ebfd146a | 5277 | && TREE_CODE (scalar_dest) != INDIRECT_REF |
e9dbe7bb IR |
5278 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
5279 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
5280 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
5281 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
5282 | return false; |
5283 | ||
5284 | gcc_assert (gimple_assign_single_p (stmt)); | |
465c8c19 | 5285 | |
f4d09712 | 5286 | tree vectype = STMT_VINFO_VECTYPE (stmt_info), rhs_vectype = NULL_TREE; |
465c8c19 JJ |
5287 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
5288 | ||
5289 | if (loop_vinfo) | |
5290 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
5291 | ||
5292 | /* Multiple types in SLP are handled by creating the appropriate number of | |
5293 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5294 | case of SLP. */ | |
5295 | if (slp || PURE_SLP_STMT (stmt_info)) | |
5296 | ncopies = 1; | |
5297 | else | |
5298 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
5299 | ||
5300 | gcc_assert (ncopies >= 1); | |
5301 | ||
5302 | /* FORNOW. This restriction should be relaxed. */ | |
5303 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) | |
5304 | { | |
5305 | if (dump_enabled_p ()) | |
5306 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5307 | "multiple types in nested loop.\n"); | |
5308 | return false; | |
5309 | } | |
5310 | ||
ebfd146a | 5311 | op = gimple_assign_rhs1 (stmt); |
f4d09712 KY |
5312 | |
5313 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt, &rhs_vectype)) | |
ebfd146a | 5314 | { |
73fbfcad | 5315 | if (dump_enabled_p ()) |
78c60e3d | 5316 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5317 | "use not simple.\n"); |
ebfd146a IR |
5318 | return false; |
5319 | } | |
5320 | ||
f4d09712 KY |
5321 | if (rhs_vectype && !useless_type_conversion_p (vectype, rhs_vectype)) |
5322 | return false; | |
5323 | ||
272c6793 | 5324 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 5325 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 5326 | |
ebfd146a IR |
5327 | /* FORNOW. In some cases can vectorize even if data-type not supported |
5328 | (e.g. - array initialization with 0). */ | |
947131ba | 5329 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
5330 | return false; |
5331 | ||
5332 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
5333 | return false; | |
5334 | ||
f2e2a985 | 5335 | if (!STMT_VINFO_STRIDED_P (stmt_info)) |
09dfa495 | 5336 | { |
f2e2a985 MM |
5337 | negative = |
5338 | tree_int_cst_compare (loop && nested_in_vect_loop_p (loop, stmt) | |
5339 | ? STMT_VINFO_DR_STEP (stmt_info) : DR_STEP (dr), | |
5340 | size_zero_node) < 0; | |
5341 | if (negative && ncopies > 1) | |
09dfa495 BM |
5342 | { |
5343 | if (dump_enabled_p ()) | |
5344 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
f2e2a985 | 5345 | "multiple types with negative step.\n"); |
09dfa495 BM |
5346 | return false; |
5347 | } | |
f2e2a985 | 5348 | if (negative) |
09dfa495 | 5349 | { |
f2e2a985 MM |
5350 | gcc_assert (!grouped_store); |
5351 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
5352 | if (alignment_support_scheme != dr_aligned | |
5353 | && alignment_support_scheme != dr_unaligned_supported) | |
5354 | { | |
5355 | if (dump_enabled_p ()) | |
5356 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5357 | "negative step but alignment required.\n"); | |
5358 | return false; | |
5359 | } | |
5360 | if (dt != vect_constant_def | |
5361 | && dt != vect_external_def | |
5362 | && !perm_mask_for_reverse (vectype)) | |
5363 | { | |
5364 | if (dump_enabled_p ()) | |
5365 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5366 | "negative step and reversing not supported.\n"); | |
5367 | return false; | |
5368 | } | |
09dfa495 BM |
5369 | } |
5370 | } | |
5371 | ||
0d0293ac | 5372 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 5373 | { |
0d0293ac | 5374 | grouped_store = true; |
e14c1050 | 5375 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
cee62fee MM |
5376 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
5377 | if (!slp | |
5378 | && !PURE_SLP_STMT (stmt_info) | |
5379 | && !STMT_VINFO_STRIDED_P (stmt_info)) | |
b602d918 | 5380 | { |
272c6793 RS |
5381 | if (vect_store_lanes_supported (vectype, group_size)) |
5382 | store_lanes_p = true; | |
0d0293ac | 5383 | else if (!vect_grouped_store_supported (vectype, group_size)) |
b602d918 RS |
5384 | return false; |
5385 | } | |
b8698a0f | 5386 | |
cee62fee MM |
5387 | if (STMT_VINFO_STRIDED_P (stmt_info) |
5388 | && (slp || PURE_SLP_STMT (stmt_info)) | |
5389 | && (group_size > nunits | |
5390 | || nunits % group_size != 0)) | |
5391 | { | |
5392 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5393 | "unhandled strided group store\n"); | |
5394 | return false; | |
5395 | } | |
5396 | ||
ebfd146a IR |
5397 | if (first_stmt == stmt) |
5398 | { | |
5399 | /* STMT is the leader of the group. Check the operands of all the | |
5400 | stmts of the group. */ | |
e14c1050 | 5401 | next_stmt = GROUP_NEXT_ELEMENT (stmt_info); |
ebfd146a IR |
5402 | while (next_stmt) |
5403 | { | |
5404 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
5405 | op = gimple_assign_rhs1 (next_stmt); | |
81c40241 | 5406 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt)) |
ebfd146a | 5407 | { |
73fbfcad | 5408 | if (dump_enabled_p ()) |
78c60e3d | 5409 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5410 | "use not simple.\n"); |
ebfd146a IR |
5411 | return false; |
5412 | } | |
e14c1050 | 5413 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
5414 | } |
5415 | } | |
5416 | } | |
5417 | ||
3bab6342 AT |
5418 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5419 | { | |
355fe088 | 5420 | gimple *def_stmt; |
3bab6342 AT |
5421 | scatter_decl = vect_check_gather_scatter (stmt, loop_vinfo, &scatter_base, |
5422 | &scatter_off, &scatter_scale); | |
5423 | gcc_assert (scatter_decl); | |
81c40241 RB |
5424 | if (!vect_is_simple_use (scatter_off, vinfo, &def_stmt, &scatter_idx_dt, |
5425 | &scatter_off_vectype)) | |
3bab6342 AT |
5426 | { |
5427 | if (dump_enabled_p ()) | |
5428 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5429 | "scatter index use not simple."); | |
5430 | return false; | |
5431 | } | |
5432 | } | |
5433 | ||
ebfd146a IR |
5434 | if (!vec_stmt) /* transformation not required. */ |
5435 | { | |
5436 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
2e8ab70c RB |
5437 | /* The SLP costs are calculated during SLP analysis. */ |
5438 | if (!PURE_SLP_STMT (stmt_info)) | |
5439 | vect_model_store_cost (stmt_info, ncopies, store_lanes_p, dt, | |
5440 | NULL, NULL, NULL); | |
ebfd146a IR |
5441 | return true; |
5442 | } | |
5443 | ||
5444 | /** Transform. **/ | |
5445 | ||
c716e67f XDL |
5446 | ensure_base_align (stmt_info, dr); |
5447 | ||
3bab6342 AT |
5448 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5449 | { | |
5450 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE, op, src; | |
5451 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (scatter_decl)); | |
5452 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
5453 | tree ptr, mask, var, scale, perm_mask = NULL_TREE; | |
5454 | edge pe = loop_preheader_edge (loop); | |
5455 | gimple_seq seq; | |
5456 | basic_block new_bb; | |
5457 | enum { NARROW, NONE, WIDEN } modifier; | |
5458 | int scatter_off_nunits = TYPE_VECTOR_SUBPARTS (scatter_off_vectype); | |
5459 | ||
5460 | if (nunits == (unsigned int) scatter_off_nunits) | |
5461 | modifier = NONE; | |
5462 | else if (nunits == (unsigned int) scatter_off_nunits / 2) | |
5463 | { | |
5464 | unsigned char *sel = XALLOCAVEC (unsigned char, scatter_off_nunits); | |
5465 | modifier = WIDEN; | |
5466 | ||
5467 | for (i = 0; i < (unsigned int) scatter_off_nunits; ++i) | |
5468 | sel[i] = i | nunits; | |
5469 | ||
5470 | perm_mask = vect_gen_perm_mask_checked (scatter_off_vectype, sel); | |
5471 | gcc_assert (perm_mask != NULL_TREE); | |
5472 | } | |
5473 | else if (nunits == (unsigned int) scatter_off_nunits * 2) | |
5474 | { | |
5475 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
5476 | modifier = NARROW; | |
5477 | ||
5478 | for (i = 0; i < (unsigned int) nunits; ++i) | |
5479 | sel[i] = i | scatter_off_nunits; | |
5480 | ||
5481 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); | |
5482 | gcc_assert (perm_mask != NULL_TREE); | |
5483 | ncopies *= 2; | |
5484 | } | |
5485 | else | |
5486 | gcc_unreachable (); | |
5487 | ||
5488 | rettype = TREE_TYPE (TREE_TYPE (scatter_decl)); | |
5489 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5490 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5491 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5492 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5493 | scaletype = TREE_VALUE (arglist); | |
5494 | ||
5495 | gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE | |
5496 | && TREE_CODE (rettype) == VOID_TYPE); | |
5497 | ||
5498 | ptr = fold_convert (ptrtype, scatter_base); | |
5499 | if (!is_gimple_min_invariant (ptr)) | |
5500 | { | |
5501 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
5502 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
5503 | gcc_assert (!new_bb); | |
5504 | } | |
5505 | ||
5506 | /* Currently we support only unconditional scatter stores, | |
5507 | so mask should be all ones. */ | |
5508 | mask = build_int_cst (masktype, -1); | |
5509 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
5510 | ||
5511 | scale = build_int_cst (scaletype, scatter_scale); | |
5512 | ||
5513 | prev_stmt_info = NULL; | |
5514 | for (j = 0; j < ncopies; ++j) | |
5515 | { | |
5516 | if (j == 0) | |
5517 | { | |
5518 | src = vec_oprnd1 | |
81c40241 | 5519 | = vect_get_vec_def_for_operand (gimple_assign_rhs1 (stmt), stmt); |
3bab6342 | 5520 | op = vec_oprnd0 |
81c40241 | 5521 | = vect_get_vec_def_for_operand (scatter_off, stmt); |
3bab6342 AT |
5522 | } |
5523 | else if (modifier != NONE && (j & 1)) | |
5524 | { | |
5525 | if (modifier == WIDEN) | |
5526 | { | |
5527 | src = vec_oprnd1 | |
5528 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5529 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, perm_mask, | |
5530 | stmt, gsi); | |
5531 | } | |
5532 | else if (modifier == NARROW) | |
5533 | { | |
5534 | src = permute_vec_elements (vec_oprnd1, vec_oprnd1, perm_mask, | |
5535 | stmt, gsi); | |
5536 | op = vec_oprnd0 | |
5537 | = vect_get_vec_def_for_stmt_copy (scatter_idx_dt, vec_oprnd0); | |
5538 | } | |
5539 | else | |
5540 | gcc_unreachable (); | |
5541 | } | |
5542 | else | |
5543 | { | |
5544 | src = vec_oprnd1 | |
5545 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5546 | op = vec_oprnd0 | |
5547 | = vect_get_vec_def_for_stmt_copy (scatter_idx_dt, vec_oprnd0); | |
5548 | } | |
5549 | ||
5550 | if (!useless_type_conversion_p (srctype, TREE_TYPE (src))) | |
5551 | { | |
5552 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)) | |
5553 | == TYPE_VECTOR_SUBPARTS (srctype)); | |
0e22bb5a | 5554 | var = vect_get_new_ssa_name (srctype, vect_simple_var); |
3bab6342 AT |
5555 | src = build1 (VIEW_CONVERT_EXPR, srctype, src); |
5556 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, src); | |
5557 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5558 | src = var; | |
5559 | } | |
5560 | ||
5561 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
5562 | { | |
5563 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
5564 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 5565 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
3bab6342 AT |
5566 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
5567 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
5568 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5569 | op = var; | |
5570 | } | |
5571 | ||
5572 | new_stmt | |
5573 | = gimple_build_call (scatter_decl, 5, ptr, mask, op, src, scale); | |
5574 | ||
5575 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5576 | ||
5577 | if (prev_stmt_info == NULL) | |
5578 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5579 | else | |
5580 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5581 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5582 | } | |
5583 | return true; | |
5584 | } | |
5585 | ||
0d0293ac | 5586 | if (grouped_store) |
ebfd146a IR |
5587 | { |
5588 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 5589 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 5590 | |
e14c1050 | 5591 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
5592 | |
5593 | /* FORNOW */ | |
a70d6342 | 5594 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
5595 | |
5596 | /* We vectorize all the stmts of the interleaving group when we | |
5597 | reach the last stmt in the group. */ | |
e14c1050 IR |
5598 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
5599 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
5600 | && !slp) |
5601 | { | |
5602 | *vec_stmt = NULL; | |
5603 | return true; | |
5604 | } | |
5605 | ||
5606 | if (slp) | |
4b5caab7 | 5607 | { |
0d0293ac | 5608 | grouped_store = false; |
4b5caab7 IR |
5609 | /* VEC_NUM is the number of vect stmts to be created for this |
5610 | group. */ | |
5611 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 5612 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
52eab378 | 5613 | gcc_assert (GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_stmt)) == first_stmt); |
4b5caab7 | 5614 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
d092494c | 5615 | op = gimple_assign_rhs1 (first_stmt); |
4b5caab7 | 5616 | } |
ebfd146a | 5617 | else |
4b5caab7 IR |
5618 | /* VEC_NUM is the number of vect stmts to be created for this |
5619 | group. */ | |
ebfd146a IR |
5620 | vec_num = group_size; |
5621 | } | |
b8698a0f | 5622 | else |
ebfd146a IR |
5623 | { |
5624 | first_stmt = stmt; | |
5625 | first_dr = dr; | |
5626 | group_size = vec_num = 1; | |
ebfd146a | 5627 | } |
b8698a0f | 5628 | |
73fbfcad | 5629 | if (dump_enabled_p ()) |
78c60e3d | 5630 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5631 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 5632 | |
f2e2a985 MM |
5633 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
5634 | { | |
5635 | gimple_stmt_iterator incr_gsi; | |
5636 | bool insert_after; | |
355fe088 | 5637 | gimple *incr; |
f2e2a985 MM |
5638 | tree offvar; |
5639 | tree ivstep; | |
5640 | tree running_off; | |
5641 | gimple_seq stmts = NULL; | |
5642 | tree stride_base, stride_step, alias_off; | |
5643 | tree vec_oprnd; | |
f502d50e | 5644 | unsigned int g; |
f2e2a985 MM |
5645 | |
5646 | gcc_assert (!nested_in_vect_loop_p (loop, stmt)); | |
5647 | ||
5648 | stride_base | |
5649 | = fold_build_pointer_plus | |
f502d50e | 5650 | (unshare_expr (DR_BASE_ADDRESS (first_dr)), |
f2e2a985 | 5651 | size_binop (PLUS_EXPR, |
f502d50e MM |
5652 | convert_to_ptrofftype (unshare_expr (DR_OFFSET (first_dr))), |
5653 | convert_to_ptrofftype (DR_INIT(first_dr)))); | |
5654 | stride_step = fold_convert (sizetype, unshare_expr (DR_STEP (first_dr))); | |
f2e2a985 MM |
5655 | |
5656 | /* For a store with loop-invariant (but other than power-of-2) | |
5657 | stride (i.e. not a grouped access) like so: | |
5658 | ||
5659 | for (i = 0; i < n; i += stride) | |
5660 | array[i] = ...; | |
5661 | ||
5662 | we generate a new induction variable and new stores from | |
5663 | the components of the (vectorized) rhs: | |
5664 | ||
5665 | for (j = 0; ; j += VF*stride) | |
5666 | vectemp = ...; | |
5667 | tmp1 = vectemp[0]; | |
5668 | array[j] = tmp1; | |
5669 | tmp2 = vectemp[1]; | |
5670 | array[j + stride] = tmp2; | |
5671 | ... | |
5672 | */ | |
5673 | ||
cee62fee MM |
5674 | unsigned nstores = nunits; |
5675 | tree ltype = elem_type; | |
5676 | if (slp) | |
5677 | { | |
5678 | nstores = nunits / group_size; | |
5679 | if (group_size < nunits) | |
5680 | ltype = build_vector_type (elem_type, group_size); | |
5681 | else | |
5682 | ltype = vectype; | |
5683 | ltype = build_aligned_type (ltype, TYPE_ALIGN (elem_type)); | |
5684 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
f502d50e | 5685 | group_size = 1; |
cee62fee MM |
5686 | } |
5687 | ||
f2e2a985 MM |
5688 | ivstep = stride_step; |
5689 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
5690 | build_int_cst (TREE_TYPE (ivstep), | |
cee62fee | 5691 | ncopies * nstores)); |
f2e2a985 MM |
5692 | |
5693 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
5694 | ||
5695 | create_iv (stride_base, ivstep, NULL, | |
5696 | loop, &incr_gsi, insert_after, | |
5697 | &offvar, NULL); | |
5698 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 5699 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
f2e2a985 MM |
5700 | |
5701 | stride_step = force_gimple_operand (stride_step, &stmts, true, NULL_TREE); | |
5702 | if (stmts) | |
5703 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
5704 | ||
5705 | prev_stmt_info = NULL; | |
f502d50e MM |
5706 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (first_dr)), 0); |
5707 | next_stmt = first_stmt; | |
5708 | for (g = 0; g < group_size; g++) | |
f2e2a985 | 5709 | { |
f502d50e MM |
5710 | running_off = offvar; |
5711 | if (g) | |
f2e2a985 | 5712 | { |
f502d50e MM |
5713 | tree size = TYPE_SIZE_UNIT (ltype); |
5714 | tree pos = fold_build2 (MULT_EXPR, sizetype, size_int (g), | |
f2e2a985 | 5715 | size); |
f502d50e | 5716 | tree newoff = copy_ssa_name (running_off, NULL); |
f2e2a985 | 5717 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
f502d50e | 5718 | running_off, pos); |
f2e2a985 | 5719 | vect_finish_stmt_generation (stmt, incr, gsi); |
f2e2a985 | 5720 | running_off = newoff; |
f502d50e MM |
5721 | } |
5722 | for (j = 0; j < ncopies; j++) | |
5723 | { | |
5724 | /* We've set op and dt above, from gimple_assign_rhs1(stmt), | |
5725 | and first_stmt == stmt. */ | |
5726 | if (j == 0) | |
5727 | { | |
5728 | if (slp) | |
5729 | { | |
5730 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, NULL, | |
5731 | slp_node, -1); | |
5732 | vec_oprnd = vec_oprnds[0]; | |
5733 | } | |
5734 | else | |
5735 | { | |
5736 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
5737 | op = gimple_assign_rhs1 (next_stmt); | |
81c40241 | 5738 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
f502d50e MM |
5739 | } |
5740 | } | |
f2e2a985 | 5741 | else |
f502d50e MM |
5742 | { |
5743 | if (slp) | |
5744 | vec_oprnd = vec_oprnds[j]; | |
5745 | else | |
c079cbac | 5746 | { |
81c40241 | 5747 | vect_is_simple_use (vec_oprnd, vinfo, &def_stmt, &dt); |
c079cbac RB |
5748 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); |
5749 | } | |
f502d50e MM |
5750 | } |
5751 | ||
5752 | for (i = 0; i < nstores; i++) | |
5753 | { | |
5754 | tree newref, newoff; | |
355fe088 | 5755 | gimple *incr, *assign; |
f502d50e MM |
5756 | tree size = TYPE_SIZE (ltype); |
5757 | /* Extract the i'th component. */ | |
5758 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, | |
5759 | bitsize_int (i), size); | |
5760 | tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd, | |
5761 | size, pos); | |
5762 | ||
5763 | elem = force_gimple_operand_gsi (gsi, elem, true, | |
5764 | NULL_TREE, true, | |
5765 | GSI_SAME_STMT); | |
5766 | ||
5767 | newref = build2 (MEM_REF, ltype, | |
5768 | running_off, alias_off); | |
5769 | ||
5770 | /* And store it to *running_off. */ | |
5771 | assign = gimple_build_assign (newref, elem); | |
5772 | vect_finish_stmt_generation (stmt, assign, gsi); | |
5773 | ||
5774 | newoff = copy_ssa_name (running_off, NULL); | |
5775 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
5776 | running_off, stride_step); | |
5777 | vect_finish_stmt_generation (stmt, incr, gsi); | |
5778 | ||
5779 | running_off = newoff; | |
225ce44b RB |
5780 | if (g == group_size - 1 |
5781 | && !slp) | |
f502d50e MM |
5782 | { |
5783 | if (j == 0 && i == 0) | |
225ce44b RB |
5784 | STMT_VINFO_VEC_STMT (stmt_info) |
5785 | = *vec_stmt = assign; | |
f502d50e MM |
5786 | else |
5787 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = assign; | |
5788 | prev_stmt_info = vinfo_for_stmt (assign); | |
5789 | } | |
5790 | } | |
f2e2a985 | 5791 | } |
f502d50e | 5792 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
f2e2a985 MM |
5793 | } |
5794 | return true; | |
5795 | } | |
5796 | ||
9771b263 DN |
5797 | dr_chain.create (group_size); |
5798 | oprnds.create (group_size); | |
ebfd146a | 5799 | |
720f5239 | 5800 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 5801 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
5802 | /* Targets with store-lane instructions must not require explicit |
5803 | realignment. */ | |
5804 | gcc_assert (!store_lanes_p | |
5805 | || alignment_support_scheme == dr_aligned | |
5806 | || alignment_support_scheme == dr_unaligned_supported); | |
5807 | ||
09dfa495 BM |
5808 | if (negative) |
5809 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
5810 | ||
272c6793 RS |
5811 | if (store_lanes_p) |
5812 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
5813 | else | |
5814 | aggr_type = vectype; | |
ebfd146a IR |
5815 | |
5816 | /* In case the vectorization factor (VF) is bigger than the number | |
5817 | of elements that we can fit in a vectype (nunits), we have to generate | |
5818 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 5819 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
5820 | vect_get_vec_def_for_copy_stmt. */ |
5821 | ||
0d0293ac | 5822 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
5823 | |
5824 | S1: &base + 2 = x2 | |
5825 | S2: &base = x0 | |
5826 | S3: &base + 1 = x1 | |
5827 | S4: &base + 3 = x3 | |
5828 | ||
5829 | We create vectorized stores starting from base address (the access of the | |
5830 | first stmt in the chain (S2 in the above example), when the last store stmt | |
5831 | of the chain (S4) is reached: | |
5832 | ||
5833 | VS1: &base = vx2 | |
5834 | VS2: &base + vec_size*1 = vx0 | |
5835 | VS3: &base + vec_size*2 = vx1 | |
5836 | VS4: &base + vec_size*3 = vx3 | |
5837 | ||
5838 | Then permutation statements are generated: | |
5839 | ||
3fcc1b55 JJ |
5840 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
5841 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 5842 | ... |
b8698a0f | 5843 | |
ebfd146a IR |
5844 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
5845 | (the order of the data-refs in the output of vect_permute_store_chain | |
5846 | corresponds to the order of scalar stmts in the interleaving chain - see | |
5847 | the documentation of vect_permute_store_chain()). | |
5848 | ||
5849 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 5850 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 5851 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 5852 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
5853 | */ |
5854 | ||
5855 | prev_stmt_info = NULL; | |
5856 | for (j = 0; j < ncopies; j++) | |
5857 | { | |
ebfd146a IR |
5858 | |
5859 | if (j == 0) | |
5860 | { | |
5861 | if (slp) | |
5862 | { | |
5863 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c IR |
5864 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
5865 | NULL, slp_node, -1); | |
ebfd146a | 5866 | |
9771b263 | 5867 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
5868 | } |
5869 | else | |
5870 | { | |
b8698a0f L |
5871 | /* For interleaved stores we collect vectorized defs for all the |
5872 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
5873 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
5874 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
5875 | ||
0d0293ac | 5876 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 5877 | OPRNDS are of size 1. */ |
b8698a0f | 5878 | next_stmt = first_stmt; |
ebfd146a IR |
5879 | for (i = 0; i < group_size; i++) |
5880 | { | |
b8698a0f L |
5881 | /* Since gaps are not supported for interleaved stores, |
5882 | GROUP_SIZE is the exact number of stmts in the chain. | |
5883 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
5884 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
5885 | iteration of the loop will be executed. */ |
5886 | gcc_assert (next_stmt | |
5887 | && gimple_assign_single_p (next_stmt)); | |
5888 | op = gimple_assign_rhs1 (next_stmt); | |
5889 | ||
81c40241 | 5890 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
9771b263 DN |
5891 | dr_chain.quick_push (vec_oprnd); |
5892 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 5893 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
5894 | } |
5895 | } | |
5896 | ||
5897 | /* We should have catched mismatched types earlier. */ | |
5898 | gcc_assert (useless_type_conversion_p (vectype, | |
5899 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
5900 | bool simd_lane_access_p |
5901 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
5902 | if (simd_lane_access_p | |
5903 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
5904 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
5905 | && integer_zerop (DR_OFFSET (first_dr)) | |
5906 | && integer_zerop (DR_INIT (first_dr)) | |
5907 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
5908 | get_alias_set (DR_REF (first_dr)))) | |
5909 | { | |
5910 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
5911 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
5912 | (DR_REF (first_dr)), 0); | |
8928eff3 | 5913 | inv_p = false; |
74bf76ed JJ |
5914 | } |
5915 | else | |
5916 | dataref_ptr | |
5917 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
5918 | simd_lane_access_p ? loop : NULL, | |
09dfa495 | 5919 | offset, &dummy, gsi, &ptr_incr, |
74bf76ed | 5920 | simd_lane_access_p, &inv_p); |
a70d6342 | 5921 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 5922 | } |
b8698a0f | 5923 | else |
ebfd146a | 5924 | { |
b8698a0f L |
5925 | /* For interleaved stores we created vectorized defs for all the |
5926 | defs stored in OPRNDS in the previous iteration (previous copy). | |
5927 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
5928 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
5929 | next copy. | |
0d0293ac | 5930 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
5931 | OPRNDS are of size 1. */ |
5932 | for (i = 0; i < group_size; i++) | |
5933 | { | |
9771b263 | 5934 | op = oprnds[i]; |
81c40241 | 5935 | vect_is_simple_use (op, vinfo, &def_stmt, &dt); |
b8698a0f | 5936 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
9771b263 DN |
5937 | dr_chain[i] = vec_oprnd; |
5938 | oprnds[i] = vec_oprnd; | |
ebfd146a | 5939 | } |
74bf76ed JJ |
5940 | if (dataref_offset) |
5941 | dataref_offset | |
5942 | = int_const_binop (PLUS_EXPR, dataref_offset, | |
5943 | TYPE_SIZE_UNIT (aggr_type)); | |
5944 | else | |
5945 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
5946 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
5947 | } |
5948 | ||
272c6793 | 5949 | if (store_lanes_p) |
ebfd146a | 5950 | { |
272c6793 | 5951 | tree vec_array; |
267d3070 | 5952 | |
272c6793 RS |
5953 | /* Combine all the vectors into an array. */ |
5954 | vec_array = create_vector_array (vectype, vec_num); | |
5955 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 5956 | { |
9771b263 | 5957 | vec_oprnd = dr_chain[i]; |
272c6793 | 5958 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 5959 | } |
b8698a0f | 5960 | |
272c6793 RS |
5961 | /* Emit: |
5962 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
5963 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
5964 | new_stmt = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array); | |
5965 | gimple_call_set_lhs (new_stmt, data_ref); | |
267d3070 | 5966 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
5967 | } |
5968 | else | |
5969 | { | |
5970 | new_stmt = NULL; | |
0d0293ac | 5971 | if (grouped_store) |
272c6793 | 5972 | { |
b6b9227d JJ |
5973 | if (j == 0) |
5974 | result_chain.create (group_size); | |
272c6793 RS |
5975 | /* Permute. */ |
5976 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
5977 | &result_chain); | |
5978 | } | |
c2d7ab2a | 5979 | |
272c6793 RS |
5980 | next_stmt = first_stmt; |
5981 | for (i = 0; i < vec_num; i++) | |
5982 | { | |
644ffefd | 5983 | unsigned align, misalign; |
272c6793 RS |
5984 | |
5985 | if (i > 0) | |
5986 | /* Bump the vector pointer. */ | |
5987 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
5988 | stmt, NULL_TREE); | |
5989 | ||
5990 | if (slp) | |
9771b263 | 5991 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
5992 | else if (grouped_store) |
5993 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 5994 | vect_permute_store_chain(). */ |
9771b263 | 5995 | vec_oprnd = result_chain[i]; |
272c6793 | 5996 | |
aed93b23 RB |
5997 | data_ref = fold_build2 (MEM_REF, TREE_TYPE (vec_oprnd), |
5998 | dataref_ptr, | |
5999 | dataref_offset | |
6000 | ? dataref_offset | |
6001 | : build_int_cst (reference_alias_ptr_type | |
6002 | (DR_REF (first_dr)), 0)); | |
644ffefd | 6003 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 | 6004 | if (aligned_access_p (first_dr)) |
644ffefd | 6005 | misalign = 0; |
272c6793 RS |
6006 | else if (DR_MISALIGNMENT (first_dr) == -1) |
6007 | { | |
52639a61 RB |
6008 | if (DR_VECT_AUX (first_dr)->base_element_aligned) |
6009 | align = TYPE_ALIGN_UNIT (elem_type); | |
6010 | else | |
6011 | align = get_object_alignment (DR_REF (first_dr)) | |
6012 | / BITS_PER_UNIT; | |
6013 | misalign = 0; | |
272c6793 RS |
6014 | TREE_TYPE (data_ref) |
6015 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 6016 | align * BITS_PER_UNIT); |
272c6793 RS |
6017 | } |
6018 | else | |
6019 | { | |
6020 | TREE_TYPE (data_ref) | |
6021 | = build_aligned_type (TREE_TYPE (data_ref), | |
6022 | TYPE_ALIGN (elem_type)); | |
644ffefd | 6023 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 6024 | } |
aed93b23 RB |
6025 | if (dataref_offset == NULL_TREE |
6026 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
6027 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
6028 | misalign); | |
c2d7ab2a | 6029 | |
f234d260 BM |
6030 | if (negative |
6031 | && dt != vect_constant_def | |
6032 | && dt != vect_external_def) | |
09dfa495 BM |
6033 | { |
6034 | tree perm_mask = perm_mask_for_reverse (vectype); | |
6035 | tree perm_dest | |
6036 | = vect_create_destination_var (gimple_assign_rhs1 (stmt), | |
6037 | vectype); | |
b731b390 | 6038 | tree new_temp = make_ssa_name (perm_dest); |
09dfa495 BM |
6039 | |
6040 | /* Generate the permute statement. */ | |
355fe088 | 6041 | gimple *perm_stmt |
0d0e4a03 JJ |
6042 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
6043 | vec_oprnd, perm_mask); | |
09dfa495 BM |
6044 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
6045 | ||
6046 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); | |
6047 | vec_oprnd = new_temp; | |
6048 | } | |
6049 | ||
272c6793 RS |
6050 | /* Arguments are ready. Create the new vector stmt. */ |
6051 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
6052 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
6053 | |
6054 | if (slp) | |
6055 | continue; | |
6056 | ||
e14c1050 | 6057 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
6058 | if (!next_stmt) |
6059 | break; | |
6060 | } | |
ebfd146a | 6061 | } |
1da0876c RS |
6062 | if (!slp) |
6063 | { | |
6064 | if (j == 0) | |
6065 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6066 | else | |
6067 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6068 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6069 | } | |
ebfd146a IR |
6070 | } |
6071 | ||
9771b263 DN |
6072 | dr_chain.release (); |
6073 | oprnds.release (); | |
6074 | result_chain.release (); | |
6075 | vec_oprnds.release (); | |
ebfd146a IR |
6076 | |
6077 | return true; | |
6078 | } | |
6079 | ||
557be5a8 AL |
6080 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
6081 | VECTOR_CST mask. No checks are made that the target platform supports the | |
6082 | mask, so callers may wish to test can_vec_perm_p separately, or use | |
6083 | vect_gen_perm_mask_checked. */ | |
a1e53f3f | 6084 | |
3fcc1b55 | 6085 | tree |
557be5a8 | 6086 | vect_gen_perm_mask_any (tree vectype, const unsigned char *sel) |
a1e53f3f | 6087 | { |
d2a12ae7 | 6088 | tree mask_elt_type, mask_type, mask_vec, *mask_elts; |
2635892a | 6089 | int i, nunits; |
a1e53f3f | 6090 | |
22e4dee7 | 6091 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
22e4dee7 | 6092 | |
96f9265a RG |
6093 | mask_elt_type = lang_hooks.types.type_for_mode |
6094 | (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype))), 1); | |
22e4dee7 | 6095 | mask_type = get_vectype_for_scalar_type (mask_elt_type); |
a1e53f3f | 6096 | |
d2a12ae7 | 6097 | mask_elts = XALLOCAVEC (tree, nunits); |
aec7ae7d | 6098 | for (i = nunits - 1; i >= 0; i--) |
d2a12ae7 RG |
6099 | mask_elts[i] = build_int_cst (mask_elt_type, sel[i]); |
6100 | mask_vec = build_vector (mask_type, mask_elts); | |
a1e53f3f | 6101 | |
2635892a | 6102 | return mask_vec; |
a1e53f3f L |
6103 | } |
6104 | ||
cf7aa6a3 AL |
6105 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_p, |
6106 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ | |
557be5a8 AL |
6107 | |
6108 | tree | |
6109 | vect_gen_perm_mask_checked (tree vectype, const unsigned char *sel) | |
6110 | { | |
6111 | gcc_assert (can_vec_perm_p (TYPE_MODE (vectype), false, sel)); | |
6112 | return vect_gen_perm_mask_any (vectype, sel); | |
6113 | } | |
6114 | ||
aec7ae7d JJ |
6115 | /* Given a vector variable X and Y, that was generated for the scalar |
6116 | STMT, generate instructions to permute the vector elements of X and Y | |
6117 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
6118 | permuted vector variable. */ | |
a1e53f3f L |
6119 | |
6120 | static tree | |
355fe088 | 6121 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple *stmt, |
aec7ae7d | 6122 | gimple_stmt_iterator *gsi) |
a1e53f3f L |
6123 | { |
6124 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 6125 | tree perm_dest, data_ref; |
355fe088 | 6126 | gimple *perm_stmt; |
a1e53f3f | 6127 | |
acdcd61b | 6128 | perm_dest = vect_create_destination_var (gimple_get_lhs (stmt), vectype); |
b731b390 | 6129 | data_ref = make_ssa_name (perm_dest); |
a1e53f3f L |
6130 | |
6131 | /* Generate the permute statement. */ | |
0d0e4a03 | 6132 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
a1e53f3f L |
6133 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
6134 | ||
6135 | return data_ref; | |
6136 | } | |
6137 | ||
6b916b36 RB |
6138 | /* Hoist the definitions of all SSA uses on STMT out of the loop LOOP, |
6139 | inserting them on the loops preheader edge. Returns true if we | |
6140 | were successful in doing so (and thus STMT can be moved then), | |
6141 | otherwise returns false. */ | |
6142 | ||
6143 | static bool | |
355fe088 | 6144 | hoist_defs_of_uses (gimple *stmt, struct loop *loop) |
6b916b36 RB |
6145 | { |
6146 | ssa_op_iter i; | |
6147 | tree op; | |
6148 | bool any = false; | |
6149 | ||
6150 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6151 | { | |
355fe088 | 6152 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6153 | if (!gimple_nop_p (def_stmt) |
6154 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6155 | { | |
6156 | /* Make sure we don't need to recurse. While we could do | |
6157 | so in simple cases when there are more complex use webs | |
6158 | we don't have an easy way to preserve stmt order to fulfil | |
6159 | dependencies within them. */ | |
6160 | tree op2; | |
6161 | ssa_op_iter i2; | |
d1417442 JJ |
6162 | if (gimple_code (def_stmt) == GIMPLE_PHI) |
6163 | return false; | |
6b916b36 RB |
6164 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
6165 | { | |
355fe088 | 6166 | gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2); |
6b916b36 RB |
6167 | if (!gimple_nop_p (def_stmt2) |
6168 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2))) | |
6169 | return false; | |
6170 | } | |
6171 | any = true; | |
6172 | } | |
6173 | } | |
6174 | ||
6175 | if (!any) | |
6176 | return true; | |
6177 | ||
6178 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6179 | { | |
355fe088 | 6180 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6181 | if (!gimple_nop_p (def_stmt) |
6182 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6183 | { | |
6184 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); | |
6185 | gsi_remove (&gsi, false); | |
6186 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); | |
6187 | } | |
6188 | } | |
6189 | ||
6190 | return true; | |
6191 | } | |
6192 | ||
ebfd146a IR |
6193 | /* vectorizable_load. |
6194 | ||
b8698a0f L |
6195 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
6196 | can be vectorized. | |
6197 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
6198 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
6199 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
6200 | ||
6201 | static bool | |
355fe088 | 6202 | vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 6203 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a IR |
6204 | { |
6205 | tree scalar_dest; | |
6206 | tree vec_dest = NULL; | |
6207 | tree data_ref = NULL; | |
6208 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 6209 | stmt_vec_info prev_stmt_info; |
ebfd146a | 6210 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 6211 | struct loop *loop = NULL; |
ebfd146a | 6212 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 6213 | bool nested_in_vect_loop = false; |
c716e67f | 6214 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
272c6793 | 6215 | tree elem_type; |
ebfd146a | 6216 | tree new_temp; |
ef4bddc2 | 6217 | machine_mode mode; |
355fe088 | 6218 | gimple *new_stmt = NULL; |
ebfd146a IR |
6219 | tree dummy; |
6220 | enum dr_alignment_support alignment_support_scheme; | |
6221 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 6222 | tree dataref_offset = NULL_TREE; |
355fe088 | 6223 | gimple *ptr_incr = NULL; |
ebfd146a | 6224 | int ncopies; |
9b999e8c | 6225 | int i, j, group_size = -1, group_gap_adj; |
ebfd146a IR |
6226 | tree msq = NULL_TREE, lsq; |
6227 | tree offset = NULL_TREE; | |
356bbc4c | 6228 | tree byte_offset = NULL_TREE; |
ebfd146a | 6229 | tree realignment_token = NULL_TREE; |
538dd0b7 | 6230 | gphi *phi = NULL; |
6e1aa848 | 6231 | vec<tree> dr_chain = vNULL; |
0d0293ac | 6232 | bool grouped_load = false; |
272c6793 | 6233 | bool load_lanes_p = false; |
355fe088 | 6234 | gimple *first_stmt; |
4f0a0218 | 6235 | gimple *first_stmt_for_drptr = NULL; |
ebfd146a | 6236 | bool inv_p; |
319e6439 | 6237 | bool negative = false; |
ebfd146a IR |
6238 | bool compute_in_loop = false; |
6239 | struct loop *at_loop; | |
6240 | int vec_num; | |
6241 | bool slp = (slp_node != NULL); | |
6242 | bool slp_perm = false; | |
6243 | enum tree_code code; | |
a70d6342 IR |
6244 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
6245 | int vf; | |
272c6793 | 6246 | tree aggr_type; |
aec7ae7d JJ |
6247 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; |
6248 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
6249 | int gather_scale = 1; | |
6250 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
310213d4 | 6251 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 6252 | |
465c8c19 JJ |
6253 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
6254 | return false; | |
6255 | ||
66c16fd9 RB |
6256 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
6257 | && ! vec_stmt) | |
465c8c19 JJ |
6258 | return false; |
6259 | ||
6260 | /* Is vectorizable load? */ | |
6261 | if (!is_gimple_assign (stmt)) | |
6262 | return false; | |
6263 | ||
6264 | scalar_dest = gimple_assign_lhs (stmt); | |
6265 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
6266 | return false; | |
6267 | ||
6268 | code = gimple_assign_rhs_code (stmt); | |
6269 | if (code != ARRAY_REF | |
6270 | && code != BIT_FIELD_REF | |
6271 | && code != INDIRECT_REF | |
6272 | && code != COMPONENT_REF | |
6273 | && code != IMAGPART_EXPR | |
6274 | && code != REALPART_EXPR | |
6275 | && code != MEM_REF | |
6276 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
6277 | return false; | |
6278 | ||
6279 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
6280 | return false; | |
6281 | ||
6282 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
6283 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
6284 | ||
a70d6342 IR |
6285 | if (loop_vinfo) |
6286 | { | |
6287 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
6288 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
6289 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
6290 | } | |
6291 | else | |
3533e503 | 6292 | vf = 1; |
ebfd146a IR |
6293 | |
6294 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 6295 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 6296 | case of SLP. */ |
437f4a00 | 6297 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
6298 | ncopies = 1; |
6299 | else | |
6300 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
6301 | ||
6302 | gcc_assert (ncopies >= 1); | |
6303 | ||
6304 | /* FORNOW. This restriction should be relaxed. */ | |
6305 | if (nested_in_vect_loop && ncopies > 1) | |
6306 | { | |
73fbfcad | 6307 | if (dump_enabled_p ()) |
78c60e3d | 6308 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6309 | "multiple types in nested loop.\n"); |
ebfd146a IR |
6310 | return false; |
6311 | } | |
6312 | ||
f2556b68 RB |
6313 | /* Invalidate assumptions made by dependence analysis when vectorization |
6314 | on the unrolled body effectively re-orders stmts. */ | |
6315 | if (ncopies > 1 | |
6316 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
6317 | && ((unsigned)LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
6318 | > STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
6319 | { | |
6320 | if (dump_enabled_p ()) | |
6321 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6322 | "cannot perform implicit CSE when unrolling " | |
6323 | "with negative dependence distance\n"); | |
6324 | return false; | |
6325 | } | |
6326 | ||
7b7b1813 | 6327 | elem_type = TREE_TYPE (vectype); |
947131ba | 6328 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
6329 | |
6330 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
6331 | (e.g. - data copies). */ | |
947131ba | 6332 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 6333 | { |
73fbfcad | 6334 | if (dump_enabled_p ()) |
78c60e3d | 6335 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6336 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
6337 | return false; |
6338 | } | |
6339 | ||
ebfd146a | 6340 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 6341 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 6342 | { |
0d0293ac | 6343 | grouped_load = true; |
ebfd146a | 6344 | /* FORNOW */ |
3bab6342 | 6345 | gcc_assert (!nested_in_vect_loop && !STMT_VINFO_GATHER_SCATTER_P (stmt_info)); |
ebfd146a | 6346 | |
e14c1050 | 6347 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
d5f035ea RB |
6348 | |
6349 | /* If this is single-element interleaving with an element distance | |
6350 | that leaves unused vector loads around punt - we at least create | |
6351 | very sub-optimal code in that case (and blow up memory, | |
6352 | see PR65518). */ | |
72c0f643 | 6353 | bool force_peeling = false; |
d5f035ea | 6354 | if (first_stmt == stmt |
72c0f643 RB |
6355 | && !GROUP_NEXT_ELEMENT (stmt_info)) |
6356 | { | |
6357 | if (GROUP_SIZE (stmt_info) > TYPE_VECTOR_SUBPARTS (vectype)) | |
6358 | { | |
6359 | if (dump_enabled_p ()) | |
6360 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6361 | "single-element interleaving not supported " | |
6362 | "for not adjacent vector loads\n"); | |
6363 | return false; | |
6364 | } | |
6365 | ||
6366 | /* Single-element interleaving requires peeling for gaps. */ | |
6367 | force_peeling = true; | |
6368 | } | |
6369 | ||
6370 | /* If there is a gap in the end of the group or the group size cannot | |
6371 | be made a multiple of the vector element count then we access excess | |
6372 | elements in the last iteration and thus need to peel that off. */ | |
6373 | if (loop_vinfo | |
6374 | && ! STMT_VINFO_STRIDED_P (stmt_info) | |
6375 | && (force_peeling | |
6376 | || GROUP_GAP (vinfo_for_stmt (first_stmt)) != 0 | |
6377 | || (!slp && vf % GROUP_SIZE (vinfo_for_stmt (first_stmt)) != 0))) | |
d5f035ea RB |
6378 | { |
6379 | if (dump_enabled_p ()) | |
6380 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
72c0f643 RB |
6381 | "Data access with gaps requires scalar " |
6382 | "epilogue loop\n"); | |
6383 | if (loop->inner) | |
6384 | { | |
6385 | if (dump_enabled_p ()) | |
6386 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6387 | "Peeling for outer loop is not supported\n"); | |
6388 | return false; | |
6389 | } | |
6390 | ||
6391 | LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true; | |
d5f035ea RB |
6392 | } |
6393 | ||
b1af7da6 RB |
6394 | if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
6395 | slp_perm = true; | |
6396 | ||
7b5fc413 | 6397 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
47d3fdb2 RB |
6398 | |
6399 | /* ??? The following is overly pessimistic (as well as the loop | |
6400 | case above) in the case we can statically determine the excess | |
6401 | elements loaded are within the bounds of a decl that is accessed. | |
6402 | Likewise for BB vectorizations using masked loads is a possibility. */ | |
6403 | if (bb_vinfo && slp_perm && group_size % nunits != 0) | |
6404 | { | |
6405 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6406 | "BB vectorization with gaps at the end of a load " | |
6407 | "is not supported\n"); | |
6408 | return false; | |
6409 | } | |
6410 | ||
7b5fc413 RB |
6411 | if (!slp |
6412 | && !PURE_SLP_STMT (stmt_info) | |
f2e2a985 | 6413 | && !STMT_VINFO_STRIDED_P (stmt_info)) |
b602d918 | 6414 | { |
272c6793 RS |
6415 | if (vect_load_lanes_supported (vectype, group_size)) |
6416 | load_lanes_p = true; | |
0d0293ac | 6417 | else if (!vect_grouped_load_supported (vectype, group_size)) |
b602d918 RS |
6418 | return false; |
6419 | } | |
f2556b68 RB |
6420 | |
6421 | /* Invalidate assumptions made by dependence analysis when vectorization | |
6422 | on the unrolled body effectively re-orders stmts. */ | |
6423 | if (!PURE_SLP_STMT (stmt_info) | |
6424 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
6425 | && ((unsigned)LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
6426 | > STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
6427 | { | |
6428 | if (dump_enabled_p ()) | |
6429 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6430 | "cannot perform implicit CSE when performing " | |
6431 | "group loads with negative dependence distance\n"); | |
6432 | return false; | |
6433 | } | |
96bb56b2 RB |
6434 | |
6435 | /* Similarly when the stmt is a load that is both part of a SLP | |
6436 | instance and a loop vectorized stmt via the same-dr mechanism | |
6437 | we have to give up. */ | |
6438 | if (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info) | |
6439 | && (STMT_SLP_TYPE (stmt_info) | |
6440 | != STMT_SLP_TYPE (vinfo_for_stmt | |
6441 | (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info))))) | |
6442 | { | |
6443 | if (dump_enabled_p ()) | |
6444 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6445 | "conflicting SLP types for CSEd load\n"); | |
6446 | return false; | |
6447 | } | |
ebfd146a IR |
6448 | } |
6449 | ||
a1e53f3f | 6450 | |
3bab6342 | 6451 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
aec7ae7d | 6452 | { |
355fe088 | 6453 | gimple *def_stmt; |
3bab6342 AT |
6454 | gather_decl = vect_check_gather_scatter (stmt, loop_vinfo, &gather_base, |
6455 | &gather_off, &gather_scale); | |
aec7ae7d | 6456 | gcc_assert (gather_decl); |
81c40241 RB |
6457 | if (!vect_is_simple_use (gather_off, vinfo, &def_stmt, &gather_dt, |
6458 | &gather_off_vectype)) | |
aec7ae7d | 6459 | { |
73fbfcad | 6460 | if (dump_enabled_p ()) |
78c60e3d | 6461 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6462 | "gather index use not simple.\n"); |
aec7ae7d JJ |
6463 | return false; |
6464 | } | |
6465 | } | |
f2e2a985 | 6466 | else if (STMT_VINFO_STRIDED_P (stmt_info)) |
7b5fc413 RB |
6467 | { |
6468 | if ((grouped_load | |
6469 | && (slp || PURE_SLP_STMT (stmt_info))) | |
6470 | && (group_size > nunits | |
b266b968 | 6471 | || nunits % group_size != 0)) |
7b5fc413 RB |
6472 | { |
6473 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6474 | "unhandled strided group load\n"); | |
6475 | return false; | |
6476 | } | |
6477 | } | |
319e6439 RG |
6478 | else |
6479 | { | |
6480 | negative = tree_int_cst_compare (nested_in_vect_loop | |
6481 | ? STMT_VINFO_DR_STEP (stmt_info) | |
6482 | : DR_STEP (dr), | |
6483 | size_zero_node) < 0; | |
6484 | if (negative && ncopies > 1) | |
6485 | { | |
73fbfcad | 6486 | if (dump_enabled_p ()) |
78c60e3d | 6487 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6488 | "multiple types with negative step.\n"); |
319e6439 RG |
6489 | return false; |
6490 | } | |
6491 | ||
6492 | if (negative) | |
6493 | { | |
08940f33 RB |
6494 | if (grouped_load) |
6495 | { | |
6496 | if (dump_enabled_p ()) | |
6497 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 TJ |
6498 | "negative step for group load not supported" |
6499 | "\n"); | |
08940f33 RB |
6500 | return false; |
6501 | } | |
319e6439 RG |
6502 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
6503 | if (alignment_support_scheme != dr_aligned | |
6504 | && alignment_support_scheme != dr_unaligned_supported) | |
6505 | { | |
73fbfcad | 6506 | if (dump_enabled_p ()) |
78c60e3d | 6507 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6508 | "negative step but alignment required.\n"); |
319e6439 RG |
6509 | return false; |
6510 | } | |
6511 | if (!perm_mask_for_reverse (vectype)) | |
6512 | { | |
73fbfcad | 6513 | if (dump_enabled_p ()) |
78c60e3d | 6514 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
6515 | "negative step and reversing not supported." |
6516 | "\n"); | |
319e6439 RG |
6517 | return false; |
6518 | } | |
6519 | } | |
7d75abc8 | 6520 | } |
aec7ae7d | 6521 | |
ebfd146a IR |
6522 | if (!vec_stmt) /* transformation not required. */ |
6523 | { | |
6524 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
2e8ab70c RB |
6525 | /* The SLP costs are calculated during SLP analysis. */ |
6526 | if (!PURE_SLP_STMT (stmt_info)) | |
6527 | vect_model_load_cost (stmt_info, ncopies, load_lanes_p, | |
6528 | NULL, NULL, NULL); | |
ebfd146a IR |
6529 | return true; |
6530 | } | |
6531 | ||
73fbfcad | 6532 | if (dump_enabled_p ()) |
78c60e3d | 6533 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6534 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a IR |
6535 | |
6536 | /** Transform. **/ | |
6537 | ||
c716e67f XDL |
6538 | ensure_base_align (stmt_info, dr); |
6539 | ||
3bab6342 | 6540 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
aec7ae7d JJ |
6541 | { |
6542 | tree vec_oprnd0 = NULL_TREE, op; | |
6543 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
6544 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
d3c2fee0 | 6545 | tree ptr, mask, var, scale, merge, perm_mask = NULL_TREE, prev_res = NULL_TREE; |
aec7ae7d JJ |
6546 | edge pe = loop_preheader_edge (loop); |
6547 | gimple_seq seq; | |
6548 | basic_block new_bb; | |
6549 | enum { NARROW, NONE, WIDEN } modifier; | |
6550 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
6551 | ||
6552 | if (nunits == gather_off_nunits) | |
6553 | modifier = NONE; | |
6554 | else if (nunits == gather_off_nunits / 2) | |
6555 | { | |
6556 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
6557 | modifier = WIDEN; | |
6558 | ||
6559 | for (i = 0; i < gather_off_nunits; ++i) | |
6560 | sel[i] = i | nunits; | |
6561 | ||
557be5a8 | 6562 | perm_mask = vect_gen_perm_mask_checked (gather_off_vectype, sel); |
aec7ae7d JJ |
6563 | } |
6564 | else if (nunits == gather_off_nunits * 2) | |
6565 | { | |
6566 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
6567 | modifier = NARROW; | |
6568 | ||
6569 | for (i = 0; i < nunits; ++i) | |
6570 | sel[i] = i < gather_off_nunits | |
6571 | ? i : i + nunits - gather_off_nunits; | |
6572 | ||
557be5a8 | 6573 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); |
aec7ae7d JJ |
6574 | ncopies *= 2; |
6575 | } | |
6576 | else | |
6577 | gcc_unreachable (); | |
6578 | ||
6579 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); | |
6580 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6581 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6582 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6583 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6584 | scaletype = TREE_VALUE (arglist); | |
d3c2fee0 | 6585 | gcc_checking_assert (types_compatible_p (srctype, rettype)); |
aec7ae7d JJ |
6586 | |
6587 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
6588 | ||
6589 | ptr = fold_convert (ptrtype, gather_base); | |
6590 | if (!is_gimple_min_invariant (ptr)) | |
6591 | { | |
6592 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
6593 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
6594 | gcc_assert (!new_bb); | |
6595 | } | |
6596 | ||
6597 | /* Currently we support only unconditional gather loads, | |
6598 | so mask should be all ones. */ | |
d3c2fee0 AI |
6599 | if (TREE_CODE (masktype) == INTEGER_TYPE) |
6600 | mask = build_int_cst (masktype, -1); | |
6601 | else if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
6602 | { | |
6603 | mask = build_int_cst (TREE_TYPE (masktype), -1); | |
6604 | mask = build_vector_from_val (masktype, mask); | |
03b9e8e4 | 6605 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
d3c2fee0 | 6606 | } |
aec7ae7d JJ |
6607 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) |
6608 | { | |
6609 | REAL_VALUE_TYPE r; | |
6610 | long tmp[6]; | |
6611 | for (j = 0; j < 6; ++j) | |
6612 | tmp[j] = -1; | |
6613 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
6614 | mask = build_real (TREE_TYPE (masktype), r); | |
d3c2fee0 | 6615 | mask = build_vector_from_val (masktype, mask); |
03b9e8e4 | 6616 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
aec7ae7d JJ |
6617 | } |
6618 | else | |
6619 | gcc_unreachable (); | |
aec7ae7d JJ |
6620 | |
6621 | scale = build_int_cst (scaletype, gather_scale); | |
6622 | ||
d3c2fee0 AI |
6623 | if (TREE_CODE (TREE_TYPE (rettype)) == INTEGER_TYPE) |
6624 | merge = build_int_cst (TREE_TYPE (rettype), 0); | |
6625 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (rettype))) | |
6626 | { | |
6627 | REAL_VALUE_TYPE r; | |
6628 | long tmp[6]; | |
6629 | for (j = 0; j < 6; ++j) | |
6630 | tmp[j] = 0; | |
6631 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (rettype))); | |
6632 | merge = build_real (TREE_TYPE (rettype), r); | |
6633 | } | |
6634 | else | |
6635 | gcc_unreachable (); | |
6636 | merge = build_vector_from_val (rettype, merge); | |
6637 | merge = vect_init_vector (stmt, merge, rettype, NULL); | |
6638 | ||
aec7ae7d JJ |
6639 | prev_stmt_info = NULL; |
6640 | for (j = 0; j < ncopies; ++j) | |
6641 | { | |
6642 | if (modifier == WIDEN && (j & 1)) | |
6643 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
6644 | perm_mask, stmt, gsi); | |
6645 | else if (j == 0) | |
6646 | op = vec_oprnd0 | |
81c40241 | 6647 | = vect_get_vec_def_for_operand (gather_off, stmt); |
aec7ae7d JJ |
6648 | else |
6649 | op = vec_oprnd0 | |
6650 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
6651 | ||
6652 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
6653 | { | |
6654 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
6655 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 6656 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
aec7ae7d JJ |
6657 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
6658 | new_stmt | |
0d0e4a03 | 6659 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
6660 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6661 | op = var; | |
6662 | } | |
6663 | ||
6664 | new_stmt | |
d3c2fee0 | 6665 | = gimple_build_call (gather_decl, 5, merge, ptr, op, mask, scale); |
aec7ae7d JJ |
6666 | |
6667 | if (!useless_type_conversion_p (vectype, rettype)) | |
6668 | { | |
6669 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
6670 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 6671 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
aec7ae7d JJ |
6672 | gimple_call_set_lhs (new_stmt, op); |
6673 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 6674 | var = make_ssa_name (vec_dest); |
aec7ae7d JJ |
6675 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
6676 | new_stmt | |
0d0e4a03 | 6677 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
6678 | } |
6679 | else | |
6680 | { | |
6681 | var = make_ssa_name (vec_dest, new_stmt); | |
6682 | gimple_call_set_lhs (new_stmt, var); | |
6683 | } | |
6684 | ||
6685 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6686 | ||
6687 | if (modifier == NARROW) | |
6688 | { | |
6689 | if ((j & 1) == 0) | |
6690 | { | |
6691 | prev_res = var; | |
6692 | continue; | |
6693 | } | |
6694 | var = permute_vec_elements (prev_res, var, | |
6695 | perm_mask, stmt, gsi); | |
6696 | new_stmt = SSA_NAME_DEF_STMT (var); | |
6697 | } | |
6698 | ||
6699 | if (prev_stmt_info == NULL) | |
6700 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6701 | else | |
6702 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6703 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6704 | } | |
6705 | return true; | |
6706 | } | |
f2e2a985 | 6707 | else if (STMT_VINFO_STRIDED_P (stmt_info)) |
7d75abc8 MM |
6708 | { |
6709 | gimple_stmt_iterator incr_gsi; | |
6710 | bool insert_after; | |
355fe088 | 6711 | gimple *incr; |
7d75abc8 | 6712 | tree offvar; |
7d75abc8 MM |
6713 | tree ivstep; |
6714 | tree running_off; | |
9771b263 | 6715 | vec<constructor_elt, va_gc> *v = NULL; |
7d75abc8 | 6716 | gimple_seq stmts = NULL; |
14ac6aa2 RB |
6717 | tree stride_base, stride_step, alias_off; |
6718 | ||
6719 | gcc_assert (!nested_in_vect_loop); | |
7d75abc8 | 6720 | |
f502d50e | 6721 | if (slp && grouped_load) |
ab313a8c RB |
6722 | first_dr = STMT_VINFO_DATA_REF |
6723 | (vinfo_for_stmt (GROUP_FIRST_ELEMENT (stmt_info))); | |
6724 | else | |
6725 | first_dr = dr; | |
6726 | ||
14ac6aa2 RB |
6727 | stride_base |
6728 | = fold_build_pointer_plus | |
ab313a8c | 6729 | (DR_BASE_ADDRESS (first_dr), |
14ac6aa2 | 6730 | size_binop (PLUS_EXPR, |
ab313a8c RB |
6731 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
6732 | convert_to_ptrofftype (DR_INIT (first_dr)))); | |
6733 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); | |
7d75abc8 MM |
6734 | |
6735 | /* For a load with loop-invariant (but other than power-of-2) | |
6736 | stride (i.e. not a grouped access) like so: | |
6737 | ||
6738 | for (i = 0; i < n; i += stride) | |
6739 | ... = array[i]; | |
6740 | ||
6741 | we generate a new induction variable and new accesses to | |
6742 | form a new vector (or vectors, depending on ncopies): | |
6743 | ||
6744 | for (j = 0; ; j += VF*stride) | |
6745 | tmp1 = array[j]; | |
6746 | tmp2 = array[j + stride]; | |
6747 | ... | |
6748 | vectemp = {tmp1, tmp2, ...} | |
6749 | */ | |
6750 | ||
ab313a8c RB |
6751 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step, |
6752 | build_int_cst (TREE_TYPE (stride_step), vf)); | |
7d75abc8 MM |
6753 | |
6754 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
6755 | ||
ab313a8c | 6756 | create_iv (unshare_expr (stride_base), unshare_expr (ivstep), NULL, |
7d75abc8 MM |
6757 | loop, &incr_gsi, insert_after, |
6758 | &offvar, NULL); | |
6759 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 6760 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
7d75abc8 | 6761 | |
ab313a8c RB |
6762 | stride_step = force_gimple_operand (unshare_expr (stride_step), |
6763 | &stmts, true, NULL_TREE); | |
7d75abc8 MM |
6764 | if (stmts) |
6765 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
6766 | ||
6767 | prev_stmt_info = NULL; | |
6768 | running_off = offvar; | |
ab313a8c | 6769 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (first_dr)), 0); |
7b5fc413 RB |
6770 | int nloads = nunits; |
6771 | tree ltype = TREE_TYPE (vectype); | |
b266b968 | 6772 | auto_vec<tree> dr_chain; |
7b5fc413 RB |
6773 | if (slp) |
6774 | { | |
6775 | nloads = nunits / group_size; | |
6776 | if (group_size < nunits) | |
6777 | ltype = build_vector_type (TREE_TYPE (vectype), group_size); | |
6778 | else | |
6779 | ltype = vectype; | |
6780 | ltype = build_aligned_type (ltype, TYPE_ALIGN (TREE_TYPE (vectype))); | |
66c16fd9 RB |
6781 | /* For SLP permutation support we need to load the whole group, |
6782 | not only the number of vector stmts the permutation result | |
6783 | fits in. */ | |
b266b968 | 6784 | if (slp_perm) |
66c16fd9 RB |
6785 | { |
6786 | ncopies = (group_size * vf + nunits - 1) / nunits; | |
6787 | dr_chain.create (ncopies); | |
6788 | } | |
6789 | else | |
6790 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
7b5fc413 | 6791 | } |
7d75abc8 MM |
6792 | for (j = 0; j < ncopies; j++) |
6793 | { | |
6794 | tree vec_inv; | |
6795 | ||
7b5fc413 RB |
6796 | if (nloads > 1) |
6797 | { | |
6798 | vec_alloc (v, nloads); | |
6799 | for (i = 0; i < nloads; i++) | |
6800 | { | |
6801 | tree newref, newoff; | |
355fe088 | 6802 | gimple *incr; |
7b5fc413 RB |
6803 | newref = build2 (MEM_REF, ltype, running_off, alias_off); |
6804 | ||
6805 | newref = force_gimple_operand_gsi (gsi, newref, true, | |
6806 | NULL_TREE, true, | |
6807 | GSI_SAME_STMT); | |
6808 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, newref); | |
6809 | newoff = copy_ssa_name (running_off); | |
6810 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
6811 | running_off, stride_step); | |
6812 | vect_finish_stmt_generation (stmt, incr, gsi); | |
6813 | ||
6814 | running_off = newoff; | |
6815 | } | |
6816 | ||
6817 | vec_inv = build_constructor (vectype, v); | |
6818 | new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi); | |
6819 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
6820 | } | |
6821 | else | |
7d75abc8 | 6822 | { |
7b5fc413 RB |
6823 | new_stmt = gimple_build_assign (make_ssa_name (ltype), |
6824 | build2 (MEM_REF, ltype, | |
6825 | running_off, alias_off)); | |
6826 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6827 | ||
6828 | tree newoff = copy_ssa_name (running_off); | |
355fe088 | 6829 | gimple *incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
0d0e4a03 | 6830 | running_off, stride_step); |
7d75abc8 MM |
6831 | vect_finish_stmt_generation (stmt, incr, gsi); |
6832 | ||
6833 | running_off = newoff; | |
6834 | } | |
6835 | ||
7b5fc413 | 6836 | if (slp) |
b266b968 | 6837 | { |
b266b968 RB |
6838 | if (slp_perm) |
6839 | dr_chain.quick_push (gimple_assign_lhs (new_stmt)); | |
66c16fd9 RB |
6840 | else |
6841 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
b266b968 | 6842 | } |
7d75abc8 | 6843 | else |
225ce44b RB |
6844 | { |
6845 | if (j == 0) | |
6846 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6847 | else | |
6848 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6849 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6850 | } | |
7d75abc8 | 6851 | } |
b266b968 RB |
6852 | if (slp_perm) |
6853 | vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, | |
6854 | slp_node_instance, false); | |
7d75abc8 MM |
6855 | return true; |
6856 | } | |
aec7ae7d | 6857 | |
0d0293ac | 6858 | if (grouped_load) |
ebfd146a | 6859 | { |
e14c1050 | 6860 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
4f0a0218 | 6861 | /* For SLP vectorization we directly vectorize a subchain |
52eab378 RB |
6862 | without permutation. */ |
6863 | if (slp && ! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
4f0a0218 RB |
6864 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
6865 | /* For BB vectorization always use the first stmt to base | |
6866 | the data ref pointer on. */ | |
6867 | if (bb_vinfo) | |
6868 | first_stmt_for_drptr = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 6869 | |
ebfd146a | 6870 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
6871 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
6872 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
6873 | ??? But we can only do so if there is exactly one | |
6874 | as we have no way to get at the rest. Leave the CSE | |
6875 | opportunity alone. | |
6876 | ??? With the group load eventually participating | |
6877 | in multiple different permutations (having multiple | |
6878 | slp nodes which refer to the same group) the CSE | |
6879 | is even wrong code. See PR56270. */ | |
6880 | && !slp) | |
ebfd146a IR |
6881 | { |
6882 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
6883 | return true; | |
6884 | } | |
6885 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 6886 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
9b999e8c | 6887 | group_gap_adj = 0; |
ebfd146a IR |
6888 | |
6889 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
6890 | if (slp) | |
6891 | { | |
0d0293ac | 6892 | grouped_load = false; |
91ff1504 RB |
6893 | /* For SLP permutation support we need to load the whole group, |
6894 | not only the number of vector stmts the permutation result | |
6895 | fits in. */ | |
6896 | if (slp_perm) | |
6897 | vec_num = (group_size * vf + nunits - 1) / nunits; | |
6898 | else | |
6899 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9b999e8c | 6900 | group_gap_adj = vf * group_size - nunits * vec_num; |
a70d6342 | 6901 | } |
ebfd146a | 6902 | else |
9b999e8c | 6903 | vec_num = group_size; |
ebfd146a IR |
6904 | } |
6905 | else | |
6906 | { | |
6907 | first_stmt = stmt; | |
6908 | first_dr = dr; | |
6909 | group_size = vec_num = 1; | |
9b999e8c | 6910 | group_gap_adj = 0; |
ebfd146a IR |
6911 | } |
6912 | ||
720f5239 | 6913 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 6914 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
6915 | /* Targets with load-lane instructions must not require explicit |
6916 | realignment. */ | |
6917 | gcc_assert (!load_lanes_p | |
6918 | || alignment_support_scheme == dr_aligned | |
6919 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
6920 | |
6921 | /* In case the vectorization factor (VF) is bigger than the number | |
6922 | of elements that we can fit in a vectype (nunits), we have to generate | |
6923 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 6924 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 6925 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 6926 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 6927 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
6928 | stmts that use the defs of the current stmt. The example below |
6929 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
6930 | need to create 4 vectorized stmts): | |
ebfd146a IR |
6931 | |
6932 | before vectorization: | |
6933 | RELATED_STMT VEC_STMT | |
6934 | S1: x = memref - - | |
6935 | S2: z = x + 1 - - | |
6936 | ||
6937 | step 1: vectorize stmt S1: | |
6938 | We first create the vector stmt VS1_0, and, as usual, record a | |
6939 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
6940 | Next, we create the vector stmt VS1_1, and record a pointer to | |
6941 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 6942 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
6943 | stmts and pointers: |
6944 | RELATED_STMT VEC_STMT | |
6945 | VS1_0: vx0 = memref0 VS1_1 - | |
6946 | VS1_1: vx1 = memref1 VS1_2 - | |
6947 | VS1_2: vx2 = memref2 VS1_3 - | |
6948 | VS1_3: vx3 = memref3 - - | |
6949 | S1: x = load - VS1_0 | |
6950 | S2: z = x + 1 - - | |
6951 | ||
b8698a0f L |
6952 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
6953 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
6954 | stmt S2. */ |
6955 | ||
0d0293ac | 6956 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
6957 | |
6958 | S1: x2 = &base + 2 | |
6959 | S2: x0 = &base | |
6960 | S3: x1 = &base + 1 | |
6961 | S4: x3 = &base + 3 | |
6962 | ||
b8698a0f | 6963 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
6964 | starting from the access of the first stmt of the chain: |
6965 | ||
6966 | VS1: vx0 = &base | |
6967 | VS2: vx1 = &base + vec_size*1 | |
6968 | VS3: vx3 = &base + vec_size*2 | |
6969 | VS4: vx4 = &base + vec_size*3 | |
6970 | ||
6971 | Then permutation statements are generated: | |
6972 | ||
e2c83630 RH |
6973 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
6974 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
6975 | ... |
6976 | ||
6977 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
6978 | (the order of the data-refs in the output of vect_permute_load_chain | |
6979 | corresponds to the order of scalar stmts in the interleaving chain - see | |
6980 | the documentation of vect_permute_load_chain()). | |
6981 | The generation of permutation stmts and recording them in | |
0d0293ac | 6982 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 6983 | |
b8698a0f | 6984 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
6985 | permutation stmts above are created for every copy. The result vector |
6986 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
6987 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
6988 | |
6989 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
6990 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
6991 | we generate the following code: | |
6992 | p = initial_addr; | |
6993 | indx = 0; | |
6994 | loop { | |
6995 | p = p + indx * vectype_size; | |
6996 | vec_dest = *(p); | |
6997 | indx = indx + 1; | |
6998 | } | |
6999 | ||
7000 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 7001 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
7002 | then generate the following code, in which the data in each iteration is |
7003 | obtained by two vector loads, one from the previous iteration, and one | |
7004 | from the current iteration: | |
7005 | p1 = initial_addr; | |
7006 | msq_init = *(floor(p1)) | |
7007 | p2 = initial_addr + VS - 1; | |
7008 | realignment_token = call target_builtin; | |
7009 | indx = 0; | |
7010 | loop { | |
7011 | p2 = p2 + indx * vectype_size | |
7012 | lsq = *(floor(p2)) | |
7013 | vec_dest = realign_load (msq, lsq, realignment_token) | |
7014 | indx = indx + 1; | |
7015 | msq = lsq; | |
7016 | } */ | |
7017 | ||
7018 | /* If the misalignment remains the same throughout the execution of the | |
7019 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 7020 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
7021 | This can only occur when vectorizing memory accesses in the inner-loop |
7022 | nested within an outer-loop that is being vectorized. */ | |
7023 | ||
d1e4b493 | 7024 | if (nested_in_vect_loop |
211bea38 | 7025 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
ebfd146a IR |
7026 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) |
7027 | { | |
7028 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
7029 | compute_in_loop = true; | |
7030 | } | |
7031 | ||
7032 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
7033 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 7034 | && !compute_in_loop) |
ebfd146a IR |
7035 | { |
7036 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
7037 | alignment_support_scheme, NULL_TREE, | |
7038 | &at_loop); | |
7039 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7040 | { | |
538dd0b7 | 7041 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (msq)); |
356bbc4c JJ |
7042 | byte_offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
7043 | size_one_node); | |
ebfd146a IR |
7044 | } |
7045 | } | |
7046 | else | |
7047 | at_loop = loop; | |
7048 | ||
a1e53f3f L |
7049 | if (negative) |
7050 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
7051 | ||
272c6793 RS |
7052 | if (load_lanes_p) |
7053 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
7054 | else | |
7055 | aggr_type = vectype; | |
7056 | ||
ebfd146a IR |
7057 | prev_stmt_info = NULL; |
7058 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 7059 | { |
272c6793 | 7060 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 7061 | if (j == 0) |
74bf76ed JJ |
7062 | { |
7063 | bool simd_lane_access_p | |
7064 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
7065 | if (simd_lane_access_p | |
7066 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
7067 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
7068 | && integer_zerop (DR_OFFSET (first_dr)) | |
7069 | && integer_zerop (DR_INIT (first_dr)) | |
7070 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
7071 | get_alias_set (DR_REF (first_dr))) | |
7072 | && (alignment_support_scheme == dr_aligned | |
7073 | || alignment_support_scheme == dr_unaligned_supported)) | |
7074 | { | |
7075 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
7076 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
7077 | (DR_REF (first_dr)), 0); | |
8928eff3 | 7078 | inv_p = false; |
74bf76ed | 7079 | } |
4f0a0218 RB |
7080 | else if (first_stmt_for_drptr |
7081 | && first_stmt != first_stmt_for_drptr) | |
7082 | { | |
7083 | dataref_ptr | |
7084 | = vect_create_data_ref_ptr (first_stmt_for_drptr, aggr_type, | |
7085 | at_loop, offset, &dummy, gsi, | |
7086 | &ptr_incr, simd_lane_access_p, | |
7087 | &inv_p, byte_offset); | |
7088 | /* Adjust the pointer by the difference to first_stmt. */ | |
7089 | data_reference_p ptrdr | |
7090 | = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt_for_drptr)); | |
7091 | tree diff = fold_convert (sizetype, | |
7092 | size_binop (MINUS_EXPR, | |
7093 | DR_INIT (first_dr), | |
7094 | DR_INIT (ptrdr))); | |
7095 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
7096 | stmt, diff); | |
7097 | } | |
74bf76ed JJ |
7098 | else |
7099 | dataref_ptr | |
7100 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
7101 | offset, &dummy, gsi, &ptr_incr, | |
356bbc4c JJ |
7102 | simd_lane_access_p, &inv_p, |
7103 | byte_offset); | |
74bf76ed JJ |
7104 | } |
7105 | else if (dataref_offset) | |
7106 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
7107 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 7108 | else |
272c6793 RS |
7109 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
7110 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 7111 | |
0d0293ac | 7112 | if (grouped_load || slp_perm) |
9771b263 | 7113 | dr_chain.create (vec_num); |
5ce1ee7f | 7114 | |
272c6793 | 7115 | if (load_lanes_p) |
ebfd146a | 7116 | { |
272c6793 RS |
7117 | tree vec_array; |
7118 | ||
7119 | vec_array = create_vector_array (vectype, vec_num); | |
7120 | ||
7121 | /* Emit: | |
7122 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
7123 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
7124 | new_stmt = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
7125 | gimple_call_set_lhs (new_stmt, vec_array); | |
7126 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a | 7127 | |
272c6793 RS |
7128 | /* Extract each vector into an SSA_NAME. */ |
7129 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 7130 | { |
272c6793 RS |
7131 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
7132 | vec_array, i); | |
9771b263 | 7133 | dr_chain.quick_push (new_temp); |
272c6793 RS |
7134 | } |
7135 | ||
7136 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 7137 | vect_record_grouped_load_vectors (stmt, dr_chain); |
272c6793 RS |
7138 | } |
7139 | else | |
7140 | { | |
7141 | for (i = 0; i < vec_num; i++) | |
7142 | { | |
7143 | if (i > 0) | |
7144 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
7145 | stmt, NULL_TREE); | |
7146 | ||
7147 | /* 2. Create the vector-load in the loop. */ | |
7148 | switch (alignment_support_scheme) | |
7149 | { | |
7150 | case dr_aligned: | |
7151 | case dr_unaligned_supported: | |
be1ac4ec | 7152 | { |
644ffefd MJ |
7153 | unsigned int align, misalign; |
7154 | ||
272c6793 | 7155 | data_ref |
aed93b23 RB |
7156 | = fold_build2 (MEM_REF, vectype, dataref_ptr, |
7157 | dataref_offset | |
7158 | ? dataref_offset | |
7159 | : build_int_cst (reference_alias_ptr_type | |
7160 | (DR_REF (first_dr)), 0)); | |
644ffefd | 7161 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 RS |
7162 | if (alignment_support_scheme == dr_aligned) |
7163 | { | |
7164 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 7165 | misalign = 0; |
272c6793 RS |
7166 | } |
7167 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
7168 | { | |
52639a61 RB |
7169 | if (DR_VECT_AUX (first_dr)->base_element_aligned) |
7170 | align = TYPE_ALIGN_UNIT (elem_type); | |
7171 | else | |
7172 | align = (get_object_alignment (DR_REF (first_dr)) | |
7173 | / BITS_PER_UNIT); | |
7174 | misalign = 0; | |
272c6793 RS |
7175 | TREE_TYPE (data_ref) |
7176 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 7177 | align * BITS_PER_UNIT); |
272c6793 RS |
7178 | } |
7179 | else | |
7180 | { | |
7181 | TREE_TYPE (data_ref) | |
7182 | = build_aligned_type (TREE_TYPE (data_ref), | |
7183 | TYPE_ALIGN (elem_type)); | |
644ffefd | 7184 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 7185 | } |
aed93b23 RB |
7186 | if (dataref_offset == NULL_TREE |
7187 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
7188 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), |
7189 | align, misalign); | |
272c6793 | 7190 | break; |
be1ac4ec | 7191 | } |
272c6793 | 7192 | case dr_explicit_realign: |
267d3070 | 7193 | { |
272c6793 | 7194 | tree ptr, bump; |
272c6793 | 7195 | |
d88981fc | 7196 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
272c6793 RS |
7197 | |
7198 | if (compute_in_loop) | |
7199 | msq = vect_setup_realignment (first_stmt, gsi, | |
7200 | &realignment_token, | |
7201 | dr_explicit_realign, | |
7202 | dataref_ptr, NULL); | |
7203 | ||
aed93b23 RB |
7204 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
7205 | ptr = copy_ssa_name (dataref_ptr); | |
7206 | else | |
7207 | ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
0d0e4a03 JJ |
7208 | new_stmt = gimple_build_assign |
7209 | (ptr, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
7210 | build_int_cst |
7211 | (TREE_TYPE (dataref_ptr), | |
7212 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
7213 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7214 | data_ref | |
7215 | = build2 (MEM_REF, vectype, ptr, | |
7216 | build_int_cst (reference_alias_ptr_type | |
7217 | (DR_REF (first_dr)), 0)); | |
7218 | vec_dest = vect_create_destination_var (scalar_dest, | |
7219 | vectype); | |
7220 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
7221 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
7222 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7223 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
7224 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
7225 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7226 | msq = new_temp; | |
7227 | ||
d88981fc | 7228 | bump = size_binop (MULT_EXPR, vs, |
7b7b1813 | 7229 | TYPE_SIZE_UNIT (elem_type)); |
d88981fc | 7230 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
272c6793 | 7231 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
0d0e4a03 JJ |
7232 | new_stmt = gimple_build_assign |
7233 | (NULL_TREE, BIT_AND_EXPR, ptr, | |
272c6793 RS |
7234 | build_int_cst |
7235 | (TREE_TYPE (ptr), | |
7236 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
aed93b23 | 7237 | ptr = copy_ssa_name (ptr, new_stmt); |
272c6793 RS |
7238 | gimple_assign_set_lhs (new_stmt, ptr); |
7239 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7240 | data_ref | |
7241 | = build2 (MEM_REF, vectype, ptr, | |
7242 | build_int_cst (reference_alias_ptr_type | |
7243 | (DR_REF (first_dr)), 0)); | |
7244 | break; | |
267d3070 | 7245 | } |
272c6793 | 7246 | case dr_explicit_realign_optimized: |
aed93b23 RB |
7247 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
7248 | new_temp = copy_ssa_name (dataref_ptr); | |
7249 | else | |
7250 | new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
0d0e4a03 JJ |
7251 | new_stmt = gimple_build_assign |
7252 | (new_temp, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
7253 | build_int_cst |
7254 | (TREE_TYPE (dataref_ptr), | |
7255 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
7256 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7257 | data_ref | |
7258 | = build2 (MEM_REF, vectype, new_temp, | |
7259 | build_int_cst (reference_alias_ptr_type | |
7260 | (DR_REF (first_dr)), 0)); | |
7261 | break; | |
7262 | default: | |
7263 | gcc_unreachable (); | |
7264 | } | |
ebfd146a | 7265 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 7266 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
7267 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7268 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7269 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7270 | ||
272c6793 RS |
7271 | /* 3. Handle explicit realignment if necessary/supported. |
7272 | Create in loop: | |
7273 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
7274 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
7275 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 7276 | { |
272c6793 RS |
7277 | lsq = gimple_assign_lhs (new_stmt); |
7278 | if (!realignment_token) | |
7279 | realignment_token = dataref_ptr; | |
7280 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
0d0e4a03 JJ |
7281 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, |
7282 | msq, lsq, realignment_token); | |
272c6793 RS |
7283 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7284 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7285 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7286 | ||
7287 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7288 | { | |
7289 | gcc_assert (phi); | |
7290 | if (i == vec_num - 1 && j == ncopies - 1) | |
7291 | add_phi_arg (phi, lsq, | |
7292 | loop_latch_edge (containing_loop), | |
9e227d60 | 7293 | UNKNOWN_LOCATION); |
272c6793 RS |
7294 | msq = lsq; |
7295 | } | |
ebfd146a | 7296 | } |
ebfd146a | 7297 | |
59fd17e3 RB |
7298 | /* 4. Handle invariant-load. */ |
7299 | if (inv_p && !bb_vinfo) | |
7300 | { | |
59fd17e3 | 7301 | gcc_assert (!grouped_load); |
d1417442 JJ |
7302 | /* If we have versioned for aliasing or the loop doesn't |
7303 | have any data dependencies that would preclude this, | |
7304 | then we are sure this is a loop invariant load and | |
7305 | thus we can insert it on the preheader edge. */ | |
7306 | if (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) | |
7307 | && !nested_in_vect_loop | |
6b916b36 | 7308 | && hoist_defs_of_uses (stmt, loop)) |
a0e35eb0 RB |
7309 | { |
7310 | if (dump_enabled_p ()) | |
7311 | { | |
7312 | dump_printf_loc (MSG_NOTE, vect_location, | |
7313 | "hoisting out of the vectorized " | |
7314 | "loop: "); | |
7315 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
a0e35eb0 | 7316 | } |
b731b390 | 7317 | tree tem = copy_ssa_name (scalar_dest); |
a0e35eb0 RB |
7318 | gsi_insert_on_edge_immediate |
7319 | (loop_preheader_edge (loop), | |
7320 | gimple_build_assign (tem, | |
7321 | unshare_expr | |
7322 | (gimple_assign_rhs1 (stmt)))); | |
7323 | new_temp = vect_init_vector (stmt, tem, vectype, NULL); | |
34cd48e5 RB |
7324 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
7325 | set_vinfo_for_stmt (new_stmt, | |
7326 | new_stmt_vec_info (new_stmt, vinfo)); | |
a0e35eb0 RB |
7327 | } |
7328 | else | |
7329 | { | |
7330 | gimple_stmt_iterator gsi2 = *gsi; | |
7331 | gsi_next (&gsi2); | |
7332 | new_temp = vect_init_vector (stmt, scalar_dest, | |
7333 | vectype, &gsi2); | |
34cd48e5 | 7334 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
a0e35eb0 | 7335 | } |
59fd17e3 RB |
7336 | } |
7337 | ||
272c6793 RS |
7338 | if (negative) |
7339 | { | |
aec7ae7d JJ |
7340 | tree perm_mask = perm_mask_for_reverse (vectype); |
7341 | new_temp = permute_vec_elements (new_temp, new_temp, | |
7342 | perm_mask, stmt, gsi); | |
ebfd146a IR |
7343 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
7344 | } | |
267d3070 | 7345 | |
272c6793 | 7346 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
7347 | vect_transform_grouped_load (). */ |
7348 | if (grouped_load || slp_perm) | |
9771b263 | 7349 | dr_chain.quick_push (new_temp); |
267d3070 | 7350 | |
272c6793 RS |
7351 | /* Store vector loads in the corresponding SLP_NODE. */ |
7352 | if (slp && !slp_perm) | |
9771b263 | 7353 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
272c6793 | 7354 | } |
9b999e8c RB |
7355 | /* Bump the vector pointer to account for a gap or for excess |
7356 | elements loaded for a permuted SLP load. */ | |
7357 | if (group_gap_adj != 0) | |
a64b9c26 | 7358 | { |
9b999e8c RB |
7359 | bool ovf; |
7360 | tree bump | |
7361 | = wide_int_to_tree (sizetype, | |
7362 | wi::smul (TYPE_SIZE_UNIT (elem_type), | |
7363 | group_gap_adj, &ovf)); | |
a64b9c26 RB |
7364 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
7365 | stmt, bump); | |
7366 | } | |
ebfd146a IR |
7367 | } |
7368 | ||
7369 | if (slp && !slp_perm) | |
7370 | continue; | |
7371 | ||
7372 | if (slp_perm) | |
7373 | { | |
01d8bf07 | 7374 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
ebfd146a IR |
7375 | slp_node_instance, false)) |
7376 | { | |
9771b263 | 7377 | dr_chain.release (); |
ebfd146a IR |
7378 | return false; |
7379 | } | |
7380 | } | |
7381 | else | |
7382 | { | |
0d0293ac | 7383 | if (grouped_load) |
ebfd146a | 7384 | { |
272c6793 | 7385 | if (!load_lanes_p) |
0d0293ac | 7386 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 7387 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
7388 | } |
7389 | else | |
7390 | { | |
7391 | if (j == 0) | |
7392 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7393 | else | |
7394 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7395 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7396 | } | |
7397 | } | |
9771b263 | 7398 | dr_chain.release (); |
ebfd146a IR |
7399 | } |
7400 | ||
ebfd146a IR |
7401 | return true; |
7402 | } | |
7403 | ||
7404 | /* Function vect_is_simple_cond. | |
b8698a0f | 7405 | |
ebfd146a IR |
7406 | Input: |
7407 | LOOP - the loop that is being vectorized. | |
7408 | COND - Condition that is checked for simple use. | |
7409 | ||
e9e1d143 RG |
7410 | Output: |
7411 | *COMP_VECTYPE - the vector type for the comparison. | |
7412 | ||
ebfd146a IR |
7413 | Returns whether a COND can be vectorized. Checks whether |
7414 | condition operands are supportable using vec_is_simple_use. */ | |
7415 | ||
87aab9b2 | 7416 | static bool |
81c40241 | 7417 | vect_is_simple_cond (tree cond, vec_info *vinfo, tree *comp_vectype) |
ebfd146a IR |
7418 | { |
7419 | tree lhs, rhs; | |
ebfd146a | 7420 | enum vect_def_type dt; |
e9e1d143 | 7421 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a | 7422 | |
a414c77f IE |
7423 | /* Mask case. */ |
7424 | if (TREE_CODE (cond) == SSA_NAME | |
7425 | && TREE_CODE (TREE_TYPE (cond)) == BOOLEAN_TYPE) | |
7426 | { | |
7427 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (cond); | |
7428 | if (!vect_is_simple_use (cond, vinfo, &lhs_def_stmt, | |
7429 | &dt, comp_vectype) | |
7430 | || !*comp_vectype | |
7431 | || !VECTOR_BOOLEAN_TYPE_P (*comp_vectype)) | |
7432 | return false; | |
7433 | return true; | |
7434 | } | |
7435 | ||
ebfd146a IR |
7436 | if (!COMPARISON_CLASS_P (cond)) |
7437 | return false; | |
7438 | ||
7439 | lhs = TREE_OPERAND (cond, 0); | |
7440 | rhs = TREE_OPERAND (cond, 1); | |
7441 | ||
7442 | if (TREE_CODE (lhs) == SSA_NAME) | |
7443 | { | |
355fe088 | 7444 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); |
81c40241 | 7445 | if (!vect_is_simple_use (lhs, vinfo, &lhs_def_stmt, &dt, &vectype1)) |
ebfd146a IR |
7446 | return false; |
7447 | } | |
7448 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
7449 | && TREE_CODE (lhs) != FIXED_CST) | |
7450 | return false; | |
7451 | ||
7452 | if (TREE_CODE (rhs) == SSA_NAME) | |
7453 | { | |
355fe088 | 7454 | gimple *rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); |
81c40241 | 7455 | if (!vect_is_simple_use (rhs, vinfo, &rhs_def_stmt, &dt, &vectype2)) |
ebfd146a IR |
7456 | return false; |
7457 | } | |
f7e531cf | 7458 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST |
ebfd146a IR |
7459 | && TREE_CODE (rhs) != FIXED_CST) |
7460 | return false; | |
7461 | ||
28b33016 IE |
7462 | if (vectype1 && vectype2 |
7463 | && TYPE_VECTOR_SUBPARTS (vectype1) != TYPE_VECTOR_SUBPARTS (vectype2)) | |
7464 | return false; | |
7465 | ||
e9e1d143 | 7466 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
ebfd146a IR |
7467 | return true; |
7468 | } | |
7469 | ||
7470 | /* vectorizable_condition. | |
7471 | ||
b8698a0f L |
7472 | Check if STMT is conditional modify expression that can be vectorized. |
7473 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
7474 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
7475 | at GSI. |
7476 | ||
7477 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
7478 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
0ad23163 | 7479 | else clause if it is 2). |
ebfd146a IR |
7480 | |
7481 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7482 | ||
4bbe8262 | 7483 | bool |
355fe088 TS |
7484 | vectorizable_condition (gimple *stmt, gimple_stmt_iterator *gsi, |
7485 | gimple **vec_stmt, tree reduc_def, int reduc_index, | |
f7e531cf | 7486 | slp_tree slp_node) |
ebfd146a IR |
7487 | { |
7488 | tree scalar_dest = NULL_TREE; | |
7489 | tree vec_dest = NULL_TREE; | |
ebfd146a IR |
7490 | tree cond_expr, then_clause, else_clause; |
7491 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
df11cc78 | 7492 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
7493 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
7494 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
5958f9e2 | 7495 | tree vec_compare; |
ebfd146a IR |
7496 | tree new_temp; |
7497 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a855b1b1 | 7498 | enum vect_def_type dt, dts[4]; |
f7e531cf | 7499 | int ncopies; |
ebfd146a | 7500 | enum tree_code code; |
a855b1b1 | 7501 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
7502 | int i, j; |
7503 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
7504 | vec<tree> vec_oprnds0 = vNULL; |
7505 | vec<tree> vec_oprnds1 = vNULL; | |
7506 | vec<tree> vec_oprnds2 = vNULL; | |
7507 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 7508 | tree vec_cmp_type; |
a414c77f | 7509 | bool masked = false; |
b8698a0f | 7510 | |
f7e531cf IR |
7511 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
7512 | return false; | |
7513 | ||
af29617a AH |
7514 | if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) == TREE_CODE_REDUCTION) |
7515 | { | |
7516 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
7517 | return false; | |
ebfd146a | 7518 | |
af29617a AH |
7519 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7520 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
7521 | && reduc_def)) | |
7522 | return false; | |
ebfd146a | 7523 | |
af29617a AH |
7524 | /* FORNOW: not yet supported. */ |
7525 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
7526 | { | |
7527 | if (dump_enabled_p ()) | |
7528 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7529 | "value used after loop.\n"); | |
7530 | return false; | |
7531 | } | |
ebfd146a IR |
7532 | } |
7533 | ||
7534 | /* Is vectorizable conditional operation? */ | |
7535 | if (!is_gimple_assign (stmt)) | |
7536 | return false; | |
7537 | ||
7538 | code = gimple_assign_rhs_code (stmt); | |
7539 | ||
7540 | if (code != COND_EXPR) | |
7541 | return false; | |
7542 | ||
465c8c19 JJ |
7543 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
7544 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
2947d3b2 | 7545 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
465c8c19 JJ |
7546 | |
7547 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
7548 | ncopies = 1; | |
7549 | else | |
7550 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
7551 | ||
7552 | gcc_assert (ncopies >= 1); | |
7553 | if (reduc_index && ncopies > 1) | |
7554 | return false; /* FORNOW */ | |
7555 | ||
4e71066d RG |
7556 | cond_expr = gimple_assign_rhs1 (stmt); |
7557 | then_clause = gimple_assign_rhs2 (stmt); | |
7558 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 7559 | |
81c40241 | 7560 | if (!vect_is_simple_cond (cond_expr, stmt_info->vinfo, &comp_vectype) |
e9e1d143 | 7561 | || !comp_vectype) |
ebfd146a IR |
7562 | return false; |
7563 | ||
81c40241 | 7564 | gimple *def_stmt; |
2947d3b2 IE |
7565 | if (!vect_is_simple_use (then_clause, stmt_info->vinfo, &def_stmt, &dt, |
7566 | &vectype1)) | |
7567 | return false; | |
7568 | if (!vect_is_simple_use (else_clause, stmt_info->vinfo, &def_stmt, &dt, | |
7569 | &vectype2)) | |
ebfd146a | 7570 | return false; |
2947d3b2 IE |
7571 | |
7572 | if (vectype1 && !useless_type_conversion_p (vectype, vectype1)) | |
7573 | return false; | |
7574 | ||
7575 | if (vectype2 && !useless_type_conversion_p (vectype, vectype2)) | |
ebfd146a IR |
7576 | return false; |
7577 | ||
28b33016 IE |
7578 | masked = !COMPARISON_CLASS_P (cond_expr); |
7579 | vec_cmp_type = build_same_sized_truth_vector_type (comp_vectype); | |
7580 | ||
74946978 MP |
7581 | if (vec_cmp_type == NULL_TREE) |
7582 | return false; | |
784fb9b3 | 7583 | |
b8698a0f | 7584 | if (!vec_stmt) |
ebfd146a IR |
7585 | { |
7586 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
e9e1d143 | 7587 | return expand_vec_cond_expr_p (vectype, comp_vectype); |
ebfd146a IR |
7588 | } |
7589 | ||
f7e531cf IR |
7590 | /* Transform. */ |
7591 | ||
7592 | if (!slp_node) | |
7593 | { | |
9771b263 DN |
7594 | vec_oprnds0.create (1); |
7595 | vec_oprnds1.create (1); | |
7596 | vec_oprnds2.create (1); | |
7597 | vec_oprnds3.create (1); | |
f7e531cf | 7598 | } |
ebfd146a IR |
7599 | |
7600 | /* Handle def. */ | |
7601 | scalar_dest = gimple_assign_lhs (stmt); | |
7602 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
7603 | ||
7604 | /* Handle cond expr. */ | |
a855b1b1 MM |
7605 | for (j = 0; j < ncopies; j++) |
7606 | { | |
538dd0b7 | 7607 | gassign *new_stmt = NULL; |
a855b1b1 MM |
7608 | if (j == 0) |
7609 | { | |
f7e531cf IR |
7610 | if (slp_node) |
7611 | { | |
00f96dc9 TS |
7612 | auto_vec<tree, 4> ops; |
7613 | auto_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 7614 | |
a414c77f IE |
7615 | if (masked) |
7616 | ops.safe_push (cond_expr); | |
7617 | else | |
7618 | { | |
7619 | ops.safe_push (TREE_OPERAND (cond_expr, 0)); | |
7620 | ops.safe_push (TREE_OPERAND (cond_expr, 1)); | |
7621 | } | |
9771b263 DN |
7622 | ops.safe_push (then_clause); |
7623 | ops.safe_push (else_clause); | |
f7e531cf | 7624 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); |
37b5ec8f JJ |
7625 | vec_oprnds3 = vec_defs.pop (); |
7626 | vec_oprnds2 = vec_defs.pop (); | |
a414c77f IE |
7627 | if (!masked) |
7628 | vec_oprnds1 = vec_defs.pop (); | |
37b5ec8f | 7629 | vec_oprnds0 = vec_defs.pop (); |
f7e531cf | 7630 | |
9771b263 DN |
7631 | ops.release (); |
7632 | vec_defs.release (); | |
f7e531cf IR |
7633 | } |
7634 | else | |
7635 | { | |
355fe088 | 7636 | gimple *gtemp; |
a414c77f IE |
7637 | if (masked) |
7638 | { | |
7639 | vec_cond_lhs | |
7640 | = vect_get_vec_def_for_operand (cond_expr, stmt, | |
7641 | comp_vectype); | |
7642 | vect_is_simple_use (cond_expr, stmt_info->vinfo, | |
7643 | >emp, &dts[0]); | |
7644 | } | |
7645 | else | |
7646 | { | |
7647 | vec_cond_lhs = | |
7648 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), | |
7649 | stmt, comp_vectype); | |
7650 | vect_is_simple_use (TREE_OPERAND (cond_expr, 0), | |
7651 | loop_vinfo, >emp, &dts[0]); | |
7652 | ||
7653 | vec_cond_rhs = | |
7654 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), | |
7655 | stmt, comp_vectype); | |
7656 | vect_is_simple_use (TREE_OPERAND (cond_expr, 1), | |
7657 | loop_vinfo, >emp, &dts[1]); | |
7658 | } | |
f7e531cf IR |
7659 | if (reduc_index == 1) |
7660 | vec_then_clause = reduc_def; | |
7661 | else | |
7662 | { | |
7663 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
81c40241 RB |
7664 | stmt); |
7665 | vect_is_simple_use (then_clause, loop_vinfo, | |
7666 | >emp, &dts[2]); | |
f7e531cf IR |
7667 | } |
7668 | if (reduc_index == 2) | |
7669 | vec_else_clause = reduc_def; | |
7670 | else | |
7671 | { | |
7672 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
81c40241 RB |
7673 | stmt); |
7674 | vect_is_simple_use (else_clause, loop_vinfo, >emp, &dts[3]); | |
f7e531cf | 7675 | } |
a855b1b1 MM |
7676 | } |
7677 | } | |
7678 | else | |
7679 | { | |
a414c77f IE |
7680 | vec_cond_lhs |
7681 | = vect_get_vec_def_for_stmt_copy (dts[0], | |
7682 | vec_oprnds0.pop ()); | |
7683 | if (!masked) | |
7684 | vec_cond_rhs | |
7685 | = vect_get_vec_def_for_stmt_copy (dts[1], | |
7686 | vec_oprnds1.pop ()); | |
7687 | ||
a855b1b1 | 7688 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 7689 | vec_oprnds2.pop ()); |
a855b1b1 | 7690 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 7691 | vec_oprnds3.pop ()); |
f7e531cf IR |
7692 | } |
7693 | ||
7694 | if (!slp_node) | |
7695 | { | |
9771b263 | 7696 | vec_oprnds0.quick_push (vec_cond_lhs); |
a414c77f IE |
7697 | if (!masked) |
7698 | vec_oprnds1.quick_push (vec_cond_rhs); | |
9771b263 DN |
7699 | vec_oprnds2.quick_push (vec_then_clause); |
7700 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
7701 | } |
7702 | ||
9dc3f7de | 7703 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 7704 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 7705 | { |
9771b263 DN |
7706 | vec_then_clause = vec_oprnds2[i]; |
7707 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 7708 | |
a414c77f IE |
7709 | if (masked) |
7710 | vec_compare = vec_cond_lhs; | |
7711 | else | |
7712 | { | |
7713 | vec_cond_rhs = vec_oprnds1[i]; | |
7714 | vec_compare = build2 (TREE_CODE (cond_expr), vec_cmp_type, | |
7715 | vec_cond_lhs, vec_cond_rhs); | |
7716 | } | |
5958f9e2 JJ |
7717 | new_temp = make_ssa_name (vec_dest); |
7718 | new_stmt = gimple_build_assign (new_temp, VEC_COND_EXPR, | |
7719 | vec_compare, vec_then_clause, | |
7720 | vec_else_clause); | |
f7e531cf IR |
7721 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7722 | if (slp_node) | |
9771b263 | 7723 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
7724 | } |
7725 | ||
7726 | if (slp_node) | |
7727 | continue; | |
7728 | ||
7729 | if (j == 0) | |
7730 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7731 | else | |
7732 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7733 | ||
7734 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 7735 | } |
b8698a0f | 7736 | |
9771b263 DN |
7737 | vec_oprnds0.release (); |
7738 | vec_oprnds1.release (); | |
7739 | vec_oprnds2.release (); | |
7740 | vec_oprnds3.release (); | |
f7e531cf | 7741 | |
ebfd146a IR |
7742 | return true; |
7743 | } | |
7744 | ||
42fd8198 IE |
7745 | /* vectorizable_comparison. |
7746 | ||
7747 | Check if STMT is comparison expression that can be vectorized. | |
7748 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
7749 | comparison, put it in VEC_STMT, and insert it at GSI. | |
7750 | ||
7751 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7752 | ||
7753 | bool | |
7754 | vectorizable_comparison (gimple *stmt, gimple_stmt_iterator *gsi, | |
7755 | gimple **vec_stmt, tree reduc_def, | |
7756 | slp_tree slp_node) | |
7757 | { | |
7758 | tree lhs, rhs1, rhs2; | |
7759 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
7760 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; | |
7761 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
7762 | tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE; | |
7763 | tree new_temp; | |
7764 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
7765 | enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
7766 | unsigned nunits; | |
7767 | int ncopies; | |
7768 | enum tree_code code; | |
7769 | stmt_vec_info prev_stmt_info = NULL; | |
7770 | int i, j; | |
7771 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
7772 | vec<tree> vec_oprnds0 = vNULL; | |
7773 | vec<tree> vec_oprnds1 = vNULL; | |
7774 | gimple *def_stmt; | |
7775 | tree mask_type; | |
7776 | tree mask; | |
7777 | ||
c245362b IE |
7778 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
7779 | return false; | |
7780 | ||
30480bcd | 7781 | if (!vectype || !VECTOR_BOOLEAN_TYPE_P (vectype)) |
42fd8198 IE |
7782 | return false; |
7783 | ||
7784 | mask_type = vectype; | |
7785 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
7786 | ||
7787 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
7788 | ncopies = 1; | |
7789 | else | |
7790 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
7791 | ||
7792 | gcc_assert (ncopies >= 1); | |
42fd8198 IE |
7793 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7794 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
7795 | && reduc_def)) | |
7796 | return false; | |
7797 | ||
7798 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
7799 | { | |
7800 | if (dump_enabled_p ()) | |
7801 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7802 | "value used after loop.\n"); | |
7803 | return false; | |
7804 | } | |
7805 | ||
7806 | if (!is_gimple_assign (stmt)) | |
7807 | return false; | |
7808 | ||
7809 | code = gimple_assign_rhs_code (stmt); | |
7810 | ||
7811 | if (TREE_CODE_CLASS (code) != tcc_comparison) | |
7812 | return false; | |
7813 | ||
7814 | rhs1 = gimple_assign_rhs1 (stmt); | |
7815 | rhs2 = gimple_assign_rhs2 (stmt); | |
7816 | ||
7817 | if (!vect_is_simple_use (rhs1, stmt_info->vinfo, &def_stmt, | |
7818 | &dts[0], &vectype1)) | |
7819 | return false; | |
7820 | ||
7821 | if (!vect_is_simple_use (rhs2, stmt_info->vinfo, &def_stmt, | |
7822 | &dts[1], &vectype2)) | |
7823 | return false; | |
7824 | ||
7825 | if (vectype1 && vectype2 | |
7826 | && TYPE_VECTOR_SUBPARTS (vectype1) != TYPE_VECTOR_SUBPARTS (vectype2)) | |
7827 | return false; | |
7828 | ||
7829 | vectype = vectype1 ? vectype1 : vectype2; | |
7830 | ||
7831 | /* Invariant comparison. */ | |
7832 | if (!vectype) | |
7833 | { | |
7834 | vectype = build_vector_type (TREE_TYPE (rhs1), nunits); | |
7835 | if (tree_to_shwi (TYPE_SIZE_UNIT (vectype)) != current_vector_size) | |
7836 | return false; | |
7837 | } | |
7838 | else if (nunits != TYPE_VECTOR_SUBPARTS (vectype)) | |
7839 | return false; | |
7840 | ||
7841 | if (!vec_stmt) | |
7842 | { | |
7843 | STMT_VINFO_TYPE (stmt_info) = comparison_vec_info_type; | |
7844 | vect_model_simple_cost (stmt_info, ncopies, dts, NULL, NULL); | |
7845 | return expand_vec_cmp_expr_p (vectype, mask_type); | |
7846 | } | |
7847 | ||
7848 | /* Transform. */ | |
7849 | if (!slp_node) | |
7850 | { | |
7851 | vec_oprnds0.create (1); | |
7852 | vec_oprnds1.create (1); | |
7853 | } | |
7854 | ||
7855 | /* Handle def. */ | |
7856 | lhs = gimple_assign_lhs (stmt); | |
7857 | mask = vect_create_destination_var (lhs, mask_type); | |
7858 | ||
7859 | /* Handle cmp expr. */ | |
7860 | for (j = 0; j < ncopies; j++) | |
7861 | { | |
7862 | gassign *new_stmt = NULL; | |
7863 | if (j == 0) | |
7864 | { | |
7865 | if (slp_node) | |
7866 | { | |
7867 | auto_vec<tree, 2> ops; | |
7868 | auto_vec<vec<tree>, 2> vec_defs; | |
7869 | ||
7870 | ops.safe_push (rhs1); | |
7871 | ops.safe_push (rhs2); | |
7872 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); | |
7873 | vec_oprnds1 = vec_defs.pop (); | |
7874 | vec_oprnds0 = vec_defs.pop (); | |
7875 | } | |
7876 | else | |
7877 | { | |
e4af0bc4 IE |
7878 | vec_rhs1 = vect_get_vec_def_for_operand (rhs1, stmt, vectype); |
7879 | vec_rhs2 = vect_get_vec_def_for_operand (rhs2, stmt, vectype); | |
42fd8198 IE |
7880 | } |
7881 | } | |
7882 | else | |
7883 | { | |
7884 | vec_rhs1 = vect_get_vec_def_for_stmt_copy (dts[0], | |
7885 | vec_oprnds0.pop ()); | |
7886 | vec_rhs2 = vect_get_vec_def_for_stmt_copy (dts[1], | |
7887 | vec_oprnds1.pop ()); | |
7888 | } | |
7889 | ||
7890 | if (!slp_node) | |
7891 | { | |
7892 | vec_oprnds0.quick_push (vec_rhs1); | |
7893 | vec_oprnds1.quick_push (vec_rhs2); | |
7894 | } | |
7895 | ||
7896 | /* Arguments are ready. Create the new vector stmt. */ | |
7897 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1) | |
7898 | { | |
7899 | vec_rhs2 = vec_oprnds1[i]; | |
7900 | ||
7901 | new_temp = make_ssa_name (mask); | |
7902 | new_stmt = gimple_build_assign (new_temp, code, vec_rhs1, vec_rhs2); | |
7903 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7904 | if (slp_node) | |
7905 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
7906 | } | |
7907 | ||
7908 | if (slp_node) | |
7909 | continue; | |
7910 | ||
7911 | if (j == 0) | |
7912 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7913 | else | |
7914 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7915 | ||
7916 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7917 | } | |
7918 | ||
7919 | vec_oprnds0.release (); | |
7920 | vec_oprnds1.release (); | |
7921 | ||
7922 | return true; | |
7923 | } | |
ebfd146a | 7924 | |
8644a673 | 7925 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
7926 | |
7927 | bool | |
355fe088 | 7928 | vect_analyze_stmt (gimple *stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 7929 | { |
8644a673 | 7930 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 7931 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 7932 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 7933 | bool ok; |
a70d6342 | 7934 | tree scalar_type, vectype; |
355fe088 | 7935 | gimple *pattern_stmt; |
363477c0 | 7936 | gimple_seq pattern_def_seq; |
ebfd146a | 7937 | |
73fbfcad | 7938 | if (dump_enabled_p ()) |
ebfd146a | 7939 | { |
78c60e3d SS |
7940 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
7941 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
8644a673 | 7942 | } |
ebfd146a | 7943 | |
1825a1f3 | 7944 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 7945 | { |
73fbfcad | 7946 | if (dump_enabled_p ()) |
78c60e3d | 7947 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7948 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
7949 | |
7950 | return false; | |
7951 | } | |
b8698a0f L |
7952 | |
7953 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
7954 | to include: |
7955 | - the COND_EXPR which is the loop exit condition | |
7956 | - any LABEL_EXPRs in the loop | |
b8698a0f | 7957 | - computations that are used only for array indexing or loop control. |
8644a673 | 7958 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 7959 | instance, therefore, all the statements are relevant. |
ebfd146a | 7960 | |
d092494c | 7961 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 7962 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
7963 | statements. In basic blocks we are called from some SLP instance |
7964 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
7965 | already will be part of SLP instance. */ | |
83197f37 IR |
7966 | |
7967 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 7968 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 7969 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 7970 | { |
9d5e7640 | 7971 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 7972 | && pattern_stmt |
9d5e7640 IR |
7973 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
7974 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
7975 | { | |
83197f37 | 7976 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
7977 | stmt = pattern_stmt; |
7978 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 7979 | if (dump_enabled_p ()) |
9d5e7640 | 7980 | { |
78c60e3d SS |
7981 | dump_printf_loc (MSG_NOTE, vect_location, |
7982 | "==> examining pattern statement: "); | |
7983 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
9d5e7640 IR |
7984 | } |
7985 | } | |
7986 | else | |
7987 | { | |
73fbfcad | 7988 | if (dump_enabled_p ()) |
e645e942 | 7989 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 7990 | |
9d5e7640 IR |
7991 | return true; |
7992 | } | |
8644a673 | 7993 | } |
83197f37 | 7994 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 7995 | && node == NULL |
83197f37 IR |
7996 | && pattern_stmt |
7997 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
7998 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
7999 | { | |
8000 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 8001 | if (dump_enabled_p ()) |
83197f37 | 8002 | { |
78c60e3d SS |
8003 | dump_printf_loc (MSG_NOTE, vect_location, |
8004 | "==> examining pattern statement: "); | |
8005 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
83197f37 IR |
8006 | } |
8007 | ||
8008 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node)) | |
8009 | return false; | |
8010 | } | |
ebfd146a | 8011 | |
1107f3ae | 8012 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 8013 | && node == NULL |
363477c0 | 8014 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 8015 | { |
363477c0 | 8016 | gimple_stmt_iterator si; |
1107f3ae | 8017 | |
363477c0 JJ |
8018 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
8019 | { | |
355fe088 | 8020 | gimple *pattern_def_stmt = gsi_stmt (si); |
363477c0 JJ |
8021 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) |
8022 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
8023 | { | |
8024 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 8025 | if (dump_enabled_p ()) |
363477c0 | 8026 | { |
78c60e3d SS |
8027 | dump_printf_loc (MSG_NOTE, vect_location, |
8028 | "==> examining pattern def statement: "); | |
8029 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
363477c0 | 8030 | } |
1107f3ae | 8031 | |
363477c0 JJ |
8032 | if (!vect_analyze_stmt (pattern_def_stmt, |
8033 | need_to_vectorize, node)) | |
8034 | return false; | |
8035 | } | |
8036 | } | |
8037 | } | |
1107f3ae | 8038 | |
8644a673 IR |
8039 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
8040 | { | |
8041 | case vect_internal_def: | |
8042 | break; | |
ebfd146a | 8043 | |
8644a673 | 8044 | case vect_reduction_def: |
7c5222ff | 8045 | case vect_nested_cycle: |
14a61437 RB |
8046 | gcc_assert (!bb_vinfo |
8047 | && (relevance == vect_used_in_outer | |
8048 | || relevance == vect_used_in_outer_by_reduction | |
8049 | || relevance == vect_used_by_reduction | |
8050 | || relevance == vect_unused_in_scope)); | |
8644a673 IR |
8051 | break; |
8052 | ||
8053 | case vect_induction_def: | |
8054 | case vect_constant_def: | |
8055 | case vect_external_def: | |
8056 | case vect_unknown_def_type: | |
8057 | default: | |
8058 | gcc_unreachable (); | |
8059 | } | |
ebfd146a | 8060 | |
a70d6342 IR |
8061 | if (bb_vinfo) |
8062 | { | |
8063 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
8064 | ||
b690cc0f | 8065 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
73fbfcad | 8066 | if (dump_enabled_p ()) |
a70d6342 | 8067 | { |
78c60e3d SS |
8068 | dump_printf_loc (MSG_NOTE, vect_location, |
8069 | "get vectype for scalar type: "); | |
8070 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
e645e942 | 8071 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
8072 | } |
8073 | ||
8074 | vectype = get_vectype_for_scalar_type (scalar_type); | |
8075 | if (!vectype) | |
8076 | { | |
73fbfcad | 8077 | if (dump_enabled_p ()) |
a70d6342 | 8078 | { |
78c60e3d SS |
8079 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8080 | "not SLPed: unsupported data-type "); | |
8081 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
8082 | scalar_type); | |
e645e942 | 8083 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
a70d6342 IR |
8084 | } |
8085 | return false; | |
8086 | } | |
8087 | ||
73fbfcad | 8088 | if (dump_enabled_p ()) |
a70d6342 | 8089 | { |
78c60e3d SS |
8090 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); |
8091 | dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); | |
e645e942 | 8092 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
8093 | } |
8094 | ||
8095 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
8096 | } | |
8097 | ||
8644a673 | 8098 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 8099 | { |
8644a673 | 8100 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
0136f8f0 AH |
8101 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
8102 | || (is_gimple_call (stmt) | |
8103 | && gimple_call_lhs (stmt) == NULL_TREE)); | |
8644a673 | 8104 | *need_to_vectorize = true; |
ebfd146a IR |
8105 | } |
8106 | ||
b1af7da6 RB |
8107 | if (PURE_SLP_STMT (stmt_info) && !node) |
8108 | { | |
8109 | dump_printf_loc (MSG_NOTE, vect_location, | |
8110 | "handled only by SLP analysis\n"); | |
8111 | return true; | |
8112 | } | |
8113 | ||
8114 | ok = true; | |
8115 | if (!bb_vinfo | |
8116 | && (STMT_VINFO_RELEVANT_P (stmt_info) | |
8117 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
8118 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
8119 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
8120 | || vectorizable_shift (stmt, NULL, NULL, node) | |
8121 | || vectorizable_operation (stmt, NULL, NULL, node) | |
8122 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
8123 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
8124 | || vectorizable_call (stmt, NULL, NULL, node) | |
8125 | || vectorizable_store (stmt, NULL, NULL, node) | |
8126 | || vectorizable_reduction (stmt, NULL, NULL, node) | |
42fd8198 IE |
8127 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
8128 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 RB |
8129 | else |
8130 | { | |
8131 | if (bb_vinfo) | |
8132 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
8133 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
8134 | || vectorizable_shift (stmt, NULL, NULL, node) | |
8135 | || vectorizable_operation (stmt, NULL, NULL, node) | |
8136 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
8137 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
8138 | || vectorizable_call (stmt, NULL, NULL, node) | |
8139 | || vectorizable_store (stmt, NULL, NULL, node) | |
42fd8198 IE |
8140 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
8141 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 | 8142 | } |
8644a673 IR |
8143 | |
8144 | if (!ok) | |
ebfd146a | 8145 | { |
73fbfcad | 8146 | if (dump_enabled_p ()) |
8644a673 | 8147 | { |
78c60e3d SS |
8148 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8149 | "not vectorized: relevant stmt not "); | |
8150 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
8151 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 8152 | } |
b8698a0f | 8153 | |
ebfd146a IR |
8154 | return false; |
8155 | } | |
8156 | ||
a70d6342 IR |
8157 | if (bb_vinfo) |
8158 | return true; | |
8159 | ||
8644a673 IR |
8160 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
8161 | need extra handling, except for vectorizable reductions. */ | |
8162 | if (STMT_VINFO_LIVE_P (stmt_info) | |
8163 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
8164 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 8165 | |
8644a673 | 8166 | if (!ok) |
ebfd146a | 8167 | { |
73fbfcad | 8168 | if (dump_enabled_p ()) |
8644a673 | 8169 | { |
78c60e3d SS |
8170 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8171 | "not vectorized: live stmt not "); | |
8172 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
8173 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 8174 | } |
b8698a0f | 8175 | |
8644a673 | 8176 | return false; |
ebfd146a IR |
8177 | } |
8178 | ||
ebfd146a IR |
8179 | return true; |
8180 | } | |
8181 | ||
8182 | ||
8183 | /* Function vect_transform_stmt. | |
8184 | ||
8185 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
8186 | ||
8187 | bool | |
355fe088 | 8188 | vect_transform_stmt (gimple *stmt, gimple_stmt_iterator *gsi, |
0d0293ac | 8189 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
8190 | slp_instance slp_node_instance) |
8191 | { | |
8192 | bool is_store = false; | |
355fe088 | 8193 | gimple *vec_stmt = NULL; |
ebfd146a | 8194 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
ebfd146a | 8195 | bool done; |
ebfd146a | 8196 | |
355fe088 | 8197 | gimple *old_vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
225ce44b | 8198 | |
ebfd146a IR |
8199 | switch (STMT_VINFO_TYPE (stmt_info)) |
8200 | { | |
8201 | case type_demotion_vec_info_type: | |
ebfd146a | 8202 | case type_promotion_vec_info_type: |
ebfd146a IR |
8203 | case type_conversion_vec_info_type: |
8204 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
8205 | gcc_assert (done); | |
8206 | break; | |
8207 | ||
8208 | case induc_vec_info_type: | |
8209 | gcc_assert (!slp_node); | |
8210 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
8211 | gcc_assert (done); | |
8212 | break; | |
8213 | ||
9dc3f7de IR |
8214 | case shift_vec_info_type: |
8215 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
8216 | gcc_assert (done); | |
8217 | break; | |
8218 | ||
ebfd146a IR |
8219 | case op_vec_info_type: |
8220 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
8221 | gcc_assert (done); | |
8222 | break; | |
8223 | ||
8224 | case assignment_vec_info_type: | |
8225 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
8226 | gcc_assert (done); | |
8227 | break; | |
8228 | ||
8229 | case load_vec_info_type: | |
b8698a0f | 8230 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
8231 | slp_node_instance); |
8232 | gcc_assert (done); | |
8233 | break; | |
8234 | ||
8235 | case store_vec_info_type: | |
8236 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
8237 | gcc_assert (done); | |
0d0293ac | 8238 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
8239 | { |
8240 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 8241 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
8242 | one are skipped, and there vec_stmt_info shouldn't be freed |
8243 | meanwhile. */ | |
0d0293ac | 8244 | *grouped_store = true; |
ebfd146a IR |
8245 | if (STMT_VINFO_VEC_STMT (stmt_info)) |
8246 | is_store = true; | |
8247 | } | |
8248 | else | |
8249 | is_store = true; | |
8250 | break; | |
8251 | ||
8252 | case condition_vec_info_type: | |
f7e531cf | 8253 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
8254 | gcc_assert (done); |
8255 | break; | |
8256 | ||
42fd8198 IE |
8257 | case comparison_vec_info_type: |
8258 | done = vectorizable_comparison (stmt, gsi, &vec_stmt, NULL, slp_node); | |
8259 | gcc_assert (done); | |
8260 | break; | |
8261 | ||
ebfd146a | 8262 | case call_vec_info_type: |
190c2236 | 8263 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 8264 | stmt = gsi_stmt (*gsi); |
5ce9450f JJ |
8265 | if (is_gimple_call (stmt) |
8266 | && gimple_call_internal_p (stmt) | |
8267 | && gimple_call_internal_fn (stmt) == IFN_MASK_STORE) | |
8268 | is_store = true; | |
ebfd146a IR |
8269 | break; |
8270 | ||
0136f8f0 AH |
8271 | case call_simd_clone_vec_info_type: |
8272 | done = vectorizable_simd_clone_call (stmt, gsi, &vec_stmt, slp_node); | |
8273 | stmt = gsi_stmt (*gsi); | |
8274 | break; | |
8275 | ||
ebfd146a | 8276 | case reduc_vec_info_type: |
b5aeb3bb | 8277 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
8278 | gcc_assert (done); |
8279 | break; | |
8280 | ||
8281 | default: | |
8282 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
8283 | { | |
73fbfcad | 8284 | if (dump_enabled_p ()) |
78c60e3d | 8285 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8286 | "stmt not supported.\n"); |
ebfd146a IR |
8287 | gcc_unreachable (); |
8288 | } | |
8289 | } | |
8290 | ||
225ce44b RB |
8291 | /* Verify SLP vectorization doesn't mess with STMT_VINFO_VEC_STMT. |
8292 | This would break hybrid SLP vectorization. */ | |
8293 | if (slp_node) | |
d90f8440 RB |
8294 | gcc_assert (!vec_stmt |
8295 | && STMT_VINFO_VEC_STMT (stmt_info) == old_vec_stmt); | |
225ce44b | 8296 | |
ebfd146a IR |
8297 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that |
8298 | is being vectorized, but outside the immediately enclosing loop. */ | |
8299 | if (vec_stmt | |
a70d6342 IR |
8300 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
8301 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
8302 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
8303 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
8304 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 8305 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 8306 | vect_used_in_outer_by_reduction)) |
ebfd146a | 8307 | { |
a70d6342 IR |
8308 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
8309 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
8310 | imm_use_iterator imm_iter; |
8311 | use_operand_p use_p; | |
8312 | tree scalar_dest; | |
355fe088 | 8313 | gimple *exit_phi; |
ebfd146a | 8314 | |
73fbfcad | 8315 | if (dump_enabled_p ()) |
78c60e3d | 8316 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 8317 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
8318 | |
8319 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
8320 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
8321 | STMT). */ | |
8322 | if (gimple_code (stmt) == GIMPLE_PHI) | |
8323 | scalar_dest = PHI_RESULT (stmt); | |
8324 | else | |
8325 | scalar_dest = gimple_assign_lhs (stmt); | |
8326 | ||
8327 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
8328 | { | |
8329 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
8330 | { | |
8331 | exit_phi = USE_STMT (use_p); | |
8332 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
8333 | } | |
8334 | } | |
8335 | } | |
8336 | ||
8337 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
8338 | being vectorized. */ | |
8339 | if (STMT_VINFO_LIVE_P (stmt_info) | |
8340 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
8341 | { | |
8342 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
8343 | gcc_assert (done); | |
8344 | } | |
8345 | ||
8346 | if (vec_stmt) | |
83197f37 | 8347 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 8348 | |
b8698a0f | 8349 | return is_store; |
ebfd146a IR |
8350 | } |
8351 | ||
8352 | ||
b8698a0f | 8353 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
8354 | stmt_vec_info. */ |
8355 | ||
8356 | void | |
355fe088 | 8357 | vect_remove_stores (gimple *first_stmt) |
ebfd146a | 8358 | { |
355fe088 TS |
8359 | gimple *next = first_stmt; |
8360 | gimple *tmp; | |
ebfd146a IR |
8361 | gimple_stmt_iterator next_si; |
8362 | ||
8363 | while (next) | |
8364 | { | |
78048b1c JJ |
8365 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
8366 | ||
8367 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
8368 | if (is_pattern_stmt_p (stmt_info)) | |
8369 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
8370 | /* Free the attached stmt_vec_info and remove the stmt. */ |
8371 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 8372 | unlink_stmt_vdef (next); |
ebfd146a | 8373 | gsi_remove (&next_si, true); |
3d3f2249 | 8374 | release_defs (next); |
ebfd146a IR |
8375 | free_stmt_vec_info (next); |
8376 | next = tmp; | |
8377 | } | |
8378 | } | |
8379 | ||
8380 | ||
8381 | /* Function new_stmt_vec_info. | |
8382 | ||
8383 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
8384 | ||
8385 | stmt_vec_info | |
310213d4 | 8386 | new_stmt_vec_info (gimple *stmt, vec_info *vinfo) |
ebfd146a IR |
8387 | { |
8388 | stmt_vec_info res; | |
8389 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
8390 | ||
8391 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
8392 | STMT_VINFO_STMT (res) = stmt; | |
310213d4 | 8393 | res->vinfo = vinfo; |
8644a673 | 8394 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
8395 | STMT_VINFO_LIVE_P (res) = false; |
8396 | STMT_VINFO_VECTYPE (res) = NULL; | |
8397 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 8398 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
8399 | STMT_VINFO_IN_PATTERN_P (res) = false; |
8400 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 8401 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a | 8402 | STMT_VINFO_DATA_REF (res) = NULL; |
af29617a | 8403 | STMT_VINFO_VEC_REDUCTION_TYPE (res) = TREE_CODE_REDUCTION; |
ebfd146a IR |
8404 | |
8405 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
8406 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
8407 | STMT_VINFO_DR_INIT (res) = NULL; | |
8408 | STMT_VINFO_DR_STEP (res) = NULL; | |
8409 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
8410 | ||
8411 | if (gimple_code (stmt) == GIMPLE_PHI | |
8412 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
8413 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
8414 | else | |
8644a673 IR |
8415 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
8416 | ||
9771b263 | 8417 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 8418 | STMT_SLP_TYPE (res) = loop_vect; |
78810bd3 RB |
8419 | STMT_VINFO_NUM_SLP_USES (res) = 0; |
8420 | ||
e14c1050 IR |
8421 | GROUP_FIRST_ELEMENT (res) = NULL; |
8422 | GROUP_NEXT_ELEMENT (res) = NULL; | |
8423 | GROUP_SIZE (res) = 0; | |
8424 | GROUP_STORE_COUNT (res) = 0; | |
8425 | GROUP_GAP (res) = 0; | |
8426 | GROUP_SAME_DR_STMT (res) = NULL; | |
ebfd146a IR |
8427 | |
8428 | return res; | |
8429 | } | |
8430 | ||
8431 | ||
8432 | /* Create a hash table for stmt_vec_info. */ | |
8433 | ||
8434 | void | |
8435 | init_stmt_vec_info_vec (void) | |
8436 | { | |
9771b263 DN |
8437 | gcc_assert (!stmt_vec_info_vec.exists ()); |
8438 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
8439 | } |
8440 | ||
8441 | ||
8442 | /* Free hash table for stmt_vec_info. */ | |
8443 | ||
8444 | void | |
8445 | free_stmt_vec_info_vec (void) | |
8446 | { | |
93675444 | 8447 | unsigned int i; |
3161455c | 8448 | stmt_vec_info info; |
93675444 JJ |
8449 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) |
8450 | if (info != NULL) | |
3161455c | 8451 | free_stmt_vec_info (STMT_VINFO_STMT (info)); |
9771b263 DN |
8452 | gcc_assert (stmt_vec_info_vec.exists ()); |
8453 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
8454 | } |
8455 | ||
8456 | ||
8457 | /* Free stmt vectorization related info. */ | |
8458 | ||
8459 | void | |
355fe088 | 8460 | free_stmt_vec_info (gimple *stmt) |
ebfd146a IR |
8461 | { |
8462 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8463 | ||
8464 | if (!stmt_info) | |
8465 | return; | |
8466 | ||
78048b1c JJ |
8467 | /* Check if this statement has a related "pattern stmt" |
8468 | (introduced by the vectorizer during the pattern recognition | |
8469 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
8470 | too. */ | |
8471 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
8472 | { | |
8473 | stmt_vec_info patt_info | |
8474 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
8475 | if (patt_info) | |
8476 | { | |
363477c0 | 8477 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
355fe088 | 8478 | gimple *patt_stmt = STMT_VINFO_STMT (patt_info); |
f0281fde RB |
8479 | gimple_set_bb (patt_stmt, NULL); |
8480 | tree lhs = gimple_get_lhs (patt_stmt); | |
e6f5c25d | 8481 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde | 8482 | release_ssa_name (lhs); |
363477c0 JJ |
8483 | if (seq) |
8484 | { | |
8485 | gimple_stmt_iterator si; | |
8486 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
f0281fde | 8487 | { |
355fe088 | 8488 | gimple *seq_stmt = gsi_stmt (si); |
f0281fde | 8489 | gimple_set_bb (seq_stmt, NULL); |
7532abf2 | 8490 | lhs = gimple_get_lhs (seq_stmt); |
e6f5c25d | 8491 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde RB |
8492 | release_ssa_name (lhs); |
8493 | free_stmt_vec_info (seq_stmt); | |
8494 | } | |
363477c0 | 8495 | } |
f0281fde | 8496 | free_stmt_vec_info (patt_stmt); |
78048b1c JJ |
8497 | } |
8498 | } | |
8499 | ||
9771b263 | 8500 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
6c9e85fb | 8501 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).release (); |
ebfd146a IR |
8502 | set_vinfo_for_stmt (stmt, NULL); |
8503 | free (stmt_info); | |
8504 | } | |
8505 | ||
8506 | ||
bb67d9c7 | 8507 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 8508 | |
bb67d9c7 | 8509 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
8510 | by the target. */ |
8511 | ||
bb67d9c7 RG |
8512 | static tree |
8513 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a | 8514 | { |
ef4bddc2 RS |
8515 | machine_mode inner_mode = TYPE_MODE (scalar_type); |
8516 | machine_mode simd_mode; | |
2f816591 | 8517 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
8518 | int nunits; |
8519 | tree vectype; | |
8520 | ||
cc4b5170 | 8521 | if (nbytes == 0) |
ebfd146a IR |
8522 | return NULL_TREE; |
8523 | ||
48f2e373 RB |
8524 | if (GET_MODE_CLASS (inner_mode) != MODE_INT |
8525 | && GET_MODE_CLASS (inner_mode) != MODE_FLOAT) | |
8526 | return NULL_TREE; | |
8527 | ||
7b7b1813 RG |
8528 | /* For vector types of elements whose mode precision doesn't |
8529 | match their types precision we use a element type of mode | |
8530 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
8531 | they support the proper result truncation/extension. |
8532 | We also make sure to build vector types with INTEGER_TYPE | |
8533 | component type only. */ | |
6d7971b8 | 8534 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
8535 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
8536 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
8537 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
8538 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 8539 | |
ccbf5bb4 RG |
8540 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
8541 | When the component mode passes the above test simply use a type | |
8542 | corresponding to that mode. The theory is that any use that | |
8543 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 8544 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 8545 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
8546 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
8547 | ||
8548 | /* We can't build a vector type of elements with alignment bigger than | |
8549 | their size. */ | |
dfc2e2ac | 8550 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
8551 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
8552 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 8553 | |
dfc2e2ac RB |
8554 | /* If we felt back to using the mode fail if there was |
8555 | no scalar type for it. */ | |
8556 | if (scalar_type == NULL_TREE) | |
8557 | return NULL_TREE; | |
8558 | ||
bb67d9c7 RG |
8559 | /* If no size was supplied use the mode the target prefers. Otherwise |
8560 | lookup a vector mode of the specified size. */ | |
8561 | if (size == 0) | |
8562 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
8563 | else | |
8564 | simd_mode = mode_for_vector (inner_mode, size / nbytes); | |
cc4b5170 RG |
8565 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
8566 | if (nunits <= 1) | |
8567 | return NULL_TREE; | |
ebfd146a IR |
8568 | |
8569 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
8570 | |
8571 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
8572 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 8573 | return NULL_TREE; |
ebfd146a IR |
8574 | |
8575 | return vectype; | |
8576 | } | |
8577 | ||
bb67d9c7 RG |
8578 | unsigned int current_vector_size; |
8579 | ||
8580 | /* Function get_vectype_for_scalar_type. | |
8581 | ||
8582 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
8583 | by the target. */ | |
8584 | ||
8585 | tree | |
8586 | get_vectype_for_scalar_type (tree scalar_type) | |
8587 | { | |
8588 | tree vectype; | |
8589 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
8590 | current_vector_size); | |
8591 | if (vectype | |
8592 | && current_vector_size == 0) | |
8593 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
8594 | return vectype; | |
8595 | } | |
8596 | ||
42fd8198 IE |
8597 | /* Function get_mask_type_for_scalar_type. |
8598 | ||
8599 | Returns the mask type corresponding to a result of comparison | |
8600 | of vectors of specified SCALAR_TYPE as supported by target. */ | |
8601 | ||
8602 | tree | |
8603 | get_mask_type_for_scalar_type (tree scalar_type) | |
8604 | { | |
8605 | tree vectype = get_vectype_for_scalar_type (scalar_type); | |
8606 | ||
8607 | if (!vectype) | |
8608 | return NULL; | |
8609 | ||
8610 | return build_truth_vector_type (TYPE_VECTOR_SUBPARTS (vectype), | |
8611 | current_vector_size); | |
8612 | } | |
8613 | ||
b690cc0f RG |
8614 | /* Function get_same_sized_vectype |
8615 | ||
8616 | Returns a vector type corresponding to SCALAR_TYPE of size | |
8617 | VECTOR_TYPE if supported by the target. */ | |
8618 | ||
8619 | tree | |
bb67d9c7 | 8620 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 8621 | { |
9f47c7e5 IE |
8622 | if (TREE_CODE (scalar_type) == BOOLEAN_TYPE) |
8623 | return build_same_sized_truth_vector_type (vector_type); | |
8624 | ||
bb67d9c7 RG |
8625 | return get_vectype_for_scalar_type_and_size |
8626 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
8627 | } |
8628 | ||
ebfd146a IR |
8629 | /* Function vect_is_simple_use. |
8630 | ||
8631 | Input: | |
81c40241 RB |
8632 | VINFO - the vect info of the loop or basic block that is being vectorized. |
8633 | OPERAND - operand in the loop or bb. | |
8634 | Output: | |
8635 | DEF_STMT - the defining stmt in case OPERAND is an SSA_NAME. | |
8636 | DT - the type of definition | |
ebfd146a IR |
8637 | |
8638 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 8639 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 8640 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 8641 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
8642 | is the case in reduction/induction computations). |
8643 | For basic blocks, supportable operands are constants and bb invariants. | |
8644 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
8645 | |
8646 | bool | |
81c40241 RB |
8647 | vect_is_simple_use (tree operand, vec_info *vinfo, |
8648 | gimple **def_stmt, enum vect_def_type *dt) | |
b8698a0f | 8649 | { |
ebfd146a | 8650 | *def_stmt = NULL; |
3fc356dc | 8651 | *dt = vect_unknown_def_type; |
b8698a0f | 8652 | |
73fbfcad | 8653 | if (dump_enabled_p ()) |
ebfd146a | 8654 | { |
78c60e3d SS |
8655 | dump_printf_loc (MSG_NOTE, vect_location, |
8656 | "vect_is_simple_use: operand "); | |
8657 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
e645e942 | 8658 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 8659 | } |
b8698a0f | 8660 | |
b758f602 | 8661 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
8662 | { |
8663 | *dt = vect_constant_def; | |
8664 | return true; | |
8665 | } | |
b8698a0f | 8666 | |
ebfd146a IR |
8667 | if (is_gimple_min_invariant (operand)) |
8668 | { | |
8644a673 | 8669 | *dt = vect_external_def; |
ebfd146a IR |
8670 | return true; |
8671 | } | |
8672 | ||
ebfd146a IR |
8673 | if (TREE_CODE (operand) != SSA_NAME) |
8674 | { | |
73fbfcad | 8675 | if (dump_enabled_p ()) |
af29617a AH |
8676 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8677 | "not ssa-name.\n"); | |
ebfd146a IR |
8678 | return false; |
8679 | } | |
b8698a0f | 8680 | |
3fc356dc | 8681 | if (SSA_NAME_IS_DEFAULT_DEF (operand)) |
ebfd146a | 8682 | { |
3fc356dc RB |
8683 | *dt = vect_external_def; |
8684 | return true; | |
ebfd146a IR |
8685 | } |
8686 | ||
3fc356dc | 8687 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
73fbfcad | 8688 | if (dump_enabled_p ()) |
ebfd146a | 8689 | { |
78c60e3d SS |
8690 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
8691 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
ebfd146a IR |
8692 | } |
8693 | ||
61d371eb | 8694 | if (! vect_stmt_in_region_p (vinfo, *def_stmt)) |
8644a673 | 8695 | *dt = vect_external_def; |
ebfd146a IR |
8696 | else |
8697 | { | |
3fc356dc | 8698 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (*def_stmt); |
603cca93 | 8699 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
ebfd146a IR |
8700 | } |
8701 | ||
2e8ab70c RB |
8702 | if (dump_enabled_p ()) |
8703 | { | |
8704 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: "); | |
8705 | switch (*dt) | |
8706 | { | |
8707 | case vect_uninitialized_def: | |
8708 | dump_printf (MSG_NOTE, "uninitialized\n"); | |
8709 | break; | |
8710 | case vect_constant_def: | |
8711 | dump_printf (MSG_NOTE, "constant\n"); | |
8712 | break; | |
8713 | case vect_external_def: | |
8714 | dump_printf (MSG_NOTE, "external\n"); | |
8715 | break; | |
8716 | case vect_internal_def: | |
8717 | dump_printf (MSG_NOTE, "internal\n"); | |
8718 | break; | |
8719 | case vect_induction_def: | |
8720 | dump_printf (MSG_NOTE, "induction\n"); | |
8721 | break; | |
8722 | case vect_reduction_def: | |
8723 | dump_printf (MSG_NOTE, "reduction\n"); | |
8724 | break; | |
8725 | case vect_double_reduction_def: | |
8726 | dump_printf (MSG_NOTE, "double reduction\n"); | |
8727 | break; | |
8728 | case vect_nested_cycle: | |
8729 | dump_printf (MSG_NOTE, "nested cycle\n"); | |
8730 | break; | |
8731 | case vect_unknown_def_type: | |
8732 | dump_printf (MSG_NOTE, "unknown\n"); | |
8733 | break; | |
8734 | } | |
8735 | } | |
8736 | ||
81c40241 | 8737 | if (*dt == vect_unknown_def_type) |
ebfd146a | 8738 | { |
73fbfcad | 8739 | if (dump_enabled_p ()) |
78c60e3d | 8740 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8741 | "Unsupported pattern.\n"); |
ebfd146a IR |
8742 | return false; |
8743 | } | |
8744 | ||
ebfd146a IR |
8745 | switch (gimple_code (*def_stmt)) |
8746 | { | |
8747 | case GIMPLE_PHI: | |
ebfd146a | 8748 | case GIMPLE_ASSIGN: |
ebfd146a | 8749 | case GIMPLE_CALL: |
81c40241 | 8750 | break; |
ebfd146a | 8751 | default: |
73fbfcad | 8752 | if (dump_enabled_p ()) |
78c60e3d | 8753 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8754 | "unsupported defining stmt:\n"); |
ebfd146a IR |
8755 | return false; |
8756 | } | |
8757 | ||
8758 | return true; | |
8759 | } | |
8760 | ||
81c40241 | 8761 | /* Function vect_is_simple_use. |
b690cc0f | 8762 | |
81c40241 | 8763 | Same as vect_is_simple_use but also determines the vector operand |
b690cc0f RG |
8764 | type of OPERAND and stores it to *VECTYPE. If the definition of |
8765 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
8766 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
8767 | is responsible to compute the best suited vector type for the | |
8768 | scalar operand. */ | |
8769 | ||
8770 | bool | |
81c40241 RB |
8771 | vect_is_simple_use (tree operand, vec_info *vinfo, |
8772 | gimple **def_stmt, enum vect_def_type *dt, tree *vectype) | |
b690cc0f | 8773 | { |
81c40241 | 8774 | if (!vect_is_simple_use (operand, vinfo, def_stmt, dt)) |
b690cc0f RG |
8775 | return false; |
8776 | ||
8777 | /* Now get a vector type if the def is internal, otherwise supply | |
8778 | NULL_TREE and leave it up to the caller to figure out a proper | |
8779 | type for the use stmt. */ | |
8780 | if (*dt == vect_internal_def | |
8781 | || *dt == vect_induction_def | |
8782 | || *dt == vect_reduction_def | |
8783 | || *dt == vect_double_reduction_def | |
8784 | || *dt == vect_nested_cycle) | |
8785 | { | |
8786 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
8787 | |
8788 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
8789 | && !STMT_VINFO_RELEVANT (stmt_info) | |
8790 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 8791 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 8792 | |
b690cc0f RG |
8793 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
8794 | gcc_assert (*vectype != NULL_TREE); | |
8795 | } | |
8796 | else if (*dt == vect_uninitialized_def | |
8797 | || *dt == vect_constant_def | |
8798 | || *dt == vect_external_def) | |
8799 | *vectype = NULL_TREE; | |
8800 | else | |
8801 | gcc_unreachable (); | |
8802 | ||
8803 | return true; | |
8804 | } | |
8805 | ||
ebfd146a IR |
8806 | |
8807 | /* Function supportable_widening_operation | |
8808 | ||
b8698a0f L |
8809 | Check whether an operation represented by the code CODE is a |
8810 | widening operation that is supported by the target platform in | |
b690cc0f RG |
8811 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
8812 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 8813 | |
ebfd146a IR |
8814 | Widening operations we currently support are NOP (CONVERT), FLOAT |
8815 | and WIDEN_MULT. This function checks if these operations are supported | |
8816 | by the target platform either directly (via vector tree-codes), or via | |
8817 | target builtins. | |
8818 | ||
8819 | Output: | |
b8698a0f L |
8820 | - CODE1 and CODE2 are codes of vector operations to be used when |
8821 | vectorizing the operation, if available. | |
ebfd146a IR |
8822 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
8823 | case of multi-step conversion (like char->short->int - in that case | |
8824 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
8825 | - INTERM_TYPES contains the intermediate type required to perform the |
8826 | widening operation (short in the above example). */ | |
ebfd146a IR |
8827 | |
8828 | bool | |
355fe088 | 8829 | supportable_widening_operation (enum tree_code code, gimple *stmt, |
b690cc0f | 8830 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
8831 | enum tree_code *code1, enum tree_code *code2, |
8832 | int *multi_step_cvt, | |
9771b263 | 8833 | vec<tree> *interm_types) |
ebfd146a IR |
8834 | { |
8835 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8836 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 8837 | struct loop *vect_loop = NULL; |
ef4bddc2 | 8838 | machine_mode vec_mode; |
81f40b79 | 8839 | enum insn_code icode1, icode2; |
ebfd146a | 8840 | optab optab1, optab2; |
b690cc0f RG |
8841 | tree vectype = vectype_in; |
8842 | tree wide_vectype = vectype_out; | |
ebfd146a | 8843 | enum tree_code c1, c2; |
4a00c761 JJ |
8844 | int i; |
8845 | tree prev_type, intermediate_type; | |
ef4bddc2 | 8846 | machine_mode intermediate_mode, prev_mode; |
4a00c761 | 8847 | optab optab3, optab4; |
ebfd146a | 8848 | |
4a00c761 | 8849 | *multi_step_cvt = 0; |
4ef69dfc IR |
8850 | if (loop_info) |
8851 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
8852 | ||
ebfd146a IR |
8853 | switch (code) |
8854 | { | |
8855 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
8856 | /* The result of a vectorized widening operation usually requires |
8857 | two vectors (because the widened results do not fit into one vector). | |
8858 | The generated vector results would normally be expected to be | |
8859 | generated in the same order as in the original scalar computation, | |
8860 | i.e. if 8 results are generated in each vector iteration, they are | |
8861 | to be organized as follows: | |
8862 | vect1: [res1,res2,res3,res4], | |
8863 | vect2: [res5,res6,res7,res8]. | |
8864 | ||
8865 | However, in the special case that the result of the widening | |
8866 | operation is used in a reduction computation only, the order doesn't | |
8867 | matter (because when vectorizing a reduction we change the order of | |
8868 | the computation). Some targets can take advantage of this and | |
8869 | generate more efficient code. For example, targets like Altivec, | |
8870 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
8871 | generate the following vectors: | |
8872 | vect1: [res1,res3,res5,res7], | |
8873 | vect2: [res2,res4,res6,res8]. | |
8874 | ||
8875 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
8876 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
8877 | iterations in parallel). We therefore don't allow to change the | |
8878 | order of the computation in the inner-loop during outer-loop | |
8879 | vectorization. */ | |
8880 | /* TODO: Another case in which order doesn't *really* matter is when we | |
8881 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
8882 | Normally, pack_trunc performs an even/odd permute, whereas the | |
8883 | repack from an even/odd expansion would be an interleave, which | |
8884 | would be significantly simpler for e.g. AVX2. */ | |
8885 | /* In any case, in order to avoid duplicating the code below, recurse | |
8886 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
8887 | are properly set up for the caller. If we fail, we'll continue with | |
8888 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
8889 | if (vect_loop | |
8890 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
8891 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
8892 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
8893 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
8894 | code1, code2, multi_step_cvt, |
8895 | interm_types)) | |
ebc047a2 CH |
8896 | { |
8897 | /* Elements in a vector with vect_used_by_reduction property cannot | |
8898 | be reordered if the use chain with this property does not have the | |
8899 | same operation. One such an example is s += a * b, where elements | |
8900 | in a and b cannot be reordered. Here we check if the vector defined | |
8901 | by STMT is only directly used in the reduction statement. */ | |
8902 | tree lhs = gimple_assign_lhs (stmt); | |
8903 | use_operand_p dummy; | |
355fe088 | 8904 | gimple *use_stmt; |
ebc047a2 CH |
8905 | stmt_vec_info use_stmt_info = NULL; |
8906 | if (single_imm_use (lhs, &dummy, &use_stmt) | |
8907 | && (use_stmt_info = vinfo_for_stmt (use_stmt)) | |
8908 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) | |
8909 | return true; | |
8910 | } | |
4a00c761 JJ |
8911 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
8912 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
8913 | break; |
8914 | ||
81c40241 RB |
8915 | case DOT_PROD_EXPR: |
8916 | c1 = DOT_PROD_EXPR; | |
8917 | c2 = DOT_PROD_EXPR; | |
8918 | break; | |
8919 | ||
8920 | case SAD_EXPR: | |
8921 | c1 = SAD_EXPR; | |
8922 | c2 = SAD_EXPR; | |
8923 | break; | |
8924 | ||
6ae6116f RH |
8925 | case VEC_WIDEN_MULT_EVEN_EXPR: |
8926 | /* Support the recursion induced just above. */ | |
8927 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
8928 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
8929 | break; | |
8930 | ||
36ba4aae | 8931 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
8932 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
8933 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
8934 | break; |
8935 | ||
ebfd146a | 8936 | CASE_CONVERT: |
4a00c761 JJ |
8937 | c1 = VEC_UNPACK_LO_EXPR; |
8938 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
8939 | break; |
8940 | ||
8941 | case FLOAT_EXPR: | |
4a00c761 JJ |
8942 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
8943 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
8944 | break; |
8945 | ||
8946 | case FIX_TRUNC_EXPR: | |
8947 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
8948 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
8949 | computing the operation. */ | |
8950 | return false; | |
8951 | ||
8952 | default: | |
8953 | gcc_unreachable (); | |
8954 | } | |
8955 | ||
6ae6116f | 8956 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
6b4db501 | 8957 | std::swap (c1, c2); |
4a00c761 | 8958 | |
ebfd146a IR |
8959 | if (code == FIX_TRUNC_EXPR) |
8960 | { | |
8961 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
8962 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
8963 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
8964 | } |
8965 | else | |
8966 | { | |
8967 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
8968 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
8969 | } | |
8970 | ||
8971 | if (!optab1 || !optab2) | |
8972 | return false; | |
8973 | ||
8974 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
8975 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
8976 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
8977 | return false; |
8978 | ||
4a00c761 JJ |
8979 | *code1 = c1; |
8980 | *code2 = c2; | |
8981 | ||
8982 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
8983 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
8984 | return true; | |
8985 | ||
b8698a0f | 8986 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 8987 | types. */ |
ebfd146a | 8988 | |
4a00c761 JJ |
8989 | prev_type = vectype; |
8990 | prev_mode = vec_mode; | |
b8698a0f | 8991 | |
4a00c761 JJ |
8992 | if (!CONVERT_EXPR_CODE_P (code)) |
8993 | return false; | |
b8698a0f | 8994 | |
4a00c761 JJ |
8995 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
8996 | intermediate steps in promotion sequence. We try | |
8997 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
8998 | not. */ | |
9771b263 | 8999 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
9000 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
9001 | { | |
9002 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
9003 | intermediate_type | |
9004 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
9005 | TYPE_UNSIGNED (prev_type)); | |
9006 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); | |
9007 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
9008 | ||
9009 | if (!optab3 || !optab4 | |
9010 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
9011 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
9012 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
9013 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
9014 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
9015 | == CODE_FOR_nothing) | |
9016 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
9017 | == CODE_FOR_nothing)) | |
9018 | break; | |
ebfd146a | 9019 | |
9771b263 | 9020 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
9021 | (*multi_step_cvt)++; |
9022 | ||
9023 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
9024 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
9025 | return true; | |
9026 | ||
9027 | prev_type = intermediate_type; | |
9028 | prev_mode = intermediate_mode; | |
ebfd146a IR |
9029 | } |
9030 | ||
9771b263 | 9031 | interm_types->release (); |
4a00c761 | 9032 | return false; |
ebfd146a IR |
9033 | } |
9034 | ||
9035 | ||
9036 | /* Function supportable_narrowing_operation | |
9037 | ||
b8698a0f L |
9038 | Check whether an operation represented by the code CODE is a |
9039 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
9040 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
9041 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 9042 | |
ebfd146a | 9043 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 9044 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
9045 | the target platform directly via vector tree-codes. |
9046 | ||
9047 | Output: | |
b8698a0f L |
9048 | - CODE1 is the code of a vector operation to be used when |
9049 | vectorizing the operation, if available. | |
ebfd146a IR |
9050 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
9051 | case of multi-step conversion (like int->short->char - in that case | |
9052 | MULTI_STEP_CVT will be 1). | |
9053 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 9054 | narrowing operation (short in the above example). */ |
ebfd146a IR |
9055 | |
9056 | bool | |
9057 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 9058 | tree vectype_out, tree vectype_in, |
ebfd146a | 9059 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 9060 | vec<tree> *interm_types) |
ebfd146a | 9061 | { |
ef4bddc2 | 9062 | machine_mode vec_mode; |
ebfd146a IR |
9063 | enum insn_code icode1; |
9064 | optab optab1, interm_optab; | |
b690cc0f RG |
9065 | tree vectype = vectype_in; |
9066 | tree narrow_vectype = vectype_out; | |
ebfd146a | 9067 | enum tree_code c1; |
4a00c761 | 9068 | tree intermediate_type; |
ef4bddc2 | 9069 | machine_mode intermediate_mode, prev_mode; |
ebfd146a | 9070 | int i; |
4a00c761 | 9071 | bool uns; |
ebfd146a | 9072 | |
4a00c761 | 9073 | *multi_step_cvt = 0; |
ebfd146a IR |
9074 | switch (code) |
9075 | { | |
9076 | CASE_CONVERT: | |
9077 | c1 = VEC_PACK_TRUNC_EXPR; | |
9078 | break; | |
9079 | ||
9080 | case FIX_TRUNC_EXPR: | |
9081 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
9082 | break; | |
9083 | ||
9084 | case FLOAT_EXPR: | |
9085 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
9086 | tree code and optabs used for computing the operation. */ | |
9087 | return false; | |
9088 | ||
9089 | default: | |
9090 | gcc_unreachable (); | |
9091 | } | |
9092 | ||
9093 | if (code == FIX_TRUNC_EXPR) | |
9094 | /* The signedness is determined from output operand. */ | |
b690cc0f | 9095 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
9096 | else |
9097 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
9098 | ||
9099 | if (!optab1) | |
9100 | return false; | |
9101 | ||
9102 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 9103 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
9104 | return false; |
9105 | ||
4a00c761 JJ |
9106 | *code1 = c1; |
9107 | ||
9108 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
9109 | return true; | |
9110 | ||
ebfd146a IR |
9111 | /* Check if it's a multi-step conversion that can be done using intermediate |
9112 | types. */ | |
4a00c761 JJ |
9113 | prev_mode = vec_mode; |
9114 | if (code == FIX_TRUNC_EXPR) | |
9115 | uns = TYPE_UNSIGNED (vectype_out); | |
9116 | else | |
9117 | uns = TYPE_UNSIGNED (vectype); | |
9118 | ||
9119 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
9120 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
9121 | costly than signed. */ | |
9122 | if (code == FIX_TRUNC_EXPR && uns) | |
9123 | { | |
9124 | enum insn_code icode2; | |
9125 | ||
9126 | intermediate_type | |
9127 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
9128 | interm_optab | |
9129 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 9130 | if (interm_optab != unknown_optab |
4a00c761 JJ |
9131 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
9132 | && insn_data[icode1].operand[0].mode | |
9133 | == insn_data[icode2].operand[0].mode) | |
9134 | { | |
9135 | uns = false; | |
9136 | optab1 = interm_optab; | |
9137 | icode1 = icode2; | |
9138 | } | |
9139 | } | |
ebfd146a | 9140 | |
4a00c761 JJ |
9141 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
9142 | intermediate steps in promotion sequence. We try | |
9143 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 9144 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
9145 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
9146 | { | |
9147 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
9148 | intermediate_type | |
9149 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
9150 | interm_optab | |
9151 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
9152 | optab_default); | |
9153 | if (!interm_optab | |
9154 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
9155 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
9156 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
9157 | == CODE_FOR_nothing)) | |
9158 | break; | |
9159 | ||
9771b263 | 9160 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
9161 | (*multi_step_cvt)++; |
9162 | ||
9163 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
9164 | return true; | |
9165 | ||
9166 | prev_mode = intermediate_mode; | |
9167 | optab1 = interm_optab; | |
ebfd146a IR |
9168 | } |
9169 | ||
9771b263 | 9170 | interm_types->release (); |
4a00c761 | 9171 | return false; |
ebfd146a | 9172 | } |