]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
d1e082c2 | 2 | Copyright (C) 2003-2013 Free Software Foundation, Inc. |
b8698a0f | 3 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
4 | and Ira Rosen <irar@il.ibm.com> |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 3, or (at your option) any later | |
11 | version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
78c60e3d | 25 | #include "dumpfile.h" |
ebfd146a IR |
26 | #include "tm.h" |
27 | #include "ggc.h" | |
28 | #include "tree.h" | |
29 | #include "target.h" | |
30 | #include "basic-block.h" | |
cf835838 | 31 | #include "gimple-pretty-print.h" |
ebfd146a | 32 | #include "tree-flow.h" |
ebfd146a | 33 | #include "cfgloop.h" |
ebfd146a | 34 | #include "expr.h" |
7ee2468b | 35 | #include "recog.h" /* FIXME: for insn_data */ |
ebfd146a | 36 | #include "optabs.h" |
718f9c0f | 37 | #include "diagnostic-core.h" |
ebfd146a | 38 | #include "tree-vectorizer.h" |
7ee2468b | 39 | #include "dumpfile.h" |
ebfd146a | 40 | |
7ee2468b SB |
41 | /* For lang_hooks.types.type_for_mode. */ |
42 | #include "langhooks.h" | |
ebfd146a | 43 | |
c3e7ee41 BS |
44 | /* Return the vectorized type for the given statement. */ |
45 | ||
46 | tree | |
47 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
48 | { | |
49 | return STMT_VINFO_VECTYPE (stmt_info); | |
50 | } | |
51 | ||
52 | /* Return TRUE iff the given statement is in an inner loop relative to | |
53 | the loop being vectorized. */ | |
54 | bool | |
55 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
56 | { | |
57 | gimple stmt = STMT_VINFO_STMT (stmt_info); | |
58 | basic_block bb = gimple_bb (stmt); | |
59 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
60 | struct loop* loop; | |
61 | ||
62 | if (!loop_vinfo) | |
63 | return false; | |
64 | ||
65 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
66 | ||
67 | return (bb->loop_father == loop->inner); | |
68 | } | |
69 | ||
70 | /* Record the cost of a statement, either by directly informing the | |
71 | target model or by saving it in a vector for later processing. | |
72 | Return a preliminary estimate of the statement's cost. */ | |
73 | ||
74 | unsigned | |
92345349 | 75 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 76 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 77 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 78 | { |
92345349 | 79 | if (body_cost_vec) |
c3e7ee41 | 80 | { |
92345349 BS |
81 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
82 | add_stmt_info_to_vec (body_cost_vec, count, kind, | |
83 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
84 | misalign); | |
c3e7ee41 | 85 | return (unsigned) |
92345349 | 86 | (builtin_vectorization_cost (kind, vectype, misalign) * count); |
c3e7ee41 BS |
87 | |
88 | } | |
89 | else | |
90 | { | |
91 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
92 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
93 | void *target_cost_data; | |
94 | ||
95 | if (loop_vinfo) | |
96 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
97 | else | |
98 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
99 | ||
92345349 BS |
100 | return add_stmt_cost (target_cost_data, count, kind, stmt_info, |
101 | misalign, where); | |
c3e7ee41 BS |
102 | } |
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 | |
120 | read_vector_array (gimple stmt, gimple_stmt_iterator *gsi, tree scalar_dest, | |
121 | tree array, unsigned HOST_WIDE_INT n) | |
122 | { | |
123 | tree vect_type, vect, vect_name, array_ref; | |
124 | gimple new_stmt; | |
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 | |
146 | write_vector_array (gimple stmt, gimple_stmt_iterator *gsi, tree vect, | |
147 | tree array, unsigned HOST_WIDE_INT n) | |
148 | { | |
149 | tree array_ref; | |
150 | gimple new_stmt; | |
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 | |
9771b263 | 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); | |
83197f37 | 190 | gimple pattern_stmt; |
ebfd146a | 191 | |
73fbfcad | 192 | if (dump_enabled_p ()) |
78c60e3d SS |
193 | dump_printf_loc (MSG_NOTE, vect_location, |
194 | "mark relevant %d, live %d.", relevant, live_p); | |
ebfd146a | 195 | |
83197f37 IR |
196 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
197 | related pattern stmt instead of the original stmt. However, such stmts | |
198 | may have their own uses that are not in any pattern, in such cases the | |
199 | stmt itself should be marked. */ | |
ebfd146a IR |
200 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
201 | { | |
83197f37 IR |
202 | bool found = false; |
203 | if (!used_in_pattern) | |
204 | { | |
205 | imm_use_iterator imm_iter; | |
206 | use_operand_p use_p; | |
207 | gimple use_stmt; | |
208 | tree lhs; | |
13c931c9 JJ |
209 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
210 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a | 211 | |
83197f37 IR |
212 | if (is_gimple_assign (stmt)) |
213 | lhs = gimple_assign_lhs (stmt); | |
214 | else | |
215 | lhs = gimple_call_lhs (stmt); | |
ebfd146a | 216 | |
83197f37 IR |
217 | /* This use is out of pattern use, if LHS has other uses that are |
218 | pattern uses, we should mark the stmt itself, and not the pattern | |
219 | stmt. */ | |
ab0ef706 JJ |
220 | if (TREE_CODE (lhs) == SSA_NAME) |
221 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs) | |
222 | { | |
223 | if (is_gimple_debug (USE_STMT (use_p))) | |
224 | continue; | |
225 | use_stmt = USE_STMT (use_p); | |
226 | ||
13c931c9 JJ |
227 | if (!flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) |
228 | continue; | |
229 | ||
ab0ef706 JJ |
230 | if (vinfo_for_stmt (use_stmt) |
231 | && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (use_stmt))) | |
232 | { | |
233 | found = true; | |
234 | break; | |
235 | } | |
236 | } | |
83197f37 IR |
237 | } |
238 | ||
239 | if (!found) | |
240 | { | |
241 | /* This is the last stmt in a sequence that was detected as a | |
242 | pattern that can potentially be vectorized. Don't mark the stmt | |
243 | as relevant/live because it's not going to be vectorized. | |
244 | Instead mark the pattern-stmt that replaces it. */ | |
245 | ||
246 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
247 | ||
73fbfcad | 248 | if (dump_enabled_p ()) |
78c60e3d SS |
249 | dump_printf_loc (MSG_NOTE, vect_location, |
250 | "last stmt in pattern. don't mark" | |
251 | " relevant/live."); | |
83197f37 IR |
252 | stmt_info = vinfo_for_stmt (pattern_stmt); |
253 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
254 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
255 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
256 | stmt = pattern_stmt; | |
257 | } | |
ebfd146a IR |
258 | } |
259 | ||
260 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
261 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
262 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
263 | ||
264 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
265 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
266 | { | |
73fbfcad | 267 | if (dump_enabled_p ()) |
78c60e3d SS |
268 | dump_printf_loc (MSG_NOTE, vect_location, |
269 | "already marked relevant/live."); | |
ebfd146a IR |
270 | return; |
271 | } | |
272 | ||
9771b263 | 273 | worklist->safe_push (stmt); |
ebfd146a IR |
274 | } |
275 | ||
276 | ||
277 | /* Function vect_stmt_relevant_p. | |
278 | ||
279 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
280 | "relevant for vectorization". | |
281 | ||
282 | A stmt is considered "relevant for vectorization" if: | |
283 | - it has uses outside the loop. | |
284 | - it has vdefs (it alters memory). | |
285 | - control stmts in the loop (except for the exit condition). | |
286 | ||
287 | CHECKME: what other side effects would the vectorizer allow? */ | |
288 | ||
289 | static bool | |
290 | vect_stmt_relevant_p (gimple stmt, loop_vec_info loop_vinfo, | |
291 | enum vect_relevant *relevant, bool *live_p) | |
292 | { | |
293 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
294 | ssa_op_iter op_iter; | |
295 | imm_use_iterator imm_iter; | |
296 | use_operand_p use_p; | |
297 | def_operand_p def_p; | |
298 | ||
8644a673 | 299 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
300 | *live_p = false; |
301 | ||
302 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
303 | if (is_ctrl_stmt (stmt) |
304 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
305 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 306 | *relevant = vect_used_in_scope; |
ebfd146a IR |
307 | |
308 | /* changing memory. */ | |
309 | if (gimple_code (stmt) != GIMPLE_PHI) | |
5006671f | 310 | if (gimple_vdef (stmt)) |
ebfd146a | 311 | { |
73fbfcad | 312 | if (dump_enabled_p ()) |
78c60e3d SS |
313 | dump_printf_loc (MSG_NOTE, vect_location, |
314 | "vec_stmt_relevant_p: stmt has vdefs."); | |
8644a673 | 315 | *relevant = vect_used_in_scope; |
ebfd146a IR |
316 | } |
317 | ||
318 | /* uses outside the loop. */ | |
319 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
320 | { | |
321 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
322 | { | |
323 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
324 | if (!flow_bb_inside_loop_p (loop, bb)) | |
325 | { | |
73fbfcad | 326 | if (dump_enabled_p ()) |
78c60e3d SS |
327 | dump_printf_loc (MSG_NOTE, vect_location, |
328 | "vec_stmt_relevant_p: used out of loop."); | |
ebfd146a | 329 | |
3157b0c2 AO |
330 | if (is_gimple_debug (USE_STMT (use_p))) |
331 | continue; | |
332 | ||
ebfd146a IR |
333 | /* We expect all such uses to be in the loop exit phis |
334 | (because of loop closed form) */ | |
335 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
336 | gcc_assert (bb == single_exit (loop)->dest); | |
337 | ||
338 | *live_p = true; | |
339 | } | |
340 | } | |
341 | } | |
342 | ||
343 | return (*live_p || *relevant); | |
344 | } | |
345 | ||
346 | ||
b8698a0f | 347 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 348 | |
ff802fa1 | 349 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
350 | used in STMT for anything other than indexing an array. */ |
351 | ||
352 | static bool | |
353 | exist_non_indexing_operands_for_use_p (tree use, gimple stmt) | |
354 | { | |
355 | tree operand; | |
356 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 357 | |
ff802fa1 | 358 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
359 | reference in STMT, then any operand that corresponds to USE |
360 | is not indexing an array. */ | |
361 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
362 | return true; | |
59a05b0c | 363 | |
ebfd146a IR |
364 | /* STMT has a data_ref. FORNOW this means that its of one of |
365 | the following forms: | |
366 | -1- ARRAY_REF = var | |
367 | -2- var = ARRAY_REF | |
368 | (This should have been verified in analyze_data_refs). | |
369 | ||
370 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 371 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
372 | for array indexing. |
373 | ||
374 | Therefore, all we need to check is if STMT falls into the | |
375 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
376 | |
377 | if (!gimple_assign_copy_p (stmt)) | |
378 | return false; | |
59a05b0c EB |
379 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
380 | return false; | |
ebfd146a | 381 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
382 | if (TREE_CODE (operand) != SSA_NAME) |
383 | return false; | |
384 | ||
385 | if (operand == use) | |
386 | return true; | |
387 | ||
388 | return false; | |
389 | } | |
390 | ||
391 | ||
b8698a0f | 392 | /* |
ebfd146a IR |
393 | Function process_use. |
394 | ||
395 | Inputs: | |
396 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b8698a0f | 397 | - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt |
ff802fa1 | 398 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 399 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
400 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
401 | be performed. | |
ebfd146a IR |
402 | |
403 | Outputs: | |
404 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
405 | relevance info of the DEF_STMT of this USE: | |
406 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
407 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
408 | Exceptions: | |
409 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 410 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 411 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
412 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
413 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
414 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
415 | be modified accordingly. | |
416 | ||
417 | Return true if everything is as expected. Return false otherwise. */ | |
418 | ||
419 | static bool | |
b8698a0f | 420 | process_use (gimple stmt, tree use, loop_vec_info loop_vinfo, bool live_p, |
9771b263 | 421 | enum vect_relevant relevant, vec<gimple> *worklist, |
aec7ae7d | 422 | bool force) |
ebfd146a IR |
423 | { |
424 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
425 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
426 | stmt_vec_info dstmt_vinfo; | |
427 | basic_block bb, def_bb; | |
428 | tree def; | |
429 | gimple def_stmt; | |
430 | enum vect_def_type dt; | |
431 | ||
b8698a0f | 432 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 433 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 434 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
435 | return true; |
436 | ||
24ee1384 | 437 | if (!vect_is_simple_use (use, stmt, loop_vinfo, NULL, &def_stmt, &def, &dt)) |
b8698a0f | 438 | { |
73fbfcad | 439 | if (dump_enabled_p ()) |
78c60e3d SS |
440 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
441 | "not vectorized: unsupported use in stmt."); | |
ebfd146a IR |
442 | return false; |
443 | } | |
444 | ||
445 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
446 | return true; | |
447 | ||
448 | def_bb = gimple_bb (def_stmt); | |
449 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
450 | { | |
73fbfcad | 451 | if (dump_enabled_p ()) |
78c60e3d | 452 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop."); |
ebfd146a IR |
453 | return true; |
454 | } | |
455 | ||
b8698a0f L |
456 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
457 | DEF_STMT must have already been processed, because this should be the | |
458 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
459 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
460 | check that everything is as expected, and we are done. */ |
461 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
462 | bb = gimple_bb (stmt); | |
463 | if (gimple_code (stmt) == GIMPLE_PHI | |
464 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
465 | && gimple_code (def_stmt) != GIMPLE_PHI | |
466 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
467 | && bb->loop_father == def_bb->loop_father) | |
468 | { | |
73fbfcad | 469 | if (dump_enabled_p ()) |
78c60e3d SS |
470 | dump_printf_loc (MSG_NOTE, vect_location, |
471 | "reduc-stmt defining reduc-phi in the same nest."); | |
ebfd146a IR |
472 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
473 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
474 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 475 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 476 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
477 | return true; |
478 | } | |
479 | ||
480 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
481 | outer-loop-header-bb: | |
482 | d = def_stmt | |
483 | inner-loop: | |
484 | stmt # use (d) | |
485 | outer-loop-tail-bb: | |
486 | ... */ | |
487 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
488 | { | |
73fbfcad | 489 | if (dump_enabled_p ()) |
78c60e3d SS |
490 | dump_printf_loc (MSG_NOTE, vect_location, |
491 | "outer-loop def-stmt defining inner-loop stmt."); | |
7c5222ff | 492 | |
ebfd146a IR |
493 | switch (relevant) |
494 | { | |
8644a673 | 495 | case vect_unused_in_scope: |
7c5222ff IR |
496 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
497 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 498 | break; |
7c5222ff | 499 | |
ebfd146a | 500 | case vect_used_in_outer_by_reduction: |
7c5222ff | 501 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
502 | relevant = vect_used_by_reduction; |
503 | break; | |
7c5222ff | 504 | |
ebfd146a | 505 | case vect_used_in_outer: |
7c5222ff | 506 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 507 | relevant = vect_used_in_scope; |
ebfd146a | 508 | break; |
7c5222ff | 509 | |
8644a673 | 510 | case vect_used_in_scope: |
ebfd146a IR |
511 | break; |
512 | ||
513 | default: | |
514 | gcc_unreachable (); | |
b8698a0f | 515 | } |
ebfd146a IR |
516 | } |
517 | ||
518 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
519 | outer-loop-header-bb: | |
520 | ... | |
521 | inner-loop: | |
522 | d = def_stmt | |
06066f92 | 523 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
524 | stmt # use (d) */ |
525 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
526 | { | |
73fbfcad | 527 | if (dump_enabled_p ()) |
78c60e3d SS |
528 | dump_printf_loc (MSG_NOTE, vect_location, |
529 | "inner-loop def-stmt defining outer-loop stmt."); | |
7c5222ff | 530 | |
ebfd146a IR |
531 | switch (relevant) |
532 | { | |
8644a673 | 533 | case vect_unused_in_scope: |
b8698a0f | 534 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 535 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 536 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
537 | break; |
538 | ||
ebfd146a IR |
539 | case vect_used_by_reduction: |
540 | relevant = vect_used_in_outer_by_reduction; | |
541 | break; | |
542 | ||
8644a673 | 543 | case vect_used_in_scope: |
ebfd146a IR |
544 | relevant = vect_used_in_outer; |
545 | break; | |
546 | ||
547 | default: | |
548 | gcc_unreachable (); | |
549 | } | |
550 | } | |
551 | ||
83197f37 IR |
552 | vect_mark_relevant (worklist, def_stmt, relevant, live_p, |
553 | is_pattern_stmt_p (stmt_vinfo)); | |
ebfd146a IR |
554 | return true; |
555 | } | |
556 | ||
557 | ||
558 | /* Function vect_mark_stmts_to_be_vectorized. | |
559 | ||
560 | Not all stmts in the loop need to be vectorized. For example: | |
561 | ||
562 | for i... | |
563 | for j... | |
564 | 1. T0 = i + j | |
565 | 2. T1 = a[T0] | |
566 | ||
567 | 3. j = j + 1 | |
568 | ||
569 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
570 | addressing of vectorized data-refs are handled differently. | |
571 | ||
572 | This pass detects such stmts. */ | |
573 | ||
574 | bool | |
575 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
576 | { | |
9771b263 | 577 | vec<gimple> worklist; |
ebfd146a IR |
578 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
579 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
580 | unsigned int nbbs = loop->num_nodes; | |
581 | gimple_stmt_iterator si; | |
582 | gimple stmt; | |
583 | unsigned int i; | |
584 | stmt_vec_info stmt_vinfo; | |
585 | basic_block bb; | |
586 | gimple phi; | |
587 | bool live_p; | |
06066f92 IR |
588 | enum vect_relevant relevant, tmp_relevant; |
589 | enum vect_def_type def_type; | |
ebfd146a | 590 | |
73fbfcad | 591 | if (dump_enabled_p ()) |
78c60e3d SS |
592 | dump_printf_loc (MSG_NOTE, vect_location, |
593 | "=== vect_mark_stmts_to_be_vectorized ==="); | |
ebfd146a | 594 | |
9771b263 | 595 | worklist.create (64); |
ebfd146a IR |
596 | |
597 | /* 1. Init worklist. */ | |
598 | for (i = 0; i < nbbs; i++) | |
599 | { | |
600 | bb = bbs[i]; | |
601 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 602 | { |
ebfd146a | 603 | phi = gsi_stmt (si); |
73fbfcad | 604 | if (dump_enabled_p ()) |
ebfd146a | 605 | { |
78c60e3d SS |
606 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
607 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
608 | } |
609 | ||
610 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 611 | vect_mark_relevant (&worklist, phi, relevant, live_p, false); |
ebfd146a IR |
612 | } |
613 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
614 | { | |
615 | stmt = gsi_stmt (si); | |
73fbfcad | 616 | if (dump_enabled_p ()) |
ebfd146a | 617 | { |
78c60e3d SS |
618 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
619 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
b8698a0f | 620 | } |
ebfd146a IR |
621 | |
622 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 623 | vect_mark_relevant (&worklist, stmt, relevant, live_p, false); |
ebfd146a IR |
624 | } |
625 | } | |
626 | ||
627 | /* 2. Process_worklist */ | |
9771b263 | 628 | while (worklist.length () > 0) |
ebfd146a IR |
629 | { |
630 | use_operand_p use_p; | |
631 | ssa_op_iter iter; | |
632 | ||
9771b263 | 633 | stmt = worklist.pop (); |
73fbfcad | 634 | if (dump_enabled_p ()) |
ebfd146a | 635 | { |
78c60e3d SS |
636 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
637 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
ebfd146a IR |
638 | } |
639 | ||
b8698a0f L |
640 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
641 | (DEF_STMT) as relevant/irrelevant and live/dead according to the | |
ebfd146a IR |
642 | liveness and relevance properties of STMT. */ |
643 | stmt_vinfo = vinfo_for_stmt (stmt); | |
644 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
645 | live_p = STMT_VINFO_LIVE_P (stmt_vinfo); | |
646 | ||
647 | /* Generally, the liveness and relevance properties of STMT are | |
648 | propagated as is to the DEF_STMTs of its USEs: | |
649 | live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO) | |
650 | relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO) | |
651 | ||
652 | One exception is when STMT has been identified as defining a reduction | |
653 | variable; in this case we set the liveness/relevance as follows: | |
654 | live_p = false | |
655 | relevant = vect_used_by_reduction | |
656 | This is because we distinguish between two kinds of relevant stmts - | |
b8698a0f | 657 | those that are used by a reduction computation, and those that are |
ff802fa1 | 658 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 659 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 660 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 661 | |
06066f92 IR |
662 | def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
663 | tmp_relevant = relevant; | |
664 | switch (def_type) | |
ebfd146a | 665 | { |
06066f92 IR |
666 | case vect_reduction_def: |
667 | switch (tmp_relevant) | |
668 | { | |
669 | case vect_unused_in_scope: | |
670 | relevant = vect_used_by_reduction; | |
671 | break; | |
672 | ||
673 | case vect_used_by_reduction: | |
674 | if (gimple_code (stmt) == GIMPLE_PHI) | |
675 | break; | |
676 | /* fall through */ | |
677 | ||
678 | default: | |
73fbfcad | 679 | if (dump_enabled_p ()) |
78c60e3d SS |
680 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
681 | "unsupported use of reduction."); | |
9771b263 | 682 | worklist.release (); |
06066f92 IR |
683 | return false; |
684 | } | |
685 | ||
b8698a0f | 686 | live_p = false; |
06066f92 | 687 | break; |
b8698a0f | 688 | |
06066f92 IR |
689 | case vect_nested_cycle: |
690 | if (tmp_relevant != vect_unused_in_scope | |
691 | && tmp_relevant != vect_used_in_outer_by_reduction | |
692 | && tmp_relevant != vect_used_in_outer) | |
693 | { | |
73fbfcad | 694 | if (dump_enabled_p ()) |
78c60e3d SS |
695 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
696 | "unsupported use of nested cycle."); | |
7c5222ff | 697 | |
9771b263 | 698 | worklist.release (); |
06066f92 IR |
699 | return false; |
700 | } | |
7c5222ff | 701 | |
b8698a0f L |
702 | live_p = false; |
703 | break; | |
704 | ||
06066f92 IR |
705 | case vect_double_reduction_def: |
706 | if (tmp_relevant != vect_unused_in_scope | |
707 | && tmp_relevant != vect_used_by_reduction) | |
708 | { | |
73fbfcad | 709 | if (dump_enabled_p ()) |
78c60e3d SS |
710 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
711 | "unsupported use of double reduction."); | |
7c5222ff | 712 | |
9771b263 | 713 | worklist.release (); |
7c5222ff | 714 | return false; |
06066f92 IR |
715 | } |
716 | ||
717 | live_p = false; | |
b8698a0f | 718 | break; |
7c5222ff | 719 | |
06066f92 IR |
720 | default: |
721 | break; | |
7c5222ff | 722 | } |
b8698a0f | 723 | |
aec7ae7d | 724 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
725 | { |
726 | /* Pattern statements are not inserted into the code, so | |
727 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
728 | have to scan the RHS or function arguments instead. */ | |
729 | if (is_gimple_assign (stmt)) | |
730 | { | |
69d2aade JJ |
731 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
732 | tree op = gimple_assign_rhs1 (stmt); | |
733 | ||
734 | i = 1; | |
735 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
736 | { | |
737 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
aec7ae7d | 738 | live_p, relevant, &worklist, false) |
69d2aade | 739 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
aec7ae7d | 740 | live_p, relevant, &worklist, false)) |
69d2aade | 741 | { |
9771b263 | 742 | worklist.release (); |
69d2aade JJ |
743 | return false; |
744 | } | |
745 | i = 2; | |
746 | } | |
747 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 748 | { |
69d2aade | 749 | op = gimple_op (stmt, i); |
9d5e7640 | 750 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, |
aec7ae7d | 751 | &worklist, false)) |
9d5e7640 | 752 | { |
9771b263 | 753 | worklist.release (); |
9d5e7640 IR |
754 | return false; |
755 | } | |
756 | } | |
757 | } | |
758 | else if (is_gimple_call (stmt)) | |
759 | { | |
760 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
761 | { | |
762 | tree arg = gimple_call_arg (stmt, i); | |
763 | if (!process_use (stmt, arg, loop_vinfo, live_p, relevant, | |
aec7ae7d | 764 | &worklist, false)) |
9d5e7640 | 765 | { |
9771b263 | 766 | worklist.release (); |
9d5e7640 IR |
767 | return false; |
768 | } | |
769 | } | |
770 | } | |
771 | } | |
772 | else | |
773 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
774 | { | |
775 | tree op = USE_FROM_PTR (use_p); | |
776 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, | |
aec7ae7d | 777 | &worklist, false)) |
9d5e7640 | 778 | { |
9771b263 | 779 | worklist.release (); |
9d5e7640 IR |
780 | return false; |
781 | } | |
782 | } | |
aec7ae7d JJ |
783 | |
784 | if (STMT_VINFO_GATHER_P (stmt_vinfo)) | |
785 | { | |
786 | tree off; | |
787 | tree decl = vect_check_gather (stmt, loop_vinfo, NULL, &off, NULL); | |
788 | gcc_assert (decl); | |
789 | if (!process_use (stmt, off, loop_vinfo, live_p, relevant, | |
790 | &worklist, true)) | |
791 | { | |
9771b263 | 792 | worklist.release (); |
aec7ae7d JJ |
793 | return false; |
794 | } | |
795 | } | |
ebfd146a IR |
796 | } /* while worklist */ |
797 | ||
9771b263 | 798 | worklist.release (); |
ebfd146a IR |
799 | return true; |
800 | } | |
801 | ||
802 | ||
b8698a0f | 803 | /* Function vect_model_simple_cost. |
ebfd146a | 804 | |
b8698a0f | 805 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
806 | single op. Right now, this does not account for multiple insns that could |
807 | be generated for the single vector op. We will handle that shortly. */ | |
808 | ||
809 | void | |
b8698a0f | 810 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 BS |
811 | enum vect_def_type *dt, |
812 | stmt_vector_for_cost *prologue_cost_vec, | |
813 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
814 | { |
815 | int i; | |
92345349 | 816 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
817 | |
818 | /* The SLP costs were already calculated during SLP tree build. */ | |
819 | if (PURE_SLP_STMT (stmt_info)) | |
820 | return; | |
821 | ||
ebfd146a IR |
822 | /* FORNOW: Assuming maximum 2 args per stmts. */ |
823 | for (i = 0; i < 2; i++) | |
92345349 BS |
824 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
825 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, vector_stmt, | |
826 | stmt_info, 0, vect_prologue); | |
c3e7ee41 BS |
827 | |
828 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
92345349 BS |
829 | inside_cost = record_stmt_cost (body_cost_vec, ncopies, vector_stmt, |
830 | stmt_info, 0, vect_body); | |
c3e7ee41 | 831 | |
73fbfcad | 832 | if (dump_enabled_p ()) |
78c60e3d SS |
833 | dump_printf_loc (MSG_NOTE, vect_location, |
834 | "vect_model_simple_cost: inside_cost = %d, " | |
835 | "prologue_cost = %d .", inside_cost, prologue_cost); | |
ebfd146a IR |
836 | } |
837 | ||
838 | ||
8bd37302 BS |
839 | /* Model cost for type demotion and promotion operations. PWR is normally |
840 | zero for single-step promotions and demotions. It will be one if | |
841 | two-step promotion/demotion is required, and so on. Each additional | |
842 | step doubles the number of instructions required. */ | |
843 | ||
844 | static void | |
845 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
846 | enum vect_def_type *dt, int pwr) | |
847 | { | |
848 | int i, tmp; | |
92345349 | 849 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 BS |
850 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
851 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
852 | void *target_cost_data; | |
8bd37302 BS |
853 | |
854 | /* The SLP costs were already calculated during SLP tree build. */ | |
855 | if (PURE_SLP_STMT (stmt_info)) | |
856 | return; | |
857 | ||
c3e7ee41 BS |
858 | if (loop_vinfo) |
859 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
860 | else | |
861 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
862 | ||
8bd37302 BS |
863 | for (i = 0; i < pwr + 1; i++) |
864 | { | |
865 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
866 | (i + 1) : i; | |
c3e7ee41 | 867 | inside_cost += add_stmt_cost (target_cost_data, vect_pow2 (tmp), |
92345349 BS |
868 | vec_promote_demote, stmt_info, 0, |
869 | vect_body); | |
8bd37302 BS |
870 | } |
871 | ||
872 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
873 | for (i = 0; i < 2; i++) | |
92345349 BS |
874 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
875 | prologue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt, | |
876 | stmt_info, 0, vect_prologue); | |
8bd37302 | 877 | |
73fbfcad | 878 | if (dump_enabled_p ()) |
78c60e3d SS |
879 | dump_printf_loc (MSG_NOTE, vect_location, |
880 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
881 | "prologue_cost = %d .", inside_cost, prologue_cost); | |
8bd37302 BS |
882 | } |
883 | ||
0d0293ac | 884 | /* Function vect_cost_group_size |
b8698a0f | 885 | |
0d0293ac | 886 | For grouped load or store, return the group_size only if it is the first |
ebfd146a IR |
887 | load or store of a group, else return 1. This ensures that group size is |
888 | only returned once per group. */ | |
889 | ||
890 | static int | |
0d0293ac | 891 | vect_cost_group_size (stmt_vec_info stmt_info) |
ebfd146a | 892 | { |
e14c1050 | 893 | gimple first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a IR |
894 | |
895 | if (first_stmt == STMT_VINFO_STMT (stmt_info)) | |
e14c1050 | 896 | return GROUP_SIZE (stmt_info); |
ebfd146a IR |
897 | |
898 | return 1; | |
899 | } | |
900 | ||
901 | ||
902 | /* Function vect_model_store_cost | |
903 | ||
0d0293ac MM |
904 | Models cost for stores. In the case of grouped accesses, one access |
905 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a IR |
906 | |
907 | void | |
b8698a0f | 908 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
272c6793 | 909 | bool store_lanes_p, enum vect_def_type dt, |
92345349 BS |
910 | slp_tree slp_node, |
911 | stmt_vector_for_cost *prologue_cost_vec, | |
912 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
913 | { |
914 | int group_size; | |
92345349 | 915 | unsigned int inside_cost = 0, prologue_cost = 0; |
720f5239 IR |
916 | struct data_reference *first_dr; |
917 | gimple first_stmt; | |
ebfd146a IR |
918 | |
919 | /* The SLP costs were already calculated during SLP tree build. */ | |
920 | if (PURE_SLP_STMT (stmt_info)) | |
921 | return; | |
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. */ | |
954 | if (!store_lanes_p && group_size > 1) | |
ebfd146a IR |
955 | { |
956 | /* Uses a high and low interleave operation for each needed permute. */ | |
c3e7ee41 BS |
957 | |
958 | int nstmts = ncopies * exact_log2 (group_size) * group_size; | |
92345349 BS |
959 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
960 | stmt_info, 0, vect_body); | |
ebfd146a | 961 | |
73fbfcad | 962 | if (dump_enabled_p ()) |
78c60e3d SS |
963 | dump_printf_loc (MSG_NOTE, vect_location, |
964 | "vect_model_store_cost: strided group_size = %d .", | |
965 | group_size); | |
ebfd146a IR |
966 | } |
967 | ||
968 | /* Costs of the stores. */ | |
92345349 | 969 | vect_get_store_cost (first_dr, ncopies, &inside_cost, body_cost_vec); |
ebfd146a | 970 | |
73fbfcad | 971 | if (dump_enabled_p ()) |
78c60e3d SS |
972 | dump_printf_loc (MSG_NOTE, vect_location, |
973 | "vect_model_store_cost: inside_cost = %d, " | |
974 | "prologue_cost = %d .", inside_cost, prologue_cost); | |
ebfd146a IR |
975 | } |
976 | ||
977 | ||
720f5239 IR |
978 | /* Calculate cost of DR's memory access. */ |
979 | void | |
980 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 981 | unsigned int *inside_cost, |
92345349 | 982 | stmt_vector_for_cost *body_cost_vec) |
720f5239 IR |
983 | { |
984 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
c3e7ee41 BS |
985 | gimple stmt = DR_STMT (dr); |
986 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
720f5239 IR |
987 | |
988 | switch (alignment_support_scheme) | |
989 | { | |
990 | case dr_aligned: | |
991 | { | |
92345349 BS |
992 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
993 | vector_store, stmt_info, 0, | |
994 | vect_body); | |
720f5239 | 995 | |
73fbfcad | 996 | if (dump_enabled_p ()) |
78c60e3d SS |
997 | dump_printf_loc (MSG_NOTE, vect_location, |
998 | "vect_model_store_cost: aligned."); | |
720f5239 IR |
999 | break; |
1000 | } | |
1001 | ||
1002 | case dr_unaligned_supported: | |
1003 | { | |
720f5239 | 1004 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1005 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1006 | unaligned_store, stmt_info, |
92345349 | 1007 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1008 | if (dump_enabled_p ()) |
78c60e3d SS |
1009 | dump_printf_loc (MSG_NOTE, vect_location, |
1010 | "vect_model_store_cost: unaligned supported by " | |
1011 | "hardware."); | |
720f5239 IR |
1012 | break; |
1013 | } | |
1014 | ||
38eec4c6 UW |
1015 | case dr_unaligned_unsupported: |
1016 | { | |
1017 | *inside_cost = VECT_MAX_COST; | |
1018 | ||
73fbfcad | 1019 | if (dump_enabled_p ()) |
78c60e3d SS |
1020 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1021 | "vect_model_store_cost: unsupported access."); | |
38eec4c6 UW |
1022 | break; |
1023 | } | |
1024 | ||
720f5239 IR |
1025 | default: |
1026 | gcc_unreachable (); | |
1027 | } | |
1028 | } | |
1029 | ||
1030 | ||
ebfd146a IR |
1031 | /* Function vect_model_load_cost |
1032 | ||
0d0293ac MM |
1033 | Models cost for loads. In the case of grouped accesses, the last access |
1034 | has the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1035 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1036 | access scheme chosen. */ |
1037 | ||
1038 | void | |
92345349 BS |
1039 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, |
1040 | bool load_lanes_p, slp_tree slp_node, | |
1041 | stmt_vector_for_cost *prologue_cost_vec, | |
1042 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
1043 | { |
1044 | int group_size; | |
ebfd146a IR |
1045 | gimple first_stmt; |
1046 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; | |
92345349 | 1047 | unsigned int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
1048 | |
1049 | /* The SLP costs were already calculated during SLP tree build. */ | |
1050 | if (PURE_SLP_STMT (stmt_info)) | |
1051 | return; | |
1052 | ||
0d0293ac | 1053 | /* Grouped accesses? */ |
e14c1050 | 1054 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 1055 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && first_stmt && !slp_node) |
ebfd146a | 1056 | { |
0d0293ac | 1057 | group_size = vect_cost_group_size (stmt_info); |
ebfd146a IR |
1058 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
1059 | } | |
0d0293ac | 1060 | /* Not a grouped access. */ |
ebfd146a IR |
1061 | else |
1062 | { | |
1063 | group_size = 1; | |
1064 | first_dr = dr; | |
1065 | } | |
1066 | ||
272c6793 | 1067 | /* We assume that the cost of a single load-lanes instruction is |
0d0293ac | 1068 | equivalent to the cost of GROUP_SIZE separate loads. If a grouped |
272c6793 RS |
1069 | access is instead being provided by a load-and-permute operation, |
1070 | include the cost of the permutes. */ | |
1071 | if (!load_lanes_p && group_size > 1) | |
ebfd146a IR |
1072 | { |
1073 | /* Uses an even and odd extract operations for each needed permute. */ | |
c3e7ee41 | 1074 | int nstmts = ncopies * exact_log2 (group_size) * group_size; |
92345349 BS |
1075 | inside_cost += record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
1076 | stmt_info, 0, vect_body); | |
ebfd146a | 1077 | |
73fbfcad | 1078 | if (dump_enabled_p ()) |
78c60e3d SS |
1079 | dump_printf_loc (MSG_NOTE, vect_location, |
1080 | "vect_model_load_cost: strided group_size = %d .", | |
1081 | group_size); | |
ebfd146a IR |
1082 | } |
1083 | ||
1084 | /* The loads themselves. */ | |
a82960aa RG |
1085 | if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
1086 | { | |
a21892ad BS |
1087 | /* N scalar loads plus gathering them into a vector. */ |
1088 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
92345349 | 1089 | inside_cost += record_stmt_cost (body_cost_vec, |
c3e7ee41 | 1090 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), |
92345349 BS |
1091 | scalar_load, stmt_info, 0, vect_body); |
1092 | inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct, | |
1093 | stmt_info, 0, vect_body); | |
a82960aa RG |
1094 | } |
1095 | else | |
1096 | vect_get_load_cost (first_dr, ncopies, | |
1097 | ((!STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1098 | || group_size > 1 || slp_node), | |
92345349 BS |
1099 | &inside_cost, &prologue_cost, |
1100 | prologue_cost_vec, body_cost_vec, true); | |
720f5239 | 1101 | |
73fbfcad | 1102 | if (dump_enabled_p ()) |
78c60e3d SS |
1103 | dump_printf_loc (MSG_NOTE, vect_location, |
1104 | "vect_model_load_cost: inside_cost = %d, " | |
1105 | "prologue_cost = %d .", inside_cost, prologue_cost); | |
720f5239 IR |
1106 | } |
1107 | ||
1108 | ||
1109 | /* Calculate cost of DR's memory access. */ | |
1110 | void | |
1111 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1112 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1113 | unsigned int *prologue_cost, |
1114 | stmt_vector_for_cost *prologue_cost_vec, | |
1115 | stmt_vector_for_cost *body_cost_vec, | |
1116 | bool record_prologue_costs) | |
720f5239 IR |
1117 | { |
1118 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
c3e7ee41 BS |
1119 | gimple stmt = DR_STMT (dr); |
1120 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
720f5239 IR |
1121 | |
1122 | switch (alignment_support_scheme) | |
ebfd146a IR |
1123 | { |
1124 | case dr_aligned: | |
1125 | { | |
92345349 BS |
1126 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1127 | stmt_info, 0, vect_body); | |
ebfd146a | 1128 | |
73fbfcad | 1129 | if (dump_enabled_p ()) |
78c60e3d SS |
1130 | dump_printf_loc (MSG_NOTE, vect_location, |
1131 | "vect_model_load_cost: aligned."); | |
ebfd146a IR |
1132 | |
1133 | break; | |
1134 | } | |
1135 | case dr_unaligned_supported: | |
1136 | { | |
720f5239 | 1137 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1138 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1139 | unaligned_load, stmt_info, |
92345349 | 1140 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1141 | |
73fbfcad | 1142 | if (dump_enabled_p ()) |
78c60e3d SS |
1143 | dump_printf_loc (MSG_NOTE, vect_location, |
1144 | "vect_model_load_cost: unaligned supported by " | |
1145 | "hardware."); | |
ebfd146a IR |
1146 | |
1147 | break; | |
1148 | } | |
1149 | case dr_explicit_realign: | |
1150 | { | |
92345349 BS |
1151 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1152 | vector_load, stmt_info, 0, vect_body); | |
1153 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1154 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1155 | |
1156 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1157 | the containing loop, the following cost should be added to the | |
92345349 | 1158 | prologue costs. */ |
ebfd146a | 1159 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1160 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1161 | stmt_info, 0, vect_body); | |
ebfd146a | 1162 | |
73fbfcad | 1163 | if (dump_enabled_p ()) |
78c60e3d SS |
1164 | dump_printf_loc (MSG_NOTE, vect_location, |
1165 | "vect_model_load_cost: explicit realign"); | |
8bd37302 | 1166 | |
ebfd146a IR |
1167 | break; |
1168 | } | |
1169 | case dr_explicit_realign_optimized: | |
1170 | { | |
73fbfcad | 1171 | if (dump_enabled_p ()) |
78c60e3d SS |
1172 | dump_printf_loc (MSG_NOTE, vect_location, |
1173 | "vect_model_load_cost: unaligned software " | |
1174 | "pipelined."); | |
ebfd146a IR |
1175 | |
1176 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1177 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1178 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1179 | access, then the above cost should only be considered for one |
ff802fa1 | 1180 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1181 | and a realignment op. */ |
1182 | ||
92345349 | 1183 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1184 | { |
92345349 BS |
1185 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1186 | vector_stmt, stmt_info, | |
1187 | 0, vect_prologue); | |
ebfd146a | 1188 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1189 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1190 | vector_stmt, stmt_info, | |
1191 | 0, vect_prologue); | |
ebfd146a IR |
1192 | } |
1193 | ||
92345349 BS |
1194 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1195 | stmt_info, 0, vect_body); | |
1196 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1197 | stmt_info, 0, vect_body); | |
8bd37302 | 1198 | |
73fbfcad | 1199 | if (dump_enabled_p ()) |
78c60e3d SS |
1200 | dump_printf_loc (MSG_NOTE, vect_location, |
1201 | "vect_model_load_cost: explicit realign optimized"); | |
8bd37302 | 1202 | |
ebfd146a IR |
1203 | break; |
1204 | } | |
1205 | ||
38eec4c6 UW |
1206 | case dr_unaligned_unsupported: |
1207 | { | |
1208 | *inside_cost = VECT_MAX_COST; | |
1209 | ||
73fbfcad | 1210 | if (dump_enabled_p ()) |
78c60e3d SS |
1211 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1212 | "vect_model_load_cost: unsupported access."); | |
38eec4c6 UW |
1213 | break; |
1214 | } | |
1215 | ||
ebfd146a IR |
1216 | default: |
1217 | gcc_unreachable (); | |
1218 | } | |
ebfd146a IR |
1219 | } |
1220 | ||
418b7df3 RG |
1221 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1222 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1223 | |
418b7df3 RG |
1224 | static void |
1225 | vect_init_vector_1 (gimple stmt, gimple new_stmt, gimple_stmt_iterator *gsi) | |
ebfd146a | 1226 | { |
ebfd146a | 1227 | if (gsi) |
418b7df3 | 1228 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1229 | else |
1230 | { | |
418b7df3 | 1231 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1232 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1233 | |
a70d6342 IR |
1234 | if (loop_vinfo) |
1235 | { | |
1236 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1237 | basic_block new_bb; |
1238 | edge pe; | |
a70d6342 IR |
1239 | |
1240 | if (nested_in_vect_loop_p (loop, stmt)) | |
1241 | loop = loop->inner; | |
b8698a0f | 1242 | |
a70d6342 | 1243 | pe = loop_preheader_edge (loop); |
418b7df3 | 1244 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1245 | gcc_assert (!new_bb); |
1246 | } | |
1247 | else | |
1248 | { | |
1249 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1250 | basic_block bb; | |
1251 | gimple_stmt_iterator gsi_bb_start; | |
1252 | ||
1253 | gcc_assert (bb_vinfo); | |
1254 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1255 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1256 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1257 | } |
ebfd146a IR |
1258 | } |
1259 | ||
73fbfcad | 1260 | if (dump_enabled_p ()) |
ebfd146a | 1261 | { |
78c60e3d SS |
1262 | dump_printf_loc (MSG_NOTE, vect_location, |
1263 | "created new init_stmt: "); | |
1264 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
ebfd146a | 1265 | } |
418b7df3 RG |
1266 | } |
1267 | ||
1268 | /* Function vect_init_vector. | |
ebfd146a | 1269 | |
5467ee52 RG |
1270 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1271 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1272 | vector type a vector with all elements equal to VAL is created first. | |
1273 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1274 | initialization at the loop preheader. | |
418b7df3 RG |
1275 | Return the DEF of INIT_STMT. |
1276 | It will be used in the vectorization of STMT. */ | |
1277 | ||
1278 | tree | |
5467ee52 | 1279 | vect_init_vector (gimple stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 RG |
1280 | { |
1281 | tree new_var; | |
1282 | gimple init_stmt; | |
1283 | tree vec_oprnd; | |
1284 | tree new_temp; | |
1285 | ||
5467ee52 RG |
1286 | if (TREE_CODE (type) == VECTOR_TYPE |
1287 | && TREE_CODE (TREE_TYPE (val)) != VECTOR_TYPE) | |
418b7df3 | 1288 | { |
5467ee52 | 1289 | if (!types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) |
418b7df3 | 1290 | { |
5467ee52 RG |
1291 | if (CONSTANT_CLASS_P (val)) |
1292 | val = fold_unary (VIEW_CONVERT_EXPR, TREE_TYPE (type), val); | |
418b7df3 RG |
1293 | else |
1294 | { | |
83d5977e | 1295 | new_temp = make_ssa_name (TREE_TYPE (type), NULL); |
418b7df3 | 1296 | init_stmt = gimple_build_assign_with_ops (NOP_EXPR, |
83d5977e | 1297 | new_temp, val, |
418b7df3 | 1298 | NULL_TREE); |
418b7df3 | 1299 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1300 | val = new_temp; |
418b7df3 RG |
1301 | } |
1302 | } | |
5467ee52 | 1303 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1304 | } |
1305 | ||
5467ee52 | 1306 | new_var = vect_get_new_vect_var (type, vect_simple_var, "cst_"); |
5467ee52 | 1307 | init_stmt = gimple_build_assign (new_var, val); |
418b7df3 RG |
1308 | new_temp = make_ssa_name (new_var, init_stmt); |
1309 | gimple_assign_set_lhs (init_stmt, new_temp); | |
1310 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
ebfd146a IR |
1311 | vec_oprnd = gimple_assign_lhs (init_stmt); |
1312 | return vec_oprnd; | |
1313 | } | |
1314 | ||
a70d6342 | 1315 | |
ebfd146a IR |
1316 | /* Function vect_get_vec_def_for_operand. |
1317 | ||
ff802fa1 | 1318 | OP is an operand in STMT. This function returns a (vector) def that will be |
ebfd146a IR |
1319 | used in the vectorized stmt for STMT. |
1320 | ||
1321 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1322 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1323 | ||
1324 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1325 | needs to be introduced. */ | |
1326 | ||
1327 | tree | |
1328 | vect_get_vec_def_for_operand (tree op, gimple stmt, tree *scalar_def) | |
1329 | { | |
1330 | tree vec_oprnd; | |
1331 | gimple vec_stmt; | |
1332 | gimple def_stmt; | |
1333 | stmt_vec_info def_stmt_info = NULL; | |
1334 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
9dc3f7de | 1335 | unsigned int nunits; |
ebfd146a | 1336 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
ebfd146a | 1337 | tree def; |
ebfd146a IR |
1338 | enum vect_def_type dt; |
1339 | bool is_simple_use; | |
1340 | tree vector_type; | |
1341 | ||
73fbfcad | 1342 | if (dump_enabled_p ()) |
ebfd146a | 1343 | { |
78c60e3d SS |
1344 | dump_printf_loc (MSG_NOTE, vect_location, |
1345 | "vect_get_vec_def_for_operand: "); | |
1346 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
ebfd146a IR |
1347 | } |
1348 | ||
24ee1384 IR |
1349 | is_simple_use = vect_is_simple_use (op, stmt, loop_vinfo, NULL, |
1350 | &def_stmt, &def, &dt); | |
ebfd146a | 1351 | gcc_assert (is_simple_use); |
73fbfcad | 1352 | if (dump_enabled_p ()) |
ebfd146a | 1353 | { |
78c60e3d | 1354 | int loc_printed = 0; |
ebfd146a IR |
1355 | if (def) |
1356 | { | |
78c60e3d SS |
1357 | dump_printf_loc (MSG_NOTE, vect_location, "def = "); |
1358 | loc_printed = 1; | |
1359 | dump_generic_expr (MSG_NOTE, TDF_SLIM, def); | |
ebfd146a IR |
1360 | } |
1361 | if (def_stmt) | |
1362 | { | |
78c60e3d SS |
1363 | if (loc_printed) |
1364 | dump_printf (MSG_NOTE, " def_stmt = "); | |
1365 | else | |
1366 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1367 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
ebfd146a IR |
1368 | } |
1369 | } | |
1370 | ||
1371 | switch (dt) | |
1372 | { | |
1373 | /* Case 1: operand is a constant. */ | |
1374 | case vect_constant_def: | |
1375 | { | |
7569a6cc RG |
1376 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); |
1377 | gcc_assert (vector_type); | |
9dc3f7de | 1378 | nunits = TYPE_VECTOR_SUBPARTS (vector_type); |
7569a6cc | 1379 | |
b8698a0f | 1380 | if (scalar_def) |
ebfd146a IR |
1381 | *scalar_def = op; |
1382 | ||
1383 | /* Create 'vect_cst_ = {cst,cst,...,cst}' */ | |
73fbfcad | 1384 | if (dump_enabled_p ()) |
78c60e3d SS |
1385 | dump_printf_loc (MSG_NOTE, vect_location, |
1386 | "Create vector_cst. nunits = %d", nunits); | |
ebfd146a | 1387 | |
418b7df3 | 1388 | return vect_init_vector (stmt, op, vector_type, NULL); |
ebfd146a IR |
1389 | } |
1390 | ||
1391 | /* Case 2: operand is defined outside the loop - loop invariant. */ | |
8644a673 | 1392 | case vect_external_def: |
ebfd146a IR |
1393 | { |
1394 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (def)); | |
1395 | gcc_assert (vector_type); | |
ebfd146a | 1396 | |
b8698a0f | 1397 | if (scalar_def) |
ebfd146a IR |
1398 | *scalar_def = def; |
1399 | ||
1400 | /* Create 'vec_inv = {inv,inv,..,inv}' */ | |
73fbfcad | 1401 | if (dump_enabled_p ()) |
78c60e3d | 1402 | dump_printf_loc (MSG_NOTE, vect_location, "Create vector_inv."); |
ebfd146a | 1403 | |
418b7df3 | 1404 | return vect_init_vector (stmt, def, vector_type, NULL); |
ebfd146a IR |
1405 | } |
1406 | ||
1407 | /* Case 3: operand is defined inside the loop. */ | |
8644a673 | 1408 | case vect_internal_def: |
ebfd146a | 1409 | { |
b8698a0f | 1410 | if (scalar_def) |
ebfd146a IR |
1411 | *scalar_def = NULL/* FIXME tuples: def_stmt*/; |
1412 | ||
1413 | /* Get the def from the vectorized stmt. */ | |
1414 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1415 | |
ebfd146a | 1416 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1417 | /* Get vectorized pattern statement. */ |
1418 | if (!vec_stmt | |
1419 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1420 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1421 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1422 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1423 | gcc_assert (vec_stmt); |
1424 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1425 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1426 | else if (is_gimple_call (vec_stmt)) | |
1427 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1428 | else | |
1429 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1430 | return vec_oprnd; | |
1431 | } | |
1432 | ||
1433 | /* Case 4: operand is defined by a loop header phi - reduction */ | |
1434 | case vect_reduction_def: | |
06066f92 | 1435 | case vect_double_reduction_def: |
7c5222ff | 1436 | case vect_nested_cycle: |
ebfd146a IR |
1437 | { |
1438 | struct loop *loop; | |
1439 | ||
1440 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
b8698a0f | 1441 | loop = (gimple_bb (def_stmt))->loop_father; |
ebfd146a IR |
1442 | |
1443 | /* Get the def before the loop */ | |
1444 | op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop)); | |
1445 | return get_initial_def_for_reduction (stmt, op, scalar_def); | |
1446 | } | |
1447 | ||
1448 | /* Case 5: operand is defined by loop-header phi - induction. */ | |
1449 | case vect_induction_def: | |
1450 | { | |
1451 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1452 | ||
1453 | /* Get the def from the vectorized stmt. */ | |
1454 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1455 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1456 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1457 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1458 | else | |
1459 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1460 | return vec_oprnd; |
1461 | } | |
1462 | ||
1463 | default: | |
1464 | gcc_unreachable (); | |
1465 | } | |
1466 | } | |
1467 | ||
1468 | ||
1469 | /* Function vect_get_vec_def_for_stmt_copy | |
1470 | ||
ff802fa1 | 1471 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1472 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1473 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1474 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1475 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1476 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1477 | DT is the type of the vector def VEC_OPRND. |
1478 | ||
1479 | Context: | |
1480 | In case the vectorization factor (VF) is bigger than the number | |
1481 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1482 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1483 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1484 | smallest data-type determines the VF, and as a result, when vectorizing |
1485 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1486 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1487 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1488 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1489 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1490 | ||
1491 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1492 | |
ebfd146a IR |
1493 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1494 | VS1.1: vx.1 = memref1 VS1.2 | |
1495 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1496 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1497 | |
1498 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1499 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1500 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1501 | VSnew.3: vz3 = vx.3 + ... | |
1502 | ||
1503 | The vectorization of S1 is explained in vectorizable_load. | |
1504 | The vectorization of S2: | |
b8698a0f L |
1505 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1506 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1507 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1508 | returns the vector-def 'vx.0'. |
1509 | ||
b8698a0f L |
1510 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1511 | function is called to get the relevant vector-def for each operand. It is | |
1512 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1513 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1514 | ||
b8698a0f L |
1515 | For example, to obtain the vector-def 'vx.1' in order to create the |
1516 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1517 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1518 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1519 | and return its def ('vx.1'). | |
1520 | Overall, to create the above sequence this function will be called 3 times: | |
1521 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1522 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1523 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1524 | ||
1525 | tree | |
1526 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1527 | { | |
1528 | gimple vec_stmt_for_operand; | |
1529 | stmt_vec_info def_stmt_info; | |
1530 | ||
1531 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1532 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1533 | return vec_oprnd; |
1534 | ||
1535 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1536 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1537 | gcc_assert (def_stmt_info); | |
1538 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1539 | gcc_assert (vec_stmt_for_operand); | |
1540 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1541 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) | |
1542 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1543 | else | |
1544 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1545 | return vec_oprnd; | |
1546 | } | |
1547 | ||
1548 | ||
1549 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1550 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a IR |
1551 | |
1552 | static void | |
b8698a0f | 1553 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1554 | vec<tree> *vec_oprnds0, |
1555 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1556 | { |
9771b263 | 1557 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1558 | |
1559 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1560 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1561 | |
9771b263 | 1562 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1563 | { |
9771b263 | 1564 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1565 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1566 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1567 | } |
1568 | } | |
1569 | ||
1570 | ||
d092494c IR |
1571 | /* Get vectorized definitions for OP0 and OP1. |
1572 | REDUC_INDEX is the index of reduction operand in case of reduction, | |
1573 | and -1 otherwise. */ | |
ebfd146a | 1574 | |
d092494c | 1575 | void |
ebfd146a | 1576 | vect_get_vec_defs (tree op0, tree op1, gimple stmt, |
9771b263 DN |
1577 | vec<tree> *vec_oprnds0, |
1578 | vec<tree> *vec_oprnds1, | |
d092494c | 1579 | slp_tree slp_node, int reduc_index) |
ebfd146a IR |
1580 | { |
1581 | if (slp_node) | |
d092494c IR |
1582 | { |
1583 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
9771b263 DN |
1584 | vec<tree> ops; |
1585 | ops.create (nops); | |
37b5ec8f | 1586 | vec<vec<tree> > vec_defs; |
9771b263 | 1587 | vec_defs.create (nops); |
d092494c | 1588 | |
9771b263 | 1589 | ops.quick_push (op0); |
d092494c | 1590 | if (op1) |
9771b263 | 1591 | ops.quick_push (op1); |
d092494c IR |
1592 | |
1593 | vect_get_slp_defs (ops, slp_node, &vec_defs, reduc_index); | |
1594 | ||
37b5ec8f | 1595 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1596 | if (op1) |
37b5ec8f | 1597 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1598 | |
9771b263 DN |
1599 | ops.release (); |
1600 | vec_defs.release (); | |
d092494c | 1601 | } |
ebfd146a IR |
1602 | else |
1603 | { | |
1604 | tree vec_oprnd; | |
1605 | ||
9771b263 | 1606 | vec_oprnds0->create (1); |
b8698a0f | 1607 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt, NULL); |
9771b263 | 1608 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1609 | |
1610 | if (op1) | |
1611 | { | |
9771b263 | 1612 | vec_oprnds1->create (1); |
b8698a0f | 1613 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt, NULL); |
9771b263 | 1614 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1615 | } |
1616 | } | |
1617 | } | |
1618 | ||
1619 | ||
1620 | /* Function vect_finish_stmt_generation. | |
1621 | ||
1622 | Insert a new stmt. */ | |
1623 | ||
1624 | void | |
1625 | vect_finish_stmt_generation (gimple stmt, gimple vec_stmt, | |
1626 | gimple_stmt_iterator *gsi) | |
1627 | { | |
1628 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1629 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 1630 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
ebfd146a IR |
1631 | |
1632 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1633 | ||
54e8e2c3 RG |
1634 | if (!gsi_end_p (*gsi) |
1635 | && gimple_has_mem_ops (vec_stmt)) | |
1636 | { | |
1637 | gimple at_stmt = gsi_stmt (*gsi); | |
1638 | tree vuse = gimple_vuse (at_stmt); | |
1639 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1640 | { | |
1641 | tree vdef = gimple_vdef (at_stmt); | |
1642 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1643 | /* If we have an SSA vuse and insert a store, update virtual | |
1644 | SSA form to avoid triggering the renamer. Do so only | |
1645 | if we can easily see all uses - which is what almost always | |
1646 | happens with the way vectorized stmts are inserted. */ | |
1647 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1648 | && ((is_gimple_assign (vec_stmt) | |
1649 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1650 | || (is_gimple_call (vec_stmt) | |
1651 | && !(gimple_call_flags (vec_stmt) | |
1652 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1653 | { | |
1654 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1655 | gimple_set_vdef (vec_stmt, new_vdef); | |
1656 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1657 | } | |
1658 | } | |
1659 | } | |
ebfd146a IR |
1660 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1661 | ||
b8698a0f | 1662 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, loop_vinfo, |
a70d6342 | 1663 | bb_vinfo)); |
ebfd146a | 1664 | |
73fbfcad | 1665 | if (dump_enabled_p ()) |
ebfd146a | 1666 | { |
78c60e3d SS |
1667 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); |
1668 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
ebfd146a IR |
1669 | } |
1670 | ||
ad885386 | 1671 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
ebfd146a IR |
1672 | } |
1673 | ||
1674 | /* Checks if CALL can be vectorized in type VECTYPE. Returns | |
1675 | a function declaration if the target has a vectorized version | |
1676 | of the function, or NULL_TREE if the function cannot be vectorized. */ | |
1677 | ||
1678 | tree | |
1679 | vectorizable_function (gimple call, tree vectype_out, tree vectype_in) | |
1680 | { | |
1681 | tree fndecl = gimple_call_fndecl (call); | |
ebfd146a IR |
1682 | |
1683 | /* We only handle functions that do not read or clobber memory -- i.e. | |
1684 | const or novops ones. */ | |
1685 | if (!(gimple_call_flags (call) & (ECF_CONST | ECF_NOVOPS))) | |
1686 | return NULL_TREE; | |
1687 | ||
1688 | if (!fndecl | |
1689 | || TREE_CODE (fndecl) != FUNCTION_DECL | |
1690 | || !DECL_BUILT_IN (fndecl)) | |
1691 | return NULL_TREE; | |
1692 | ||
62f7fd21 | 1693 | return targetm.vectorize.builtin_vectorized_function (fndecl, vectype_out, |
ebfd146a IR |
1694 | vectype_in); |
1695 | } | |
1696 | ||
1697 | /* Function vectorizable_call. | |
1698 | ||
b8698a0f L |
1699 | Check if STMT performs a function call that can be vectorized. |
1700 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
1701 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
1702 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1703 | ||
1704 | static bool | |
190c2236 JJ |
1705 | vectorizable_call (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, |
1706 | slp_tree slp_node) | |
ebfd146a IR |
1707 | { |
1708 | tree vec_dest; | |
1709 | tree scalar_dest; | |
1710 | tree op, type; | |
1711 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
1712 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
1713 | tree vectype_out, vectype_in; | |
1714 | int nunits_in; | |
1715 | int nunits_out; | |
1716 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
190c2236 | 1717 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b690cc0f | 1718 | tree fndecl, new_temp, def, rhs_type; |
ebfd146a | 1719 | gimple def_stmt; |
0502fb85 UB |
1720 | enum vect_def_type dt[3] |
1721 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
63827fb8 | 1722 | gimple new_stmt = NULL; |
ebfd146a | 1723 | int ncopies, j; |
6e1aa848 | 1724 | vec<tree> vargs = vNULL; |
ebfd146a IR |
1725 | enum { NARROW, NONE, WIDEN } modifier; |
1726 | size_t i, nargs; | |
9d5e7640 | 1727 | tree lhs; |
ebfd146a | 1728 | |
190c2236 | 1729 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
1730 | return false; |
1731 | ||
8644a673 | 1732 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
1733 | return false; |
1734 | ||
ebfd146a IR |
1735 | /* Is STMT a vectorizable call? */ |
1736 | if (!is_gimple_call (stmt)) | |
1737 | return false; | |
1738 | ||
1739 | if (TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
1740 | return false; | |
1741 | ||
822ba6d7 | 1742 | if (stmt_can_throw_internal (stmt)) |
5a2c1986 IR |
1743 | return false; |
1744 | ||
b690cc0f RG |
1745 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
1746 | ||
ebfd146a IR |
1747 | /* Process function arguments. */ |
1748 | rhs_type = NULL_TREE; | |
b690cc0f | 1749 | vectype_in = NULL_TREE; |
ebfd146a IR |
1750 | nargs = gimple_call_num_args (stmt); |
1751 | ||
1b1562a5 MM |
1752 | /* Bail out if the function has more than three arguments, we do not have |
1753 | interesting builtin functions to vectorize with more than two arguments | |
1754 | except for fma. No arguments is also not good. */ | |
1755 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
1756 | return false; |
1757 | ||
74bf76ed JJ |
1758 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
1759 | if (gimple_call_internal_p (stmt) | |
1760 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
1761 | { | |
1762 | nargs = 0; | |
1763 | rhs_type = unsigned_type_node; | |
1764 | } | |
1765 | ||
ebfd146a IR |
1766 | for (i = 0; i < nargs; i++) |
1767 | { | |
b690cc0f RG |
1768 | tree opvectype; |
1769 | ||
ebfd146a IR |
1770 | op = gimple_call_arg (stmt, i); |
1771 | ||
1772 | /* We can only handle calls with arguments of the same type. */ | |
1773 | if (rhs_type | |
8533c9d8 | 1774 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 1775 | { |
73fbfcad | 1776 | if (dump_enabled_p ()) |
78c60e3d SS |
1777 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1778 | "argument types differ."); | |
ebfd146a IR |
1779 | return false; |
1780 | } | |
b690cc0f RG |
1781 | if (!rhs_type) |
1782 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 1783 | |
24ee1384 | 1784 | if (!vect_is_simple_use_1 (op, stmt, loop_vinfo, bb_vinfo, |
b690cc0f | 1785 | &def_stmt, &def, &dt[i], &opvectype)) |
ebfd146a | 1786 | { |
73fbfcad | 1787 | if (dump_enabled_p ()) |
78c60e3d SS |
1788 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1789 | "use not simple."); | |
ebfd146a IR |
1790 | return false; |
1791 | } | |
ebfd146a | 1792 | |
b690cc0f RG |
1793 | if (!vectype_in) |
1794 | vectype_in = opvectype; | |
1795 | else if (opvectype | |
1796 | && opvectype != vectype_in) | |
1797 | { | |
73fbfcad | 1798 | if (dump_enabled_p ()) |
78c60e3d SS |
1799 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1800 | "argument vector types differ."); | |
b690cc0f RG |
1801 | return false; |
1802 | } | |
1803 | } | |
1804 | /* If all arguments are external or constant defs use a vector type with | |
1805 | the same size as the output vector type. */ | |
ebfd146a | 1806 | if (!vectype_in) |
b690cc0f | 1807 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
1808 | if (vec_stmt) |
1809 | gcc_assert (vectype_in); | |
1810 | if (!vectype_in) | |
1811 | { | |
73fbfcad | 1812 | if (dump_enabled_p ()) |
7d8930a0 | 1813 | { |
78c60e3d SS |
1814 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1815 | "no vectype for scalar type "); | |
1816 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
7d8930a0 IR |
1817 | } |
1818 | ||
1819 | return false; | |
1820 | } | |
ebfd146a IR |
1821 | |
1822 | /* FORNOW */ | |
b690cc0f RG |
1823 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
1824 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
1825 | if (nunits_in == nunits_out / 2) |
1826 | modifier = NARROW; | |
1827 | else if (nunits_out == nunits_in) | |
1828 | modifier = NONE; | |
1829 | else if (nunits_out == nunits_in / 2) | |
1830 | modifier = WIDEN; | |
1831 | else | |
1832 | return false; | |
1833 | ||
1834 | /* For now, we only vectorize functions if a target specific builtin | |
1835 | is available. TODO -- in some cases, it might be profitable to | |
1836 | insert the calls for pieces of the vector, in order to be able | |
1837 | to vectorize other operations in the loop. */ | |
1838 | fndecl = vectorizable_function (stmt, vectype_out, vectype_in); | |
1839 | if (fndecl == NULL_TREE) | |
1840 | { | |
74bf76ed JJ |
1841 | if (gimple_call_internal_p (stmt) |
1842 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE | |
1843 | && !slp_node | |
1844 | && loop_vinfo | |
1845 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
1846 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
1847 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
1848 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
1849 | { | |
1850 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
1851 | { 0, 1, 2, ... vf - 1 } vector. */ | |
1852 | gcc_assert (nargs == 0); | |
1853 | } | |
1854 | else | |
1855 | { | |
1856 | if (dump_enabled_p ()) | |
1857 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1858 | "function is not vectorizable."); | |
1859 | return false; | |
1860 | } | |
ebfd146a IR |
1861 | } |
1862 | ||
5006671f | 1863 | gcc_assert (!gimple_vuse (stmt)); |
ebfd146a | 1864 | |
190c2236 JJ |
1865 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
1866 | ncopies = 1; | |
1867 | else if (modifier == NARROW) | |
ebfd146a IR |
1868 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; |
1869 | else | |
1870 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
1871 | ||
1872 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
1873 | needs to be generated. */ | |
1874 | gcc_assert (ncopies >= 1); | |
1875 | ||
1876 | if (!vec_stmt) /* transformation not required. */ | |
1877 | { | |
1878 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 1879 | if (dump_enabled_p ()) |
78c60e3d | 1880 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ==="); |
c3e7ee41 | 1881 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
1882 | return true; |
1883 | } | |
1884 | ||
1885 | /** Transform. **/ | |
1886 | ||
73fbfcad | 1887 | if (dump_enabled_p ()) |
78c60e3d | 1888 | dump_printf_loc (MSG_NOTE, vect_location, "transform call."); |
ebfd146a IR |
1889 | |
1890 | /* Handle def. */ | |
1891 | scalar_dest = gimple_call_lhs (stmt); | |
1892 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
1893 | ||
1894 | prev_stmt_info = NULL; | |
1895 | switch (modifier) | |
1896 | { | |
1897 | case NONE: | |
1898 | for (j = 0; j < ncopies; ++j) | |
1899 | { | |
1900 | /* Build argument list for the vectorized call. */ | |
1901 | if (j == 0) | |
9771b263 | 1902 | vargs.create (nargs); |
ebfd146a | 1903 | else |
9771b263 | 1904 | vargs.truncate (0); |
ebfd146a | 1905 | |
190c2236 JJ |
1906 | if (slp_node) |
1907 | { | |
37b5ec8f | 1908 | vec<vec<tree> > vec_defs; |
9771b263 DN |
1909 | vec_defs.create (nargs); |
1910 | vec<tree> vec_oprnds0; | |
190c2236 JJ |
1911 | |
1912 | for (i = 0; i < nargs; i++) | |
9771b263 | 1913 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 1914 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 1915 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
1916 | |
1917 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 1918 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
1919 | { |
1920 | size_t k; | |
1921 | for (k = 0; k < nargs; k++) | |
1922 | { | |
37b5ec8f | 1923 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 1924 | vargs[k] = vec_oprndsk[i]; |
190c2236 JJ |
1925 | } |
1926 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1927 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1928 | gimple_call_set_lhs (new_stmt, new_temp); | |
1929 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 1930 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
1931 | } |
1932 | ||
1933 | for (i = 0; i < nargs; i++) | |
1934 | { | |
37b5ec8f | 1935 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 1936 | vec_oprndsi.release (); |
190c2236 | 1937 | } |
9771b263 | 1938 | vec_defs.release (); |
190c2236 JJ |
1939 | continue; |
1940 | } | |
1941 | ||
ebfd146a IR |
1942 | for (i = 0; i < nargs; i++) |
1943 | { | |
1944 | op = gimple_call_arg (stmt, i); | |
1945 | if (j == 0) | |
1946 | vec_oprnd0 | |
1947 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
1948 | else | |
63827fb8 IR |
1949 | { |
1950 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
1951 | vec_oprnd0 | |
1952 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
1953 | } | |
ebfd146a | 1954 | |
9771b263 | 1955 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
1956 | } |
1957 | ||
74bf76ed JJ |
1958 | if (gimple_call_internal_p (stmt) |
1959 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
1960 | { | |
1961 | tree *v = XALLOCAVEC (tree, nunits_out); | |
1962 | int k; | |
1963 | for (k = 0; k < nunits_out; ++k) | |
1964 | v[k] = build_int_cst (unsigned_type_node, j * nunits_out + k); | |
1965 | tree cst = build_vector (vectype_out, v); | |
1966 | tree new_var | |
1967 | = vect_get_new_vect_var (vectype_out, vect_simple_var, "cst_"); | |
1968 | gimple init_stmt = gimple_build_assign (new_var, cst); | |
1969 | new_temp = make_ssa_name (new_var, init_stmt); | |
1970 | gimple_assign_set_lhs (init_stmt, new_temp); | |
1971 | vect_init_vector_1 (stmt, init_stmt, NULL); | |
1972 | new_temp = make_ssa_name (vec_dest, NULL); | |
1973 | new_stmt = gimple_build_assign (new_temp, | |
1974 | gimple_assign_lhs (init_stmt)); | |
1975 | } | |
1976 | else | |
1977 | { | |
1978 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
1979 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
1980 | gimple_call_set_lhs (new_stmt, new_temp); | |
1981 | } | |
ebfd146a IR |
1982 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1983 | ||
1984 | if (j == 0) | |
1985 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1986 | else | |
1987 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1988 | ||
1989 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1990 | } | |
1991 | ||
1992 | break; | |
1993 | ||
1994 | case NARROW: | |
1995 | for (j = 0; j < ncopies; ++j) | |
1996 | { | |
1997 | /* Build argument list for the vectorized call. */ | |
1998 | if (j == 0) | |
9771b263 | 1999 | vargs.create (nargs * 2); |
ebfd146a | 2000 | else |
9771b263 | 2001 | vargs.truncate (0); |
ebfd146a | 2002 | |
190c2236 JJ |
2003 | if (slp_node) |
2004 | { | |
37b5ec8f | 2005 | vec<vec<tree> > vec_defs; |
9771b263 DN |
2006 | vec_defs.create (nargs); |
2007 | vec<tree> vec_oprnds0; | |
190c2236 JJ |
2008 | |
2009 | for (i = 0; i < nargs; i++) | |
9771b263 | 2010 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2011 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2012 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2013 | |
2014 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2015 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
2016 | { |
2017 | size_t k; | |
9771b263 | 2018 | vargs.truncate (0); |
190c2236 JJ |
2019 | for (k = 0; k < nargs; k++) |
2020 | { | |
37b5ec8f | 2021 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
2022 | vargs.quick_push (vec_oprndsk[i]); |
2023 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 JJ |
2024 | } |
2025 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2026 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2027 | gimple_call_set_lhs (new_stmt, new_temp); | |
2028 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 2029 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2030 | } |
2031 | ||
2032 | for (i = 0; i < nargs; i++) | |
2033 | { | |
37b5ec8f | 2034 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2035 | vec_oprndsi.release (); |
190c2236 | 2036 | } |
9771b263 | 2037 | vec_defs.release (); |
190c2236 JJ |
2038 | continue; |
2039 | } | |
2040 | ||
ebfd146a IR |
2041 | for (i = 0; i < nargs; i++) |
2042 | { | |
2043 | op = gimple_call_arg (stmt, i); | |
2044 | if (j == 0) | |
2045 | { | |
2046 | vec_oprnd0 | |
2047 | = vect_get_vec_def_for_operand (op, stmt, NULL); | |
2048 | vec_oprnd1 | |
63827fb8 | 2049 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2050 | } |
2051 | else | |
2052 | { | |
336ecb65 | 2053 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 2054 | vec_oprnd0 |
63827fb8 | 2055 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 2056 | vec_oprnd1 |
63827fb8 | 2057 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2058 | } |
2059 | ||
9771b263 DN |
2060 | vargs.quick_push (vec_oprnd0); |
2061 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
2062 | } |
2063 | ||
2064 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2065 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2066 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
2067 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2068 | ||
2069 | if (j == 0) | |
2070 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2071 | else | |
2072 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2073 | ||
2074 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2075 | } | |
2076 | ||
2077 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
2078 | ||
2079 | break; | |
2080 | ||
2081 | case WIDEN: | |
2082 | /* No current target implements this case. */ | |
2083 | return false; | |
2084 | } | |
2085 | ||
9771b263 | 2086 | vargs.release (); |
ebfd146a IR |
2087 | |
2088 | /* Update the exception handling table with the vector stmt if necessary. */ | |
2089 | if (maybe_clean_or_replace_eh_stmt (stmt, *vec_stmt)) | |
2090 | gimple_purge_dead_eh_edges (gimple_bb (stmt)); | |
2091 | ||
2092 | /* The call in STMT might prevent it from being removed in dce. | |
2093 | We however cannot remove it here, due to the way the ssa name | |
2094 | it defines is mapped to the new definition. So just replace | |
2095 | rhs of the statement with something harmless. */ | |
2096 | ||
dd34c087 JJ |
2097 | if (slp_node) |
2098 | return true; | |
2099 | ||
ebfd146a | 2100 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
2101 | if (is_pattern_stmt_p (stmt_info)) |
2102 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
2103 | else | |
2104 | lhs = gimple_call_lhs (stmt); | |
2105 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
ebfd146a | 2106 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 2107 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
2108 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
2109 | gsi_replace (gsi, new_stmt, false); | |
2110 | SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt)) = new_stmt; | |
2111 | ||
2112 | return true; | |
2113 | } | |
2114 | ||
2115 | ||
2116 | /* Function vect_gen_widened_results_half | |
2117 | ||
2118 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 2119 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 2120 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
2121 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
2122 | needs to be created (DECL is a function-decl of a target-builtin). | |
2123 | STMT is the original scalar stmt that we are vectorizing. */ | |
2124 | ||
2125 | static gimple | |
2126 | vect_gen_widened_results_half (enum tree_code code, | |
2127 | tree decl, | |
2128 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
2129 | tree vec_dest, gimple_stmt_iterator *gsi, | |
2130 | gimple stmt) | |
b8698a0f | 2131 | { |
ebfd146a | 2132 | gimple new_stmt; |
b8698a0f L |
2133 | tree new_temp; |
2134 | ||
2135 | /* Generate half of the widened result: */ | |
2136 | if (code == CALL_EXPR) | |
2137 | { | |
2138 | /* Target specific support */ | |
ebfd146a IR |
2139 | if (op_type == binary_op) |
2140 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
2141 | else | |
2142 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
2143 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2144 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
2145 | } |
2146 | else | |
ebfd146a | 2147 | { |
b8698a0f L |
2148 | /* Generic support */ |
2149 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
2150 | if (op_type != binary_op) |
2151 | vec_oprnd1 = NULL; | |
2152 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vec_oprnd0, | |
2153 | vec_oprnd1); | |
2154 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2155 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 2156 | } |
ebfd146a IR |
2157 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2158 | ||
ebfd146a IR |
2159 | return new_stmt; |
2160 | } | |
2161 | ||
4a00c761 JJ |
2162 | |
2163 | /* Get vectorized definitions for loop-based vectorization. For the first | |
2164 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
2165 | scalar operand), and for the rest we get a copy with | |
2166 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
2167 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
2168 | The vectors are collected into VEC_OPRNDS. */ | |
2169 | ||
2170 | static void | |
2171 | vect_get_loop_based_defs (tree *oprnd, gimple stmt, enum vect_def_type dt, | |
9771b263 | 2172 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
2173 | { |
2174 | tree vec_oprnd; | |
2175 | ||
2176 | /* Get first vector operand. */ | |
2177 | /* All the vector operands except the very first one (that is scalar oprnd) | |
2178 | are stmt copies. */ | |
2179 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
2180 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt, NULL); | |
2181 | else | |
2182 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
2183 | ||
9771b263 | 2184 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
2185 | |
2186 | /* Get second vector operand. */ | |
2187 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 2188 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
2189 | |
2190 | *oprnd = vec_oprnd; | |
2191 | ||
2192 | /* For conversion in multiple steps, continue to get operands | |
2193 | recursively. */ | |
2194 | if (multi_step_cvt) | |
2195 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
2196 | } | |
2197 | ||
2198 | ||
2199 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
2200 | For multi-step conversions store the resulting vectors and call the function | |
2201 | recursively. */ | |
2202 | ||
2203 | static void | |
9771b263 | 2204 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
4a00c761 | 2205 | int multi_step_cvt, gimple stmt, |
9771b263 | 2206 | vec<tree> vec_dsts, |
4a00c761 JJ |
2207 | gimple_stmt_iterator *gsi, |
2208 | slp_tree slp_node, enum tree_code code, | |
2209 | stmt_vec_info *prev_stmt_info) | |
2210 | { | |
2211 | unsigned int i; | |
2212 | tree vop0, vop1, new_tmp, vec_dest; | |
2213 | gimple new_stmt; | |
2214 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2215 | ||
9771b263 | 2216 | vec_dest = vec_dsts.pop (); |
4a00c761 | 2217 | |
9771b263 | 2218 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
2219 | { |
2220 | /* Create demotion operation. */ | |
9771b263 DN |
2221 | vop0 = (*vec_oprnds)[i]; |
2222 | vop1 = (*vec_oprnds)[i + 1]; | |
4a00c761 JJ |
2223 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); |
2224 | new_tmp = make_ssa_name (vec_dest, new_stmt); | |
2225 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
2226 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2227 | ||
2228 | if (multi_step_cvt) | |
2229 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 2230 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
2231 | else |
2232 | { | |
2233 | /* This is the last step of the conversion sequence. Store the | |
2234 | vectors in SLP_NODE or in vector info of the scalar statement | |
2235 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
2236 | if (slp_node) | |
9771b263 | 2237 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
2238 | else |
2239 | { | |
2240 | if (!*prev_stmt_info) | |
2241 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2242 | else | |
2243 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
2244 | ||
2245 | *prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2246 | } | |
2247 | } | |
2248 | } | |
2249 | ||
2250 | /* For multi-step demotion operations we first generate demotion operations | |
2251 | from the source type to the intermediate types, and then combine the | |
2252 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
2253 | type. */ | |
2254 | if (multi_step_cvt) | |
2255 | { | |
2256 | /* At each level of recursion we have half of the operands we had at the | |
2257 | previous level. */ | |
9771b263 | 2258 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
2259 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
2260 | stmt, vec_dsts, gsi, slp_node, | |
2261 | VEC_PACK_TRUNC_EXPR, | |
2262 | prev_stmt_info); | |
2263 | } | |
2264 | ||
9771b263 | 2265 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
2266 | } |
2267 | ||
2268 | ||
2269 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
2270 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
2271 | the resulting vectors and call the function recursively. */ | |
2272 | ||
2273 | static void | |
9771b263 DN |
2274 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
2275 | vec<tree> *vec_oprnds1, | |
4a00c761 JJ |
2276 | gimple stmt, tree vec_dest, |
2277 | gimple_stmt_iterator *gsi, | |
2278 | enum tree_code code1, | |
2279 | enum tree_code code2, tree decl1, | |
2280 | tree decl2, int op_type) | |
2281 | { | |
2282 | int i; | |
2283 | tree vop0, vop1, new_tmp1, new_tmp2; | |
2284 | gimple new_stmt1, new_stmt2; | |
6e1aa848 | 2285 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 2286 | |
9771b263 DN |
2287 | vec_tmp.create (vec_oprnds0->length () * 2); |
2288 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
2289 | { |
2290 | if (op_type == binary_op) | |
9771b263 | 2291 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
2292 | else |
2293 | vop1 = NULL_TREE; | |
2294 | ||
2295 | /* Generate the two halves of promotion operation. */ | |
2296 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
2297 | op_type, vec_dest, gsi, stmt); | |
2298 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
2299 | op_type, vec_dest, gsi, stmt); | |
2300 | if (is_gimple_call (new_stmt1)) | |
2301 | { | |
2302 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
2303 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
2304 | } | |
2305 | else | |
2306 | { | |
2307 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
2308 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
2309 | } | |
2310 | ||
2311 | /* Store the results for the next step. */ | |
9771b263 DN |
2312 | vec_tmp.quick_push (new_tmp1); |
2313 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
2314 | } |
2315 | ||
689eaba3 | 2316 | vec_oprnds0->release (); |
4a00c761 JJ |
2317 | *vec_oprnds0 = vec_tmp; |
2318 | } | |
2319 | ||
2320 | ||
b8698a0f L |
2321 | /* Check if STMT performs a conversion operation, that can be vectorized. |
2322 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 2323 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
2324 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
2325 | ||
2326 | static bool | |
2327 | vectorizable_conversion (gimple stmt, gimple_stmt_iterator *gsi, | |
2328 | gimple *vec_stmt, slp_tree slp_node) | |
2329 | { | |
2330 | tree vec_dest; | |
2331 | tree scalar_dest; | |
4a00c761 | 2332 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
2333 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
2334 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2335 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2336 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 2337 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
2338 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
2339 | tree new_temp; | |
2340 | tree def; | |
2341 | gimple def_stmt; | |
2342 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
2343 | gimple new_stmt = NULL; | |
2344 | stmt_vec_info prev_stmt_info; | |
2345 | int nunits_in; | |
2346 | int nunits_out; | |
2347 | tree vectype_out, vectype_in; | |
4a00c761 JJ |
2348 | int ncopies, i, j; |
2349 | tree lhs_type, rhs_type; | |
ebfd146a | 2350 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
2351 | vec<tree> vec_oprnds0 = vNULL; |
2352 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 2353 | tree vop0; |
4a00c761 JJ |
2354 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
2355 | int multi_step_cvt = 0; | |
6e1aa848 DN |
2356 | vec<tree> vec_dsts = vNULL; |
2357 | vec<tree> interm_types = vNULL; | |
4a00c761 JJ |
2358 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
2359 | int op_type; | |
2360 | enum machine_mode rhs_mode; | |
2361 | unsigned short fltsz; | |
ebfd146a IR |
2362 | |
2363 | /* Is STMT a vectorizable conversion? */ | |
2364 | ||
4a00c761 | 2365 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2366 | return false; |
2367 | ||
8644a673 | 2368 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2369 | return false; |
2370 | ||
2371 | if (!is_gimple_assign (stmt)) | |
2372 | return false; | |
2373 | ||
2374 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
2375 | return false; | |
2376 | ||
2377 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
2378 | if (!CONVERT_EXPR_CODE_P (code) |
2379 | && code != FIX_TRUNC_EXPR | |
2380 | && code != FLOAT_EXPR | |
2381 | && code != WIDEN_MULT_EXPR | |
2382 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
2383 | return false; |
2384 | ||
4a00c761 JJ |
2385 | op_type = TREE_CODE_LENGTH (code); |
2386 | ||
ebfd146a | 2387 | /* Check types of lhs and rhs. */ |
b690cc0f | 2388 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 2389 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
2390 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
2391 | ||
ebfd146a IR |
2392 | op0 = gimple_assign_rhs1 (stmt); |
2393 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
2394 | |
2395 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
2396 | && !((INTEGRAL_TYPE_P (lhs_type) | |
2397 | && INTEGRAL_TYPE_P (rhs_type)) | |
2398 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
2399 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
2400 | return false; | |
2401 | ||
2402 | if ((INTEGRAL_TYPE_P (lhs_type) | |
2403 | && (TYPE_PRECISION (lhs_type) | |
2404 | != GET_MODE_PRECISION (TYPE_MODE (lhs_type)))) | |
2405 | || (INTEGRAL_TYPE_P (rhs_type) | |
2406 | && (TYPE_PRECISION (rhs_type) | |
2407 | != GET_MODE_PRECISION (TYPE_MODE (rhs_type))))) | |
2408 | { | |
73fbfcad | 2409 | if (dump_enabled_p ()) |
78c60e3d SS |
2410 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2411 | "type conversion to/from bit-precision unsupported."); | |
4a00c761 JJ |
2412 | return false; |
2413 | } | |
2414 | ||
b690cc0f | 2415 | /* Check the operands of the operation. */ |
24ee1384 | 2416 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
b690cc0f RG |
2417 | &def_stmt, &def, &dt[0], &vectype_in)) |
2418 | { | |
73fbfcad | 2419 | if (dump_enabled_p ()) |
78c60e3d SS |
2420 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2421 | "use not simple."); | |
b690cc0f RG |
2422 | return false; |
2423 | } | |
4a00c761 JJ |
2424 | if (op_type == binary_op) |
2425 | { | |
2426 | bool ok; | |
2427 | ||
2428 | op1 = gimple_assign_rhs2 (stmt); | |
2429 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
2430 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
2431 | OP1. */ | |
2432 | if (CONSTANT_CLASS_P (op0)) | |
f5709183 | 2433 | ok = vect_is_simple_use_1 (op1, stmt, loop_vinfo, bb_vinfo, |
4a00c761 JJ |
2434 | &def_stmt, &def, &dt[1], &vectype_in); |
2435 | else | |
f5709183 | 2436 | ok = vect_is_simple_use (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
24ee1384 | 2437 | &def, &dt[1]); |
4a00c761 JJ |
2438 | |
2439 | if (!ok) | |
2440 | { | |
73fbfcad | 2441 | if (dump_enabled_p ()) |
78c60e3d SS |
2442 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2443 | "use not simple."); | |
4a00c761 JJ |
2444 | return false; |
2445 | } | |
2446 | } | |
2447 | ||
b690cc0f RG |
2448 | /* If op0 is an external or constant defs use a vector type of |
2449 | the same size as the output vector type. */ | |
ebfd146a | 2450 | if (!vectype_in) |
b690cc0f | 2451 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
2452 | if (vec_stmt) |
2453 | gcc_assert (vectype_in); | |
2454 | if (!vectype_in) | |
2455 | { | |
73fbfcad | 2456 | if (dump_enabled_p ()) |
4a00c761 | 2457 | { |
78c60e3d SS |
2458 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2459 | "no vectype for scalar type "); | |
2460 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
4a00c761 | 2461 | } |
7d8930a0 IR |
2462 | |
2463 | return false; | |
2464 | } | |
ebfd146a | 2465 | |
b690cc0f RG |
2466 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
2467 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4a00c761 | 2468 | if (nunits_in < nunits_out) |
ebfd146a IR |
2469 | modifier = NARROW; |
2470 | else if (nunits_out == nunits_in) | |
2471 | modifier = NONE; | |
ebfd146a | 2472 | else |
4a00c761 | 2473 | modifier = WIDEN; |
ebfd146a | 2474 | |
ff802fa1 IR |
2475 | /* Multiple types in SLP are handled by creating the appropriate number of |
2476 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2477 | case of SLP. */ | |
437f4a00 | 2478 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a | 2479 | ncopies = 1; |
4a00c761 JJ |
2480 | else if (modifier == NARROW) |
2481 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
2482 | else | |
2483 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
b8698a0f | 2484 | |
ebfd146a IR |
2485 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
2486 | needs to be generated. */ | |
2487 | gcc_assert (ncopies >= 1); | |
2488 | ||
ebfd146a | 2489 | /* Supportable by target? */ |
4a00c761 | 2490 | switch (modifier) |
ebfd146a | 2491 | { |
4a00c761 JJ |
2492 | case NONE: |
2493 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
2494 | return false; | |
2495 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
2496 | &decl1, &code1)) | |
2497 | break; | |
2498 | /* FALLTHRU */ | |
2499 | unsupported: | |
73fbfcad | 2500 | if (dump_enabled_p ()) |
78c60e3d SS |
2501 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2502 | "conversion not supported by target."); | |
ebfd146a | 2503 | return false; |
ebfd146a | 2504 | |
4a00c761 JJ |
2505 | case WIDEN: |
2506 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
2507 | &code1, &code2, &multi_step_cvt, |
2508 | &interm_types)) | |
4a00c761 JJ |
2509 | { |
2510 | /* Binary widening operation can only be supported directly by the | |
2511 | architecture. */ | |
2512 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
2513 | break; | |
2514 | } | |
2515 | ||
2516 | if (code != FLOAT_EXPR | |
2517 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
2518 | <= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
2519 | goto unsupported; | |
2520 | ||
2521 | rhs_mode = TYPE_MODE (rhs_type); | |
2522 | fltsz = GET_MODE_SIZE (TYPE_MODE (lhs_type)); | |
2523 | for (rhs_mode = GET_MODE_2XWIDER_MODE (TYPE_MODE (rhs_type)); | |
2524 | rhs_mode != VOIDmode && GET_MODE_SIZE (rhs_mode) <= fltsz; | |
2525 | rhs_mode = GET_MODE_2XWIDER_MODE (rhs_mode)) | |
2526 | { | |
2527 | cvt_type | |
2528 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
2529 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
2530 | if (cvt_type == NULL_TREE) | |
2531 | goto unsupported; | |
2532 | ||
2533 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
2534 | { | |
2535 | if (!supportable_convert_operation (code, vectype_out, | |
2536 | cvt_type, &decl1, &codecvt1)) | |
2537 | goto unsupported; | |
2538 | } | |
2539 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
2540 | cvt_type, &codecvt1, |
2541 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
2542 | &interm_types)) |
2543 | continue; | |
2544 | else | |
2545 | gcc_assert (multi_step_cvt == 0); | |
2546 | ||
2547 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
2548 | vectype_in, &code1, &code2, |
2549 | &multi_step_cvt, &interm_types)) | |
4a00c761 JJ |
2550 | break; |
2551 | } | |
2552 | ||
2553 | if (rhs_mode == VOIDmode || GET_MODE_SIZE (rhs_mode) > fltsz) | |
2554 | goto unsupported; | |
2555 | ||
2556 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
2557 | codecvt2 = ERROR_MARK; | |
2558 | else | |
2559 | { | |
2560 | multi_step_cvt++; | |
9771b263 | 2561 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
2562 | cvt_type = NULL_TREE; |
2563 | } | |
2564 | break; | |
2565 | ||
2566 | case NARROW: | |
2567 | gcc_assert (op_type == unary_op); | |
2568 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
2569 | &code1, &multi_step_cvt, | |
2570 | &interm_types)) | |
2571 | break; | |
2572 | ||
2573 | if (code != FIX_TRUNC_EXPR | |
2574 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
2575 | >= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
2576 | goto unsupported; | |
2577 | ||
2578 | rhs_mode = TYPE_MODE (rhs_type); | |
2579 | cvt_type | |
2580 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
2581 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
2582 | if (cvt_type == NULL_TREE) | |
2583 | goto unsupported; | |
2584 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
2585 | &decl1, &codecvt1)) | |
2586 | goto unsupported; | |
2587 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
2588 | &code1, &multi_step_cvt, | |
2589 | &interm_types)) | |
2590 | break; | |
2591 | goto unsupported; | |
2592 | ||
2593 | default: | |
2594 | gcc_unreachable (); | |
ebfd146a IR |
2595 | } |
2596 | ||
2597 | if (!vec_stmt) /* transformation not required. */ | |
2598 | { | |
73fbfcad | 2599 | if (dump_enabled_p ()) |
78c60e3d SS |
2600 | dump_printf_loc (MSG_NOTE, vect_location, |
2601 | "=== vectorizable_conversion ==="); | |
4a00c761 | 2602 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
2603 | { |
2604 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
c3e7ee41 | 2605 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
8bd37302 | 2606 | } |
4a00c761 JJ |
2607 | else if (modifier == NARROW) |
2608 | { | |
2609 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
8bd37302 | 2610 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 JJ |
2611 | } |
2612 | else | |
2613 | { | |
2614 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
8bd37302 | 2615 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 | 2616 | } |
9771b263 | 2617 | interm_types.release (); |
ebfd146a IR |
2618 | return true; |
2619 | } | |
2620 | ||
2621 | /** Transform. **/ | |
73fbfcad | 2622 | if (dump_enabled_p ()) |
78c60e3d SS |
2623 | dump_printf_loc (MSG_NOTE, vect_location, |
2624 | "transform conversion. ncopies = %d.", ncopies); | |
ebfd146a | 2625 | |
4a00c761 JJ |
2626 | if (op_type == binary_op) |
2627 | { | |
2628 | if (CONSTANT_CLASS_P (op0)) | |
2629 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
2630 | else if (CONSTANT_CLASS_P (op1)) | |
2631 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
2632 | } | |
2633 | ||
2634 | /* In case of multi-step conversion, we first generate conversion operations | |
2635 | to the intermediate types, and then from that types to the final one. | |
2636 | We create vector destinations for the intermediate type (TYPES) received | |
2637 | from supportable_*_operation, and store them in the correct order | |
2638 | for future use in vect_create_vectorized_*_stmts (). */ | |
9771b263 | 2639 | vec_dsts.create (multi_step_cvt + 1); |
82294ec1 JJ |
2640 | vec_dest = vect_create_destination_var (scalar_dest, |
2641 | (cvt_type && modifier == WIDEN) | |
2642 | ? cvt_type : vectype_out); | |
9771b263 | 2643 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
2644 | |
2645 | if (multi_step_cvt) | |
2646 | { | |
9771b263 DN |
2647 | for (i = interm_types.length () - 1; |
2648 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
2649 | { |
2650 | vec_dest = vect_create_destination_var (scalar_dest, | |
2651 | intermediate_type); | |
9771b263 | 2652 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
2653 | } |
2654 | } | |
ebfd146a | 2655 | |
4a00c761 | 2656 | if (cvt_type) |
82294ec1 JJ |
2657 | vec_dest = vect_create_destination_var (scalar_dest, |
2658 | modifier == WIDEN | |
2659 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
2660 | |
2661 | if (!slp_node) | |
2662 | { | |
30862efc | 2663 | if (modifier == WIDEN) |
4a00c761 | 2664 | { |
9771b263 | 2665 | vec_oprnds0.create (multi_step_cvt ? vect_pow2(multi_step_cvt) : 1); |
4a00c761 | 2666 | if (op_type == binary_op) |
9771b263 | 2667 | vec_oprnds1.create (1); |
4a00c761 | 2668 | } |
30862efc | 2669 | else if (modifier == NARROW) |
9771b263 DN |
2670 | vec_oprnds0.create ( |
2671 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
2672 | } |
2673 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 2674 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 2675 | |
4a00c761 | 2676 | last_oprnd = op0; |
ebfd146a IR |
2677 | prev_stmt_info = NULL; |
2678 | switch (modifier) | |
2679 | { | |
2680 | case NONE: | |
2681 | for (j = 0; j < ncopies; j++) | |
2682 | { | |
ebfd146a | 2683 | if (j == 0) |
d092494c IR |
2684 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node, |
2685 | -1); | |
ebfd146a IR |
2686 | else |
2687 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
2688 | ||
9771b263 | 2689 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
2690 | { |
2691 | /* Arguments are ready, create the new vector stmt. */ | |
2692 | if (code1 == CALL_EXPR) | |
2693 | { | |
2694 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
2695 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2696 | gimple_call_set_lhs (new_stmt, new_temp); | |
2697 | } | |
2698 | else | |
2699 | { | |
2700 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
2701 | new_stmt = gimple_build_assign_with_ops (code1, vec_dest, | |
2702 | vop0, NULL); | |
2703 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2704 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2705 | } | |
2706 | ||
2707 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2708 | if (slp_node) | |
9771b263 | 2709 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
2710 | } |
2711 | ||
ebfd146a IR |
2712 | if (j == 0) |
2713 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2714 | else | |
2715 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2716 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2717 | } | |
2718 | break; | |
2719 | ||
2720 | case WIDEN: | |
2721 | /* In case the vectorization factor (VF) is bigger than the number | |
2722 | of elements that we can fit in a vectype (nunits), we have to | |
2723 | generate more than one vector stmt - i.e - we need to "unroll" | |
2724 | the vector stmt by a factor VF/nunits. */ | |
2725 | for (j = 0; j < ncopies; j++) | |
2726 | { | |
4a00c761 | 2727 | /* Handle uses. */ |
ebfd146a | 2728 | if (j == 0) |
4a00c761 JJ |
2729 | { |
2730 | if (slp_node) | |
2731 | { | |
2732 | if (code == WIDEN_LSHIFT_EXPR) | |
2733 | { | |
2734 | unsigned int k; | |
ebfd146a | 2735 | |
4a00c761 JJ |
2736 | vec_oprnd1 = op1; |
2737 | /* Store vec_oprnd1 for every vector stmt to be created | |
2738 | for SLP_NODE. We check during the analysis that all | |
2739 | the shift arguments are the same. */ | |
2740 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 2741 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
2742 | |
2743 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
2744 | slp_node, -1); | |
2745 | } | |
2746 | else | |
2747 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
2748 | &vec_oprnds1, slp_node, -1); | |
2749 | } | |
2750 | else | |
2751 | { | |
2752 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); | |
9771b263 | 2753 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
2754 | if (op_type == binary_op) |
2755 | { | |
2756 | if (code == WIDEN_LSHIFT_EXPR) | |
2757 | vec_oprnd1 = op1; | |
2758 | else | |
2759 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, | |
2760 | NULL); | |
9771b263 | 2761 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
2762 | } |
2763 | } | |
2764 | } | |
ebfd146a | 2765 | else |
4a00c761 JJ |
2766 | { |
2767 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
2768 | vec_oprnds0.truncate (0); |
2769 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
2770 | if (op_type == binary_op) |
2771 | { | |
2772 | if (code == WIDEN_LSHIFT_EXPR) | |
2773 | vec_oprnd1 = op1; | |
2774 | else | |
2775 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
2776 | vec_oprnd1); | |
9771b263 DN |
2777 | vec_oprnds1.truncate (0); |
2778 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
2779 | } |
2780 | } | |
ebfd146a | 2781 | |
4a00c761 JJ |
2782 | /* Arguments are ready. Create the new vector stmts. */ |
2783 | for (i = multi_step_cvt; i >= 0; i--) | |
2784 | { | |
9771b263 | 2785 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
2786 | enum tree_code c1 = code1, c2 = code2; |
2787 | if (i == 0 && codecvt2 != ERROR_MARK) | |
2788 | { | |
2789 | c1 = codecvt1; | |
2790 | c2 = codecvt2; | |
2791 | } | |
2792 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
2793 | &vec_oprnds1, | |
2794 | stmt, this_dest, gsi, | |
2795 | c1, c2, decl1, decl2, | |
2796 | op_type); | |
2797 | } | |
2798 | ||
9771b263 | 2799 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
2800 | { |
2801 | if (cvt_type) | |
2802 | { | |
2803 | if (codecvt1 == CALL_EXPR) | |
2804 | { | |
2805 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
2806 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2807 | gimple_call_set_lhs (new_stmt, new_temp); | |
2808 | } | |
2809 | else | |
2810 | { | |
2811 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
2812 | new_temp = make_ssa_name (vec_dest, NULL); | |
2813 | new_stmt = gimple_build_assign_with_ops (codecvt1, | |
2814 | new_temp, | |
2815 | vop0, NULL); | |
2816 | } | |
2817 | ||
2818 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2819 | } | |
2820 | else | |
2821 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
2822 | ||
2823 | if (slp_node) | |
9771b263 | 2824 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 JJ |
2825 | else |
2826 | { | |
2827 | if (!prev_stmt_info) | |
2828 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2829 | else | |
2830 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2831 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2832 | } | |
2833 | } | |
ebfd146a | 2834 | } |
4a00c761 JJ |
2835 | |
2836 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
2837 | break; |
2838 | ||
2839 | case NARROW: | |
2840 | /* In case the vectorization factor (VF) is bigger than the number | |
2841 | of elements that we can fit in a vectype (nunits), we have to | |
2842 | generate more than one vector stmt - i.e - we need to "unroll" | |
2843 | the vector stmt by a factor VF/nunits. */ | |
2844 | for (j = 0; j < ncopies; j++) | |
2845 | { | |
2846 | /* Handle uses. */ | |
4a00c761 JJ |
2847 | if (slp_node) |
2848 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
2849 | slp_node, -1); | |
ebfd146a IR |
2850 | else |
2851 | { | |
9771b263 | 2852 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
2853 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
2854 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
2855 | } |
2856 | ||
4a00c761 JJ |
2857 | /* Arguments are ready. Create the new vector stmts. */ |
2858 | if (cvt_type) | |
9771b263 | 2859 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
2860 | { |
2861 | if (codecvt1 == CALL_EXPR) | |
2862 | { | |
2863 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
2864 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2865 | gimple_call_set_lhs (new_stmt, new_temp); | |
2866 | } | |
2867 | else | |
2868 | { | |
2869 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
2870 | new_temp = make_ssa_name (vec_dest, NULL); | |
2871 | new_stmt = gimple_build_assign_with_ops (codecvt1, new_temp, | |
2872 | vop0, NULL); | |
2873 | } | |
ebfd146a | 2874 | |
4a00c761 | 2875 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 2876 | vec_oprnds0[i] = new_temp; |
4a00c761 | 2877 | } |
ebfd146a | 2878 | |
4a00c761 JJ |
2879 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
2880 | stmt, vec_dsts, gsi, | |
2881 | slp_node, code1, | |
2882 | &prev_stmt_info); | |
ebfd146a IR |
2883 | } |
2884 | ||
2885 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 2886 | break; |
ebfd146a IR |
2887 | } |
2888 | ||
9771b263 DN |
2889 | vec_oprnds0.release (); |
2890 | vec_oprnds1.release (); | |
2891 | vec_dsts.release (); | |
2892 | interm_types.release (); | |
ebfd146a IR |
2893 | |
2894 | return true; | |
2895 | } | |
ff802fa1 IR |
2896 | |
2897 | ||
ebfd146a IR |
2898 | /* Function vectorizable_assignment. |
2899 | ||
b8698a0f L |
2900 | Check if STMT performs an assignment (copy) that can be vectorized. |
2901 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
2902 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2903 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2904 | ||
2905 | static bool | |
2906 | vectorizable_assignment (gimple stmt, gimple_stmt_iterator *gsi, | |
2907 | gimple *vec_stmt, slp_tree slp_node) | |
2908 | { | |
2909 | tree vec_dest; | |
2910 | tree scalar_dest; | |
2911 | tree op; | |
2912 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2913 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2914 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2915 | tree new_temp; | |
2916 | tree def; | |
2917 | gimple def_stmt; | |
2918 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
fde9c428 | 2919 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
ebfd146a | 2920 | int ncopies; |
f18b55bd | 2921 | int i, j; |
6e1aa848 | 2922 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 2923 | tree vop; |
a70d6342 | 2924 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
f18b55bd IR |
2925 | gimple new_stmt = NULL; |
2926 | stmt_vec_info prev_stmt_info = NULL; | |
fde9c428 RG |
2927 | enum tree_code code; |
2928 | tree vectype_in; | |
ebfd146a IR |
2929 | |
2930 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2931 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2932 | case of SLP. */ | |
437f4a00 | 2933 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
2934 | ncopies = 1; |
2935 | else | |
2936 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
2937 | ||
2938 | gcc_assert (ncopies >= 1); | |
ebfd146a | 2939 | |
a70d6342 | 2940 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2941 | return false; |
2942 | ||
8644a673 | 2943 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2944 | return false; |
2945 | ||
2946 | /* Is vectorizable assignment? */ | |
2947 | if (!is_gimple_assign (stmt)) | |
2948 | return false; | |
2949 | ||
2950 | scalar_dest = gimple_assign_lhs (stmt); | |
2951 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
2952 | return false; | |
2953 | ||
fde9c428 | 2954 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 2955 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
2956 | || code == PAREN_EXPR |
2957 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
2958 | op = gimple_assign_rhs1 (stmt); |
2959 | else | |
2960 | return false; | |
2961 | ||
7b7ec6c5 RG |
2962 | if (code == VIEW_CONVERT_EXPR) |
2963 | op = TREE_OPERAND (op, 0); | |
2964 | ||
24ee1384 | 2965 | if (!vect_is_simple_use_1 (op, stmt, loop_vinfo, bb_vinfo, |
fde9c428 | 2966 | &def_stmt, &def, &dt[0], &vectype_in)) |
ebfd146a | 2967 | { |
73fbfcad | 2968 | if (dump_enabled_p ()) |
78c60e3d SS |
2969 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2970 | "use not simple."); | |
ebfd146a IR |
2971 | return false; |
2972 | } | |
2973 | ||
fde9c428 RG |
2974 | /* We can handle NOP_EXPR conversions that do not change the number |
2975 | of elements or the vector size. */ | |
7b7ec6c5 RG |
2976 | if ((CONVERT_EXPR_CODE_P (code) |
2977 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 RG |
2978 | && (!vectype_in |
2979 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
2980 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
2981 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
2982 | return false; | |
2983 | ||
7b7b1813 RG |
2984 | /* We do not handle bit-precision changes. */ |
2985 | if ((CONVERT_EXPR_CODE_P (code) | |
2986 | || code == VIEW_CONVERT_EXPR) | |
2987 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
2988 | && ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
2989 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
2990 | || ((TYPE_PRECISION (TREE_TYPE (op)) | |
2991 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))) | |
2992 | /* But a conversion that does not change the bit-pattern is ok. */ | |
2993 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
2994 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2995 | && TYPE_UNSIGNED (TREE_TYPE (op)))) | |
2996 | { | |
73fbfcad | 2997 | if (dump_enabled_p ()) |
78c60e3d SS |
2998 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2999 | "type conversion to/from bit-precision " | |
3000 | "unsupported."); | |
7b7b1813 RG |
3001 | return false; |
3002 | } | |
3003 | ||
ebfd146a IR |
3004 | if (!vec_stmt) /* transformation not required. */ |
3005 | { | |
3006 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 3007 | if (dump_enabled_p ()) |
78c60e3d SS |
3008 | dump_printf_loc (MSG_NOTE, vect_location, |
3009 | "=== vectorizable_assignment ==="); | |
c3e7ee41 | 3010 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
3011 | return true; |
3012 | } | |
3013 | ||
3014 | /** Transform. **/ | |
73fbfcad | 3015 | if (dump_enabled_p ()) |
78c60e3d | 3016 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment."); |
ebfd146a IR |
3017 | |
3018 | /* Handle def. */ | |
3019 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3020 | ||
3021 | /* Handle use. */ | |
f18b55bd | 3022 | for (j = 0; j < ncopies; j++) |
ebfd146a | 3023 | { |
f18b55bd IR |
3024 | /* Handle uses. */ |
3025 | if (j == 0) | |
d092494c | 3026 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node, -1); |
f18b55bd IR |
3027 | else |
3028 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
3029 | ||
3030 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 3031 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 3032 | { |
7b7ec6c5 RG |
3033 | if (CONVERT_EXPR_CODE_P (code) |
3034 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 3035 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
3036 | new_stmt = gimple_build_assign (vec_dest, vop); |
3037 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3038 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3039 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3040 | if (slp_node) | |
9771b263 | 3041 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 3042 | } |
ebfd146a IR |
3043 | |
3044 | if (slp_node) | |
f18b55bd IR |
3045 | continue; |
3046 | ||
3047 | if (j == 0) | |
3048 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3049 | else | |
3050 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3051 | ||
3052 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3053 | } | |
b8698a0f | 3054 | |
9771b263 | 3055 | vec_oprnds.release (); |
ebfd146a IR |
3056 | return true; |
3057 | } | |
3058 | ||
9dc3f7de | 3059 | |
1107f3ae IR |
3060 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
3061 | either as shift by a scalar or by a vector. */ | |
3062 | ||
3063 | bool | |
3064 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
3065 | { | |
3066 | ||
3067 | enum machine_mode vec_mode; | |
3068 | optab optab; | |
3069 | int icode; | |
3070 | tree vectype; | |
3071 | ||
3072 | vectype = get_vectype_for_scalar_type (scalar_type); | |
3073 | if (!vectype) | |
3074 | return false; | |
3075 | ||
3076 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
3077 | if (!optab | |
3078 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
3079 | { | |
3080 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
3081 | if (!optab | |
3082 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
3083 | == CODE_FOR_nothing)) | |
3084 | return false; | |
3085 | } | |
3086 | ||
3087 | vec_mode = TYPE_MODE (vectype); | |
3088 | icode = (int) optab_handler (optab, vec_mode); | |
3089 | if (icode == CODE_FOR_nothing) | |
3090 | return false; | |
3091 | ||
3092 | return true; | |
3093 | } | |
3094 | ||
3095 | ||
9dc3f7de IR |
3096 | /* Function vectorizable_shift. |
3097 | ||
3098 | Check if STMT performs a shift operation that can be vectorized. | |
3099 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3100 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3101 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3102 | ||
3103 | static bool | |
3104 | vectorizable_shift (gimple stmt, gimple_stmt_iterator *gsi, | |
3105 | gimple *vec_stmt, slp_tree slp_node) | |
3106 | { | |
3107 | tree vec_dest; | |
3108 | tree scalar_dest; | |
3109 | tree op0, op1 = NULL; | |
3110 | tree vec_oprnd1 = NULL_TREE; | |
3111 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3112 | tree vectype; | |
3113 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3114 | enum tree_code code; | |
3115 | enum machine_mode vec_mode; | |
3116 | tree new_temp; | |
3117 | optab optab; | |
3118 | int icode; | |
3119 | enum machine_mode optab_op2_mode; | |
3120 | tree def; | |
3121 | gimple def_stmt; | |
3122 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
3123 | gimple new_stmt = NULL; | |
3124 | stmt_vec_info prev_stmt_info; | |
3125 | int nunits_in; | |
3126 | int nunits_out; | |
3127 | tree vectype_out; | |
cede2577 | 3128 | tree op1_vectype; |
9dc3f7de IR |
3129 | int ncopies; |
3130 | int j, i; | |
6e1aa848 DN |
3131 | vec<tree> vec_oprnds0 = vNULL; |
3132 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
3133 | tree vop0, vop1; |
3134 | unsigned int k; | |
49eab32e | 3135 | bool scalar_shift_arg = true; |
9dc3f7de IR |
3136 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3137 | int vf; | |
3138 | ||
3139 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
3140 | return false; | |
3141 | ||
3142 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
3143 | return false; | |
3144 | ||
3145 | /* Is STMT a vectorizable binary/unary operation? */ | |
3146 | if (!is_gimple_assign (stmt)) | |
3147 | return false; | |
3148 | ||
3149 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3150 | return false; | |
3151 | ||
3152 | code = gimple_assign_rhs_code (stmt); | |
3153 | ||
3154 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
3155 | || code == RROTATE_EXPR)) | |
3156 | return false; | |
3157 | ||
3158 | scalar_dest = gimple_assign_lhs (stmt); | |
3159 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
7b7b1813 RG |
3160 | if (TYPE_PRECISION (TREE_TYPE (scalar_dest)) |
3161 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
3162 | { | |
73fbfcad | 3163 | if (dump_enabled_p ()) |
78c60e3d SS |
3164 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3165 | "bit-precision shifts not supported."); | |
7b7b1813 RG |
3166 | return false; |
3167 | } | |
9dc3f7de IR |
3168 | |
3169 | op0 = gimple_assign_rhs1 (stmt); | |
24ee1384 | 3170 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
9dc3f7de IR |
3171 | &def_stmt, &def, &dt[0], &vectype)) |
3172 | { | |
73fbfcad | 3173 | if (dump_enabled_p ()) |
78c60e3d SS |
3174 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3175 | "use not simple."); | |
9dc3f7de IR |
3176 | return false; |
3177 | } | |
3178 | /* If op0 is an external or constant def use a vector type with | |
3179 | the same size as the output vector type. */ | |
3180 | if (!vectype) | |
3181 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
3182 | if (vec_stmt) | |
3183 | gcc_assert (vectype); | |
3184 | if (!vectype) | |
3185 | { | |
73fbfcad | 3186 | if (dump_enabled_p ()) |
78c60e3d SS |
3187 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3188 | "no vectype for scalar type "); | |
9dc3f7de IR |
3189 | return false; |
3190 | } | |
3191 | ||
3192 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
3193 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
3194 | if (nunits_out != nunits_in) | |
3195 | return false; | |
3196 | ||
3197 | op1 = gimple_assign_rhs2 (stmt); | |
24ee1384 IR |
3198 | if (!vect_is_simple_use_1 (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3199 | &def, &dt[1], &op1_vectype)) | |
9dc3f7de | 3200 | { |
73fbfcad | 3201 | if (dump_enabled_p ()) |
78c60e3d SS |
3202 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3203 | "use not simple."); | |
9dc3f7de IR |
3204 | return false; |
3205 | } | |
3206 | ||
3207 | if (loop_vinfo) | |
3208 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
3209 | else | |
3210 | vf = 1; | |
3211 | ||
3212 | /* Multiple types in SLP are handled by creating the appropriate number of | |
3213 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3214 | case of SLP. */ | |
437f4a00 | 3215 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
9dc3f7de IR |
3216 | ncopies = 1; |
3217 | else | |
3218 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
3219 | ||
3220 | gcc_assert (ncopies >= 1); | |
3221 | ||
3222 | /* Determine whether the shift amount is a vector, or scalar. If the | |
3223 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
3224 | ||
49eab32e JJ |
3225 | if (dt[1] == vect_internal_def && !slp_node) |
3226 | scalar_shift_arg = false; | |
3227 | else if (dt[1] == vect_constant_def | |
3228 | || dt[1] == vect_external_def | |
3229 | || dt[1] == vect_internal_def) | |
3230 | { | |
3231 | /* In SLP, need to check whether the shift count is the same, | |
3232 | in loops if it is a constant or invariant, it is always | |
3233 | a scalar shift. */ | |
3234 | if (slp_node) | |
3235 | { | |
9771b263 | 3236 | vec<gimple> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
49eab32e JJ |
3237 | gimple slpstmt; |
3238 | ||
9771b263 | 3239 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
3240 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
3241 | scalar_shift_arg = false; | |
3242 | } | |
3243 | } | |
3244 | else | |
3245 | { | |
73fbfcad | 3246 | if (dump_enabled_p ()) |
78c60e3d SS |
3247 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3248 | "operand mode requires invariant argument."); | |
49eab32e JJ |
3249 | return false; |
3250 | } | |
3251 | ||
9dc3f7de | 3252 | /* Vector shifted by vector. */ |
49eab32e | 3253 | if (!scalar_shift_arg) |
9dc3f7de IR |
3254 | { |
3255 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 3256 | if (dump_enabled_p ()) |
78c60e3d SS |
3257 | dump_printf_loc (MSG_NOTE, vect_location, |
3258 | "vector/vector shift/rotate found."); | |
3259 | ||
aa948027 JJ |
3260 | if (!op1_vectype) |
3261 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
3262 | if (op1_vectype == NULL_TREE | |
3263 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 3264 | { |
73fbfcad | 3265 | if (dump_enabled_p ()) |
78c60e3d SS |
3266 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3267 | "unusable type for last operand in" | |
3268 | " vector/vector shift/rotate."); | |
cede2577 JJ |
3269 | return false; |
3270 | } | |
9dc3f7de IR |
3271 | } |
3272 | /* See if the machine has a vector shifted by scalar insn and if not | |
3273 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 3274 | else |
9dc3f7de IR |
3275 | { |
3276 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
3277 | if (optab | |
3278 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
3279 | { | |
73fbfcad | 3280 | if (dump_enabled_p ()) |
78c60e3d SS |
3281 | dump_printf_loc (MSG_NOTE, vect_location, |
3282 | "vector/scalar shift/rotate found."); | |
9dc3f7de IR |
3283 | } |
3284 | else | |
3285 | { | |
3286 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
3287 | if (optab | |
3288 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
3289 | != CODE_FOR_nothing)) | |
3290 | { | |
49eab32e JJ |
3291 | scalar_shift_arg = false; |
3292 | ||
73fbfcad | 3293 | if (dump_enabled_p ()) |
78c60e3d SS |
3294 | dump_printf_loc (MSG_NOTE, vect_location, |
3295 | "vector/vector shift/rotate found."); | |
9dc3f7de IR |
3296 | |
3297 | /* Unlike the other binary operators, shifts/rotates have | |
3298 | the rhs being int, instead of the same type as the lhs, | |
3299 | so make sure the scalar is the right type if we are | |
aa948027 | 3300 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
3301 | if (dt[1] == vect_constant_def) |
3302 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
3303 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
3304 | TREE_TYPE (op1))) | |
3305 | { | |
3306 | if (slp_node | |
3307 | && TYPE_MODE (TREE_TYPE (vectype)) | |
3308 | != TYPE_MODE (TREE_TYPE (op1))) | |
3309 | { | |
73fbfcad | 3310 | if (dump_enabled_p ()) |
78c60e3d SS |
3311 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3312 | "unusable type for last operand in" | |
3313 | " vector/vector shift/rotate."); | |
aa948027 JJ |
3314 | return false; |
3315 | } | |
3316 | if (vec_stmt && !slp_node) | |
3317 | { | |
3318 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
3319 | op1 = vect_init_vector (stmt, op1, | |
3320 | TREE_TYPE (vectype), NULL); | |
3321 | } | |
3322 | } | |
9dc3f7de IR |
3323 | } |
3324 | } | |
3325 | } | |
9dc3f7de IR |
3326 | |
3327 | /* Supportable by target? */ | |
3328 | if (!optab) | |
3329 | { | |
73fbfcad | 3330 | if (dump_enabled_p ()) |
78c60e3d SS |
3331 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3332 | "no optab."); | |
9dc3f7de IR |
3333 | return false; |
3334 | } | |
3335 | vec_mode = TYPE_MODE (vectype); | |
3336 | icode = (int) optab_handler (optab, vec_mode); | |
3337 | if (icode == CODE_FOR_nothing) | |
3338 | { | |
73fbfcad | 3339 | if (dump_enabled_p ()) |
78c60e3d SS |
3340 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3341 | "op not supported by target."); | |
9dc3f7de IR |
3342 | /* Check only during analysis. */ |
3343 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
3344 | || (vf < vect_min_worthwhile_factor (code) | |
3345 | && !vec_stmt)) | |
3346 | return false; | |
73fbfcad | 3347 | if (dump_enabled_p ()) |
78c60e3d | 3348 | dump_printf_loc (MSG_NOTE, vect_location, "proceeding using word mode."); |
9dc3f7de IR |
3349 | } |
3350 | ||
3351 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
3352 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
3353 | && vf < vect_min_worthwhile_factor (code) | |
3354 | && !vec_stmt) | |
3355 | { | |
73fbfcad | 3356 | if (dump_enabled_p ()) |
78c60e3d SS |
3357 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3358 | "not worthwhile without SIMD support."); | |
9dc3f7de IR |
3359 | return false; |
3360 | } | |
3361 | ||
3362 | if (!vec_stmt) /* transformation not required. */ | |
3363 | { | |
3364 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 3365 | if (dump_enabled_p ()) |
78c60e3d | 3366 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_shift ==="); |
c3e7ee41 | 3367 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
9dc3f7de IR |
3368 | return true; |
3369 | } | |
3370 | ||
3371 | /** Transform. **/ | |
3372 | ||
73fbfcad | 3373 | if (dump_enabled_p ()) |
78c60e3d SS |
3374 | dump_printf_loc (MSG_NOTE, vect_location, |
3375 | "transform binary/unary operation."); | |
9dc3f7de IR |
3376 | |
3377 | /* Handle def. */ | |
3378 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3379 | ||
9dc3f7de IR |
3380 | prev_stmt_info = NULL; |
3381 | for (j = 0; j < ncopies; j++) | |
3382 | { | |
3383 | /* Handle uses. */ | |
3384 | if (j == 0) | |
3385 | { | |
3386 | if (scalar_shift_arg) | |
3387 | { | |
3388 | /* Vector shl and shr insn patterns can be defined with scalar | |
3389 | operand 2 (shift operand). In this case, use constant or loop | |
3390 | invariant op1 directly, without extending it to vector mode | |
3391 | first. */ | |
3392 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
3393 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
3394 | { | |
73fbfcad | 3395 | if (dump_enabled_p ()) |
78c60e3d SS |
3396 | dump_printf_loc (MSG_NOTE, vect_location, |
3397 | "operand 1 using scalar mode."); | |
9dc3f7de | 3398 | vec_oprnd1 = op1; |
8930f723 | 3399 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 3400 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
3401 | if (slp_node) |
3402 | { | |
3403 | /* Store vec_oprnd1 for every vector stmt to be created | |
3404 | for SLP_NODE. We check during the analysis that all | |
3405 | the shift arguments are the same. | |
3406 | TODO: Allow different constants for different vector | |
3407 | stmts generated for an SLP instance. */ | |
3408 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 3409 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
3410 | } |
3411 | } | |
3412 | } | |
3413 | ||
3414 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
3415 | (a special case for certain kind of vector shifts); otherwise, | |
3416 | operand 1 should be of a vector type (the usual case). */ | |
3417 | if (vec_oprnd1) | |
3418 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
d092494c | 3419 | slp_node, -1); |
9dc3f7de IR |
3420 | else |
3421 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
d092494c | 3422 | slp_node, -1); |
9dc3f7de IR |
3423 | } |
3424 | else | |
3425 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
3426 | ||
3427 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3428 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 3429 | { |
9771b263 | 3430 | vop1 = vec_oprnds1[i]; |
9dc3f7de IR |
3431 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, vop0, vop1); |
3432 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3433 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3434 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3435 | if (slp_node) | |
9771b263 | 3436 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
3437 | } |
3438 | ||
3439 | if (slp_node) | |
3440 | continue; | |
3441 | ||
3442 | if (j == 0) | |
3443 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3444 | else | |
3445 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3446 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3447 | } | |
3448 | ||
9771b263 DN |
3449 | vec_oprnds0.release (); |
3450 | vec_oprnds1.release (); | |
9dc3f7de IR |
3451 | |
3452 | return true; | |
3453 | } | |
3454 | ||
3455 | ||
5deb57cb JJ |
3456 | static tree permute_vec_elements (tree, tree, tree, gimple, |
3457 | gimple_stmt_iterator *); | |
3458 | ||
3459 | ||
ebfd146a IR |
3460 | /* Function vectorizable_operation. |
3461 | ||
16949072 RG |
3462 | Check if STMT performs a binary, unary or ternary operation that can |
3463 | be vectorized. | |
b8698a0f | 3464 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
3465 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3466 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3467 | ||
3468 | static bool | |
3469 | vectorizable_operation (gimple stmt, gimple_stmt_iterator *gsi, | |
3470 | gimple *vec_stmt, slp_tree slp_node) | |
3471 | { | |
00f07b86 | 3472 | tree vec_dest; |
ebfd146a | 3473 | tree scalar_dest; |
16949072 | 3474 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 3475 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 3476 | tree vectype; |
ebfd146a IR |
3477 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
3478 | enum tree_code code; | |
3479 | enum machine_mode vec_mode; | |
3480 | tree new_temp; | |
3481 | int op_type; | |
00f07b86 | 3482 | optab optab; |
ebfd146a | 3483 | int icode; |
ebfd146a IR |
3484 | tree def; |
3485 | gimple def_stmt; | |
16949072 RG |
3486 | enum vect_def_type dt[3] |
3487 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
ebfd146a IR |
3488 | gimple new_stmt = NULL; |
3489 | stmt_vec_info prev_stmt_info; | |
b690cc0f | 3490 | int nunits_in; |
ebfd146a IR |
3491 | int nunits_out; |
3492 | tree vectype_out; | |
3493 | int ncopies; | |
3494 | int j, i; | |
6e1aa848 DN |
3495 | vec<tree> vec_oprnds0 = vNULL; |
3496 | vec<tree> vec_oprnds1 = vNULL; | |
3497 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 3498 | tree vop0, vop1, vop2; |
a70d6342 IR |
3499 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
3500 | int vf; | |
3501 | ||
a70d6342 | 3502 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3503 | return false; |
3504 | ||
8644a673 | 3505 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3506 | return false; |
3507 | ||
3508 | /* Is STMT a vectorizable binary/unary operation? */ | |
3509 | if (!is_gimple_assign (stmt)) | |
3510 | return false; | |
3511 | ||
3512 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3513 | return false; | |
3514 | ||
ebfd146a IR |
3515 | code = gimple_assign_rhs_code (stmt); |
3516 | ||
3517 | /* For pointer addition, we should use the normal plus for | |
3518 | the vector addition. */ | |
3519 | if (code == POINTER_PLUS_EXPR) | |
3520 | code = PLUS_EXPR; | |
3521 | ||
3522 | /* Support only unary or binary operations. */ | |
3523 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 3524 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 3525 | { |
73fbfcad | 3526 | if (dump_enabled_p ()) |
78c60e3d SS |
3527 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3528 | "num. args = %d (not unary/binary/ternary op).", | |
3529 | op_type); | |
ebfd146a IR |
3530 | return false; |
3531 | } | |
3532 | ||
b690cc0f RG |
3533 | scalar_dest = gimple_assign_lhs (stmt); |
3534 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
3535 | ||
7b7b1813 RG |
3536 | /* Most operations cannot handle bit-precision types without extra |
3537 | truncations. */ | |
3538 | if ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
3539 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
3540 | /* Exception are bitwise binary operations. */ | |
3541 | && code != BIT_IOR_EXPR | |
3542 | && code != BIT_XOR_EXPR | |
3543 | && code != BIT_AND_EXPR) | |
3544 | { | |
73fbfcad | 3545 | if (dump_enabled_p ()) |
78c60e3d SS |
3546 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3547 | "bit-precision arithmetic not supported."); | |
7b7b1813 RG |
3548 | return false; |
3549 | } | |
3550 | ||
ebfd146a | 3551 | op0 = gimple_assign_rhs1 (stmt); |
24ee1384 | 3552 | if (!vect_is_simple_use_1 (op0, stmt, loop_vinfo, bb_vinfo, |
b690cc0f | 3553 | &def_stmt, &def, &dt[0], &vectype)) |
ebfd146a | 3554 | { |
73fbfcad | 3555 | if (dump_enabled_p ()) |
78c60e3d SS |
3556 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3557 | "use not simple."); | |
ebfd146a IR |
3558 | return false; |
3559 | } | |
b690cc0f RG |
3560 | /* If op0 is an external or constant def use a vector type with |
3561 | the same size as the output vector type. */ | |
3562 | if (!vectype) | |
3563 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
7d8930a0 IR |
3564 | if (vec_stmt) |
3565 | gcc_assert (vectype); | |
3566 | if (!vectype) | |
3567 | { | |
73fbfcad | 3568 | if (dump_enabled_p ()) |
7d8930a0 | 3569 | { |
78c60e3d SS |
3570 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3571 | "no vectype for scalar type "); | |
3572 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
3573 | TREE_TYPE (op0)); | |
7d8930a0 IR |
3574 | } |
3575 | ||
3576 | return false; | |
3577 | } | |
b690cc0f RG |
3578 | |
3579 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
3580 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
3581 | if (nunits_out != nunits_in) | |
3582 | return false; | |
ebfd146a | 3583 | |
16949072 | 3584 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
3585 | { |
3586 | op1 = gimple_assign_rhs2 (stmt); | |
24ee1384 IR |
3587 | if (!vect_is_simple_use (op1, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3588 | &def, &dt[1])) | |
ebfd146a | 3589 | { |
73fbfcad | 3590 | if (dump_enabled_p ()) |
78c60e3d SS |
3591 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3592 | "use not simple."); | |
ebfd146a IR |
3593 | return false; |
3594 | } | |
3595 | } | |
16949072 RG |
3596 | if (op_type == ternary_op) |
3597 | { | |
3598 | op2 = gimple_assign_rhs3 (stmt); | |
24ee1384 IR |
3599 | if (!vect_is_simple_use (op2, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3600 | &def, &dt[2])) | |
16949072 | 3601 | { |
73fbfcad | 3602 | if (dump_enabled_p ()) |
78c60e3d SS |
3603 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3604 | "use not simple."); | |
16949072 RG |
3605 | return false; |
3606 | } | |
3607 | } | |
ebfd146a | 3608 | |
b690cc0f RG |
3609 | if (loop_vinfo) |
3610 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
3611 | else | |
3612 | vf = 1; | |
3613 | ||
3614 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 3615 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 3616 | case of SLP. */ |
437f4a00 | 3617 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
b690cc0f RG |
3618 | ncopies = 1; |
3619 | else | |
3620 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
3621 | ||
3622 | gcc_assert (ncopies >= 1); | |
3623 | ||
9dc3f7de | 3624 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
3625 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
3626 | || code == RROTATE_EXPR) | |
9dc3f7de | 3627 | return false; |
ebfd146a | 3628 | |
ebfd146a | 3629 | /* Supportable by target? */ |
00f07b86 RH |
3630 | |
3631 | vec_mode = TYPE_MODE (vectype); | |
3632 | if (code == MULT_HIGHPART_EXPR) | |
ebfd146a | 3633 | { |
00f07b86 | 3634 | if (can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype))) |
dee54b6e | 3635 | icode = LAST_INSN_CODE; |
00f07b86 RH |
3636 | else |
3637 | icode = CODE_FOR_nothing; | |
ebfd146a | 3638 | } |
00f07b86 RH |
3639 | else |
3640 | { | |
3641 | optab = optab_for_tree_code (code, vectype, optab_default); | |
3642 | if (!optab) | |
5deb57cb | 3643 | { |
73fbfcad | 3644 | if (dump_enabled_p ()) |
78c60e3d SS |
3645 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3646 | "no optab."); | |
00f07b86 | 3647 | return false; |
5deb57cb | 3648 | } |
00f07b86 | 3649 | icode = (int) optab_handler (optab, vec_mode); |
5deb57cb JJ |
3650 | } |
3651 | ||
ebfd146a IR |
3652 | if (icode == CODE_FOR_nothing) |
3653 | { | |
73fbfcad | 3654 | if (dump_enabled_p ()) |
78c60e3d SS |
3655 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3656 | "op not supported by target."); | |
ebfd146a IR |
3657 | /* Check only during analysis. */ |
3658 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
5deb57cb | 3659 | || (!vec_stmt && vf < vect_min_worthwhile_factor (code))) |
ebfd146a | 3660 | return false; |
73fbfcad | 3661 | if (dump_enabled_p ()) |
78c60e3d | 3662 | dump_printf_loc (MSG_NOTE, vect_location, "proceeding using word mode."); |
383d9c83 IR |
3663 | } |
3664 | ||
4a00c761 | 3665 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
3666 | if (!VECTOR_MODE_P (vec_mode) |
3667 | && !vec_stmt | |
3668 | && vf < vect_min_worthwhile_factor (code)) | |
7d8930a0 | 3669 | { |
73fbfcad | 3670 | if (dump_enabled_p ()) |
78c60e3d SS |
3671 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3672 | "not worthwhile without SIMD support."); | |
e34842c6 | 3673 | return false; |
7d8930a0 | 3674 | } |
ebfd146a | 3675 | |
ebfd146a IR |
3676 | if (!vec_stmt) /* transformation not required. */ |
3677 | { | |
4a00c761 | 3678 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 3679 | if (dump_enabled_p ()) |
78c60e3d SS |
3680 | dump_printf_loc (MSG_NOTE, vect_location, |
3681 | "=== vectorizable_operation ==="); | |
c3e7ee41 | 3682 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
3683 | return true; |
3684 | } | |
3685 | ||
3686 | /** Transform. **/ | |
3687 | ||
73fbfcad | 3688 | if (dump_enabled_p ()) |
78c60e3d SS |
3689 | dump_printf_loc (MSG_NOTE, vect_location, |
3690 | "transform binary/unary operation."); | |
383d9c83 | 3691 | |
ebfd146a | 3692 | /* Handle def. */ |
00f07b86 | 3693 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 3694 | |
ebfd146a IR |
3695 | /* In case the vectorization factor (VF) is bigger than the number |
3696 | of elements that we can fit in a vectype (nunits), we have to generate | |
3697 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
3698 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
3699 | from one copy of the vector stmt to the next, in the field | |
3700 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
3701 | stages to find the correct vector defs to be used when vectorizing | |
3702 | stmts that use the defs of the current stmt. The example below | |
3703 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
3704 | we need to create 4 vectorized stmts): | |
3705 | ||
3706 | before vectorization: | |
3707 | RELATED_STMT VEC_STMT | |
3708 | S1: x = memref - - | |
3709 | S2: z = x + 1 - - | |
3710 | ||
3711 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
3712 | there): | |
3713 | RELATED_STMT VEC_STMT | |
3714 | VS1_0: vx0 = memref0 VS1_1 - | |
3715 | VS1_1: vx1 = memref1 VS1_2 - | |
3716 | VS1_2: vx2 = memref2 VS1_3 - | |
3717 | VS1_3: vx3 = memref3 - - | |
3718 | S1: x = load - VS1_0 | |
3719 | S2: z = x + 1 - - | |
3720 | ||
3721 | step2: vectorize stmt S2 (done here): | |
3722 | To vectorize stmt S2 we first need to find the relevant vector | |
3723 | def for the first operand 'x'. This is, as usual, obtained from | |
3724 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
3725 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
3726 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
3727 | the vector stmt VS2_0, and as usual, record it in the | |
3728 | STMT_VINFO_VEC_STMT of stmt S2. | |
3729 | When creating the second copy (VS2_1), we obtain the relevant vector | |
3730 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
3731 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
3732 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
3733 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
3734 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
3735 | chain of stmts and pointers: | |
3736 | RELATED_STMT VEC_STMT | |
3737 | VS1_0: vx0 = memref0 VS1_1 - | |
3738 | VS1_1: vx1 = memref1 VS1_2 - | |
3739 | VS1_2: vx2 = memref2 VS1_3 - | |
3740 | VS1_3: vx3 = memref3 - - | |
3741 | S1: x = load - VS1_0 | |
3742 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
3743 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
3744 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
3745 | VS2_3: vz3 = vx3 + v1 - - | |
3746 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
3747 | |
3748 | prev_stmt_info = NULL; | |
3749 | for (j = 0; j < ncopies; j++) | |
3750 | { | |
3751 | /* Handle uses. */ | |
3752 | if (j == 0) | |
4a00c761 JJ |
3753 | { |
3754 | if (op_type == binary_op || op_type == ternary_op) | |
3755 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
3756 | slp_node, -1); | |
3757 | else | |
3758 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
3759 | slp_node, -1); | |
3760 | if (op_type == ternary_op) | |
36ba4aae | 3761 | { |
9771b263 DN |
3762 | vec_oprnds2.create (1); |
3763 | vec_oprnds2.quick_push (vect_get_vec_def_for_operand (op2, | |
3764 | stmt, | |
3765 | NULL)); | |
36ba4aae | 3766 | } |
4a00c761 | 3767 | } |
ebfd146a | 3768 | else |
4a00c761 JJ |
3769 | { |
3770 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
3771 | if (op_type == ternary_op) | |
3772 | { | |
9771b263 DN |
3773 | tree vec_oprnd = vec_oprnds2.pop (); |
3774 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
3775 | vec_oprnd)); | |
4a00c761 JJ |
3776 | } |
3777 | } | |
3778 | ||
3779 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3780 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 3781 | { |
4a00c761 | 3782 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 3783 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 3784 | vop2 = ((op_type == ternary_op) |
9771b263 | 3785 | ? vec_oprnds2[i] : NULL_TREE); |
73804b12 RG |
3786 | new_stmt = gimple_build_assign_with_ops (code, vec_dest, |
3787 | vop0, vop1, vop2); | |
4a00c761 JJ |
3788 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3789 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3790 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3791 | if (slp_node) | |
9771b263 | 3792 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
3793 | } |
3794 | ||
4a00c761 JJ |
3795 | if (slp_node) |
3796 | continue; | |
3797 | ||
3798 | if (j == 0) | |
3799 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3800 | else | |
3801 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3802 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
3803 | } |
3804 | ||
9771b263 DN |
3805 | vec_oprnds0.release (); |
3806 | vec_oprnds1.release (); | |
3807 | vec_oprnds2.release (); | |
ebfd146a | 3808 | |
ebfd146a IR |
3809 | return true; |
3810 | } | |
3811 | ||
c716e67f XDL |
3812 | /* A helper function to ensure data reference DR's base alignment |
3813 | for STMT_INFO. */ | |
3814 | ||
3815 | static void | |
3816 | ensure_base_align (stmt_vec_info stmt_info, struct data_reference *dr) | |
3817 | { | |
3818 | if (!dr->aux) | |
3819 | return; | |
3820 | ||
3821 | if (((dataref_aux *)dr->aux)->base_misaligned) | |
3822 | { | |
3823 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3824 | tree base_decl = ((dataref_aux *)dr->aux)->base_decl; | |
3825 | ||
3826 | DECL_ALIGN (base_decl) = TYPE_ALIGN (vectype); | |
3827 | DECL_USER_ALIGN (base_decl) = 1; | |
3828 | ((dataref_aux *)dr->aux)->base_misaligned = false; | |
3829 | } | |
3830 | } | |
3831 | ||
ebfd146a IR |
3832 | |
3833 | /* Function vectorizable_store. | |
3834 | ||
b8698a0f L |
3835 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
3836 | can be vectorized. | |
3837 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
3838 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3839 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3840 | ||
3841 | static bool | |
3842 | vectorizable_store (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
c716e67f | 3843 | slp_tree slp_node) |
ebfd146a IR |
3844 | { |
3845 | tree scalar_dest; | |
3846 | tree data_ref; | |
3847 | tree op; | |
3848 | tree vec_oprnd = NULL_TREE; | |
3849 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3850 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
3851 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
272c6793 | 3852 | tree elem_type; |
ebfd146a | 3853 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 3854 | struct loop *loop = NULL; |
ebfd146a IR |
3855 | enum machine_mode vec_mode; |
3856 | tree dummy; | |
3857 | enum dr_alignment_support alignment_support_scheme; | |
3858 | tree def; | |
3859 | gimple def_stmt; | |
3860 | enum vect_def_type dt; | |
3861 | stmt_vec_info prev_stmt_info = NULL; | |
3862 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 3863 | tree dataref_offset = NULL_TREE; |
fef4d2b3 | 3864 | gimple ptr_incr = NULL; |
ebfd146a IR |
3865 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
3866 | int ncopies; | |
3867 | int j; | |
3868 | gimple next_stmt, first_stmt = NULL; | |
0d0293ac | 3869 | bool grouped_store = false; |
272c6793 | 3870 | bool store_lanes_p = false; |
ebfd146a | 3871 | unsigned int group_size, i; |
6e1aa848 DN |
3872 | vec<tree> dr_chain = vNULL; |
3873 | vec<tree> oprnds = vNULL; | |
3874 | vec<tree> result_chain = vNULL; | |
ebfd146a | 3875 | bool inv_p; |
6e1aa848 | 3876 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 3877 | bool slp = (slp_node != NULL); |
ebfd146a | 3878 | unsigned int vec_num; |
a70d6342 | 3879 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
272c6793 | 3880 | tree aggr_type; |
a70d6342 IR |
3881 | |
3882 | if (loop_vinfo) | |
3883 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
3884 | |
3885 | /* Multiple types in SLP are handled by creating the appropriate number of | |
3886 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3887 | case of SLP. */ | |
437f4a00 | 3888 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
3889 | ncopies = 1; |
3890 | else | |
3891 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3892 | ||
3893 | gcc_assert (ncopies >= 1); | |
3894 | ||
3895 | /* FORNOW. This restriction should be relaxed. */ | |
a70d6342 | 3896 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) |
ebfd146a | 3897 | { |
73fbfcad | 3898 | if (dump_enabled_p ()) |
78c60e3d SS |
3899 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3900 | "multiple types in nested loop."); | |
ebfd146a IR |
3901 | return false; |
3902 | } | |
3903 | ||
a70d6342 | 3904 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3905 | return false; |
3906 | ||
8644a673 | 3907 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3908 | return false; |
3909 | ||
3910 | /* Is vectorizable store? */ | |
3911 | ||
3912 | if (!is_gimple_assign (stmt)) | |
3913 | return false; | |
3914 | ||
3915 | scalar_dest = gimple_assign_lhs (stmt); | |
ab0ef706 JJ |
3916 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
3917 | && is_pattern_stmt_p (stmt_info)) | |
3918 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
ebfd146a | 3919 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
38000232 | 3920 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF |
ebfd146a | 3921 | && TREE_CODE (scalar_dest) != INDIRECT_REF |
e9dbe7bb IR |
3922 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
3923 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
3924 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
3925 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
3926 | return false; |
3927 | ||
3928 | gcc_assert (gimple_assign_single_p (stmt)); | |
3929 | op = gimple_assign_rhs1 (stmt); | |
24ee1384 IR |
3930 | if (!vect_is_simple_use (op, stmt, loop_vinfo, bb_vinfo, &def_stmt, |
3931 | &def, &dt)) | |
ebfd146a | 3932 | { |
73fbfcad | 3933 | if (dump_enabled_p ()) |
78c60e3d SS |
3934 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3935 | "use not simple."); | |
ebfd146a IR |
3936 | return false; |
3937 | } | |
3938 | ||
272c6793 | 3939 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 3940 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 3941 | |
ebfd146a IR |
3942 | /* FORNOW. In some cases can vectorize even if data-type not supported |
3943 | (e.g. - array initialization with 0). */ | |
947131ba | 3944 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
3945 | return false; |
3946 | ||
3947 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
3948 | return false; | |
3949 | ||
a7ce6ec3 RG |
3950 | if (tree_int_cst_compare (loop && nested_in_vect_loop_p (loop, stmt) |
3951 | ? STMT_VINFO_DR_STEP (stmt_info) : DR_STEP (dr), | |
3952 | size_zero_node) < 0) | |
a1e53f3f | 3953 | { |
73fbfcad | 3954 | if (dump_enabled_p ()) |
78c60e3d SS |
3955 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3956 | "negative step for store."); | |
a1e53f3f L |
3957 | return false; |
3958 | } | |
3959 | ||
0d0293ac | 3960 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 3961 | { |
0d0293ac | 3962 | grouped_store = true; |
e14c1050 | 3963 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
b602d918 RS |
3964 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
3965 | { | |
e14c1050 | 3966 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
3967 | if (vect_store_lanes_supported (vectype, group_size)) |
3968 | store_lanes_p = true; | |
0d0293ac | 3969 | else if (!vect_grouped_store_supported (vectype, group_size)) |
b602d918 RS |
3970 | return false; |
3971 | } | |
b8698a0f | 3972 | |
ebfd146a IR |
3973 | if (first_stmt == stmt) |
3974 | { | |
3975 | /* STMT is the leader of the group. Check the operands of all the | |
3976 | stmts of the group. */ | |
e14c1050 | 3977 | next_stmt = GROUP_NEXT_ELEMENT (stmt_info); |
ebfd146a IR |
3978 | while (next_stmt) |
3979 | { | |
3980 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
3981 | op = gimple_assign_rhs1 (next_stmt); | |
24ee1384 IR |
3982 | if (!vect_is_simple_use (op, next_stmt, loop_vinfo, bb_vinfo, |
3983 | &def_stmt, &def, &dt)) | |
ebfd146a | 3984 | { |
73fbfcad | 3985 | if (dump_enabled_p ()) |
78c60e3d SS |
3986 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3987 | "use not simple."); | |
ebfd146a IR |
3988 | return false; |
3989 | } | |
e14c1050 | 3990 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
3991 | } |
3992 | } | |
3993 | } | |
3994 | ||
3995 | if (!vec_stmt) /* transformation not required. */ | |
3996 | { | |
3997 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
92345349 BS |
3998 | vect_model_store_cost (stmt_info, ncopies, store_lanes_p, dt, |
3999 | NULL, NULL, NULL); | |
ebfd146a IR |
4000 | return true; |
4001 | } | |
4002 | ||
4003 | /** Transform. **/ | |
4004 | ||
c716e67f XDL |
4005 | ensure_base_align (stmt_info, dr); |
4006 | ||
0d0293ac | 4007 | if (grouped_store) |
ebfd146a IR |
4008 | { |
4009 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 4010 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 4011 | |
e14c1050 | 4012 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
4013 | |
4014 | /* FORNOW */ | |
a70d6342 | 4015 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
4016 | |
4017 | /* We vectorize all the stmts of the interleaving group when we | |
4018 | reach the last stmt in the group. */ | |
e14c1050 IR |
4019 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
4020 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
4021 | && !slp) |
4022 | { | |
4023 | *vec_stmt = NULL; | |
4024 | return true; | |
4025 | } | |
4026 | ||
4027 | if (slp) | |
4b5caab7 | 4028 | { |
0d0293ac | 4029 | grouped_store = false; |
4b5caab7 IR |
4030 | /* VEC_NUM is the number of vect stmts to be created for this |
4031 | group. */ | |
4032 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 4033 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
4b5caab7 | 4034 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
d092494c | 4035 | op = gimple_assign_rhs1 (first_stmt); |
4b5caab7 | 4036 | } |
ebfd146a | 4037 | else |
4b5caab7 IR |
4038 | /* VEC_NUM is the number of vect stmts to be created for this |
4039 | group. */ | |
ebfd146a IR |
4040 | vec_num = group_size; |
4041 | } | |
b8698a0f | 4042 | else |
ebfd146a IR |
4043 | { |
4044 | first_stmt = stmt; | |
4045 | first_dr = dr; | |
4046 | group_size = vec_num = 1; | |
ebfd146a | 4047 | } |
b8698a0f | 4048 | |
73fbfcad | 4049 | if (dump_enabled_p ()) |
78c60e3d SS |
4050 | dump_printf_loc (MSG_NOTE, vect_location, |
4051 | "transform store. ncopies = %d", ncopies); | |
ebfd146a | 4052 | |
9771b263 DN |
4053 | dr_chain.create (group_size); |
4054 | oprnds.create (group_size); | |
ebfd146a | 4055 | |
720f5239 | 4056 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 4057 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
4058 | /* Targets with store-lane instructions must not require explicit |
4059 | realignment. */ | |
4060 | gcc_assert (!store_lanes_p | |
4061 | || alignment_support_scheme == dr_aligned | |
4062 | || alignment_support_scheme == dr_unaligned_supported); | |
4063 | ||
4064 | if (store_lanes_p) | |
4065 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
4066 | else | |
4067 | aggr_type = vectype; | |
ebfd146a IR |
4068 | |
4069 | /* In case the vectorization factor (VF) is bigger than the number | |
4070 | of elements that we can fit in a vectype (nunits), we have to generate | |
4071 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 4072 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
4073 | vect_get_vec_def_for_copy_stmt. */ |
4074 | ||
0d0293ac | 4075 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
4076 | |
4077 | S1: &base + 2 = x2 | |
4078 | S2: &base = x0 | |
4079 | S3: &base + 1 = x1 | |
4080 | S4: &base + 3 = x3 | |
4081 | ||
4082 | We create vectorized stores starting from base address (the access of the | |
4083 | first stmt in the chain (S2 in the above example), when the last store stmt | |
4084 | of the chain (S4) is reached: | |
4085 | ||
4086 | VS1: &base = vx2 | |
4087 | VS2: &base + vec_size*1 = vx0 | |
4088 | VS3: &base + vec_size*2 = vx1 | |
4089 | VS4: &base + vec_size*3 = vx3 | |
4090 | ||
4091 | Then permutation statements are generated: | |
4092 | ||
3fcc1b55 JJ |
4093 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
4094 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 4095 | ... |
b8698a0f | 4096 | |
ebfd146a IR |
4097 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
4098 | (the order of the data-refs in the output of vect_permute_store_chain | |
4099 | corresponds to the order of scalar stmts in the interleaving chain - see | |
4100 | the documentation of vect_permute_store_chain()). | |
4101 | ||
4102 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 4103 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 4104 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 4105 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
4106 | */ |
4107 | ||
4108 | prev_stmt_info = NULL; | |
4109 | for (j = 0; j < ncopies; j++) | |
4110 | { | |
4111 | gimple new_stmt; | |
ebfd146a IR |
4112 | |
4113 | if (j == 0) | |
4114 | { | |
4115 | if (slp) | |
4116 | { | |
4117 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c IR |
4118 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
4119 | NULL, slp_node, -1); | |
ebfd146a | 4120 | |
9771b263 | 4121 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
4122 | } |
4123 | else | |
4124 | { | |
b8698a0f L |
4125 | /* For interleaved stores we collect vectorized defs for all the |
4126 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
4127 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
4128 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
4129 | ||
0d0293ac | 4130 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 4131 | OPRNDS are of size 1. */ |
b8698a0f | 4132 | next_stmt = first_stmt; |
ebfd146a IR |
4133 | for (i = 0; i < group_size; i++) |
4134 | { | |
b8698a0f L |
4135 | /* Since gaps are not supported for interleaved stores, |
4136 | GROUP_SIZE is the exact number of stmts in the chain. | |
4137 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
4138 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
4139 | iteration of the loop will be executed. */ |
4140 | gcc_assert (next_stmt | |
4141 | && gimple_assign_single_p (next_stmt)); | |
4142 | op = gimple_assign_rhs1 (next_stmt); | |
4143 | ||
b8698a0f | 4144 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt, |
ebfd146a | 4145 | NULL); |
9771b263 DN |
4146 | dr_chain.quick_push (vec_oprnd); |
4147 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 4148 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
4149 | } |
4150 | } | |
4151 | ||
4152 | /* We should have catched mismatched types earlier. */ | |
4153 | gcc_assert (useless_type_conversion_p (vectype, | |
4154 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
4155 | bool simd_lane_access_p |
4156 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
4157 | if (simd_lane_access_p | |
4158 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
4159 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
4160 | && integer_zerop (DR_OFFSET (first_dr)) | |
4161 | && integer_zerop (DR_INIT (first_dr)) | |
4162 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
4163 | get_alias_set (DR_REF (first_dr)))) | |
4164 | { | |
4165 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
4166 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
4167 | (DR_REF (first_dr)), 0); | |
4168 | } | |
4169 | else | |
4170 | dataref_ptr | |
4171 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
4172 | simd_lane_access_p ? loop : NULL, | |
4173 | NULL_TREE, &dummy, gsi, &ptr_incr, | |
4174 | simd_lane_access_p, &inv_p); | |
a70d6342 | 4175 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 4176 | } |
b8698a0f | 4177 | else |
ebfd146a | 4178 | { |
b8698a0f L |
4179 | /* For interleaved stores we created vectorized defs for all the |
4180 | defs stored in OPRNDS in the previous iteration (previous copy). | |
4181 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
4182 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
4183 | next copy. | |
0d0293ac | 4184 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
4185 | OPRNDS are of size 1. */ |
4186 | for (i = 0; i < group_size; i++) | |
4187 | { | |
9771b263 | 4188 | op = oprnds[i]; |
24ee1384 IR |
4189 | vect_is_simple_use (op, NULL, loop_vinfo, bb_vinfo, &def_stmt, |
4190 | &def, &dt); | |
b8698a0f | 4191 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
9771b263 DN |
4192 | dr_chain[i] = vec_oprnd; |
4193 | oprnds[i] = vec_oprnd; | |
ebfd146a | 4194 | } |
74bf76ed JJ |
4195 | if (dataref_offset) |
4196 | dataref_offset | |
4197 | = int_const_binop (PLUS_EXPR, dataref_offset, | |
4198 | TYPE_SIZE_UNIT (aggr_type)); | |
4199 | else | |
4200 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
4201 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
4202 | } |
4203 | ||
272c6793 | 4204 | if (store_lanes_p) |
ebfd146a | 4205 | { |
272c6793 | 4206 | tree vec_array; |
267d3070 | 4207 | |
272c6793 RS |
4208 | /* Combine all the vectors into an array. */ |
4209 | vec_array = create_vector_array (vectype, vec_num); | |
4210 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 4211 | { |
9771b263 | 4212 | vec_oprnd = dr_chain[i]; |
272c6793 | 4213 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 4214 | } |
b8698a0f | 4215 | |
272c6793 RS |
4216 | /* Emit: |
4217 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
4218 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
4219 | new_stmt = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array); | |
4220 | gimple_call_set_lhs (new_stmt, data_ref); | |
267d3070 | 4221 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
4222 | } |
4223 | else | |
4224 | { | |
4225 | new_stmt = NULL; | |
0d0293ac | 4226 | if (grouped_store) |
272c6793 | 4227 | { |
b6b9227d JJ |
4228 | if (j == 0) |
4229 | result_chain.create (group_size); | |
272c6793 RS |
4230 | /* Permute. */ |
4231 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
4232 | &result_chain); | |
4233 | } | |
c2d7ab2a | 4234 | |
272c6793 RS |
4235 | next_stmt = first_stmt; |
4236 | for (i = 0; i < vec_num; i++) | |
4237 | { | |
644ffefd | 4238 | unsigned align, misalign; |
272c6793 RS |
4239 | |
4240 | if (i > 0) | |
4241 | /* Bump the vector pointer. */ | |
4242 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
4243 | stmt, NULL_TREE); | |
4244 | ||
4245 | if (slp) | |
9771b263 | 4246 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
4247 | else if (grouped_store) |
4248 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 4249 | vect_permute_store_chain(). */ |
9771b263 | 4250 | vec_oprnd = result_chain[i]; |
272c6793 RS |
4251 | |
4252 | data_ref = build2 (MEM_REF, TREE_TYPE (vec_oprnd), dataref_ptr, | |
74bf76ed JJ |
4253 | dataref_offset |
4254 | ? dataref_offset | |
4255 | : build_int_cst (reference_alias_ptr_type | |
4256 | (DR_REF (first_dr)), 0)); | |
644ffefd | 4257 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 | 4258 | if (aligned_access_p (first_dr)) |
644ffefd | 4259 | misalign = 0; |
272c6793 RS |
4260 | else if (DR_MISALIGNMENT (first_dr) == -1) |
4261 | { | |
4262 | TREE_TYPE (data_ref) | |
4263 | = build_aligned_type (TREE_TYPE (data_ref), | |
4264 | TYPE_ALIGN (elem_type)); | |
644ffefd MJ |
4265 | align = TYPE_ALIGN_UNIT (elem_type); |
4266 | misalign = 0; | |
272c6793 RS |
4267 | } |
4268 | else | |
4269 | { | |
4270 | TREE_TYPE (data_ref) | |
4271 | = build_aligned_type (TREE_TYPE (data_ref), | |
4272 | TYPE_ALIGN (elem_type)); | |
644ffefd | 4273 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 4274 | } |
74bf76ed JJ |
4275 | if (dataref_offset == NULL_TREE) |
4276 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
4277 | misalign); | |
c2d7ab2a | 4278 | |
272c6793 RS |
4279 | /* Arguments are ready. Create the new vector stmt. */ |
4280 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
4281 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
4282 | |
4283 | if (slp) | |
4284 | continue; | |
4285 | ||
e14c1050 | 4286 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
4287 | if (!next_stmt) |
4288 | break; | |
4289 | } | |
ebfd146a | 4290 | } |
1da0876c RS |
4291 | if (!slp) |
4292 | { | |
4293 | if (j == 0) | |
4294 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4295 | else | |
4296 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4297 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4298 | } | |
ebfd146a IR |
4299 | } |
4300 | ||
9771b263 DN |
4301 | dr_chain.release (); |
4302 | oprnds.release (); | |
4303 | result_chain.release (); | |
4304 | vec_oprnds.release (); | |
ebfd146a IR |
4305 | |
4306 | return true; | |
4307 | } | |
4308 | ||
aec7ae7d JJ |
4309 | /* Given a vector type VECTYPE and permutation SEL returns |
4310 | the VECTOR_CST mask that implements the permutation of the | |
4311 | vector elements. If that is impossible to do, returns NULL. */ | |
a1e53f3f | 4312 | |
3fcc1b55 JJ |
4313 | tree |
4314 | vect_gen_perm_mask (tree vectype, unsigned char *sel) | |
a1e53f3f | 4315 | { |
d2a12ae7 | 4316 | tree mask_elt_type, mask_type, mask_vec, *mask_elts; |
2635892a | 4317 | int i, nunits; |
a1e53f3f | 4318 | |
22e4dee7 | 4319 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
22e4dee7 RH |
4320 | |
4321 | if (!can_vec_perm_p (TYPE_MODE (vectype), false, sel)) | |
a1e53f3f L |
4322 | return NULL; |
4323 | ||
96f9265a RG |
4324 | mask_elt_type = lang_hooks.types.type_for_mode |
4325 | (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype))), 1); | |
22e4dee7 | 4326 | mask_type = get_vectype_for_scalar_type (mask_elt_type); |
a1e53f3f | 4327 | |
d2a12ae7 | 4328 | mask_elts = XALLOCAVEC (tree, nunits); |
aec7ae7d | 4329 | for (i = nunits - 1; i >= 0; i--) |
d2a12ae7 RG |
4330 | mask_elts[i] = build_int_cst (mask_elt_type, sel[i]); |
4331 | mask_vec = build_vector (mask_type, mask_elts); | |
a1e53f3f | 4332 | |
2635892a | 4333 | return mask_vec; |
a1e53f3f L |
4334 | } |
4335 | ||
aec7ae7d JJ |
4336 | /* Given a vector type VECTYPE returns the VECTOR_CST mask that implements |
4337 | reversal of the vector elements. If that is impossible to do, | |
4338 | returns NULL. */ | |
4339 | ||
4340 | static tree | |
4341 | perm_mask_for_reverse (tree vectype) | |
4342 | { | |
4343 | int i, nunits; | |
4344 | unsigned char *sel; | |
4345 | ||
4346 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4347 | sel = XALLOCAVEC (unsigned char, nunits); | |
4348 | ||
4349 | for (i = 0; i < nunits; ++i) | |
4350 | sel[i] = nunits - 1 - i; | |
4351 | ||
3fcc1b55 | 4352 | return vect_gen_perm_mask (vectype, sel); |
aec7ae7d JJ |
4353 | } |
4354 | ||
4355 | /* Given a vector variable X and Y, that was generated for the scalar | |
4356 | STMT, generate instructions to permute the vector elements of X and Y | |
4357 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
4358 | permuted vector variable. */ | |
a1e53f3f L |
4359 | |
4360 | static tree | |
aec7ae7d JJ |
4361 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple stmt, |
4362 | gimple_stmt_iterator *gsi) | |
a1e53f3f L |
4363 | { |
4364 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 4365 | tree perm_dest, data_ref; |
a1e53f3f L |
4366 | gimple perm_stmt; |
4367 | ||
a1e53f3f | 4368 | perm_dest = vect_create_destination_var (gimple_assign_lhs (stmt), vectype); |
aec7ae7d | 4369 | data_ref = make_ssa_name (perm_dest, NULL); |
a1e53f3f L |
4370 | |
4371 | /* Generate the permute statement. */ | |
73804b12 RG |
4372 | perm_stmt = gimple_build_assign_with_ops (VEC_PERM_EXPR, data_ref, |
4373 | x, y, mask_vec); | |
a1e53f3f L |
4374 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
4375 | ||
4376 | return data_ref; | |
4377 | } | |
4378 | ||
ebfd146a IR |
4379 | /* vectorizable_load. |
4380 | ||
b8698a0f L |
4381 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
4382 | can be vectorized. | |
4383 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4384 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4385 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4386 | ||
4387 | static bool | |
4388 | vectorizable_load (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt, | |
c716e67f | 4389 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a IR |
4390 | { |
4391 | tree scalar_dest; | |
4392 | tree vec_dest = NULL; | |
4393 | tree data_ref = NULL; | |
4394 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 4395 | stmt_vec_info prev_stmt_info; |
ebfd146a | 4396 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 4397 | struct loop *loop = NULL; |
ebfd146a | 4398 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 4399 | bool nested_in_vect_loop = false; |
c716e67f | 4400 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
ebfd146a | 4401 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
272c6793 | 4402 | tree elem_type; |
ebfd146a | 4403 | tree new_temp; |
947131ba | 4404 | enum machine_mode mode; |
ebfd146a IR |
4405 | gimple new_stmt = NULL; |
4406 | tree dummy; | |
4407 | enum dr_alignment_support alignment_support_scheme; | |
4408 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 4409 | tree dataref_offset = NULL_TREE; |
fef4d2b3 | 4410 | gimple ptr_incr = NULL; |
ebfd146a IR |
4411 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
4412 | int ncopies; | |
a64b9c26 | 4413 | int i, j, group_size, group_gap; |
ebfd146a IR |
4414 | tree msq = NULL_TREE, lsq; |
4415 | tree offset = NULL_TREE; | |
4416 | tree realignment_token = NULL_TREE; | |
4417 | gimple phi = NULL; | |
6e1aa848 | 4418 | vec<tree> dr_chain = vNULL; |
0d0293ac | 4419 | bool grouped_load = false; |
272c6793 | 4420 | bool load_lanes_p = false; |
ebfd146a | 4421 | gimple first_stmt; |
ebfd146a | 4422 | bool inv_p; |
319e6439 | 4423 | bool negative = false; |
ebfd146a IR |
4424 | bool compute_in_loop = false; |
4425 | struct loop *at_loop; | |
4426 | int vec_num; | |
4427 | bool slp = (slp_node != NULL); | |
4428 | bool slp_perm = false; | |
4429 | enum tree_code code; | |
a70d6342 IR |
4430 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
4431 | int vf; | |
272c6793 | 4432 | tree aggr_type; |
aec7ae7d JJ |
4433 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; |
4434 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
4435 | int gather_scale = 1; | |
4436 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
a70d6342 IR |
4437 | |
4438 | if (loop_vinfo) | |
4439 | { | |
4440 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
4441 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
4442 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4443 | } | |
4444 | else | |
3533e503 | 4445 | vf = 1; |
ebfd146a IR |
4446 | |
4447 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 4448 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 4449 | case of SLP. */ |
437f4a00 | 4450 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
4451 | ncopies = 1; |
4452 | else | |
4453 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4454 | ||
4455 | gcc_assert (ncopies >= 1); | |
4456 | ||
4457 | /* FORNOW. This restriction should be relaxed. */ | |
4458 | if (nested_in_vect_loop && ncopies > 1) | |
4459 | { | |
73fbfcad | 4460 | if (dump_enabled_p ()) |
78c60e3d SS |
4461 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4462 | "multiple types in nested loop."); | |
ebfd146a IR |
4463 | return false; |
4464 | } | |
4465 | ||
a70d6342 | 4466 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4467 | return false; |
4468 | ||
8644a673 | 4469 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
4470 | return false; |
4471 | ||
4472 | /* Is vectorizable load? */ | |
4473 | if (!is_gimple_assign (stmt)) | |
4474 | return false; | |
4475 | ||
4476 | scalar_dest = gimple_assign_lhs (stmt); | |
4477 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4478 | return false; | |
4479 | ||
4480 | code = gimple_assign_rhs_code (stmt); | |
4481 | if (code != ARRAY_REF | |
38000232 | 4482 | && code != BIT_FIELD_REF |
ebfd146a | 4483 | && code != INDIRECT_REF |
e9dbe7bb IR |
4484 | && code != COMPONENT_REF |
4485 | && code != IMAGPART_EXPR | |
70f34814 | 4486 | && code != REALPART_EXPR |
42373e0b RG |
4487 | && code != MEM_REF |
4488 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
ebfd146a IR |
4489 | return false; |
4490 | ||
4491 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
4492 | return false; | |
4493 | ||
7b7b1813 | 4494 | elem_type = TREE_TYPE (vectype); |
947131ba | 4495 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
4496 | |
4497 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
4498 | (e.g. - data copies). */ | |
947131ba | 4499 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 4500 | { |
73fbfcad | 4501 | if (dump_enabled_p ()) |
78c60e3d SS |
4502 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4503 | "Aligned load, but unsupported type."); | |
ebfd146a IR |
4504 | return false; |
4505 | } | |
4506 | ||
ebfd146a | 4507 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 4508 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 4509 | { |
0d0293ac | 4510 | grouped_load = true; |
ebfd146a | 4511 | /* FORNOW */ |
aec7ae7d | 4512 | gcc_assert (! nested_in_vect_loop && !STMT_VINFO_GATHER_P (stmt_info)); |
ebfd146a | 4513 | |
e14c1050 | 4514 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
b602d918 RS |
4515 | if (!slp && !PURE_SLP_STMT (stmt_info)) |
4516 | { | |
e14c1050 | 4517 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
272c6793 RS |
4518 | if (vect_load_lanes_supported (vectype, group_size)) |
4519 | load_lanes_p = true; | |
0d0293ac | 4520 | else if (!vect_grouped_load_supported (vectype, group_size)) |
b602d918 RS |
4521 | return false; |
4522 | } | |
ebfd146a IR |
4523 | } |
4524 | ||
a1e53f3f | 4525 | |
aec7ae7d JJ |
4526 | if (STMT_VINFO_GATHER_P (stmt_info)) |
4527 | { | |
4528 | gimple def_stmt; | |
4529 | tree def; | |
4530 | gather_decl = vect_check_gather (stmt, loop_vinfo, &gather_base, | |
4531 | &gather_off, &gather_scale); | |
4532 | gcc_assert (gather_decl); | |
24ee1384 | 4533 | if (!vect_is_simple_use_1 (gather_off, NULL, loop_vinfo, bb_vinfo, |
aec7ae7d JJ |
4534 | &def_stmt, &def, &gather_dt, |
4535 | &gather_off_vectype)) | |
4536 | { | |
73fbfcad | 4537 | if (dump_enabled_p ()) |
78c60e3d SS |
4538 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4539 | "gather index use not simple."); | |
aec7ae7d JJ |
4540 | return false; |
4541 | } | |
4542 | } | |
7d75abc8 | 4543 | else if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
14ac6aa2 | 4544 | ; |
319e6439 RG |
4545 | else |
4546 | { | |
4547 | negative = tree_int_cst_compare (nested_in_vect_loop | |
4548 | ? STMT_VINFO_DR_STEP (stmt_info) | |
4549 | : DR_STEP (dr), | |
4550 | size_zero_node) < 0; | |
4551 | if (negative && ncopies > 1) | |
4552 | { | |
73fbfcad | 4553 | if (dump_enabled_p ()) |
78c60e3d SS |
4554 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4555 | "multiple types with negative step."); | |
319e6439 RG |
4556 | return false; |
4557 | } | |
4558 | ||
4559 | if (negative) | |
4560 | { | |
08940f33 RB |
4561 | if (grouped_load) |
4562 | { | |
4563 | if (dump_enabled_p ()) | |
4564 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
4565 | "negative step for group load not supported"); | |
4566 | return false; | |
4567 | } | |
319e6439 RG |
4568 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
4569 | if (alignment_support_scheme != dr_aligned | |
4570 | && alignment_support_scheme != dr_unaligned_supported) | |
4571 | { | |
73fbfcad | 4572 | if (dump_enabled_p ()) |
78c60e3d SS |
4573 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4574 | "negative step but alignment required."); | |
319e6439 RG |
4575 | return false; |
4576 | } | |
4577 | if (!perm_mask_for_reverse (vectype)) | |
4578 | { | |
73fbfcad | 4579 | if (dump_enabled_p ()) |
78c60e3d SS |
4580 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4581 | "negative step and reversing not supported."); | |
319e6439 RG |
4582 | return false; |
4583 | } | |
4584 | } | |
7d75abc8 | 4585 | } |
aec7ae7d | 4586 | |
ebfd146a IR |
4587 | if (!vec_stmt) /* transformation not required. */ |
4588 | { | |
4589 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
92345349 | 4590 | vect_model_load_cost (stmt_info, ncopies, load_lanes_p, NULL, NULL, NULL); |
ebfd146a IR |
4591 | return true; |
4592 | } | |
4593 | ||
73fbfcad | 4594 | if (dump_enabled_p ()) |
78c60e3d SS |
4595 | dump_printf_loc (MSG_NOTE, vect_location, |
4596 | "transform load. ncopies = %d", ncopies); | |
ebfd146a IR |
4597 | |
4598 | /** Transform. **/ | |
4599 | ||
c716e67f XDL |
4600 | ensure_base_align (stmt_info, dr); |
4601 | ||
aec7ae7d JJ |
4602 | if (STMT_VINFO_GATHER_P (stmt_info)) |
4603 | { | |
4604 | tree vec_oprnd0 = NULL_TREE, op; | |
4605 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
4606 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
4607 | tree ptr, mask, var, scale, perm_mask = NULL_TREE, prev_res = NULL_TREE; | |
4608 | edge pe = loop_preheader_edge (loop); | |
4609 | gimple_seq seq; | |
4610 | basic_block new_bb; | |
4611 | enum { NARROW, NONE, WIDEN } modifier; | |
4612 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
4613 | ||
4614 | if (nunits == gather_off_nunits) | |
4615 | modifier = NONE; | |
4616 | else if (nunits == gather_off_nunits / 2) | |
4617 | { | |
4618 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
4619 | modifier = WIDEN; | |
4620 | ||
4621 | for (i = 0; i < gather_off_nunits; ++i) | |
4622 | sel[i] = i | nunits; | |
4623 | ||
3fcc1b55 | 4624 | perm_mask = vect_gen_perm_mask (gather_off_vectype, sel); |
aec7ae7d JJ |
4625 | gcc_assert (perm_mask != NULL_TREE); |
4626 | } | |
4627 | else if (nunits == gather_off_nunits * 2) | |
4628 | { | |
4629 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
4630 | modifier = NARROW; | |
4631 | ||
4632 | for (i = 0; i < nunits; ++i) | |
4633 | sel[i] = i < gather_off_nunits | |
4634 | ? i : i + nunits - gather_off_nunits; | |
4635 | ||
3fcc1b55 | 4636 | perm_mask = vect_gen_perm_mask (vectype, sel); |
aec7ae7d JJ |
4637 | gcc_assert (perm_mask != NULL_TREE); |
4638 | ncopies *= 2; | |
4639 | } | |
4640 | else | |
4641 | gcc_unreachable (); | |
4642 | ||
4643 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); | |
4644 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4645 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4646 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4647 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
4648 | scaletype = TREE_VALUE (arglist); | |
4649 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
4650 | && types_compatible_p (srctype, masktype)); | |
4651 | ||
4652 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4653 | ||
4654 | ptr = fold_convert (ptrtype, gather_base); | |
4655 | if (!is_gimple_min_invariant (ptr)) | |
4656 | { | |
4657 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
4658 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
4659 | gcc_assert (!new_bb); | |
4660 | } | |
4661 | ||
4662 | /* Currently we support only unconditional gather loads, | |
4663 | so mask should be all ones. */ | |
4664 | if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
4665 | mask = build_int_cst (TREE_TYPE (masktype), -1); | |
4666 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) | |
4667 | { | |
4668 | REAL_VALUE_TYPE r; | |
4669 | long tmp[6]; | |
4670 | for (j = 0; j < 6; ++j) | |
4671 | tmp[j] = -1; | |
4672 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
4673 | mask = build_real (TREE_TYPE (masktype), r); | |
4674 | } | |
4675 | else | |
4676 | gcc_unreachable (); | |
4677 | mask = build_vector_from_val (masktype, mask); | |
4678 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
4679 | ||
4680 | scale = build_int_cst (scaletype, gather_scale); | |
4681 | ||
4682 | prev_stmt_info = NULL; | |
4683 | for (j = 0; j < ncopies; ++j) | |
4684 | { | |
4685 | if (modifier == WIDEN && (j & 1)) | |
4686 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
4687 | perm_mask, stmt, gsi); | |
4688 | else if (j == 0) | |
4689 | op = vec_oprnd0 | |
4690 | = vect_get_vec_def_for_operand (gather_off, stmt, NULL); | |
4691 | else | |
4692 | op = vec_oprnd0 | |
4693 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
4694 | ||
4695 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
4696 | { | |
4697 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
4698 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
4699 | var = vect_get_new_vect_var (idxtype, vect_simple_var, NULL); | |
aec7ae7d JJ |
4700 | var = make_ssa_name (var, NULL); |
4701 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); | |
4702 | new_stmt | |
4703 | = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var, | |
4704 | op, NULL_TREE); | |
4705 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4706 | op = var; | |
4707 | } | |
4708 | ||
4709 | new_stmt | |
4710 | = gimple_build_call (gather_decl, 5, mask, ptr, op, mask, scale); | |
4711 | ||
4712 | if (!useless_type_conversion_p (vectype, rettype)) | |
4713 | { | |
4714 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
4715 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
4716 | var = vect_get_new_vect_var (rettype, vect_simple_var, NULL); | |
aec7ae7d JJ |
4717 | op = make_ssa_name (var, new_stmt); |
4718 | gimple_call_set_lhs (new_stmt, op); | |
4719 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4720 | var = make_ssa_name (vec_dest, NULL); | |
4721 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); | |
4722 | new_stmt | |
4723 | = gimple_build_assign_with_ops (VIEW_CONVERT_EXPR, var, op, | |
4724 | NULL_TREE); | |
4725 | } | |
4726 | else | |
4727 | { | |
4728 | var = make_ssa_name (vec_dest, new_stmt); | |
4729 | gimple_call_set_lhs (new_stmt, var); | |
4730 | } | |
4731 | ||
4732 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4733 | ||
4734 | if (modifier == NARROW) | |
4735 | { | |
4736 | if ((j & 1) == 0) | |
4737 | { | |
4738 | prev_res = var; | |
4739 | continue; | |
4740 | } | |
4741 | var = permute_vec_elements (prev_res, var, | |
4742 | perm_mask, stmt, gsi); | |
4743 | new_stmt = SSA_NAME_DEF_STMT (var); | |
4744 | } | |
4745 | ||
4746 | if (prev_stmt_info == NULL) | |
4747 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4748 | else | |
4749 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4750 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4751 | } | |
4752 | return true; | |
4753 | } | |
7d75abc8 MM |
4754 | else if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) |
4755 | { | |
4756 | gimple_stmt_iterator incr_gsi; | |
4757 | bool insert_after; | |
4758 | gimple incr; | |
4759 | tree offvar; | |
7d75abc8 MM |
4760 | tree ivstep; |
4761 | tree running_off; | |
9771b263 | 4762 | vec<constructor_elt, va_gc> *v = NULL; |
7d75abc8 | 4763 | gimple_seq stmts = NULL; |
14ac6aa2 RB |
4764 | tree stride_base, stride_step, alias_off; |
4765 | ||
4766 | gcc_assert (!nested_in_vect_loop); | |
7d75abc8 | 4767 | |
14ac6aa2 RB |
4768 | stride_base |
4769 | = fold_build_pointer_plus | |
4770 | (unshare_expr (DR_BASE_ADDRESS (dr)), | |
4771 | size_binop (PLUS_EXPR, | |
4772 | convert_to_ptrofftype (unshare_expr (DR_OFFSET (dr))), | |
4773 | convert_to_ptrofftype (DR_INIT(dr)))); | |
4774 | stride_step = fold_convert (sizetype, unshare_expr (DR_STEP (dr))); | |
7d75abc8 MM |
4775 | |
4776 | /* For a load with loop-invariant (but other than power-of-2) | |
4777 | stride (i.e. not a grouped access) like so: | |
4778 | ||
4779 | for (i = 0; i < n; i += stride) | |
4780 | ... = array[i]; | |
4781 | ||
4782 | we generate a new induction variable and new accesses to | |
4783 | form a new vector (or vectors, depending on ncopies): | |
4784 | ||
4785 | for (j = 0; ; j += VF*stride) | |
4786 | tmp1 = array[j]; | |
4787 | tmp2 = array[j + stride]; | |
4788 | ... | |
4789 | vectemp = {tmp1, tmp2, ...} | |
4790 | */ | |
4791 | ||
4792 | ivstep = stride_step; | |
4793 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
4794 | build_int_cst (TREE_TYPE (ivstep), vf)); | |
4795 | ||
4796 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
4797 | ||
4798 | create_iv (stride_base, ivstep, NULL, | |
4799 | loop, &incr_gsi, insert_after, | |
4800 | &offvar, NULL); | |
4801 | incr = gsi_stmt (incr_gsi); | |
4802 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo, NULL)); | |
4803 | ||
4804 | stride_step = force_gimple_operand (stride_step, &stmts, true, NULL_TREE); | |
4805 | if (stmts) | |
4806 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
4807 | ||
4808 | prev_stmt_info = NULL; | |
4809 | running_off = offvar; | |
14ac6aa2 | 4810 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (dr)), 0); |
7d75abc8 MM |
4811 | for (j = 0; j < ncopies; j++) |
4812 | { | |
4813 | tree vec_inv; | |
4814 | ||
9771b263 | 4815 | vec_alloc (v, nunits); |
7d75abc8 MM |
4816 | for (i = 0; i < nunits; i++) |
4817 | { | |
4818 | tree newref, newoff; | |
4819 | gimple incr; | |
14ac6aa2 RB |
4820 | newref = build2 (MEM_REF, TREE_TYPE (vectype), |
4821 | running_off, alias_off); | |
7d75abc8 MM |
4822 | |
4823 | newref = force_gimple_operand_gsi (gsi, newref, true, | |
4824 | NULL_TREE, true, | |
4825 | GSI_SAME_STMT); | |
4826 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, newref); | |
070ecdfd | 4827 | newoff = copy_ssa_name (running_off, NULL); |
14ac6aa2 RB |
4828 | incr = gimple_build_assign_with_ops (POINTER_PLUS_EXPR, newoff, |
4829 | running_off, stride_step); | |
7d75abc8 MM |
4830 | vect_finish_stmt_generation (stmt, incr, gsi); |
4831 | ||
4832 | running_off = newoff; | |
4833 | } | |
4834 | ||
4835 | vec_inv = build_constructor (vectype, v); | |
4836 | new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi); | |
4837 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
7d75abc8 MM |
4838 | |
4839 | if (j == 0) | |
4840 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4841 | else | |
4842 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4843 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4844 | } | |
4845 | return true; | |
4846 | } | |
aec7ae7d | 4847 | |
0d0293ac | 4848 | if (grouped_load) |
ebfd146a | 4849 | { |
e14c1050 | 4850 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
6aa904c4 | 4851 | if (slp |
01d8bf07 | 4852 | && !SLP_TREE_LOAD_PERMUTATION (slp_node).exists () |
9771b263 DN |
4853 | && first_stmt != SLP_TREE_SCALAR_STMTS (slp_node)[0]) |
4854 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 4855 | |
ebfd146a | 4856 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
4857 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
4858 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
4859 | ??? But we can only do so if there is exactly one | |
4860 | as we have no way to get at the rest. Leave the CSE | |
4861 | opportunity alone. | |
4862 | ??? With the group load eventually participating | |
4863 | in multiple different permutations (having multiple | |
4864 | slp nodes which refer to the same group) the CSE | |
4865 | is even wrong code. See PR56270. */ | |
4866 | && !slp) | |
ebfd146a IR |
4867 | { |
4868 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4869 | return true; | |
4870 | } | |
4871 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 4872 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a IR |
4873 | |
4874 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
4875 | if (slp) | |
4876 | { | |
0d0293ac | 4877 | grouped_load = false; |
ebfd146a | 4878 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); |
01d8bf07 | 4879 | if (SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
a70d6342 | 4880 | slp_perm = true; |
a64b9c26 | 4881 | group_gap = GROUP_GAP (vinfo_for_stmt (first_stmt)); |
a70d6342 | 4882 | } |
ebfd146a | 4883 | else |
a64b9c26 RB |
4884 | { |
4885 | vec_num = group_size; | |
4886 | group_gap = 0; | |
4887 | } | |
ebfd146a IR |
4888 | } |
4889 | else | |
4890 | { | |
4891 | first_stmt = stmt; | |
4892 | first_dr = dr; | |
4893 | group_size = vec_num = 1; | |
a64b9c26 | 4894 | group_gap = 0; |
ebfd146a IR |
4895 | } |
4896 | ||
720f5239 | 4897 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 4898 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
4899 | /* Targets with load-lane instructions must not require explicit |
4900 | realignment. */ | |
4901 | gcc_assert (!load_lanes_p | |
4902 | || alignment_support_scheme == dr_aligned | |
4903 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
4904 | |
4905 | /* In case the vectorization factor (VF) is bigger than the number | |
4906 | of elements that we can fit in a vectype (nunits), we have to generate | |
4907 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 4908 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 4909 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 4910 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 4911 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
4912 | stmts that use the defs of the current stmt. The example below |
4913 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
4914 | need to create 4 vectorized stmts): | |
ebfd146a IR |
4915 | |
4916 | before vectorization: | |
4917 | RELATED_STMT VEC_STMT | |
4918 | S1: x = memref - - | |
4919 | S2: z = x + 1 - - | |
4920 | ||
4921 | step 1: vectorize stmt S1: | |
4922 | We first create the vector stmt VS1_0, and, as usual, record a | |
4923 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
4924 | Next, we create the vector stmt VS1_1, and record a pointer to | |
4925 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 4926 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
4927 | stmts and pointers: |
4928 | RELATED_STMT VEC_STMT | |
4929 | VS1_0: vx0 = memref0 VS1_1 - | |
4930 | VS1_1: vx1 = memref1 VS1_2 - | |
4931 | VS1_2: vx2 = memref2 VS1_3 - | |
4932 | VS1_3: vx3 = memref3 - - | |
4933 | S1: x = load - VS1_0 | |
4934 | S2: z = x + 1 - - | |
4935 | ||
b8698a0f L |
4936 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
4937 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
4938 | stmt S2. */ |
4939 | ||
0d0293ac | 4940 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
4941 | |
4942 | S1: x2 = &base + 2 | |
4943 | S2: x0 = &base | |
4944 | S3: x1 = &base + 1 | |
4945 | S4: x3 = &base + 3 | |
4946 | ||
b8698a0f | 4947 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
4948 | starting from the access of the first stmt of the chain: |
4949 | ||
4950 | VS1: vx0 = &base | |
4951 | VS2: vx1 = &base + vec_size*1 | |
4952 | VS3: vx3 = &base + vec_size*2 | |
4953 | VS4: vx4 = &base + vec_size*3 | |
4954 | ||
4955 | Then permutation statements are generated: | |
4956 | ||
e2c83630 RH |
4957 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
4958 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
4959 | ... |
4960 | ||
4961 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
4962 | (the order of the data-refs in the output of vect_permute_load_chain | |
4963 | corresponds to the order of scalar stmts in the interleaving chain - see | |
4964 | the documentation of vect_permute_load_chain()). | |
4965 | The generation of permutation stmts and recording them in | |
0d0293ac | 4966 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 4967 | |
b8698a0f | 4968 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
4969 | permutation stmts above are created for every copy. The result vector |
4970 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
4971 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
4972 | |
4973 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
4974 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
4975 | we generate the following code: | |
4976 | p = initial_addr; | |
4977 | indx = 0; | |
4978 | loop { | |
4979 | p = p + indx * vectype_size; | |
4980 | vec_dest = *(p); | |
4981 | indx = indx + 1; | |
4982 | } | |
4983 | ||
4984 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 4985 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
4986 | then generate the following code, in which the data in each iteration is |
4987 | obtained by two vector loads, one from the previous iteration, and one | |
4988 | from the current iteration: | |
4989 | p1 = initial_addr; | |
4990 | msq_init = *(floor(p1)) | |
4991 | p2 = initial_addr + VS - 1; | |
4992 | realignment_token = call target_builtin; | |
4993 | indx = 0; | |
4994 | loop { | |
4995 | p2 = p2 + indx * vectype_size | |
4996 | lsq = *(floor(p2)) | |
4997 | vec_dest = realign_load (msq, lsq, realignment_token) | |
4998 | indx = indx + 1; | |
4999 | msq = lsq; | |
5000 | } */ | |
5001 | ||
5002 | /* If the misalignment remains the same throughout the execution of the | |
5003 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 5004 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
5005 | This can only occur when vectorizing memory accesses in the inner-loop |
5006 | nested within an outer-loop that is being vectorized. */ | |
5007 | ||
d1e4b493 | 5008 | if (nested_in_vect_loop |
211bea38 | 5009 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
ebfd146a IR |
5010 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) |
5011 | { | |
5012 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
5013 | compute_in_loop = true; | |
5014 | } | |
5015 | ||
5016 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
5017 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 5018 | && !compute_in_loop) |
ebfd146a IR |
5019 | { |
5020 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
5021 | alignment_support_scheme, NULL_TREE, | |
5022 | &at_loop); | |
5023 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
5024 | { | |
5025 | phi = SSA_NAME_DEF_STMT (msq); | |
5026 | offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
5027 | } | |
5028 | } | |
5029 | else | |
5030 | at_loop = loop; | |
5031 | ||
a1e53f3f L |
5032 | if (negative) |
5033 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
5034 | ||
272c6793 RS |
5035 | if (load_lanes_p) |
5036 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
5037 | else | |
5038 | aggr_type = vectype; | |
5039 | ||
ebfd146a IR |
5040 | prev_stmt_info = NULL; |
5041 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 5042 | { |
272c6793 | 5043 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 5044 | if (j == 0) |
74bf76ed JJ |
5045 | { |
5046 | bool simd_lane_access_p | |
5047 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
5048 | if (simd_lane_access_p | |
5049 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
5050 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
5051 | && integer_zerop (DR_OFFSET (first_dr)) | |
5052 | && integer_zerop (DR_INIT (first_dr)) | |
5053 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
5054 | get_alias_set (DR_REF (first_dr))) | |
5055 | && (alignment_support_scheme == dr_aligned | |
5056 | || alignment_support_scheme == dr_unaligned_supported)) | |
5057 | { | |
5058 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
5059 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
5060 | (DR_REF (first_dr)), 0); | |
5061 | } | |
5062 | else | |
5063 | dataref_ptr | |
5064 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
5065 | offset, &dummy, gsi, &ptr_incr, | |
5066 | simd_lane_access_p, &inv_p); | |
5067 | } | |
5068 | else if (dataref_offset) | |
5069 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
5070 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 5071 | else |
272c6793 RS |
5072 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
5073 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 5074 | |
0d0293ac | 5075 | if (grouped_load || slp_perm) |
9771b263 | 5076 | dr_chain.create (vec_num); |
5ce1ee7f | 5077 | |
272c6793 | 5078 | if (load_lanes_p) |
ebfd146a | 5079 | { |
272c6793 RS |
5080 | tree vec_array; |
5081 | ||
5082 | vec_array = create_vector_array (vectype, vec_num); | |
5083 | ||
5084 | /* Emit: | |
5085 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
5086 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
5087 | new_stmt = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
5088 | gimple_call_set_lhs (new_stmt, vec_array); | |
5089 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a | 5090 | |
272c6793 RS |
5091 | /* Extract each vector into an SSA_NAME. */ |
5092 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 5093 | { |
272c6793 RS |
5094 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
5095 | vec_array, i); | |
9771b263 | 5096 | dr_chain.quick_push (new_temp); |
272c6793 RS |
5097 | } |
5098 | ||
5099 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 5100 | vect_record_grouped_load_vectors (stmt, dr_chain); |
272c6793 RS |
5101 | } |
5102 | else | |
5103 | { | |
5104 | for (i = 0; i < vec_num; i++) | |
5105 | { | |
5106 | if (i > 0) | |
5107 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
5108 | stmt, NULL_TREE); | |
5109 | ||
5110 | /* 2. Create the vector-load in the loop. */ | |
5111 | switch (alignment_support_scheme) | |
5112 | { | |
5113 | case dr_aligned: | |
5114 | case dr_unaligned_supported: | |
be1ac4ec | 5115 | { |
644ffefd MJ |
5116 | unsigned int align, misalign; |
5117 | ||
272c6793 RS |
5118 | data_ref |
5119 | = build2 (MEM_REF, vectype, dataref_ptr, | |
74bf76ed JJ |
5120 | dataref_offset |
5121 | ? dataref_offset | |
5122 | : build_int_cst (reference_alias_ptr_type | |
5123 | (DR_REF (first_dr)), 0)); | |
644ffefd | 5124 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 RS |
5125 | if (alignment_support_scheme == dr_aligned) |
5126 | { | |
5127 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 5128 | misalign = 0; |
272c6793 RS |
5129 | } |
5130 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
5131 | { | |
5132 | TREE_TYPE (data_ref) | |
5133 | = build_aligned_type (TREE_TYPE (data_ref), | |
5134 | TYPE_ALIGN (elem_type)); | |
644ffefd MJ |
5135 | align = TYPE_ALIGN_UNIT (elem_type); |
5136 | misalign = 0; | |
272c6793 RS |
5137 | } |
5138 | else | |
5139 | { | |
5140 | TREE_TYPE (data_ref) | |
5141 | = build_aligned_type (TREE_TYPE (data_ref), | |
5142 | TYPE_ALIGN (elem_type)); | |
644ffefd | 5143 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 5144 | } |
74bf76ed JJ |
5145 | if (dataref_offset == NULL_TREE) |
5146 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), | |
5147 | align, misalign); | |
272c6793 | 5148 | break; |
be1ac4ec | 5149 | } |
272c6793 | 5150 | case dr_explicit_realign: |
267d3070 | 5151 | { |
272c6793 RS |
5152 | tree ptr, bump; |
5153 | tree vs_minus_1; | |
5154 | ||
5155 | vs_minus_1 = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); | |
5156 | ||
5157 | if (compute_in_loop) | |
5158 | msq = vect_setup_realignment (first_stmt, gsi, | |
5159 | &realignment_token, | |
5160 | dr_explicit_realign, | |
5161 | dataref_ptr, NULL); | |
5162 | ||
070ecdfd | 5163 | ptr = copy_ssa_name (dataref_ptr, NULL); |
272c6793 | 5164 | new_stmt = gimple_build_assign_with_ops |
070ecdfd | 5165 | (BIT_AND_EXPR, ptr, dataref_ptr, |
272c6793 RS |
5166 | build_int_cst |
5167 | (TREE_TYPE (dataref_ptr), | |
5168 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
5169 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
5170 | data_ref | |
5171 | = build2 (MEM_REF, vectype, ptr, | |
5172 | build_int_cst (reference_alias_ptr_type | |
5173 | (DR_REF (first_dr)), 0)); | |
5174 | vec_dest = vect_create_destination_var (scalar_dest, | |
5175 | vectype); | |
5176 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
5177 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5178 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5179 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
5180 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
5181 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5182 | msq = new_temp; | |
5183 | ||
5184 | bump = size_binop (MULT_EXPR, vs_minus_1, | |
7b7b1813 | 5185 | TYPE_SIZE_UNIT (elem_type)); |
272c6793 RS |
5186 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
5187 | new_stmt = gimple_build_assign_with_ops | |
5188 | (BIT_AND_EXPR, NULL_TREE, ptr, | |
5189 | build_int_cst | |
5190 | (TREE_TYPE (ptr), | |
5191 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
070ecdfd | 5192 | ptr = copy_ssa_name (dataref_ptr, new_stmt); |
272c6793 RS |
5193 | gimple_assign_set_lhs (new_stmt, ptr); |
5194 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5195 | data_ref | |
5196 | = build2 (MEM_REF, vectype, ptr, | |
5197 | build_int_cst (reference_alias_ptr_type | |
5198 | (DR_REF (first_dr)), 0)); | |
5199 | break; | |
267d3070 | 5200 | } |
272c6793 | 5201 | case dr_explicit_realign_optimized: |
070ecdfd | 5202 | new_temp = copy_ssa_name (dataref_ptr, NULL); |
272c6793 | 5203 | new_stmt = gimple_build_assign_with_ops |
070ecdfd | 5204 | (BIT_AND_EXPR, new_temp, dataref_ptr, |
272c6793 RS |
5205 | build_int_cst |
5206 | (TREE_TYPE (dataref_ptr), | |
5207 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
5208 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
5209 | data_ref | |
5210 | = build2 (MEM_REF, vectype, new_temp, | |
5211 | build_int_cst (reference_alias_ptr_type | |
5212 | (DR_REF (first_dr)), 0)); | |
5213 | break; | |
5214 | default: | |
5215 | gcc_unreachable (); | |
5216 | } | |
ebfd146a | 5217 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 5218 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
5219 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5220 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5221 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5222 | ||
272c6793 RS |
5223 | /* 3. Handle explicit realignment if necessary/supported. |
5224 | Create in loop: | |
5225 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
5226 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
5227 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 5228 | { |
272c6793 RS |
5229 | lsq = gimple_assign_lhs (new_stmt); |
5230 | if (!realignment_token) | |
5231 | realignment_token = dataref_ptr; | |
5232 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5233 | new_stmt | |
73804b12 RG |
5234 | = gimple_build_assign_with_ops (REALIGN_LOAD_EXPR, |
5235 | vec_dest, msq, lsq, | |
5236 | realignment_token); | |
272c6793 RS |
5237 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5238 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5239 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5240 | ||
5241 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
5242 | { | |
5243 | gcc_assert (phi); | |
5244 | if (i == vec_num - 1 && j == ncopies - 1) | |
5245 | add_phi_arg (phi, lsq, | |
5246 | loop_latch_edge (containing_loop), | |
9e227d60 | 5247 | UNKNOWN_LOCATION); |
272c6793 RS |
5248 | msq = lsq; |
5249 | } | |
ebfd146a | 5250 | } |
ebfd146a | 5251 | |
59fd17e3 RB |
5252 | /* 4. Handle invariant-load. */ |
5253 | if (inv_p && !bb_vinfo) | |
5254 | { | |
5255 | gimple_stmt_iterator gsi2 = *gsi; | |
5256 | gcc_assert (!grouped_load); | |
5257 | gsi_next (&gsi2); | |
5258 | new_temp = vect_init_vector (stmt, scalar_dest, | |
5259 | vectype, &gsi2); | |
5260 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
5261 | } | |
5262 | ||
272c6793 RS |
5263 | if (negative) |
5264 | { | |
aec7ae7d JJ |
5265 | tree perm_mask = perm_mask_for_reverse (vectype); |
5266 | new_temp = permute_vec_elements (new_temp, new_temp, | |
5267 | perm_mask, stmt, gsi); | |
ebfd146a IR |
5268 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
5269 | } | |
267d3070 | 5270 | |
272c6793 | 5271 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
5272 | vect_transform_grouped_load (). */ |
5273 | if (grouped_load || slp_perm) | |
9771b263 | 5274 | dr_chain.quick_push (new_temp); |
267d3070 | 5275 | |
272c6793 RS |
5276 | /* Store vector loads in the corresponding SLP_NODE. */ |
5277 | if (slp && !slp_perm) | |
9771b263 | 5278 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
272c6793 | 5279 | } |
a64b9c26 RB |
5280 | /* Bump the vector pointer to account for a gap. */ |
5281 | if (slp && group_gap != 0) | |
5282 | { | |
5283 | tree bump = size_binop (MULT_EXPR, | |
5284 | TYPE_SIZE_UNIT (elem_type), | |
5285 | size_int (group_gap)); | |
5286 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
5287 | stmt, bump); | |
5288 | } | |
ebfd146a IR |
5289 | } |
5290 | ||
5291 | if (slp && !slp_perm) | |
5292 | continue; | |
5293 | ||
5294 | if (slp_perm) | |
5295 | { | |
01d8bf07 | 5296 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
ebfd146a IR |
5297 | slp_node_instance, false)) |
5298 | { | |
9771b263 | 5299 | dr_chain.release (); |
ebfd146a IR |
5300 | return false; |
5301 | } | |
5302 | } | |
5303 | else | |
5304 | { | |
0d0293ac | 5305 | if (grouped_load) |
ebfd146a | 5306 | { |
272c6793 | 5307 | if (!load_lanes_p) |
0d0293ac | 5308 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 5309 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
5310 | } |
5311 | else | |
5312 | { | |
5313 | if (j == 0) | |
5314 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5315 | else | |
5316 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5317 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5318 | } | |
5319 | } | |
9771b263 | 5320 | dr_chain.release (); |
ebfd146a IR |
5321 | } |
5322 | ||
ebfd146a IR |
5323 | return true; |
5324 | } | |
5325 | ||
5326 | /* Function vect_is_simple_cond. | |
b8698a0f | 5327 | |
ebfd146a IR |
5328 | Input: |
5329 | LOOP - the loop that is being vectorized. | |
5330 | COND - Condition that is checked for simple use. | |
5331 | ||
e9e1d143 RG |
5332 | Output: |
5333 | *COMP_VECTYPE - the vector type for the comparison. | |
5334 | ||
ebfd146a IR |
5335 | Returns whether a COND can be vectorized. Checks whether |
5336 | condition operands are supportable using vec_is_simple_use. */ | |
5337 | ||
87aab9b2 | 5338 | static bool |
24ee1384 IR |
5339 | vect_is_simple_cond (tree cond, gimple stmt, loop_vec_info loop_vinfo, |
5340 | bb_vec_info bb_vinfo, tree *comp_vectype) | |
ebfd146a IR |
5341 | { |
5342 | tree lhs, rhs; | |
5343 | tree def; | |
5344 | enum vect_def_type dt; | |
e9e1d143 | 5345 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a IR |
5346 | |
5347 | if (!COMPARISON_CLASS_P (cond)) | |
5348 | return false; | |
5349 | ||
5350 | lhs = TREE_OPERAND (cond, 0); | |
5351 | rhs = TREE_OPERAND (cond, 1); | |
5352 | ||
5353 | if (TREE_CODE (lhs) == SSA_NAME) | |
5354 | { | |
5355 | gimple lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); | |
24ee1384 IR |
5356 | if (!vect_is_simple_use_1 (lhs, stmt, loop_vinfo, bb_vinfo, |
5357 | &lhs_def_stmt, &def, &dt, &vectype1)) | |
ebfd146a IR |
5358 | return false; |
5359 | } | |
5360 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
5361 | && TREE_CODE (lhs) != FIXED_CST) | |
5362 | return false; | |
5363 | ||
5364 | if (TREE_CODE (rhs) == SSA_NAME) | |
5365 | { | |
5366 | gimple rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); | |
24ee1384 IR |
5367 | if (!vect_is_simple_use_1 (rhs, stmt, loop_vinfo, bb_vinfo, |
5368 | &rhs_def_stmt, &def, &dt, &vectype2)) | |
ebfd146a IR |
5369 | return false; |
5370 | } | |
f7e531cf | 5371 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST |
ebfd146a IR |
5372 | && TREE_CODE (rhs) != FIXED_CST) |
5373 | return false; | |
5374 | ||
e9e1d143 | 5375 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
ebfd146a IR |
5376 | return true; |
5377 | } | |
5378 | ||
5379 | /* vectorizable_condition. | |
5380 | ||
b8698a0f L |
5381 | Check if STMT is conditional modify expression that can be vectorized. |
5382 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
5383 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
5384 | at GSI. |
5385 | ||
5386 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
5387 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
5388 | else caluse if it is 2). | |
ebfd146a IR |
5389 | |
5390 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5391 | ||
4bbe8262 | 5392 | bool |
ebfd146a | 5393 | vectorizable_condition (gimple stmt, gimple_stmt_iterator *gsi, |
f7e531cf IR |
5394 | gimple *vec_stmt, tree reduc_def, int reduc_index, |
5395 | slp_tree slp_node) | |
ebfd146a IR |
5396 | { |
5397 | tree scalar_dest = NULL_TREE; | |
5398 | tree vec_dest = NULL_TREE; | |
ebfd146a IR |
5399 | tree cond_expr, then_clause, else_clause; |
5400 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5401 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
df11cc78 | 5402 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
5403 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
5404 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
ebfd146a IR |
5405 | tree vec_compare, vec_cond_expr; |
5406 | tree new_temp; | |
5407 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
ebfd146a | 5408 | tree def; |
a855b1b1 | 5409 | enum vect_def_type dt, dts[4]; |
ebfd146a | 5410 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
f7e531cf | 5411 | int ncopies; |
ebfd146a | 5412 | enum tree_code code; |
a855b1b1 | 5413 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
5414 | int i, j; |
5415 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
5416 | vec<tree> vec_oprnds0 = vNULL; |
5417 | vec<tree> vec_oprnds1 = vNULL; | |
5418 | vec<tree> vec_oprnds2 = vNULL; | |
5419 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 5420 | tree vec_cmp_type; |
b8698a0f | 5421 | |
f7e531cf IR |
5422 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
5423 | ncopies = 1; | |
5424 | else | |
5425 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
437f4a00 | 5426 | |
ebfd146a | 5427 | gcc_assert (ncopies >= 1); |
a855b1b1 | 5428 | if (reduc_index && ncopies > 1) |
ebfd146a IR |
5429 | return false; /* FORNOW */ |
5430 | ||
f7e531cf IR |
5431 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
5432 | return false; | |
5433 | ||
5434 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
ebfd146a IR |
5435 | return false; |
5436 | ||
4bbe8262 IR |
5437 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5438 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
5439 | && reduc_def)) | |
ebfd146a IR |
5440 | return false; |
5441 | ||
ebfd146a | 5442 | /* FORNOW: not yet supported. */ |
b8698a0f | 5443 | if (STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 5444 | { |
73fbfcad | 5445 | if (dump_enabled_p ()) |
78c60e3d SS |
5446 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5447 | "value used after loop."); | |
ebfd146a IR |
5448 | return false; |
5449 | } | |
5450 | ||
5451 | /* Is vectorizable conditional operation? */ | |
5452 | if (!is_gimple_assign (stmt)) | |
5453 | return false; | |
5454 | ||
5455 | code = gimple_assign_rhs_code (stmt); | |
5456 | ||
5457 | if (code != COND_EXPR) | |
5458 | return false; | |
5459 | ||
4e71066d RG |
5460 | cond_expr = gimple_assign_rhs1 (stmt); |
5461 | then_clause = gimple_assign_rhs2 (stmt); | |
5462 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 5463 | |
24ee1384 IR |
5464 | if (!vect_is_simple_cond (cond_expr, stmt, loop_vinfo, bb_vinfo, |
5465 | &comp_vectype) | |
e9e1d143 | 5466 | || !comp_vectype) |
ebfd146a IR |
5467 | return false; |
5468 | ||
5469 | if (TREE_CODE (then_clause) == SSA_NAME) | |
5470 | { | |
5471 | gimple then_def_stmt = SSA_NAME_DEF_STMT (then_clause); | |
24ee1384 | 5472 | if (!vect_is_simple_use (then_clause, stmt, loop_vinfo, bb_vinfo, |
ebfd146a IR |
5473 | &then_def_stmt, &def, &dt)) |
5474 | return false; | |
5475 | } | |
b8698a0f | 5476 | else if (TREE_CODE (then_clause) != INTEGER_CST |
ebfd146a IR |
5477 | && TREE_CODE (then_clause) != REAL_CST |
5478 | && TREE_CODE (then_clause) != FIXED_CST) | |
5479 | return false; | |
5480 | ||
5481 | if (TREE_CODE (else_clause) == SSA_NAME) | |
5482 | { | |
5483 | gimple else_def_stmt = SSA_NAME_DEF_STMT (else_clause); | |
24ee1384 | 5484 | if (!vect_is_simple_use (else_clause, stmt, loop_vinfo, bb_vinfo, |
ebfd146a IR |
5485 | &else_def_stmt, &def, &dt)) |
5486 | return false; | |
5487 | } | |
b8698a0f | 5488 | else if (TREE_CODE (else_clause) != INTEGER_CST |
ebfd146a IR |
5489 | && TREE_CODE (else_clause) != REAL_CST |
5490 | && TREE_CODE (else_clause) != FIXED_CST) | |
5491 | return false; | |
5492 | ||
74946978 MP |
5493 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (vectype))); |
5494 | /* The result of a vector comparison should be signed type. */ | |
5495 | tree cmp_type = build_nonstandard_integer_type (prec, 0); | |
5496 | vec_cmp_type = get_same_sized_vectype (cmp_type, vectype); | |
5497 | if (vec_cmp_type == NULL_TREE) | |
5498 | return false; | |
784fb9b3 | 5499 | |
b8698a0f | 5500 | if (!vec_stmt) |
ebfd146a IR |
5501 | { |
5502 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
e9e1d143 | 5503 | return expand_vec_cond_expr_p (vectype, comp_vectype); |
ebfd146a IR |
5504 | } |
5505 | ||
f7e531cf IR |
5506 | /* Transform. */ |
5507 | ||
5508 | if (!slp_node) | |
5509 | { | |
9771b263 DN |
5510 | vec_oprnds0.create (1); |
5511 | vec_oprnds1.create (1); | |
5512 | vec_oprnds2.create (1); | |
5513 | vec_oprnds3.create (1); | |
f7e531cf | 5514 | } |
ebfd146a IR |
5515 | |
5516 | /* Handle def. */ | |
5517 | scalar_dest = gimple_assign_lhs (stmt); | |
5518 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5519 | ||
5520 | /* Handle cond expr. */ | |
a855b1b1 MM |
5521 | for (j = 0; j < ncopies; j++) |
5522 | { | |
f7e531cf | 5523 | gimple new_stmt = NULL; |
a855b1b1 MM |
5524 | if (j == 0) |
5525 | { | |
f7e531cf IR |
5526 | if (slp_node) |
5527 | { | |
9771b263 DN |
5528 | vec<tree> ops; |
5529 | ops.create (4); | |
37b5ec8f | 5530 | vec<vec<tree> > vec_defs; |
9771b263 DN |
5531 | |
5532 | vec_defs.create (4); | |
5533 | ops.safe_push (TREE_OPERAND (cond_expr, 0)); | |
5534 | ops.safe_push (TREE_OPERAND (cond_expr, 1)); | |
5535 | ops.safe_push (then_clause); | |
5536 | ops.safe_push (else_clause); | |
f7e531cf | 5537 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); |
37b5ec8f JJ |
5538 | vec_oprnds3 = vec_defs.pop (); |
5539 | vec_oprnds2 = vec_defs.pop (); | |
5540 | vec_oprnds1 = vec_defs.pop (); | |
5541 | vec_oprnds0 = vec_defs.pop (); | |
f7e531cf | 5542 | |
9771b263 DN |
5543 | ops.release (); |
5544 | vec_defs.release (); | |
f7e531cf IR |
5545 | } |
5546 | else | |
5547 | { | |
5548 | gimple gtemp; | |
5549 | vec_cond_lhs = | |
a855b1b1 MM |
5550 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), |
5551 | stmt, NULL); | |
24ee1384 IR |
5552 | vect_is_simple_use (TREE_OPERAND (cond_expr, 0), stmt, |
5553 | loop_vinfo, NULL, >emp, &def, &dts[0]); | |
f7e531cf IR |
5554 | |
5555 | vec_cond_rhs = | |
5556 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), | |
5557 | stmt, NULL); | |
24ee1384 IR |
5558 | vect_is_simple_use (TREE_OPERAND (cond_expr, 1), stmt, |
5559 | loop_vinfo, NULL, >emp, &def, &dts[1]); | |
f7e531cf IR |
5560 | if (reduc_index == 1) |
5561 | vec_then_clause = reduc_def; | |
5562 | else | |
5563 | { | |
5564 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
5565 | stmt, NULL); | |
24ee1384 | 5566 | vect_is_simple_use (then_clause, stmt, loop_vinfo, |
f7e531cf IR |
5567 | NULL, >emp, &def, &dts[2]); |
5568 | } | |
5569 | if (reduc_index == 2) | |
5570 | vec_else_clause = reduc_def; | |
5571 | else | |
5572 | { | |
5573 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
a855b1b1 | 5574 | stmt, NULL); |
24ee1384 | 5575 | vect_is_simple_use (else_clause, stmt, loop_vinfo, |
a855b1b1 | 5576 | NULL, >emp, &def, &dts[3]); |
f7e531cf | 5577 | } |
a855b1b1 MM |
5578 | } |
5579 | } | |
5580 | else | |
5581 | { | |
f7e531cf | 5582 | vec_cond_lhs = vect_get_vec_def_for_stmt_copy (dts[0], |
9771b263 | 5583 | vec_oprnds0.pop ()); |
f7e531cf | 5584 | vec_cond_rhs = vect_get_vec_def_for_stmt_copy (dts[1], |
9771b263 | 5585 | vec_oprnds1.pop ()); |
a855b1b1 | 5586 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 5587 | vec_oprnds2.pop ()); |
a855b1b1 | 5588 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 5589 | vec_oprnds3.pop ()); |
f7e531cf IR |
5590 | } |
5591 | ||
5592 | if (!slp_node) | |
5593 | { | |
9771b263 DN |
5594 | vec_oprnds0.quick_push (vec_cond_lhs); |
5595 | vec_oprnds1.quick_push (vec_cond_rhs); | |
5596 | vec_oprnds2.quick_push (vec_then_clause); | |
5597 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
5598 | } |
5599 | ||
9dc3f7de | 5600 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 5601 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 5602 | { |
9771b263 DN |
5603 | vec_cond_rhs = vec_oprnds1[i]; |
5604 | vec_then_clause = vec_oprnds2[i]; | |
5605 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 5606 | |
784fb9b3 JJ |
5607 | vec_compare = build2 (TREE_CODE (cond_expr), vec_cmp_type, |
5608 | vec_cond_lhs, vec_cond_rhs); | |
f7e531cf IR |
5609 | vec_cond_expr = build3 (VEC_COND_EXPR, vectype, |
5610 | vec_compare, vec_then_clause, vec_else_clause); | |
a855b1b1 | 5611 | |
f7e531cf IR |
5612 | new_stmt = gimple_build_assign (vec_dest, vec_cond_expr); |
5613 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5614 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5615 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5616 | if (slp_node) | |
9771b263 | 5617 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
5618 | } |
5619 | ||
5620 | if (slp_node) | |
5621 | continue; | |
5622 | ||
5623 | if (j == 0) | |
5624 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5625 | else | |
5626 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5627 | ||
5628 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 5629 | } |
b8698a0f | 5630 | |
9771b263 DN |
5631 | vec_oprnds0.release (); |
5632 | vec_oprnds1.release (); | |
5633 | vec_oprnds2.release (); | |
5634 | vec_oprnds3.release (); | |
f7e531cf | 5635 | |
ebfd146a IR |
5636 | return true; |
5637 | } | |
5638 | ||
5639 | ||
8644a673 | 5640 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
5641 | |
5642 | bool | |
a70d6342 | 5643 | vect_analyze_stmt (gimple stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 5644 | { |
8644a673 | 5645 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 5646 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 5647 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 5648 | bool ok; |
a70d6342 | 5649 | tree scalar_type, vectype; |
363477c0 JJ |
5650 | gimple pattern_stmt; |
5651 | gimple_seq pattern_def_seq; | |
ebfd146a | 5652 | |
73fbfcad | 5653 | if (dump_enabled_p ()) |
ebfd146a | 5654 | { |
78c60e3d SS |
5655 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
5656 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
8644a673 | 5657 | } |
ebfd146a | 5658 | |
1825a1f3 | 5659 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 5660 | { |
73fbfcad | 5661 | if (dump_enabled_p ()) |
78c60e3d SS |
5662 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5663 | "not vectorized: stmt has volatile operands"); | |
1825a1f3 IR |
5664 | |
5665 | return false; | |
5666 | } | |
b8698a0f L |
5667 | |
5668 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
5669 | to include: |
5670 | - the COND_EXPR which is the loop exit condition | |
5671 | - any LABEL_EXPRs in the loop | |
b8698a0f | 5672 | - computations that are used only for array indexing or loop control. |
8644a673 | 5673 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 5674 | instance, therefore, all the statements are relevant. |
ebfd146a | 5675 | |
d092494c | 5676 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 5677 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
5678 | statements. In basic blocks we are called from some SLP instance |
5679 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
5680 | already will be part of SLP instance. */ | |
83197f37 IR |
5681 | |
5682 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 5683 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 5684 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 5685 | { |
9d5e7640 | 5686 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 5687 | && pattern_stmt |
9d5e7640 IR |
5688 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
5689 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
5690 | { | |
83197f37 | 5691 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
5692 | stmt = pattern_stmt; |
5693 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 5694 | if (dump_enabled_p ()) |
9d5e7640 | 5695 | { |
78c60e3d SS |
5696 | dump_printf_loc (MSG_NOTE, vect_location, |
5697 | "==> examining pattern statement: "); | |
5698 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
9d5e7640 IR |
5699 | } |
5700 | } | |
5701 | else | |
5702 | { | |
73fbfcad | 5703 | if (dump_enabled_p ()) |
78c60e3d | 5704 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant."); |
ebfd146a | 5705 | |
9d5e7640 IR |
5706 | return true; |
5707 | } | |
8644a673 | 5708 | } |
83197f37 | 5709 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 5710 | && node == NULL |
83197f37 IR |
5711 | && pattern_stmt |
5712 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
5713 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
5714 | { | |
5715 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 5716 | if (dump_enabled_p ()) |
83197f37 | 5717 | { |
78c60e3d SS |
5718 | dump_printf_loc (MSG_NOTE, vect_location, |
5719 | "==> examining pattern statement: "); | |
5720 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
83197f37 IR |
5721 | } |
5722 | ||
5723 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node)) | |
5724 | return false; | |
5725 | } | |
ebfd146a | 5726 | |
1107f3ae | 5727 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 5728 | && node == NULL |
363477c0 | 5729 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 5730 | { |
363477c0 | 5731 | gimple_stmt_iterator si; |
1107f3ae | 5732 | |
363477c0 JJ |
5733 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
5734 | { | |
5735 | gimple pattern_def_stmt = gsi_stmt (si); | |
5736 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) | |
5737 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
5738 | { | |
5739 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 5740 | if (dump_enabled_p ()) |
363477c0 | 5741 | { |
78c60e3d SS |
5742 | dump_printf_loc (MSG_NOTE, vect_location, |
5743 | "==> examining pattern def statement: "); | |
5744 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
363477c0 | 5745 | } |
1107f3ae | 5746 | |
363477c0 JJ |
5747 | if (!vect_analyze_stmt (pattern_def_stmt, |
5748 | need_to_vectorize, node)) | |
5749 | return false; | |
5750 | } | |
5751 | } | |
5752 | } | |
1107f3ae | 5753 | |
8644a673 IR |
5754 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
5755 | { | |
5756 | case vect_internal_def: | |
5757 | break; | |
ebfd146a | 5758 | |
8644a673 | 5759 | case vect_reduction_def: |
7c5222ff | 5760 | case vect_nested_cycle: |
a70d6342 | 5761 | gcc_assert (!bb_vinfo && (relevance == vect_used_in_outer |
8644a673 | 5762 | || relevance == vect_used_in_outer_by_reduction |
a70d6342 | 5763 | || relevance == vect_unused_in_scope)); |
8644a673 IR |
5764 | break; |
5765 | ||
5766 | case vect_induction_def: | |
5767 | case vect_constant_def: | |
5768 | case vect_external_def: | |
5769 | case vect_unknown_def_type: | |
5770 | default: | |
5771 | gcc_unreachable (); | |
5772 | } | |
ebfd146a | 5773 | |
a70d6342 IR |
5774 | if (bb_vinfo) |
5775 | { | |
5776 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
5777 | ||
b690cc0f | 5778 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
73fbfcad | 5779 | if (dump_enabled_p ()) |
a70d6342 | 5780 | { |
78c60e3d SS |
5781 | dump_printf_loc (MSG_NOTE, vect_location, |
5782 | "get vectype for scalar type: "); | |
5783 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
a70d6342 IR |
5784 | } |
5785 | ||
5786 | vectype = get_vectype_for_scalar_type (scalar_type); | |
5787 | if (!vectype) | |
5788 | { | |
73fbfcad | 5789 | if (dump_enabled_p ()) |
a70d6342 | 5790 | { |
78c60e3d SS |
5791 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5792 | "not SLPed: unsupported data-type "); | |
5793 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
5794 | scalar_type); | |
a70d6342 IR |
5795 | } |
5796 | return false; | |
5797 | } | |
5798 | ||
73fbfcad | 5799 | if (dump_enabled_p ()) |
a70d6342 | 5800 | { |
78c60e3d SS |
5801 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); |
5802 | dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); | |
a70d6342 IR |
5803 | } |
5804 | ||
5805 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
5806 | } | |
5807 | ||
8644a673 | 5808 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 5809 | { |
8644a673 IR |
5810 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
5811 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info)); | |
5812 | *need_to_vectorize = true; | |
ebfd146a IR |
5813 | } |
5814 | ||
8644a673 | 5815 | ok = true; |
b8698a0f | 5816 | if (!bb_vinfo |
a70d6342 IR |
5817 | && (STMT_VINFO_RELEVANT_P (stmt_info) |
5818 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
4a00c761 | 5819 | ok = (vectorizable_conversion (stmt, NULL, NULL, NULL) |
9dc3f7de | 5820 | || vectorizable_shift (stmt, NULL, NULL, NULL) |
8644a673 IR |
5821 | || vectorizable_operation (stmt, NULL, NULL, NULL) |
5822 | || vectorizable_assignment (stmt, NULL, NULL, NULL) | |
5823 | || vectorizable_load (stmt, NULL, NULL, NULL, NULL) | |
190c2236 | 5824 | || vectorizable_call (stmt, NULL, NULL, NULL) |
8644a673 | 5825 | || vectorizable_store (stmt, NULL, NULL, NULL) |
b5aeb3bb | 5826 | || vectorizable_reduction (stmt, NULL, NULL, NULL) |
f7e531cf | 5827 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, NULL)); |
a70d6342 IR |
5828 | else |
5829 | { | |
5830 | if (bb_vinfo) | |
4a00c761 JJ |
5831 | ok = (vectorizable_conversion (stmt, NULL, NULL, node) |
5832 | || vectorizable_shift (stmt, NULL, NULL, node) | |
9dc3f7de | 5833 | || vectorizable_operation (stmt, NULL, NULL, node) |
a70d6342 IR |
5834 | || vectorizable_assignment (stmt, NULL, NULL, node) |
5835 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
190c2236 | 5836 | || vectorizable_call (stmt, NULL, NULL, node) |
f7e531cf IR |
5837 | || vectorizable_store (stmt, NULL, NULL, node) |
5838 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node)); | |
b8698a0f | 5839 | } |
8644a673 IR |
5840 | |
5841 | if (!ok) | |
ebfd146a | 5842 | { |
73fbfcad | 5843 | if (dump_enabled_p ()) |
8644a673 | 5844 | { |
78c60e3d SS |
5845 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5846 | "not vectorized: relevant stmt not "); | |
5847 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
5848 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 5849 | } |
b8698a0f | 5850 | |
ebfd146a IR |
5851 | return false; |
5852 | } | |
5853 | ||
a70d6342 IR |
5854 | if (bb_vinfo) |
5855 | return true; | |
5856 | ||
8644a673 IR |
5857 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
5858 | need extra handling, except for vectorizable reductions. */ | |
5859 | if (STMT_VINFO_LIVE_P (stmt_info) | |
5860 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
5861 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 5862 | |
8644a673 | 5863 | if (!ok) |
ebfd146a | 5864 | { |
73fbfcad | 5865 | if (dump_enabled_p ()) |
8644a673 | 5866 | { |
78c60e3d SS |
5867 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5868 | "not vectorized: live stmt not "); | |
5869 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
5870 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 5871 | } |
b8698a0f | 5872 | |
8644a673 | 5873 | return false; |
ebfd146a IR |
5874 | } |
5875 | ||
ebfd146a IR |
5876 | return true; |
5877 | } | |
5878 | ||
5879 | ||
5880 | /* Function vect_transform_stmt. | |
5881 | ||
5882 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
5883 | ||
5884 | bool | |
5885 | vect_transform_stmt (gimple stmt, gimple_stmt_iterator *gsi, | |
0d0293ac | 5886 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
5887 | slp_instance slp_node_instance) |
5888 | { | |
5889 | bool is_store = false; | |
5890 | gimple vec_stmt = NULL; | |
5891 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a | 5892 | bool done; |
ebfd146a IR |
5893 | |
5894 | switch (STMT_VINFO_TYPE (stmt_info)) | |
5895 | { | |
5896 | case type_demotion_vec_info_type: | |
ebfd146a | 5897 | case type_promotion_vec_info_type: |
ebfd146a IR |
5898 | case type_conversion_vec_info_type: |
5899 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
5900 | gcc_assert (done); | |
5901 | break; | |
5902 | ||
5903 | case induc_vec_info_type: | |
5904 | gcc_assert (!slp_node); | |
5905 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
5906 | gcc_assert (done); | |
5907 | break; | |
5908 | ||
9dc3f7de IR |
5909 | case shift_vec_info_type: |
5910 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
5911 | gcc_assert (done); | |
5912 | break; | |
5913 | ||
ebfd146a IR |
5914 | case op_vec_info_type: |
5915 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
5916 | gcc_assert (done); | |
5917 | break; | |
5918 | ||
5919 | case assignment_vec_info_type: | |
5920 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
5921 | gcc_assert (done); | |
5922 | break; | |
5923 | ||
5924 | case load_vec_info_type: | |
b8698a0f | 5925 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
5926 | slp_node_instance); |
5927 | gcc_assert (done); | |
5928 | break; | |
5929 | ||
5930 | case store_vec_info_type: | |
5931 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
5932 | gcc_assert (done); | |
0d0293ac | 5933 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
5934 | { |
5935 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 5936 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
5937 | one are skipped, and there vec_stmt_info shouldn't be freed |
5938 | meanwhile. */ | |
0d0293ac | 5939 | *grouped_store = true; |
ebfd146a IR |
5940 | if (STMT_VINFO_VEC_STMT (stmt_info)) |
5941 | is_store = true; | |
5942 | } | |
5943 | else | |
5944 | is_store = true; | |
5945 | break; | |
5946 | ||
5947 | case condition_vec_info_type: | |
f7e531cf | 5948 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
5949 | gcc_assert (done); |
5950 | break; | |
5951 | ||
5952 | case call_vec_info_type: | |
190c2236 | 5953 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 5954 | stmt = gsi_stmt (*gsi); |
ebfd146a IR |
5955 | break; |
5956 | ||
5957 | case reduc_vec_info_type: | |
b5aeb3bb | 5958 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
5959 | gcc_assert (done); |
5960 | break; | |
5961 | ||
5962 | default: | |
5963 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
5964 | { | |
73fbfcad | 5965 | if (dump_enabled_p ()) |
78c60e3d SS |
5966 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5967 | "stmt not supported."); | |
ebfd146a IR |
5968 | gcc_unreachable (); |
5969 | } | |
5970 | } | |
5971 | ||
5972 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that | |
5973 | is being vectorized, but outside the immediately enclosing loop. */ | |
5974 | if (vec_stmt | |
a70d6342 IR |
5975 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
5976 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
5977 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
5978 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
5979 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 5980 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 5981 | vect_used_in_outer_by_reduction)) |
ebfd146a | 5982 | { |
a70d6342 IR |
5983 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
5984 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
5985 | imm_use_iterator imm_iter; |
5986 | use_operand_p use_p; | |
5987 | tree scalar_dest; | |
5988 | gimple exit_phi; | |
5989 | ||
73fbfcad | 5990 | if (dump_enabled_p ()) |
78c60e3d SS |
5991 | dump_printf_loc (MSG_NOTE, vect_location, |
5992 | "Record the vdef for outer-loop vectorization."); | |
ebfd146a IR |
5993 | |
5994 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
5995 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
5996 | STMT). */ | |
5997 | if (gimple_code (stmt) == GIMPLE_PHI) | |
5998 | scalar_dest = PHI_RESULT (stmt); | |
5999 | else | |
6000 | scalar_dest = gimple_assign_lhs (stmt); | |
6001 | ||
6002 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
6003 | { | |
6004 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
6005 | { | |
6006 | exit_phi = USE_STMT (use_p); | |
6007 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
6008 | } | |
6009 | } | |
6010 | } | |
6011 | ||
6012 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
6013 | being vectorized. */ | |
6014 | if (STMT_VINFO_LIVE_P (stmt_info) | |
6015 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
6016 | { | |
6017 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
6018 | gcc_assert (done); | |
6019 | } | |
6020 | ||
6021 | if (vec_stmt) | |
83197f37 | 6022 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 6023 | |
b8698a0f | 6024 | return is_store; |
ebfd146a IR |
6025 | } |
6026 | ||
6027 | ||
b8698a0f | 6028 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
6029 | stmt_vec_info. */ |
6030 | ||
6031 | void | |
6032 | vect_remove_stores (gimple first_stmt) | |
6033 | { | |
6034 | gimple next = first_stmt; | |
6035 | gimple tmp; | |
6036 | gimple_stmt_iterator next_si; | |
6037 | ||
6038 | while (next) | |
6039 | { | |
78048b1c JJ |
6040 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
6041 | ||
6042 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
6043 | if (is_pattern_stmt_p (stmt_info)) | |
6044 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
6045 | /* Free the attached stmt_vec_info and remove the stmt. */ |
6046 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 6047 | unlink_stmt_vdef (next); |
ebfd146a | 6048 | gsi_remove (&next_si, true); |
3d3f2249 | 6049 | release_defs (next); |
ebfd146a IR |
6050 | free_stmt_vec_info (next); |
6051 | next = tmp; | |
6052 | } | |
6053 | } | |
6054 | ||
6055 | ||
6056 | /* Function new_stmt_vec_info. | |
6057 | ||
6058 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
6059 | ||
6060 | stmt_vec_info | |
b8698a0f | 6061 | new_stmt_vec_info (gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 6062 | bb_vec_info bb_vinfo) |
ebfd146a IR |
6063 | { |
6064 | stmt_vec_info res; | |
6065 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
6066 | ||
6067 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
6068 | STMT_VINFO_STMT (res) = stmt; | |
6069 | STMT_VINFO_LOOP_VINFO (res) = loop_vinfo; | |
a70d6342 | 6070 | STMT_VINFO_BB_VINFO (res) = bb_vinfo; |
8644a673 | 6071 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
6072 | STMT_VINFO_LIVE_P (res) = false; |
6073 | STMT_VINFO_VECTYPE (res) = NULL; | |
6074 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 6075 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
6076 | STMT_VINFO_IN_PATTERN_P (res) = false; |
6077 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 6078 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a IR |
6079 | STMT_VINFO_DATA_REF (res) = NULL; |
6080 | ||
6081 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
6082 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
6083 | STMT_VINFO_DR_INIT (res) = NULL; | |
6084 | STMT_VINFO_DR_STEP (res) = NULL; | |
6085 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
6086 | ||
6087 | if (gimple_code (stmt) == GIMPLE_PHI | |
6088 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
6089 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
6090 | else | |
8644a673 IR |
6091 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
6092 | ||
9771b263 | 6093 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 6094 | STMT_SLP_TYPE (res) = loop_vect; |
e14c1050 IR |
6095 | GROUP_FIRST_ELEMENT (res) = NULL; |
6096 | GROUP_NEXT_ELEMENT (res) = NULL; | |
6097 | GROUP_SIZE (res) = 0; | |
6098 | GROUP_STORE_COUNT (res) = 0; | |
6099 | GROUP_GAP (res) = 0; | |
6100 | GROUP_SAME_DR_STMT (res) = NULL; | |
ebfd146a IR |
6101 | |
6102 | return res; | |
6103 | } | |
6104 | ||
6105 | ||
6106 | /* Create a hash table for stmt_vec_info. */ | |
6107 | ||
6108 | void | |
6109 | init_stmt_vec_info_vec (void) | |
6110 | { | |
9771b263 DN |
6111 | gcc_assert (!stmt_vec_info_vec.exists ()); |
6112 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
6113 | } |
6114 | ||
6115 | ||
6116 | /* Free hash table for stmt_vec_info. */ | |
6117 | ||
6118 | void | |
6119 | free_stmt_vec_info_vec (void) | |
6120 | { | |
93675444 JJ |
6121 | unsigned int i; |
6122 | vec_void_p info; | |
6123 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) | |
6124 | if (info != NULL) | |
6125 | free_stmt_vec_info (STMT_VINFO_STMT ((stmt_vec_info) info)); | |
9771b263 DN |
6126 | gcc_assert (stmt_vec_info_vec.exists ()); |
6127 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
6128 | } |
6129 | ||
6130 | ||
6131 | /* Free stmt vectorization related info. */ | |
6132 | ||
6133 | void | |
6134 | free_stmt_vec_info (gimple stmt) | |
6135 | { | |
6136 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
6137 | ||
6138 | if (!stmt_info) | |
6139 | return; | |
6140 | ||
78048b1c JJ |
6141 | /* Check if this statement has a related "pattern stmt" |
6142 | (introduced by the vectorizer during the pattern recognition | |
6143 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
6144 | too. */ | |
6145 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
6146 | { | |
6147 | stmt_vec_info patt_info | |
6148 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
6149 | if (patt_info) | |
6150 | { | |
363477c0 JJ |
6151 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
6152 | if (seq) | |
6153 | { | |
6154 | gimple_stmt_iterator si; | |
6155 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
6156 | free_stmt_vec_info (gsi_stmt (si)); | |
6157 | } | |
78048b1c JJ |
6158 | free_stmt_vec_info (STMT_VINFO_RELATED_STMT (stmt_info)); |
6159 | } | |
6160 | } | |
6161 | ||
9771b263 | 6162 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
ebfd146a IR |
6163 | set_vinfo_for_stmt (stmt, NULL); |
6164 | free (stmt_info); | |
6165 | } | |
6166 | ||
6167 | ||
bb67d9c7 | 6168 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 6169 | |
bb67d9c7 | 6170 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
6171 | by the target. */ |
6172 | ||
bb67d9c7 RG |
6173 | static tree |
6174 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a IR |
6175 | { |
6176 | enum machine_mode inner_mode = TYPE_MODE (scalar_type); | |
cc4b5170 | 6177 | enum machine_mode simd_mode; |
2f816591 | 6178 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
6179 | int nunits; |
6180 | tree vectype; | |
6181 | ||
cc4b5170 | 6182 | if (nbytes == 0) |
ebfd146a IR |
6183 | return NULL_TREE; |
6184 | ||
48f2e373 RB |
6185 | if (GET_MODE_CLASS (inner_mode) != MODE_INT |
6186 | && GET_MODE_CLASS (inner_mode) != MODE_FLOAT) | |
6187 | return NULL_TREE; | |
6188 | ||
7b7b1813 RG |
6189 | /* For vector types of elements whose mode precision doesn't |
6190 | match their types precision we use a element type of mode | |
6191 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
6192 | they support the proper result truncation/extension. |
6193 | We also make sure to build vector types with INTEGER_TYPE | |
6194 | component type only. */ | |
6d7971b8 | 6195 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
6196 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
6197 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
6198 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
6199 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 6200 | |
ccbf5bb4 RG |
6201 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
6202 | When the component mode passes the above test simply use a type | |
6203 | corresponding to that mode. The theory is that any use that | |
6204 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac RB |
6205 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
6206 | && !INTEGRAL_TYPE_P (scalar_type) | |
6207 | && !POINTER_TYPE_P (scalar_type)) | |
60b95d28 RB |
6208 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
6209 | ||
6210 | /* We can't build a vector type of elements with alignment bigger than | |
6211 | their size. */ | |
dfc2e2ac | 6212 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
6213 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
6214 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 6215 | |
dfc2e2ac RB |
6216 | /* If we felt back to using the mode fail if there was |
6217 | no scalar type for it. */ | |
6218 | if (scalar_type == NULL_TREE) | |
6219 | return NULL_TREE; | |
6220 | ||
bb67d9c7 RG |
6221 | /* If no size was supplied use the mode the target prefers. Otherwise |
6222 | lookup a vector mode of the specified size. */ | |
6223 | if (size == 0) | |
6224 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
6225 | else | |
6226 | simd_mode = mode_for_vector (inner_mode, size / nbytes); | |
cc4b5170 RG |
6227 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
6228 | if (nunits <= 1) | |
6229 | return NULL_TREE; | |
ebfd146a IR |
6230 | |
6231 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
6232 | |
6233 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
6234 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 6235 | return NULL_TREE; |
ebfd146a IR |
6236 | |
6237 | return vectype; | |
6238 | } | |
6239 | ||
bb67d9c7 RG |
6240 | unsigned int current_vector_size; |
6241 | ||
6242 | /* Function get_vectype_for_scalar_type. | |
6243 | ||
6244 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
6245 | by the target. */ | |
6246 | ||
6247 | tree | |
6248 | get_vectype_for_scalar_type (tree scalar_type) | |
6249 | { | |
6250 | tree vectype; | |
6251 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
6252 | current_vector_size); | |
6253 | if (vectype | |
6254 | && current_vector_size == 0) | |
6255 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
6256 | return vectype; | |
6257 | } | |
6258 | ||
b690cc0f RG |
6259 | /* Function get_same_sized_vectype |
6260 | ||
6261 | Returns a vector type corresponding to SCALAR_TYPE of size | |
6262 | VECTOR_TYPE if supported by the target. */ | |
6263 | ||
6264 | tree | |
bb67d9c7 | 6265 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 6266 | { |
bb67d9c7 RG |
6267 | return get_vectype_for_scalar_type_and_size |
6268 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
6269 | } |
6270 | ||
ebfd146a IR |
6271 | /* Function vect_is_simple_use. |
6272 | ||
6273 | Input: | |
a70d6342 IR |
6274 | LOOP_VINFO - the vect info of the loop that is being vectorized. |
6275 | BB_VINFO - the vect info of the basic block that is being vectorized. | |
24ee1384 | 6276 | OPERAND - operand of STMT in the loop or bb. |
ebfd146a IR |
6277 | DEF - the defining stmt in case OPERAND is an SSA_NAME. |
6278 | ||
6279 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 6280 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 6281 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 6282 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
6283 | is the case in reduction/induction computations). |
6284 | For basic blocks, supportable operands are constants and bb invariants. | |
6285 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
6286 | |
6287 | bool | |
24ee1384 | 6288 | vect_is_simple_use (tree operand, gimple stmt, loop_vec_info loop_vinfo, |
a70d6342 | 6289 | bb_vec_info bb_vinfo, gimple *def_stmt, |
ebfd146a | 6290 | tree *def, enum vect_def_type *dt) |
b8698a0f | 6291 | { |
ebfd146a IR |
6292 | basic_block bb; |
6293 | stmt_vec_info stmt_vinfo; | |
a70d6342 | 6294 | struct loop *loop = NULL; |
b8698a0f | 6295 | |
a70d6342 IR |
6296 | if (loop_vinfo) |
6297 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
6298 | |
6299 | *def_stmt = NULL; | |
6300 | *def = NULL_TREE; | |
b8698a0f | 6301 | |
73fbfcad | 6302 | if (dump_enabled_p ()) |
ebfd146a | 6303 | { |
78c60e3d SS |
6304 | dump_printf_loc (MSG_NOTE, vect_location, |
6305 | "vect_is_simple_use: operand "); | |
6306 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
ebfd146a | 6307 | } |
b8698a0f | 6308 | |
b758f602 | 6309 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
6310 | { |
6311 | *dt = vect_constant_def; | |
6312 | return true; | |
6313 | } | |
b8698a0f | 6314 | |
ebfd146a IR |
6315 | if (is_gimple_min_invariant (operand)) |
6316 | { | |
6317 | *def = operand; | |
8644a673 | 6318 | *dt = vect_external_def; |
ebfd146a IR |
6319 | return true; |
6320 | } | |
6321 | ||
6322 | if (TREE_CODE (operand) == PAREN_EXPR) | |
6323 | { | |
73fbfcad | 6324 | if (dump_enabled_p ()) |
78c60e3d | 6325 | dump_printf_loc (MSG_NOTE, vect_location, "non-associatable copy."); |
ebfd146a IR |
6326 | operand = TREE_OPERAND (operand, 0); |
6327 | } | |
b8698a0f | 6328 | |
ebfd146a IR |
6329 | if (TREE_CODE (operand) != SSA_NAME) |
6330 | { | |
73fbfcad | 6331 | if (dump_enabled_p ()) |
78c60e3d SS |
6332 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6333 | "not ssa-name."); | |
ebfd146a IR |
6334 | return false; |
6335 | } | |
b8698a0f | 6336 | |
ebfd146a IR |
6337 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
6338 | if (*def_stmt == NULL) | |
6339 | { | |
73fbfcad | 6340 | if (dump_enabled_p ()) |
78c60e3d SS |
6341 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6342 | "no def_stmt."); | |
ebfd146a IR |
6343 | return false; |
6344 | } | |
6345 | ||
73fbfcad | 6346 | if (dump_enabled_p ()) |
ebfd146a | 6347 | { |
78c60e3d SS |
6348 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
6349 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
ebfd146a IR |
6350 | } |
6351 | ||
8644a673 | 6352 | /* Empty stmt is expected only in case of a function argument. |
ebfd146a IR |
6353 | (Otherwise - we expect a phi_node or a GIMPLE_ASSIGN). */ |
6354 | if (gimple_nop_p (*def_stmt)) | |
6355 | { | |
6356 | *def = operand; | |
8644a673 | 6357 | *dt = vect_external_def; |
ebfd146a IR |
6358 | return true; |
6359 | } | |
6360 | ||
6361 | bb = gimple_bb (*def_stmt); | |
a70d6342 IR |
6362 | |
6363 | if ((loop && !flow_bb_inside_loop_p (loop, bb)) | |
6364 | || (!loop && bb != BB_VINFO_BB (bb_vinfo)) | |
b8698a0f | 6365 | || (!loop && gimple_code (*def_stmt) == GIMPLE_PHI)) |
8644a673 | 6366 | *dt = vect_external_def; |
ebfd146a IR |
6367 | else |
6368 | { | |
6369 | stmt_vinfo = vinfo_for_stmt (*def_stmt); | |
6370 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
6371 | } | |
6372 | ||
24ee1384 IR |
6373 | if (*dt == vect_unknown_def_type |
6374 | || (stmt | |
6375 | && *dt == vect_double_reduction_def | |
6376 | && gimple_code (stmt) != GIMPLE_PHI)) | |
ebfd146a | 6377 | { |
73fbfcad | 6378 | if (dump_enabled_p ()) |
78c60e3d SS |
6379 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6380 | "Unsupported pattern."); | |
ebfd146a IR |
6381 | return false; |
6382 | } | |
6383 | ||
73fbfcad | 6384 | if (dump_enabled_p ()) |
78c60e3d | 6385 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: %d.", *dt); |
ebfd146a IR |
6386 | |
6387 | switch (gimple_code (*def_stmt)) | |
6388 | { | |
6389 | case GIMPLE_PHI: | |
6390 | *def = gimple_phi_result (*def_stmt); | |
6391 | break; | |
6392 | ||
6393 | case GIMPLE_ASSIGN: | |
6394 | *def = gimple_assign_lhs (*def_stmt); | |
6395 | break; | |
6396 | ||
6397 | case GIMPLE_CALL: | |
6398 | *def = gimple_call_lhs (*def_stmt); | |
6399 | if (*def != NULL) | |
6400 | break; | |
6401 | /* FALLTHRU */ | |
6402 | default: | |
73fbfcad | 6403 | if (dump_enabled_p ()) |
78c60e3d SS |
6404 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
6405 | "unsupported defining stmt: "); | |
ebfd146a IR |
6406 | return false; |
6407 | } | |
6408 | ||
6409 | return true; | |
6410 | } | |
6411 | ||
b690cc0f RG |
6412 | /* Function vect_is_simple_use_1. |
6413 | ||
6414 | Same as vect_is_simple_use_1 but also determines the vector operand | |
6415 | type of OPERAND and stores it to *VECTYPE. If the definition of | |
6416 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
6417 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
6418 | is responsible to compute the best suited vector type for the | |
6419 | scalar operand. */ | |
6420 | ||
6421 | bool | |
24ee1384 | 6422 | vect_is_simple_use_1 (tree operand, gimple stmt, loop_vec_info loop_vinfo, |
b690cc0f RG |
6423 | bb_vec_info bb_vinfo, gimple *def_stmt, |
6424 | tree *def, enum vect_def_type *dt, tree *vectype) | |
6425 | { | |
24ee1384 IR |
6426 | if (!vect_is_simple_use (operand, stmt, loop_vinfo, bb_vinfo, def_stmt, |
6427 | def, dt)) | |
b690cc0f RG |
6428 | return false; |
6429 | ||
6430 | /* Now get a vector type if the def is internal, otherwise supply | |
6431 | NULL_TREE and leave it up to the caller to figure out a proper | |
6432 | type for the use stmt. */ | |
6433 | if (*dt == vect_internal_def | |
6434 | || *dt == vect_induction_def | |
6435 | || *dt == vect_reduction_def | |
6436 | || *dt == vect_double_reduction_def | |
6437 | || *dt == vect_nested_cycle) | |
6438 | { | |
6439 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
6440 | |
6441 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
6442 | && !STMT_VINFO_RELEVANT (stmt_info) | |
6443 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 6444 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 6445 | |
b690cc0f RG |
6446 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
6447 | gcc_assert (*vectype != NULL_TREE); | |
6448 | } | |
6449 | else if (*dt == vect_uninitialized_def | |
6450 | || *dt == vect_constant_def | |
6451 | || *dt == vect_external_def) | |
6452 | *vectype = NULL_TREE; | |
6453 | else | |
6454 | gcc_unreachable (); | |
6455 | ||
6456 | return true; | |
6457 | } | |
6458 | ||
ebfd146a IR |
6459 | |
6460 | /* Function supportable_widening_operation | |
6461 | ||
b8698a0f L |
6462 | Check whether an operation represented by the code CODE is a |
6463 | widening operation that is supported by the target platform in | |
b690cc0f RG |
6464 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
6465 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 6466 | |
ebfd146a IR |
6467 | Widening operations we currently support are NOP (CONVERT), FLOAT |
6468 | and WIDEN_MULT. This function checks if these operations are supported | |
6469 | by the target platform either directly (via vector tree-codes), or via | |
6470 | target builtins. | |
6471 | ||
6472 | Output: | |
b8698a0f L |
6473 | - CODE1 and CODE2 are codes of vector operations to be used when |
6474 | vectorizing the operation, if available. | |
ebfd146a IR |
6475 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
6476 | case of multi-step conversion (like char->short->int - in that case | |
6477 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
6478 | - INTERM_TYPES contains the intermediate type required to perform the |
6479 | widening operation (short in the above example). */ | |
ebfd146a IR |
6480 | |
6481 | bool | |
b690cc0f RG |
6482 | supportable_widening_operation (enum tree_code code, gimple stmt, |
6483 | tree vectype_out, tree vectype_in, | |
ebfd146a IR |
6484 | enum tree_code *code1, enum tree_code *code2, |
6485 | int *multi_step_cvt, | |
9771b263 | 6486 | vec<tree> *interm_types) |
ebfd146a IR |
6487 | { |
6488 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
6489 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 6490 | struct loop *vect_loop = NULL; |
ebfd146a | 6491 | enum machine_mode vec_mode; |
81f40b79 | 6492 | enum insn_code icode1, icode2; |
ebfd146a | 6493 | optab optab1, optab2; |
b690cc0f RG |
6494 | tree vectype = vectype_in; |
6495 | tree wide_vectype = vectype_out; | |
ebfd146a | 6496 | enum tree_code c1, c2; |
4a00c761 JJ |
6497 | int i; |
6498 | tree prev_type, intermediate_type; | |
6499 | enum machine_mode intermediate_mode, prev_mode; | |
6500 | optab optab3, optab4; | |
ebfd146a | 6501 | |
4a00c761 | 6502 | *multi_step_cvt = 0; |
4ef69dfc IR |
6503 | if (loop_info) |
6504 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
6505 | ||
ebfd146a IR |
6506 | switch (code) |
6507 | { | |
6508 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
6509 | /* The result of a vectorized widening operation usually requires |
6510 | two vectors (because the widened results do not fit into one vector). | |
6511 | The generated vector results would normally be expected to be | |
6512 | generated in the same order as in the original scalar computation, | |
6513 | i.e. if 8 results are generated in each vector iteration, they are | |
6514 | to be organized as follows: | |
6515 | vect1: [res1,res2,res3,res4], | |
6516 | vect2: [res5,res6,res7,res8]. | |
6517 | ||
6518 | However, in the special case that the result of the widening | |
6519 | operation is used in a reduction computation only, the order doesn't | |
6520 | matter (because when vectorizing a reduction we change the order of | |
6521 | the computation). Some targets can take advantage of this and | |
6522 | generate more efficient code. For example, targets like Altivec, | |
6523 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
6524 | generate the following vectors: | |
6525 | vect1: [res1,res3,res5,res7], | |
6526 | vect2: [res2,res4,res6,res8]. | |
6527 | ||
6528 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
6529 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
6530 | iterations in parallel). We therefore don't allow to change the | |
6531 | order of the computation in the inner-loop during outer-loop | |
6532 | vectorization. */ | |
6533 | /* TODO: Another case in which order doesn't *really* matter is when we | |
6534 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
6535 | Normally, pack_trunc performs an even/odd permute, whereas the | |
6536 | repack from an even/odd expansion would be an interleave, which | |
6537 | would be significantly simpler for e.g. AVX2. */ | |
6538 | /* In any case, in order to avoid duplicating the code below, recurse | |
6539 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
6540 | are properly set up for the caller. If we fail, we'll continue with | |
6541 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
6542 | if (vect_loop | |
6543 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
6544 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
6545 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
6546 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
6547 | code1, code2, multi_step_cvt, |
6548 | interm_types)) | |
6ae6116f | 6549 | return true; |
4a00c761 JJ |
6550 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
6551 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
6552 | break; |
6553 | ||
6ae6116f RH |
6554 | case VEC_WIDEN_MULT_EVEN_EXPR: |
6555 | /* Support the recursion induced just above. */ | |
6556 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
6557 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
6558 | break; | |
6559 | ||
36ba4aae | 6560 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
6561 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
6562 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
6563 | break; |
6564 | ||
ebfd146a | 6565 | CASE_CONVERT: |
4a00c761 JJ |
6566 | c1 = VEC_UNPACK_LO_EXPR; |
6567 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
6568 | break; |
6569 | ||
6570 | case FLOAT_EXPR: | |
4a00c761 JJ |
6571 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
6572 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
6573 | break; |
6574 | ||
6575 | case FIX_TRUNC_EXPR: | |
6576 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
6577 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
6578 | computing the operation. */ | |
6579 | return false; | |
6580 | ||
6581 | default: | |
6582 | gcc_unreachable (); | |
6583 | } | |
6584 | ||
6ae6116f | 6585 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
4a00c761 JJ |
6586 | { |
6587 | enum tree_code ctmp = c1; | |
6588 | c1 = c2; | |
6589 | c2 = ctmp; | |
6590 | } | |
6591 | ||
ebfd146a IR |
6592 | if (code == FIX_TRUNC_EXPR) |
6593 | { | |
6594 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
6595 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
6596 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
6597 | } |
6598 | else | |
6599 | { | |
6600 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
6601 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
6602 | } | |
6603 | ||
6604 | if (!optab1 || !optab2) | |
6605 | return false; | |
6606 | ||
6607 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
6608 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
6609 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
6610 | return false; |
6611 | ||
4a00c761 JJ |
6612 | *code1 = c1; |
6613 | *code2 = c2; | |
6614 | ||
6615 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
6616 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
6617 | return true; | |
6618 | ||
b8698a0f | 6619 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 6620 | types. */ |
ebfd146a | 6621 | |
4a00c761 JJ |
6622 | prev_type = vectype; |
6623 | prev_mode = vec_mode; | |
b8698a0f | 6624 | |
4a00c761 JJ |
6625 | if (!CONVERT_EXPR_CODE_P (code)) |
6626 | return false; | |
b8698a0f | 6627 | |
4a00c761 JJ |
6628 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
6629 | intermediate steps in promotion sequence. We try | |
6630 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
6631 | not. */ | |
9771b263 | 6632 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
6633 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
6634 | { | |
6635 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
6636 | intermediate_type | |
6637 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
6638 | TYPE_UNSIGNED (prev_type)); | |
6639 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); | |
6640 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
6641 | ||
6642 | if (!optab3 || !optab4 | |
6643 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
6644 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
6645 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
6646 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
6647 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
6648 | == CODE_FOR_nothing) | |
6649 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
6650 | == CODE_FOR_nothing)) | |
6651 | break; | |
ebfd146a | 6652 | |
9771b263 | 6653 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
6654 | (*multi_step_cvt)++; |
6655 | ||
6656 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
6657 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
6658 | return true; | |
6659 | ||
6660 | prev_type = intermediate_type; | |
6661 | prev_mode = intermediate_mode; | |
ebfd146a IR |
6662 | } |
6663 | ||
9771b263 | 6664 | interm_types->release (); |
4a00c761 | 6665 | return false; |
ebfd146a IR |
6666 | } |
6667 | ||
6668 | ||
6669 | /* Function supportable_narrowing_operation | |
6670 | ||
b8698a0f L |
6671 | Check whether an operation represented by the code CODE is a |
6672 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
6673 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
6674 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 6675 | |
ebfd146a | 6676 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 6677 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
6678 | the target platform directly via vector tree-codes. |
6679 | ||
6680 | Output: | |
b8698a0f L |
6681 | - CODE1 is the code of a vector operation to be used when |
6682 | vectorizing the operation, if available. | |
ebfd146a IR |
6683 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
6684 | case of multi-step conversion (like int->short->char - in that case | |
6685 | MULTI_STEP_CVT will be 1). | |
6686 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 6687 | narrowing operation (short in the above example). */ |
ebfd146a IR |
6688 | |
6689 | bool | |
6690 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 6691 | tree vectype_out, tree vectype_in, |
ebfd146a | 6692 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 6693 | vec<tree> *interm_types) |
ebfd146a IR |
6694 | { |
6695 | enum machine_mode vec_mode; | |
6696 | enum insn_code icode1; | |
6697 | optab optab1, interm_optab; | |
b690cc0f RG |
6698 | tree vectype = vectype_in; |
6699 | tree narrow_vectype = vectype_out; | |
ebfd146a | 6700 | enum tree_code c1; |
4a00c761 JJ |
6701 | tree intermediate_type; |
6702 | enum machine_mode intermediate_mode, prev_mode; | |
ebfd146a | 6703 | int i; |
4a00c761 | 6704 | bool uns; |
ebfd146a | 6705 | |
4a00c761 | 6706 | *multi_step_cvt = 0; |
ebfd146a IR |
6707 | switch (code) |
6708 | { | |
6709 | CASE_CONVERT: | |
6710 | c1 = VEC_PACK_TRUNC_EXPR; | |
6711 | break; | |
6712 | ||
6713 | case FIX_TRUNC_EXPR: | |
6714 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
6715 | break; | |
6716 | ||
6717 | case FLOAT_EXPR: | |
6718 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
6719 | tree code and optabs used for computing the operation. */ | |
6720 | return false; | |
6721 | ||
6722 | default: | |
6723 | gcc_unreachable (); | |
6724 | } | |
6725 | ||
6726 | if (code == FIX_TRUNC_EXPR) | |
6727 | /* The signedness is determined from output operand. */ | |
b690cc0f | 6728 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
6729 | else |
6730 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
6731 | ||
6732 | if (!optab1) | |
6733 | return false; | |
6734 | ||
6735 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 6736 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
6737 | return false; |
6738 | ||
4a00c761 JJ |
6739 | *code1 = c1; |
6740 | ||
6741 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
6742 | return true; | |
6743 | ||
ebfd146a IR |
6744 | /* Check if it's a multi-step conversion that can be done using intermediate |
6745 | types. */ | |
4a00c761 JJ |
6746 | prev_mode = vec_mode; |
6747 | if (code == FIX_TRUNC_EXPR) | |
6748 | uns = TYPE_UNSIGNED (vectype_out); | |
6749 | else | |
6750 | uns = TYPE_UNSIGNED (vectype); | |
6751 | ||
6752 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
6753 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
6754 | costly than signed. */ | |
6755 | if (code == FIX_TRUNC_EXPR && uns) | |
6756 | { | |
6757 | enum insn_code icode2; | |
6758 | ||
6759 | intermediate_type | |
6760 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
6761 | interm_optab | |
6762 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 6763 | if (interm_optab != unknown_optab |
4a00c761 JJ |
6764 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
6765 | && insn_data[icode1].operand[0].mode | |
6766 | == insn_data[icode2].operand[0].mode) | |
6767 | { | |
6768 | uns = false; | |
6769 | optab1 = interm_optab; | |
6770 | icode1 = icode2; | |
6771 | } | |
6772 | } | |
ebfd146a | 6773 | |
4a00c761 JJ |
6774 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
6775 | intermediate steps in promotion sequence. We try | |
6776 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 6777 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
6778 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
6779 | { | |
6780 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
6781 | intermediate_type | |
6782 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
6783 | interm_optab | |
6784 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
6785 | optab_default); | |
6786 | if (!interm_optab | |
6787 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
6788 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
6789 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
6790 | == CODE_FOR_nothing)) | |
6791 | break; | |
6792 | ||
9771b263 | 6793 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
6794 | (*multi_step_cvt)++; |
6795 | ||
6796 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
6797 | return true; | |
6798 | ||
6799 | prev_mode = intermediate_mode; | |
6800 | optab1 = interm_optab; | |
ebfd146a IR |
6801 | } |
6802 | ||
9771b263 | 6803 | interm_types->release (); |
4a00c761 | 6804 | return false; |
ebfd146a | 6805 | } |