]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
5624e564 | 2 | Copyright (C) 2003-2015 Free Software Foundation, Inc. |
b8698a0f | 3 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
4 | and Ira Rosen <irar@il.ibm.com> |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 3, or (at your option) any later | |
11 | version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
c7131fb2 | 25 | #include "backend.h" |
957060b5 AM |
26 | #include "target.h" |
27 | #include "rtl.h" | |
ebfd146a | 28 | #include "tree.h" |
c7131fb2 | 29 | #include "gimple.h" |
c7131fb2 | 30 | #include "ssa.h" |
957060b5 AM |
31 | #include "optabs-tree.h" |
32 | #include "insn-config.h" | |
33 | #include "recog.h" /* FIXME: for insn_data */ | |
34 | #include "cgraph.h" | |
957060b5 | 35 | #include "dumpfile.h" |
c7131fb2 | 36 | #include "alias.h" |
40e23961 | 37 | #include "fold-const.h" |
d8a2d370 | 38 | #include "stor-layout.h" |
2fb9a547 | 39 | #include "tree-eh.h" |
45b0be94 | 40 | #include "gimplify.h" |
5be5c238 | 41 | #include "gimple-iterator.h" |
18f429e2 | 42 | #include "gimplify-me.h" |
442b4905 | 43 | #include "tree-cfg.h" |
e28030cf | 44 | #include "tree-ssa-loop-manip.h" |
ebfd146a | 45 | #include "cfgloop.h" |
0136f8f0 AH |
46 | #include "tree-ssa-loop.h" |
47 | #include "tree-scalar-evolution.h" | |
ebfd146a | 48 | #include "tree-vectorizer.h" |
9b2b7279 | 49 | #include "builtins.h" |
ebfd146a | 50 | |
7ee2468b SB |
51 | /* For lang_hooks.types.type_for_mode. */ |
52 | #include "langhooks.h" | |
ebfd146a | 53 | |
c3e7ee41 BS |
54 | /* Return the vectorized type for the given statement. */ |
55 | ||
56 | tree | |
57 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
58 | { | |
59 | return STMT_VINFO_VECTYPE (stmt_info); | |
60 | } | |
61 | ||
62 | /* Return TRUE iff the given statement is in an inner loop relative to | |
63 | the loop being vectorized. */ | |
64 | bool | |
65 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
66 | { | |
355fe088 | 67 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
c3e7ee41 BS |
68 | basic_block bb = gimple_bb (stmt); |
69 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
70 | struct loop* loop; | |
71 | ||
72 | if (!loop_vinfo) | |
73 | return false; | |
74 | ||
75 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
76 | ||
77 | return (bb->loop_father == loop->inner); | |
78 | } | |
79 | ||
80 | /* Record the cost of a statement, either by directly informing the | |
81 | target model or by saving it in a vector for later processing. | |
82 | Return a preliminary estimate of the statement's cost. */ | |
83 | ||
84 | unsigned | |
92345349 | 85 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 86 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 87 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 88 | { |
92345349 | 89 | if (body_cost_vec) |
c3e7ee41 | 90 | { |
92345349 | 91 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
ddf56386 RB |
92 | stmt_info_for_cost si = { count, kind, |
93 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
94 | misalign }; | |
95 | body_cost_vec->safe_push (si); | |
c3e7ee41 | 96 | return (unsigned) |
92345349 | 97 | (builtin_vectorization_cost (kind, vectype, misalign) * count); |
c3e7ee41 BS |
98 | } |
99 | else | |
310213d4 RB |
100 | return add_stmt_cost (stmt_info->vinfo->target_cost_data, |
101 | count, kind, stmt_info, misalign, where); | |
c3e7ee41 BS |
102 | } |
103 | ||
272c6793 RS |
104 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
105 | ||
106 | static tree | |
107 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
108 | { | |
109 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
110 | "vect_array"); | |
111 | } | |
112 | ||
113 | /* ARRAY is an array of vectors created by create_vector_array. | |
114 | Return an SSA_NAME for the vector in index N. The reference | |
115 | is part of the vectorization of STMT and the vector is associated | |
116 | with scalar destination SCALAR_DEST. */ | |
117 | ||
118 | static tree | |
355fe088 | 119 | read_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree scalar_dest, |
272c6793 RS |
120 | tree array, unsigned HOST_WIDE_INT n) |
121 | { | |
122 | tree vect_type, vect, vect_name, array_ref; | |
355fe088 | 123 | gimple *new_stmt; |
272c6793 RS |
124 | |
125 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
126 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
127 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
128 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
129 | build_int_cst (size_type_node, n), | |
130 | NULL_TREE, NULL_TREE); | |
131 | ||
132 | new_stmt = gimple_build_assign (vect, array_ref); | |
133 | vect_name = make_ssa_name (vect, new_stmt); | |
134 | gimple_assign_set_lhs (new_stmt, vect_name); | |
135 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
136 | |
137 | return vect_name; | |
138 | } | |
139 | ||
140 | /* ARRAY is an array of vectors created by create_vector_array. | |
141 | Emit code to store SSA_NAME VECT in index N of the array. | |
142 | The store is part of the vectorization of STMT. */ | |
143 | ||
144 | static void | |
355fe088 | 145 | write_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree vect, |
272c6793 RS |
146 | tree array, unsigned HOST_WIDE_INT n) |
147 | { | |
148 | tree array_ref; | |
355fe088 | 149 | gimple *new_stmt; |
272c6793 RS |
150 | |
151 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
152 | build_int_cst (size_type_node, n), | |
153 | NULL_TREE, NULL_TREE); | |
154 | ||
155 | new_stmt = gimple_build_assign (array_ref, vect); | |
156 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
157 | } |
158 | ||
159 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
160 | of *PTR. The memory reference replaces those in FIRST_DR | |
161 | (and its group). */ | |
162 | ||
163 | static tree | |
164 | create_array_ref (tree type, tree ptr, struct data_reference *first_dr) | |
165 | { | |
272c6793 RS |
166 | tree mem_ref, alias_ptr_type; |
167 | ||
168 | alias_ptr_type = reference_alias_ptr_type (DR_REF (first_dr)); | |
169 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); | |
170 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 171 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
172 | return mem_ref; |
173 | } | |
174 | ||
ebfd146a IR |
175 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
176 | ||
177 | /* Function vect_mark_relevant. | |
178 | ||
179 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
180 | ||
181 | static void | |
355fe088 | 182 | vect_mark_relevant (vec<gimple *> *worklist, gimple *stmt, |
83197f37 IR |
183 | enum vect_relevant relevant, bool live_p, |
184 | bool used_in_pattern) | |
ebfd146a IR |
185 | { |
186 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
187 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
188 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
355fe088 | 189 | gimple *pattern_stmt; |
ebfd146a | 190 | |
73fbfcad | 191 | if (dump_enabled_p ()) |
78c60e3d | 192 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 193 | "mark relevant %d, live %d.\n", relevant, live_p); |
ebfd146a | 194 | |
83197f37 IR |
195 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
196 | related pattern stmt instead of the original stmt. However, such stmts | |
197 | may have their own uses that are not in any pattern, in such cases the | |
198 | stmt itself should be marked. */ | |
ebfd146a IR |
199 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
200 | { | |
83197f37 IR |
201 | bool found = false; |
202 | if (!used_in_pattern) | |
203 | { | |
204 | imm_use_iterator imm_iter; | |
205 | use_operand_p use_p; | |
355fe088 | 206 | gimple *use_stmt; |
83197f37 | 207 | tree lhs; |
13c931c9 JJ |
208 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
209 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a | 210 | |
83197f37 IR |
211 | if (is_gimple_assign (stmt)) |
212 | lhs = gimple_assign_lhs (stmt); | |
213 | else | |
214 | lhs = gimple_call_lhs (stmt); | |
ebfd146a | 215 | |
83197f37 IR |
216 | /* This use is out of pattern use, if LHS has other uses that are |
217 | pattern uses, we should mark the stmt itself, and not the pattern | |
218 | stmt. */ | |
5ce9450f | 219 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
ab0ef706 JJ |
220 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs) |
221 | { | |
222 | if (is_gimple_debug (USE_STMT (use_p))) | |
223 | continue; | |
224 | use_stmt = USE_STMT (use_p); | |
225 | ||
13c931c9 JJ |
226 | if (!flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) |
227 | continue; | |
228 | ||
ab0ef706 JJ |
229 | if (vinfo_for_stmt (use_stmt) |
230 | && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (use_stmt))) | |
231 | { | |
232 | found = true; | |
233 | break; | |
234 | } | |
235 | } | |
83197f37 IR |
236 | } |
237 | ||
238 | if (!found) | |
239 | { | |
240 | /* This is the last stmt in a sequence that was detected as a | |
241 | pattern that can potentially be vectorized. Don't mark the stmt | |
242 | as relevant/live because it's not going to be vectorized. | |
243 | Instead mark the pattern-stmt that replaces it. */ | |
244 | ||
245 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
246 | ||
73fbfcad | 247 | if (dump_enabled_p ()) |
78c60e3d SS |
248 | dump_printf_loc (MSG_NOTE, vect_location, |
249 | "last stmt in pattern. don't mark" | |
e645e942 | 250 | " relevant/live.\n"); |
83197f37 IR |
251 | stmt_info = vinfo_for_stmt (pattern_stmt); |
252 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
253 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
254 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
255 | stmt = pattern_stmt; | |
256 | } | |
ebfd146a IR |
257 | } |
258 | ||
259 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
260 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
261 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
262 | ||
263 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
264 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
265 | { | |
73fbfcad | 266 | if (dump_enabled_p ()) |
78c60e3d | 267 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 268 | "already marked relevant/live.\n"); |
ebfd146a IR |
269 | return; |
270 | } | |
271 | ||
9771b263 | 272 | worklist->safe_push (stmt); |
ebfd146a IR |
273 | } |
274 | ||
275 | ||
276 | /* Function vect_stmt_relevant_p. | |
277 | ||
278 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
279 | "relevant for vectorization". | |
280 | ||
281 | A stmt is considered "relevant for vectorization" if: | |
282 | - it has uses outside the loop. | |
283 | - it has vdefs (it alters memory). | |
284 | - control stmts in the loop (except for the exit condition). | |
285 | ||
286 | CHECKME: what other side effects would the vectorizer allow? */ | |
287 | ||
288 | static bool | |
355fe088 | 289 | vect_stmt_relevant_p (gimple *stmt, loop_vec_info loop_vinfo, |
ebfd146a IR |
290 | enum vect_relevant *relevant, bool *live_p) |
291 | { | |
292 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
293 | ssa_op_iter op_iter; | |
294 | imm_use_iterator imm_iter; | |
295 | use_operand_p use_p; | |
296 | def_operand_p def_p; | |
297 | ||
8644a673 | 298 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
299 | *live_p = false; |
300 | ||
301 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
302 | if (is_ctrl_stmt (stmt) |
303 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
304 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 305 | *relevant = vect_used_in_scope; |
ebfd146a IR |
306 | |
307 | /* changing memory. */ | |
308 | if (gimple_code (stmt) != GIMPLE_PHI) | |
ac6aeab4 RB |
309 | if (gimple_vdef (stmt) |
310 | && !gimple_clobber_p (stmt)) | |
ebfd146a | 311 | { |
73fbfcad | 312 | if (dump_enabled_p ()) |
78c60e3d | 313 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 314 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
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 | 327 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 328 | "vec_stmt_relevant_p: used out of loop.\n"); |
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 | |
355fe088 | 353 | exist_non_indexing_operands_for_use_p (tree use, gimple *stmt) |
ebfd146a IR |
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)) | |
5ce9450f JJ |
378 | { |
379 | if (is_gimple_call (stmt) | |
380 | && gimple_call_internal_p (stmt)) | |
381 | switch (gimple_call_internal_fn (stmt)) | |
382 | { | |
383 | case IFN_MASK_STORE: | |
384 | operand = gimple_call_arg (stmt, 3); | |
385 | if (operand == use) | |
386 | return true; | |
387 | /* FALLTHRU */ | |
388 | case IFN_MASK_LOAD: | |
389 | operand = gimple_call_arg (stmt, 2); | |
390 | if (operand == use) | |
391 | return true; | |
392 | break; | |
393 | default: | |
394 | break; | |
395 | } | |
396 | return false; | |
397 | } | |
398 | ||
59a05b0c EB |
399 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
400 | return false; | |
ebfd146a | 401 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
402 | if (TREE_CODE (operand) != SSA_NAME) |
403 | return false; | |
404 | ||
405 | if (operand == use) | |
406 | return true; | |
407 | ||
408 | return false; | |
409 | } | |
410 | ||
411 | ||
b8698a0f | 412 | /* |
ebfd146a IR |
413 | Function process_use. |
414 | ||
415 | Inputs: | |
416 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b8698a0f | 417 | - LIVE_P, RELEVANT - enum values to be set in the STMT_VINFO of the stmt |
ff802fa1 | 418 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 419 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
420 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
421 | be performed. | |
ebfd146a IR |
422 | |
423 | Outputs: | |
424 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
425 | relevance info of the DEF_STMT of this USE: | |
426 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
427 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
428 | Exceptions: | |
429 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 430 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 431 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
432 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
433 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
434 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
435 | be modified accordingly. | |
436 | ||
437 | Return true if everything is as expected. Return false otherwise. */ | |
438 | ||
439 | static bool | |
355fe088 TS |
440 | process_use (gimple *stmt, tree use, loop_vec_info loop_vinfo, bool live_p, |
441 | enum vect_relevant relevant, vec<gimple *> *worklist, | |
aec7ae7d | 442 | bool force) |
ebfd146a IR |
443 | { |
444 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
445 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
446 | stmt_vec_info dstmt_vinfo; | |
447 | basic_block bb, def_bb; | |
355fe088 | 448 | gimple *def_stmt; |
ebfd146a IR |
449 | enum vect_def_type dt; |
450 | ||
b8698a0f | 451 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 452 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 453 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
454 | return true; |
455 | ||
81c40241 | 456 | if (!vect_is_simple_use (use, loop_vinfo, &def_stmt, &dt)) |
b8698a0f | 457 | { |
73fbfcad | 458 | if (dump_enabled_p ()) |
78c60e3d | 459 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 460 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
461 | return false; |
462 | } | |
463 | ||
464 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
465 | return true; | |
466 | ||
467 | def_bb = gimple_bb (def_stmt); | |
468 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
469 | { | |
73fbfcad | 470 | if (dump_enabled_p ()) |
e645e942 | 471 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n"); |
ebfd146a IR |
472 | return true; |
473 | } | |
474 | ||
b8698a0f L |
475 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
476 | DEF_STMT must have already been processed, because this should be the | |
477 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
478 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
479 | check that everything is as expected, and we are done. */ |
480 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
481 | bb = gimple_bb (stmt); | |
482 | if (gimple_code (stmt) == GIMPLE_PHI | |
483 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
484 | && gimple_code (def_stmt) != GIMPLE_PHI | |
485 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
486 | && bb->loop_father == def_bb->loop_father) | |
487 | { | |
73fbfcad | 488 | if (dump_enabled_p ()) |
78c60e3d | 489 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 490 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a IR |
491 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
492 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
493 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 494 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 495 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
496 | return true; |
497 | } | |
498 | ||
499 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
500 | outer-loop-header-bb: | |
501 | d = def_stmt | |
502 | inner-loop: | |
503 | stmt # use (d) | |
504 | outer-loop-tail-bb: | |
505 | ... */ | |
506 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
507 | { | |
73fbfcad | 508 | if (dump_enabled_p ()) |
78c60e3d | 509 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 510 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 511 | |
ebfd146a IR |
512 | switch (relevant) |
513 | { | |
8644a673 | 514 | case vect_unused_in_scope: |
7c5222ff IR |
515 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
516 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 517 | break; |
7c5222ff | 518 | |
ebfd146a | 519 | case vect_used_in_outer_by_reduction: |
7c5222ff | 520 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
521 | relevant = vect_used_by_reduction; |
522 | break; | |
7c5222ff | 523 | |
ebfd146a | 524 | case vect_used_in_outer: |
7c5222ff | 525 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 526 | relevant = vect_used_in_scope; |
ebfd146a | 527 | break; |
7c5222ff | 528 | |
8644a673 | 529 | case vect_used_in_scope: |
ebfd146a IR |
530 | break; |
531 | ||
532 | default: | |
533 | gcc_unreachable (); | |
b8698a0f | 534 | } |
ebfd146a IR |
535 | } |
536 | ||
537 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
538 | outer-loop-header-bb: | |
539 | ... | |
540 | inner-loop: | |
541 | d = def_stmt | |
06066f92 | 542 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
543 | stmt # use (d) */ |
544 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
545 | { | |
73fbfcad | 546 | if (dump_enabled_p ()) |
78c60e3d | 547 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 548 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 549 | |
ebfd146a IR |
550 | switch (relevant) |
551 | { | |
8644a673 | 552 | case vect_unused_in_scope: |
b8698a0f | 553 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 554 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 555 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
556 | break; |
557 | ||
ebfd146a IR |
558 | case vect_used_by_reduction: |
559 | relevant = vect_used_in_outer_by_reduction; | |
560 | break; | |
561 | ||
8644a673 | 562 | case vect_used_in_scope: |
ebfd146a IR |
563 | relevant = vect_used_in_outer; |
564 | break; | |
565 | ||
566 | default: | |
567 | gcc_unreachable (); | |
568 | } | |
569 | } | |
570 | ||
83197f37 IR |
571 | vect_mark_relevant (worklist, def_stmt, relevant, live_p, |
572 | is_pattern_stmt_p (stmt_vinfo)); | |
ebfd146a IR |
573 | return true; |
574 | } | |
575 | ||
576 | ||
577 | /* Function vect_mark_stmts_to_be_vectorized. | |
578 | ||
579 | Not all stmts in the loop need to be vectorized. For example: | |
580 | ||
581 | for i... | |
582 | for j... | |
583 | 1. T0 = i + j | |
584 | 2. T1 = a[T0] | |
585 | ||
586 | 3. j = j + 1 | |
587 | ||
588 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
589 | addressing of vectorized data-refs are handled differently. | |
590 | ||
591 | This pass detects such stmts. */ | |
592 | ||
593 | bool | |
594 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
595 | { | |
ebfd146a IR |
596 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
597 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
598 | unsigned int nbbs = loop->num_nodes; | |
599 | gimple_stmt_iterator si; | |
355fe088 | 600 | gimple *stmt; |
ebfd146a IR |
601 | unsigned int i; |
602 | stmt_vec_info stmt_vinfo; | |
603 | basic_block bb; | |
355fe088 | 604 | gimple *phi; |
ebfd146a | 605 | bool live_p; |
06066f92 IR |
606 | enum vect_relevant relevant, tmp_relevant; |
607 | enum vect_def_type def_type; | |
ebfd146a | 608 | |
73fbfcad | 609 | if (dump_enabled_p ()) |
78c60e3d | 610 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 611 | "=== vect_mark_stmts_to_be_vectorized ===\n"); |
ebfd146a | 612 | |
355fe088 | 613 | auto_vec<gimple *, 64> worklist; |
ebfd146a IR |
614 | |
615 | /* 1. Init worklist. */ | |
616 | for (i = 0; i < nbbs; i++) | |
617 | { | |
618 | bb = bbs[i]; | |
619 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 620 | { |
ebfd146a | 621 | phi = gsi_stmt (si); |
73fbfcad | 622 | if (dump_enabled_p ()) |
ebfd146a | 623 | { |
78c60e3d SS |
624 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
625 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
626 | } |
627 | ||
628 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 629 | vect_mark_relevant (&worklist, phi, relevant, live_p, false); |
ebfd146a IR |
630 | } |
631 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
632 | { | |
633 | stmt = gsi_stmt (si); | |
73fbfcad | 634 | if (dump_enabled_p ()) |
ebfd146a | 635 | { |
78c60e3d SS |
636 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
637 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
b8698a0f | 638 | } |
ebfd146a IR |
639 | |
640 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
83197f37 | 641 | vect_mark_relevant (&worklist, stmt, relevant, live_p, false); |
ebfd146a IR |
642 | } |
643 | } | |
644 | ||
645 | /* 2. Process_worklist */ | |
9771b263 | 646 | while (worklist.length () > 0) |
ebfd146a IR |
647 | { |
648 | use_operand_p use_p; | |
649 | ssa_op_iter iter; | |
650 | ||
9771b263 | 651 | stmt = worklist.pop (); |
73fbfcad | 652 | if (dump_enabled_p ()) |
ebfd146a | 653 | { |
78c60e3d SS |
654 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
655 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
ebfd146a IR |
656 | } |
657 | ||
b8698a0f L |
658 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
659 | (DEF_STMT) as relevant/irrelevant and live/dead according to the | |
ebfd146a IR |
660 | liveness and relevance properties of STMT. */ |
661 | stmt_vinfo = vinfo_for_stmt (stmt); | |
662 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
663 | live_p = STMT_VINFO_LIVE_P (stmt_vinfo); | |
664 | ||
665 | /* Generally, the liveness and relevance properties of STMT are | |
666 | propagated as is to the DEF_STMTs of its USEs: | |
667 | live_p <-- STMT_VINFO_LIVE_P (STMT_VINFO) | |
668 | relevant <-- STMT_VINFO_RELEVANT (STMT_VINFO) | |
669 | ||
670 | One exception is when STMT has been identified as defining a reduction | |
671 | variable; in this case we set the liveness/relevance as follows: | |
672 | live_p = false | |
673 | relevant = vect_used_by_reduction | |
674 | This is because we distinguish between two kinds of relevant stmts - | |
b8698a0f | 675 | those that are used by a reduction computation, and those that are |
ff802fa1 | 676 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 677 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 678 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 679 | |
06066f92 IR |
680 | def_type = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
681 | tmp_relevant = relevant; | |
682 | switch (def_type) | |
ebfd146a | 683 | { |
06066f92 IR |
684 | case vect_reduction_def: |
685 | switch (tmp_relevant) | |
686 | { | |
687 | case vect_unused_in_scope: | |
688 | relevant = vect_used_by_reduction; | |
689 | break; | |
690 | ||
691 | case vect_used_by_reduction: | |
692 | if (gimple_code (stmt) == GIMPLE_PHI) | |
693 | break; | |
694 | /* fall through */ | |
695 | ||
696 | default: | |
73fbfcad | 697 | if (dump_enabled_p ()) |
78c60e3d | 698 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 699 | "unsupported use of reduction.\n"); |
06066f92 IR |
700 | return false; |
701 | } | |
702 | ||
b8698a0f | 703 | live_p = false; |
06066f92 | 704 | break; |
b8698a0f | 705 | |
06066f92 IR |
706 | case vect_nested_cycle: |
707 | if (tmp_relevant != vect_unused_in_scope | |
708 | && tmp_relevant != vect_used_in_outer_by_reduction | |
709 | && tmp_relevant != vect_used_in_outer) | |
710 | { | |
73fbfcad | 711 | if (dump_enabled_p ()) |
78c60e3d | 712 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 713 | "unsupported use of nested cycle.\n"); |
7c5222ff | 714 | |
06066f92 IR |
715 | return false; |
716 | } | |
7c5222ff | 717 | |
b8698a0f L |
718 | live_p = false; |
719 | break; | |
720 | ||
06066f92 IR |
721 | case vect_double_reduction_def: |
722 | if (tmp_relevant != vect_unused_in_scope | |
723 | && tmp_relevant != vect_used_by_reduction) | |
724 | { | |
73fbfcad | 725 | if (dump_enabled_p ()) |
78c60e3d | 726 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 727 | "unsupported use of double reduction.\n"); |
7c5222ff | 728 | |
7c5222ff | 729 | return false; |
06066f92 IR |
730 | } |
731 | ||
732 | live_p = false; | |
b8698a0f | 733 | break; |
7c5222ff | 734 | |
06066f92 IR |
735 | default: |
736 | break; | |
7c5222ff | 737 | } |
b8698a0f | 738 | |
aec7ae7d | 739 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
740 | { |
741 | /* Pattern statements are not inserted into the code, so | |
742 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
743 | have to scan the RHS or function arguments instead. */ | |
744 | if (is_gimple_assign (stmt)) | |
745 | { | |
69d2aade JJ |
746 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
747 | tree op = gimple_assign_rhs1 (stmt); | |
748 | ||
749 | i = 1; | |
750 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
751 | { | |
752 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
aec7ae7d | 753 | live_p, relevant, &worklist, false) |
69d2aade | 754 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
aec7ae7d | 755 | live_p, relevant, &worklist, false)) |
566d377a | 756 | return false; |
69d2aade JJ |
757 | i = 2; |
758 | } | |
759 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 760 | { |
69d2aade | 761 | op = gimple_op (stmt, i); |
afbe6325 RB |
762 | if (TREE_CODE (op) == SSA_NAME |
763 | && !process_use (stmt, op, loop_vinfo, live_p, relevant, | |
764 | &worklist, false)) | |
07687835 | 765 | return false; |
9d5e7640 IR |
766 | } |
767 | } | |
768 | else if (is_gimple_call (stmt)) | |
769 | { | |
770 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
771 | { | |
772 | tree arg = gimple_call_arg (stmt, i); | |
773 | if (!process_use (stmt, arg, loop_vinfo, live_p, relevant, | |
aec7ae7d | 774 | &worklist, false)) |
07687835 | 775 | return false; |
9d5e7640 IR |
776 | } |
777 | } | |
778 | } | |
779 | else | |
780 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
781 | { | |
782 | tree op = USE_FROM_PTR (use_p); | |
783 | if (!process_use (stmt, op, loop_vinfo, live_p, relevant, | |
aec7ae7d | 784 | &worklist, false)) |
07687835 | 785 | return false; |
9d5e7640 | 786 | } |
aec7ae7d | 787 | |
3bab6342 | 788 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo)) |
aec7ae7d JJ |
789 | { |
790 | tree off; | |
3bab6342 | 791 | tree decl = vect_check_gather_scatter (stmt, loop_vinfo, NULL, &off, NULL); |
aec7ae7d JJ |
792 | gcc_assert (decl); |
793 | if (!process_use (stmt, off, loop_vinfo, live_p, relevant, | |
794 | &worklist, true)) | |
566d377a | 795 | return false; |
aec7ae7d | 796 | } |
ebfd146a IR |
797 | } /* while worklist */ |
798 | ||
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, " | |
e645e942 | 835 | "prologue_cost = %d .\n", 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, " | |
e645e942 | 881 | "prologue_cost = %d .\n", 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 | { |
355fe088 | 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 | 916 | struct data_reference *first_dr; |
355fe088 | 917 | gimple *first_stmt; |
ebfd146a | 918 | |
8644a673 | 919 | if (dt == vect_constant_def || dt == vect_external_def) |
92345349 BS |
920 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
921 | stmt_info, 0, vect_prologue); | |
ebfd146a | 922 | |
0d0293ac MM |
923 | /* Grouped access? */ |
924 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
720f5239 IR |
925 | { |
926 | if (slp_node) | |
927 | { | |
9771b263 | 928 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
720f5239 IR |
929 | group_size = 1; |
930 | } | |
931 | else | |
932 | { | |
e14c1050 | 933 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 934 | group_size = vect_cost_group_size (stmt_info); |
720f5239 IR |
935 | } |
936 | ||
937 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
938 | } | |
0d0293ac | 939 | /* Not a grouped access. */ |
ebfd146a | 940 | else |
720f5239 IR |
941 | { |
942 | group_size = 1; | |
943 | first_dr = STMT_VINFO_DATA_REF (stmt_info); | |
944 | } | |
ebfd146a | 945 | |
272c6793 | 946 | /* We assume that the cost of a single store-lanes instruction is |
0d0293ac | 947 | equivalent to the cost of GROUP_SIZE separate stores. If a grouped |
272c6793 RS |
948 | access is instead being provided by a permute-and-store operation, |
949 | include the cost of the permutes. */ | |
cee62fee MM |
950 | if (!store_lanes_p && group_size > 1 |
951 | && !STMT_VINFO_STRIDED_P (stmt_info)) | |
ebfd146a | 952 | { |
e1377713 ES |
953 | /* Uses a high and low interleave or shuffle operations for each |
954 | needed permute. */ | |
955 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; | |
92345349 BS |
956 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
957 | stmt_info, 0, vect_body); | |
ebfd146a | 958 | |
73fbfcad | 959 | if (dump_enabled_p ()) |
78c60e3d | 960 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 961 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 962 | group_size); |
ebfd146a IR |
963 | } |
964 | ||
cee62fee | 965 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a | 966 | /* Costs of the stores. */ |
cee62fee MM |
967 | if (STMT_VINFO_STRIDED_P (stmt_info) |
968 | && !STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
f2e2a985 MM |
969 | { |
970 | /* N scalar stores plus extracting the elements. */ | |
f2e2a985 MM |
971 | inside_cost += record_stmt_cost (body_cost_vec, |
972 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), | |
973 | scalar_store, stmt_info, 0, vect_body); | |
f2e2a985 MM |
974 | } |
975 | else | |
976 | vect_get_store_cost (first_dr, ncopies, &inside_cost, body_cost_vec); | |
ebfd146a | 977 | |
cee62fee MM |
978 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
979 | inside_cost += record_stmt_cost (body_cost_vec, | |
980 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), | |
981 | vec_to_scalar, stmt_info, 0, vect_body); | |
982 | ||
73fbfcad | 983 | if (dump_enabled_p ()) |
78c60e3d SS |
984 | dump_printf_loc (MSG_NOTE, vect_location, |
985 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 986 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
987 | } |
988 | ||
989 | ||
720f5239 IR |
990 | /* Calculate cost of DR's memory access. */ |
991 | void | |
992 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 993 | unsigned int *inside_cost, |
92345349 | 994 | stmt_vector_for_cost *body_cost_vec) |
720f5239 IR |
995 | { |
996 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 997 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 998 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
999 | |
1000 | switch (alignment_support_scheme) | |
1001 | { | |
1002 | case dr_aligned: | |
1003 | { | |
92345349 BS |
1004 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
1005 | vector_store, stmt_info, 0, | |
1006 | vect_body); | |
720f5239 | 1007 | |
73fbfcad | 1008 | if (dump_enabled_p ()) |
78c60e3d | 1009 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1010 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
1011 | break; |
1012 | } | |
1013 | ||
1014 | case dr_unaligned_supported: | |
1015 | { | |
720f5239 | 1016 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1017 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1018 | unaligned_store, stmt_info, |
92345349 | 1019 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1020 | if (dump_enabled_p ()) |
78c60e3d SS |
1021 | dump_printf_loc (MSG_NOTE, vect_location, |
1022 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1023 | "hardware.\n"); |
720f5239 IR |
1024 | break; |
1025 | } | |
1026 | ||
38eec4c6 UW |
1027 | case dr_unaligned_unsupported: |
1028 | { | |
1029 | *inside_cost = VECT_MAX_COST; | |
1030 | ||
73fbfcad | 1031 | if (dump_enabled_p ()) |
78c60e3d | 1032 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1033 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1034 | break; |
1035 | } | |
1036 | ||
720f5239 IR |
1037 | default: |
1038 | gcc_unreachable (); | |
1039 | } | |
1040 | } | |
1041 | ||
1042 | ||
ebfd146a IR |
1043 | /* Function vect_model_load_cost |
1044 | ||
0d0293ac MM |
1045 | Models cost for loads. In the case of grouped accesses, the last access |
1046 | has the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1047 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1048 | access scheme chosen. */ |
1049 | ||
1050 | void | |
92345349 BS |
1051 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, |
1052 | bool load_lanes_p, slp_tree slp_node, | |
1053 | stmt_vector_for_cost *prologue_cost_vec, | |
1054 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
1055 | { |
1056 | int group_size; | |
355fe088 | 1057 | gimple *first_stmt; |
ebfd146a | 1058 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; |
92345349 | 1059 | unsigned int inside_cost = 0, prologue_cost = 0; |
ebfd146a | 1060 | |
0d0293ac | 1061 | /* Grouped accesses? */ |
e14c1050 | 1062 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
0d0293ac | 1063 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && first_stmt && !slp_node) |
ebfd146a | 1064 | { |
0d0293ac | 1065 | group_size = vect_cost_group_size (stmt_info); |
ebfd146a IR |
1066 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
1067 | } | |
0d0293ac | 1068 | /* Not a grouped access. */ |
ebfd146a IR |
1069 | else |
1070 | { | |
1071 | group_size = 1; | |
1072 | first_dr = dr; | |
1073 | } | |
1074 | ||
272c6793 | 1075 | /* We assume that the cost of a single load-lanes instruction is |
0d0293ac | 1076 | equivalent to the cost of GROUP_SIZE separate loads. If a grouped |
272c6793 RS |
1077 | access is instead being provided by a load-and-permute operation, |
1078 | include the cost of the permutes. */ | |
7b5fc413 | 1079 | if (!load_lanes_p && group_size > 1 |
f2e2a985 | 1080 | && !STMT_VINFO_STRIDED_P (stmt_info)) |
ebfd146a | 1081 | { |
2c23db6d ES |
1082 | /* Uses an even and odd extract operations or shuffle operations |
1083 | for each needed permute. */ | |
1084 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; | |
1085 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, | |
1086 | stmt_info, 0, vect_body); | |
ebfd146a | 1087 | |
73fbfcad | 1088 | if (dump_enabled_p ()) |
e645e942 TJ |
1089 | dump_printf_loc (MSG_NOTE, vect_location, |
1090 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1091 | group_size); |
ebfd146a IR |
1092 | } |
1093 | ||
1094 | /* The loads themselves. */ | |
f2e2a985 | 1095 | if (STMT_VINFO_STRIDED_P (stmt_info) |
7b5fc413 | 1096 | && !STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
a82960aa | 1097 | { |
a21892ad BS |
1098 | /* N scalar loads plus gathering them into a vector. */ |
1099 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
92345349 | 1100 | inside_cost += record_stmt_cost (body_cost_vec, |
c3e7ee41 | 1101 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), |
92345349 | 1102 | scalar_load, stmt_info, 0, vect_body); |
a82960aa RG |
1103 | } |
1104 | else | |
1105 | vect_get_load_cost (first_dr, ncopies, | |
1106 | ((!STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1107 | || group_size > 1 || slp_node), | |
92345349 BS |
1108 | &inside_cost, &prologue_cost, |
1109 | prologue_cost_vec, body_cost_vec, true); | |
f2e2a985 | 1110 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
7b5fc413 RB |
1111 | inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct, |
1112 | stmt_info, 0, vect_body); | |
720f5239 | 1113 | |
73fbfcad | 1114 | if (dump_enabled_p ()) |
78c60e3d SS |
1115 | dump_printf_loc (MSG_NOTE, vect_location, |
1116 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1117 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1118 | } |
1119 | ||
1120 | ||
1121 | /* Calculate cost of DR's memory access. */ | |
1122 | void | |
1123 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1124 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1125 | unsigned int *prologue_cost, |
1126 | stmt_vector_for_cost *prologue_cost_vec, | |
1127 | stmt_vector_for_cost *body_cost_vec, | |
1128 | bool record_prologue_costs) | |
720f5239 IR |
1129 | { |
1130 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 1131 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 1132 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
1133 | |
1134 | switch (alignment_support_scheme) | |
ebfd146a IR |
1135 | { |
1136 | case dr_aligned: | |
1137 | { | |
92345349 BS |
1138 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1139 | stmt_info, 0, vect_body); | |
ebfd146a | 1140 | |
73fbfcad | 1141 | if (dump_enabled_p ()) |
78c60e3d | 1142 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1143 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1144 | |
1145 | break; | |
1146 | } | |
1147 | case dr_unaligned_supported: | |
1148 | { | |
720f5239 | 1149 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1150 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1151 | unaligned_load, stmt_info, |
92345349 | 1152 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1153 | |
73fbfcad | 1154 | if (dump_enabled_p ()) |
78c60e3d SS |
1155 | dump_printf_loc (MSG_NOTE, vect_location, |
1156 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1157 | "hardware.\n"); |
ebfd146a IR |
1158 | |
1159 | break; | |
1160 | } | |
1161 | case dr_explicit_realign: | |
1162 | { | |
92345349 BS |
1163 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1164 | vector_load, stmt_info, 0, vect_body); | |
1165 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1166 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1167 | |
1168 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1169 | the containing loop, the following cost should be added to the | |
92345349 | 1170 | prologue costs. */ |
ebfd146a | 1171 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1172 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1173 | stmt_info, 0, vect_body); | |
ebfd146a | 1174 | |
73fbfcad | 1175 | if (dump_enabled_p ()) |
e645e942 TJ |
1176 | dump_printf_loc (MSG_NOTE, vect_location, |
1177 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1178 | |
ebfd146a IR |
1179 | break; |
1180 | } | |
1181 | case dr_explicit_realign_optimized: | |
1182 | { | |
73fbfcad | 1183 | if (dump_enabled_p ()) |
e645e942 | 1184 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1185 | "vect_model_load_cost: unaligned software " |
e645e942 | 1186 | "pipelined.\n"); |
ebfd146a IR |
1187 | |
1188 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1189 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1190 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1191 | access, then the above cost should only be considered for one |
ff802fa1 | 1192 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1193 | and a realignment op. */ |
1194 | ||
92345349 | 1195 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1196 | { |
92345349 BS |
1197 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1198 | vector_stmt, stmt_info, | |
1199 | 0, vect_prologue); | |
ebfd146a | 1200 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1201 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1202 | vector_stmt, stmt_info, | |
1203 | 0, vect_prologue); | |
ebfd146a IR |
1204 | } |
1205 | ||
92345349 BS |
1206 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1207 | stmt_info, 0, vect_body); | |
1208 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1209 | stmt_info, 0, vect_body); | |
8bd37302 | 1210 | |
73fbfcad | 1211 | if (dump_enabled_p ()) |
78c60e3d | 1212 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1213 | "vect_model_load_cost: explicit realign optimized" |
1214 | "\n"); | |
8bd37302 | 1215 | |
ebfd146a IR |
1216 | break; |
1217 | } | |
1218 | ||
38eec4c6 UW |
1219 | case dr_unaligned_unsupported: |
1220 | { | |
1221 | *inside_cost = VECT_MAX_COST; | |
1222 | ||
73fbfcad | 1223 | if (dump_enabled_p ()) |
78c60e3d | 1224 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1225 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1226 | break; |
1227 | } | |
1228 | ||
ebfd146a IR |
1229 | default: |
1230 | gcc_unreachable (); | |
1231 | } | |
ebfd146a IR |
1232 | } |
1233 | ||
418b7df3 RG |
1234 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1235 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1236 | |
418b7df3 | 1237 | static void |
355fe088 | 1238 | vect_init_vector_1 (gimple *stmt, gimple *new_stmt, gimple_stmt_iterator *gsi) |
ebfd146a | 1239 | { |
ebfd146a | 1240 | if (gsi) |
418b7df3 | 1241 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1242 | else |
1243 | { | |
418b7df3 | 1244 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1245 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1246 | |
a70d6342 IR |
1247 | if (loop_vinfo) |
1248 | { | |
1249 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1250 | basic_block new_bb; |
1251 | edge pe; | |
a70d6342 IR |
1252 | |
1253 | if (nested_in_vect_loop_p (loop, stmt)) | |
1254 | loop = loop->inner; | |
b8698a0f | 1255 | |
a70d6342 | 1256 | pe = loop_preheader_edge (loop); |
418b7df3 | 1257 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1258 | gcc_assert (!new_bb); |
1259 | } | |
1260 | else | |
1261 | { | |
1262 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1263 | basic_block bb; | |
1264 | gimple_stmt_iterator gsi_bb_start; | |
1265 | ||
1266 | gcc_assert (bb_vinfo); | |
1267 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1268 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1269 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1270 | } |
ebfd146a IR |
1271 | } |
1272 | ||
73fbfcad | 1273 | if (dump_enabled_p ()) |
ebfd146a | 1274 | { |
78c60e3d SS |
1275 | dump_printf_loc (MSG_NOTE, vect_location, |
1276 | "created new init_stmt: "); | |
1277 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
ebfd146a | 1278 | } |
418b7df3 RG |
1279 | } |
1280 | ||
1281 | /* Function vect_init_vector. | |
ebfd146a | 1282 | |
5467ee52 RG |
1283 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1284 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1285 | vector type a vector with all elements equal to VAL is created first. | |
1286 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1287 | initialization at the loop preheader. | |
418b7df3 RG |
1288 | Return the DEF of INIT_STMT. |
1289 | It will be used in the vectorization of STMT. */ | |
1290 | ||
1291 | tree | |
355fe088 | 1292 | vect_init_vector (gimple *stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 | 1293 | { |
355fe088 | 1294 | gimple *init_stmt; |
418b7df3 RG |
1295 | tree new_temp; |
1296 | ||
5467ee52 RG |
1297 | if (TREE_CODE (type) == VECTOR_TYPE |
1298 | && TREE_CODE (TREE_TYPE (val)) != VECTOR_TYPE) | |
418b7df3 | 1299 | { |
5467ee52 | 1300 | if (!types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) |
418b7df3 | 1301 | { |
5467ee52 | 1302 | if (CONSTANT_CLASS_P (val)) |
42fd8198 | 1303 | val = fold_convert (TREE_TYPE (type), val); |
418b7df3 RG |
1304 | else |
1305 | { | |
b731b390 | 1306 | new_temp = make_ssa_name (TREE_TYPE (type)); |
0d0e4a03 | 1307 | init_stmt = gimple_build_assign (new_temp, NOP_EXPR, val); |
418b7df3 | 1308 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1309 | val = new_temp; |
418b7df3 RG |
1310 | } |
1311 | } | |
5467ee52 | 1312 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1313 | } |
1314 | ||
0e22bb5a RB |
1315 | new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_"); |
1316 | init_stmt = gimple_build_assign (new_temp, val); | |
418b7df3 | 1317 | vect_init_vector_1 (stmt, init_stmt, gsi); |
0e22bb5a | 1318 | return new_temp; |
ebfd146a IR |
1319 | } |
1320 | ||
a70d6342 | 1321 | |
ebfd146a IR |
1322 | /* Function vect_get_vec_def_for_operand. |
1323 | ||
ff802fa1 | 1324 | OP is an operand in STMT. This function returns a (vector) def that will be |
ebfd146a IR |
1325 | used in the vectorized stmt for STMT. |
1326 | ||
1327 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1328 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1329 | ||
1330 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
42fd8198 IE |
1331 | needs to be introduced. VECTYPE may be used to specify a required type for |
1332 | vector invariant. */ | |
ebfd146a IR |
1333 | |
1334 | tree | |
42fd8198 | 1335 | vect_get_vec_def_for_operand (tree op, gimple *stmt, tree vectype) |
ebfd146a IR |
1336 | { |
1337 | tree vec_oprnd; | |
355fe088 TS |
1338 | gimple *vec_stmt; |
1339 | gimple *def_stmt; | |
ebfd146a IR |
1340 | stmt_vec_info def_stmt_info = NULL; |
1341 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
42fd8198 | 1342 | tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo); |
ebfd146a | 1343 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
ebfd146a IR |
1344 | enum vect_def_type dt; |
1345 | bool is_simple_use; | |
1346 | tree vector_type; | |
1347 | ||
73fbfcad | 1348 | if (dump_enabled_p ()) |
ebfd146a | 1349 | { |
78c60e3d SS |
1350 | dump_printf_loc (MSG_NOTE, vect_location, |
1351 | "vect_get_vec_def_for_operand: "); | |
1352 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
e645e942 | 1353 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1354 | } |
1355 | ||
81c40241 | 1356 | is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt); |
ebfd146a | 1357 | gcc_assert (is_simple_use); |
73fbfcad | 1358 | if (dump_enabled_p ()) |
ebfd146a | 1359 | { |
78c60e3d | 1360 | int loc_printed = 0; |
ebfd146a IR |
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 | { | |
81c40241 | 1373 | /* operand is a constant or a loop invariant. */ |
ebfd146a | 1374 | case vect_constant_def: |
81c40241 | 1375 | case vect_external_def: |
ebfd146a | 1376 | { |
42fd8198 IE |
1377 | if (vectype) |
1378 | vector_type = vectype; | |
1379 | else if (TREE_CODE (TREE_TYPE (op)) == BOOLEAN_TYPE | |
1380 | && VECTOR_BOOLEAN_TYPE_P (stmt_vectype)) | |
1381 | vector_type = build_same_sized_truth_vector_type (stmt_vectype); | |
1382 | else | |
1383 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
1384 | ||
7569a6cc | 1385 | gcc_assert (vector_type); |
418b7df3 | 1386 | return vect_init_vector (stmt, op, vector_type, NULL); |
ebfd146a IR |
1387 | } |
1388 | ||
81c40241 | 1389 | /* operand is defined inside the loop. */ |
8644a673 | 1390 | case vect_internal_def: |
ebfd146a | 1391 | { |
ebfd146a IR |
1392 | /* Get the def from the vectorized stmt. */ |
1393 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1394 | |
ebfd146a | 1395 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1396 | /* Get vectorized pattern statement. */ |
1397 | if (!vec_stmt | |
1398 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1399 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1400 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1401 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1402 | gcc_assert (vec_stmt); |
1403 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1404 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1405 | else if (is_gimple_call (vec_stmt)) | |
1406 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1407 | else | |
1408 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1409 | return vec_oprnd; | |
1410 | } | |
1411 | ||
81c40241 | 1412 | /* operand is defined by a loop header phi - reduction */ |
ebfd146a | 1413 | case vect_reduction_def: |
06066f92 | 1414 | case vect_double_reduction_def: |
7c5222ff | 1415 | case vect_nested_cycle: |
81c40241 RB |
1416 | /* Code should use get_initial_def_for_reduction. */ |
1417 | gcc_unreachable (); | |
ebfd146a | 1418 | |
81c40241 | 1419 | /* operand is defined by loop-header phi - induction. */ |
ebfd146a IR |
1420 | case vect_induction_def: |
1421 | { | |
1422 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1423 | ||
1424 | /* Get the def from the vectorized stmt. */ | |
1425 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1426 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1427 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1428 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1429 | else | |
1430 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1431 | return vec_oprnd; |
1432 | } | |
1433 | ||
1434 | default: | |
1435 | gcc_unreachable (); | |
1436 | } | |
1437 | } | |
1438 | ||
1439 | ||
1440 | /* Function vect_get_vec_def_for_stmt_copy | |
1441 | ||
ff802fa1 | 1442 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1443 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1444 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1445 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1446 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1447 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1448 | DT is the type of the vector def VEC_OPRND. |
1449 | ||
1450 | Context: | |
1451 | In case the vectorization factor (VF) is bigger than the number | |
1452 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1453 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1454 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1455 | smallest data-type determines the VF, and as a result, when vectorizing |
1456 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1457 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1458 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1459 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1460 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1461 | ||
1462 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1463 | |
ebfd146a IR |
1464 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1465 | VS1.1: vx.1 = memref1 VS1.2 | |
1466 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1467 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1468 | |
1469 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1470 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1471 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1472 | VSnew.3: vz3 = vx.3 + ... | |
1473 | ||
1474 | The vectorization of S1 is explained in vectorizable_load. | |
1475 | The vectorization of S2: | |
b8698a0f L |
1476 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1477 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1478 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1479 | returns the vector-def 'vx.0'. |
1480 | ||
b8698a0f L |
1481 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1482 | function is called to get the relevant vector-def for each operand. It is | |
1483 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1484 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1485 | ||
b8698a0f L |
1486 | For example, to obtain the vector-def 'vx.1' in order to create the |
1487 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1488 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1489 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1490 | and return its def ('vx.1'). | |
1491 | Overall, to create the above sequence this function will be called 3 times: | |
1492 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1493 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1494 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1495 | ||
1496 | tree | |
1497 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1498 | { | |
355fe088 | 1499 | gimple *vec_stmt_for_operand; |
ebfd146a IR |
1500 | stmt_vec_info def_stmt_info; |
1501 | ||
1502 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1503 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1504 | return vec_oprnd; |
1505 | ||
1506 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1507 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1508 | gcc_assert (def_stmt_info); | |
1509 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1510 | gcc_assert (vec_stmt_for_operand); | |
ebfd146a IR |
1511 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) |
1512 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1513 | else | |
1514 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1515 | return vec_oprnd; | |
1516 | } | |
1517 | ||
1518 | ||
1519 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1520 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a IR |
1521 | |
1522 | static void | |
b8698a0f | 1523 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1524 | vec<tree> *vec_oprnds0, |
1525 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1526 | { |
9771b263 | 1527 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1528 | |
1529 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1530 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1531 | |
9771b263 | 1532 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1533 | { |
9771b263 | 1534 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1535 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1536 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1537 | } |
1538 | } | |
1539 | ||
1540 | ||
d092494c IR |
1541 | /* Get vectorized definitions for OP0 and OP1. |
1542 | REDUC_INDEX is the index of reduction operand in case of reduction, | |
1543 | and -1 otherwise. */ | |
ebfd146a | 1544 | |
d092494c | 1545 | void |
355fe088 | 1546 | vect_get_vec_defs (tree op0, tree op1, gimple *stmt, |
9771b263 DN |
1547 | vec<tree> *vec_oprnds0, |
1548 | vec<tree> *vec_oprnds1, | |
d092494c | 1549 | slp_tree slp_node, int reduc_index) |
ebfd146a IR |
1550 | { |
1551 | if (slp_node) | |
d092494c IR |
1552 | { |
1553 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1554 | auto_vec<tree> ops (nops); |
1555 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1556 | |
9771b263 | 1557 | ops.quick_push (op0); |
d092494c | 1558 | if (op1) |
9771b263 | 1559 | ops.quick_push (op1); |
d092494c IR |
1560 | |
1561 | vect_get_slp_defs (ops, slp_node, &vec_defs, reduc_index); | |
1562 | ||
37b5ec8f | 1563 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1564 | if (op1) |
37b5ec8f | 1565 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1566 | } |
ebfd146a IR |
1567 | else |
1568 | { | |
1569 | tree vec_oprnd; | |
1570 | ||
9771b263 | 1571 | vec_oprnds0->create (1); |
81c40241 | 1572 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 1573 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1574 | |
1575 | if (op1) | |
1576 | { | |
9771b263 | 1577 | vec_oprnds1->create (1); |
81c40241 | 1578 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 1579 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1580 | } |
1581 | } | |
1582 | } | |
1583 | ||
1584 | ||
1585 | /* Function vect_finish_stmt_generation. | |
1586 | ||
1587 | Insert a new stmt. */ | |
1588 | ||
1589 | void | |
355fe088 | 1590 | vect_finish_stmt_generation (gimple *stmt, gimple *vec_stmt, |
ebfd146a IR |
1591 | gimple_stmt_iterator *gsi) |
1592 | { | |
1593 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
310213d4 | 1594 | vec_info *vinfo = stmt_info->vinfo; |
ebfd146a IR |
1595 | |
1596 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1597 | ||
54e8e2c3 RG |
1598 | if (!gsi_end_p (*gsi) |
1599 | && gimple_has_mem_ops (vec_stmt)) | |
1600 | { | |
355fe088 | 1601 | gimple *at_stmt = gsi_stmt (*gsi); |
54e8e2c3 RG |
1602 | tree vuse = gimple_vuse (at_stmt); |
1603 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1604 | { | |
1605 | tree vdef = gimple_vdef (at_stmt); | |
1606 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1607 | /* If we have an SSA vuse and insert a store, update virtual | |
1608 | SSA form to avoid triggering the renamer. Do so only | |
1609 | if we can easily see all uses - which is what almost always | |
1610 | happens with the way vectorized stmts are inserted. */ | |
1611 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1612 | && ((is_gimple_assign (vec_stmt) | |
1613 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1614 | || (is_gimple_call (vec_stmt) | |
1615 | && !(gimple_call_flags (vec_stmt) | |
1616 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1617 | { | |
1618 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1619 | gimple_set_vdef (vec_stmt, new_vdef); | |
1620 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1621 | } | |
1622 | } | |
1623 | } | |
ebfd146a IR |
1624 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1625 | ||
310213d4 | 1626 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, vinfo)); |
ebfd146a | 1627 | |
73fbfcad | 1628 | if (dump_enabled_p ()) |
ebfd146a | 1629 | { |
78c60e3d SS |
1630 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); |
1631 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
ebfd146a IR |
1632 | } |
1633 | ||
ad885386 | 1634 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
8e91d222 JJ |
1635 | |
1636 | /* While EH edges will generally prevent vectorization, stmt might | |
1637 | e.g. be in a must-not-throw region. Ensure newly created stmts | |
1638 | that could throw are part of the same region. */ | |
1639 | int lp_nr = lookup_stmt_eh_lp (stmt); | |
1640 | if (lp_nr != 0 && stmt_could_throw_p (vec_stmt)) | |
1641 | add_stmt_to_eh_lp (vec_stmt, lp_nr); | |
ebfd146a IR |
1642 | } |
1643 | ||
1644 | /* Checks if CALL can be vectorized in type VECTYPE. Returns | |
1645 | a function declaration if the target has a vectorized version | |
1646 | of the function, or NULL_TREE if the function cannot be vectorized. */ | |
1647 | ||
1648 | tree | |
538dd0b7 | 1649 | vectorizable_function (gcall *call, tree vectype_out, tree vectype_in) |
ebfd146a IR |
1650 | { |
1651 | tree fndecl = gimple_call_fndecl (call); | |
ebfd146a IR |
1652 | |
1653 | /* We only handle functions that do not read or clobber memory -- i.e. | |
1654 | const or novops ones. */ | |
1655 | if (!(gimple_call_flags (call) & (ECF_CONST | ECF_NOVOPS))) | |
1656 | return NULL_TREE; | |
1657 | ||
1658 | if (!fndecl | |
1659 | || TREE_CODE (fndecl) != FUNCTION_DECL | |
1660 | || !DECL_BUILT_IN (fndecl)) | |
1661 | return NULL_TREE; | |
1662 | ||
62f7fd21 | 1663 | return targetm.vectorize.builtin_vectorized_function (fndecl, vectype_out, |
ebfd146a IR |
1664 | vectype_in); |
1665 | } | |
1666 | ||
5ce9450f | 1667 | |
355fe088 | 1668 | static tree permute_vec_elements (tree, tree, tree, gimple *, |
5ce9450f JJ |
1669 | gimple_stmt_iterator *); |
1670 | ||
1671 | ||
1672 | /* Function vectorizable_mask_load_store. | |
1673 | ||
1674 | Check if STMT performs a conditional load or store that can be vectorized. | |
1675 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
1676 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. | |
1677 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1678 | ||
1679 | static bool | |
355fe088 TS |
1680 | vectorizable_mask_load_store (gimple *stmt, gimple_stmt_iterator *gsi, |
1681 | gimple **vec_stmt, slp_tree slp_node) | |
5ce9450f JJ |
1682 | { |
1683 | tree vec_dest = NULL; | |
1684 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1685 | stmt_vec_info prev_stmt_info; | |
1686 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1687 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1688 | bool nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
1689 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
1690 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
1691 | tree elem_type; | |
355fe088 | 1692 | gimple *new_stmt; |
5ce9450f JJ |
1693 | tree dummy; |
1694 | tree dataref_ptr = NULL_TREE; | |
355fe088 | 1695 | gimple *ptr_incr; |
5ce9450f JJ |
1696 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
1697 | int ncopies; | |
1698 | int i, j; | |
1699 | bool inv_p; | |
1700 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; | |
1701 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
1702 | int gather_scale = 1; | |
1703 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
1704 | bool is_store; | |
1705 | tree mask; | |
355fe088 | 1706 | gimple *def_stmt; |
5ce9450f JJ |
1707 | enum vect_def_type dt; |
1708 | ||
1709 | if (slp_node != NULL) | |
1710 | return false; | |
1711 | ||
1712 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
1713 | gcc_assert (ncopies >= 1); | |
1714 | ||
1715 | is_store = gimple_call_internal_fn (stmt) == IFN_MASK_STORE; | |
1716 | mask = gimple_call_arg (stmt, 2); | |
1717 | if (TYPE_PRECISION (TREE_TYPE (mask)) | |
1718 | != GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (vectype)))) | |
1719 | return false; | |
1720 | ||
1721 | /* FORNOW. This restriction should be relaxed. */ | |
1722 | if (nested_in_vect_loop && ncopies > 1) | |
1723 | { | |
1724 | if (dump_enabled_p ()) | |
1725 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1726 | "multiple types in nested loop."); | |
1727 | return false; | |
1728 | } | |
1729 | ||
1730 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) | |
1731 | return false; | |
1732 | ||
1733 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
1734 | return false; | |
1735 | ||
1736 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
1737 | return false; | |
1738 | ||
1739 | elem_type = TREE_TYPE (vectype); | |
1740 | ||
1741 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1742 | return false; | |
1743 | ||
f2e2a985 | 1744 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
5ce9450f JJ |
1745 | return false; |
1746 | ||
3bab6342 | 1747 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5ce9450f | 1748 | { |
355fe088 | 1749 | gimple *def_stmt; |
3bab6342 | 1750 | gather_decl = vect_check_gather_scatter (stmt, loop_vinfo, &gather_base, |
5ce9450f JJ |
1751 | &gather_off, &gather_scale); |
1752 | gcc_assert (gather_decl); | |
81c40241 RB |
1753 | if (!vect_is_simple_use (gather_off, loop_vinfo, &def_stmt, &gather_dt, |
1754 | &gather_off_vectype)) | |
5ce9450f JJ |
1755 | { |
1756 | if (dump_enabled_p ()) | |
1757 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1758 | "gather index use not simple."); | |
1759 | return false; | |
1760 | } | |
03b9e8e4 JJ |
1761 | |
1762 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
1763 | tree masktype | |
1764 | = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (TREE_CHAIN (arglist)))); | |
1765 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
1766 | { | |
1767 | if (dump_enabled_p ()) | |
1768 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1769 | "masked gather with integer mask not supported."); | |
1770 | return false; | |
1771 | } | |
5ce9450f JJ |
1772 | } |
1773 | else if (tree_int_cst_compare (nested_in_vect_loop | |
1774 | ? STMT_VINFO_DR_STEP (stmt_info) | |
1775 | : DR_STEP (dr), size_zero_node) <= 0) | |
1776 | return false; | |
1777 | else if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
1778 | || !can_vec_mask_load_store_p (TYPE_MODE (vectype), !is_store)) | |
1779 | return false; | |
1780 | ||
1781 | if (TREE_CODE (mask) != SSA_NAME) | |
1782 | return false; | |
1783 | ||
81c40241 | 1784 | if (!vect_is_simple_use (mask, loop_vinfo, &def_stmt, &dt)) |
5ce9450f JJ |
1785 | return false; |
1786 | ||
1787 | if (is_store) | |
1788 | { | |
1789 | tree rhs = gimple_call_arg (stmt, 3); | |
81c40241 | 1790 | if (!vect_is_simple_use (rhs, loop_vinfo, &def_stmt, &dt)) |
5ce9450f JJ |
1791 | return false; |
1792 | } | |
1793 | ||
1794 | if (!vec_stmt) /* transformation not required. */ | |
1795 | { | |
1796 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
1797 | if (is_store) | |
1798 | vect_model_store_cost (stmt_info, ncopies, false, dt, | |
1799 | NULL, NULL, NULL); | |
1800 | else | |
1801 | vect_model_load_cost (stmt_info, ncopies, false, NULL, NULL, NULL); | |
1802 | return true; | |
1803 | } | |
1804 | ||
1805 | /** Transform. **/ | |
1806 | ||
3bab6342 | 1807 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5ce9450f JJ |
1808 | { |
1809 | tree vec_oprnd0 = NULL_TREE, op; | |
1810 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
1811 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
acdcd61b | 1812 | tree ptr, vec_mask = NULL_TREE, mask_op = NULL_TREE, var, scale; |
5ce9450f | 1813 | tree perm_mask = NULL_TREE, prev_res = NULL_TREE; |
acdcd61b | 1814 | tree mask_perm_mask = NULL_TREE; |
5ce9450f JJ |
1815 | edge pe = loop_preheader_edge (loop); |
1816 | gimple_seq seq; | |
1817 | basic_block new_bb; | |
1818 | enum { NARROW, NONE, WIDEN } modifier; | |
1819 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
1820 | ||
acdcd61b JJ |
1821 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); |
1822 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1823 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1824 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1825 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1826 | scaletype = TREE_VALUE (arglist); | |
1827 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
1828 | && types_compatible_p (srctype, masktype)); | |
1829 | ||
5ce9450f JJ |
1830 | if (nunits == gather_off_nunits) |
1831 | modifier = NONE; | |
1832 | else if (nunits == gather_off_nunits / 2) | |
1833 | { | |
1834 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
1835 | modifier = WIDEN; | |
1836 | ||
1837 | for (i = 0; i < gather_off_nunits; ++i) | |
1838 | sel[i] = i | nunits; | |
1839 | ||
557be5a8 | 1840 | perm_mask = vect_gen_perm_mask_checked (gather_off_vectype, sel); |
5ce9450f JJ |
1841 | } |
1842 | else if (nunits == gather_off_nunits * 2) | |
1843 | { | |
1844 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
1845 | modifier = NARROW; | |
1846 | ||
1847 | for (i = 0; i < nunits; ++i) | |
1848 | sel[i] = i < gather_off_nunits | |
1849 | ? i : i + nunits - gather_off_nunits; | |
1850 | ||
557be5a8 | 1851 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); |
5ce9450f | 1852 | ncopies *= 2; |
acdcd61b JJ |
1853 | for (i = 0; i < nunits; ++i) |
1854 | sel[i] = i | gather_off_nunits; | |
557be5a8 | 1855 | mask_perm_mask = vect_gen_perm_mask_checked (masktype, sel); |
5ce9450f JJ |
1856 | } |
1857 | else | |
1858 | gcc_unreachable (); | |
1859 | ||
5ce9450f JJ |
1860 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); |
1861 | ||
1862 | ptr = fold_convert (ptrtype, gather_base); | |
1863 | if (!is_gimple_min_invariant (ptr)) | |
1864 | { | |
1865 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
1866 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
1867 | gcc_assert (!new_bb); | |
1868 | } | |
1869 | ||
1870 | scale = build_int_cst (scaletype, gather_scale); | |
1871 | ||
1872 | prev_stmt_info = NULL; | |
1873 | for (j = 0; j < ncopies; ++j) | |
1874 | { | |
1875 | if (modifier == WIDEN && (j & 1)) | |
1876 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
1877 | perm_mask, stmt, gsi); | |
1878 | else if (j == 0) | |
1879 | op = vec_oprnd0 | |
81c40241 | 1880 | = vect_get_vec_def_for_operand (gather_off, stmt); |
5ce9450f JJ |
1881 | else |
1882 | op = vec_oprnd0 | |
1883 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
1884 | ||
1885 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
1886 | { | |
1887 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
1888 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 1889 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
5ce9450f JJ |
1890 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
1891 | new_stmt | |
0d0e4a03 | 1892 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
1893 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1894 | op = var; | |
1895 | } | |
1896 | ||
acdcd61b JJ |
1897 | if (mask_perm_mask && (j & 1)) |
1898 | mask_op = permute_vec_elements (mask_op, mask_op, | |
1899 | mask_perm_mask, stmt, gsi); | |
5ce9450f JJ |
1900 | else |
1901 | { | |
acdcd61b | 1902 | if (j == 0) |
81c40241 | 1903 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); |
acdcd61b JJ |
1904 | else |
1905 | { | |
81c40241 | 1906 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
acdcd61b JJ |
1907 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
1908 | } | |
5ce9450f | 1909 | |
acdcd61b JJ |
1910 | mask_op = vec_mask; |
1911 | if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask))) | |
1912 | { | |
1913 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)) | |
1914 | == TYPE_VECTOR_SUBPARTS (masktype)); | |
0e22bb5a | 1915 | var = vect_get_new_ssa_name (masktype, vect_simple_var); |
acdcd61b JJ |
1916 | mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op); |
1917 | new_stmt | |
0d0e4a03 | 1918 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, mask_op); |
acdcd61b JJ |
1919 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1920 | mask_op = var; | |
1921 | } | |
5ce9450f JJ |
1922 | } |
1923 | ||
1924 | new_stmt | |
1925 | = gimple_build_call (gather_decl, 5, mask_op, ptr, op, mask_op, | |
1926 | scale); | |
1927 | ||
1928 | if (!useless_type_conversion_p (vectype, rettype)) | |
1929 | { | |
1930 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
1931 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 1932 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
5ce9450f JJ |
1933 | gimple_call_set_lhs (new_stmt, op); |
1934 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 1935 | var = make_ssa_name (vec_dest); |
5ce9450f | 1936 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
0d0e4a03 | 1937 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
1938 | } |
1939 | else | |
1940 | { | |
1941 | var = make_ssa_name (vec_dest, new_stmt); | |
1942 | gimple_call_set_lhs (new_stmt, var); | |
1943 | } | |
1944 | ||
1945 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1946 | ||
1947 | if (modifier == NARROW) | |
1948 | { | |
1949 | if ((j & 1) == 0) | |
1950 | { | |
1951 | prev_res = var; | |
1952 | continue; | |
1953 | } | |
1954 | var = permute_vec_elements (prev_res, var, | |
1955 | perm_mask, stmt, gsi); | |
1956 | new_stmt = SSA_NAME_DEF_STMT (var); | |
1957 | } | |
1958 | ||
1959 | if (prev_stmt_info == NULL) | |
1960 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1961 | else | |
1962 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1963 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1964 | } | |
3efe2e2c JJ |
1965 | |
1966 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
1967 | from the IL. */ | |
1968 | tree lhs = gimple_call_lhs (stmt); | |
1969 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
1970 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
1971 | set_vinfo_for_stmt (stmt, NULL); | |
1972 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
1973 | gsi_replace (gsi, new_stmt, true); | |
5ce9450f JJ |
1974 | return true; |
1975 | } | |
1976 | else if (is_store) | |
1977 | { | |
1978 | tree vec_rhs = NULL_TREE, vec_mask = NULL_TREE; | |
1979 | prev_stmt_info = NULL; | |
1980 | for (i = 0; i < ncopies; i++) | |
1981 | { | |
1982 | unsigned align, misalign; | |
1983 | ||
1984 | if (i == 0) | |
1985 | { | |
1986 | tree rhs = gimple_call_arg (stmt, 3); | |
81c40241 RB |
1987 | vec_rhs = vect_get_vec_def_for_operand (rhs, stmt); |
1988 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); | |
5ce9450f JJ |
1989 | /* We should have catched mismatched types earlier. */ |
1990 | gcc_assert (useless_type_conversion_p (vectype, | |
1991 | TREE_TYPE (vec_rhs))); | |
1992 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, | |
1993 | NULL_TREE, &dummy, gsi, | |
1994 | &ptr_incr, false, &inv_p); | |
1995 | gcc_assert (!inv_p); | |
1996 | } | |
1997 | else | |
1998 | { | |
81c40241 | 1999 | vect_is_simple_use (vec_rhs, loop_vinfo, &def_stmt, &dt); |
5ce9450f | 2000 | vec_rhs = vect_get_vec_def_for_stmt_copy (dt, vec_rhs); |
81c40241 | 2001 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2002 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2003 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2004 | TYPE_SIZE_UNIT (vectype)); | |
2005 | } | |
2006 | ||
2007 | align = TYPE_ALIGN_UNIT (vectype); | |
2008 | if (aligned_access_p (dr)) | |
2009 | misalign = 0; | |
2010 | else if (DR_MISALIGNMENT (dr) == -1) | |
2011 | { | |
2012 | align = TYPE_ALIGN_UNIT (elem_type); | |
2013 | misalign = 0; | |
2014 | } | |
2015 | else | |
2016 | misalign = DR_MISALIGNMENT (dr); | |
2017 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2018 | misalign); | |
2019 | new_stmt | |
2020 | = gimple_build_call_internal (IFN_MASK_STORE, 4, dataref_ptr, | |
2021 | gimple_call_arg (stmt, 1), | |
2022 | vec_mask, vec_rhs); | |
2023 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2024 | if (i == 0) | |
2025 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2026 | else | |
2027 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2028 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2029 | } | |
2030 | } | |
2031 | else | |
2032 | { | |
2033 | tree vec_mask = NULL_TREE; | |
2034 | prev_stmt_info = NULL; | |
2035 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); | |
2036 | for (i = 0; i < ncopies; i++) | |
2037 | { | |
2038 | unsigned align, misalign; | |
2039 | ||
2040 | if (i == 0) | |
2041 | { | |
81c40241 | 2042 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); |
5ce9450f JJ |
2043 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, |
2044 | NULL_TREE, &dummy, gsi, | |
2045 | &ptr_incr, false, &inv_p); | |
2046 | gcc_assert (!inv_p); | |
2047 | } | |
2048 | else | |
2049 | { | |
81c40241 | 2050 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2051 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2052 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2053 | TYPE_SIZE_UNIT (vectype)); | |
2054 | } | |
2055 | ||
2056 | align = TYPE_ALIGN_UNIT (vectype); | |
2057 | if (aligned_access_p (dr)) | |
2058 | misalign = 0; | |
2059 | else if (DR_MISALIGNMENT (dr) == -1) | |
2060 | { | |
2061 | align = TYPE_ALIGN_UNIT (elem_type); | |
2062 | misalign = 0; | |
2063 | } | |
2064 | else | |
2065 | misalign = DR_MISALIGNMENT (dr); | |
2066 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2067 | misalign); | |
2068 | new_stmt | |
2069 | = gimple_build_call_internal (IFN_MASK_LOAD, 3, dataref_ptr, | |
2070 | gimple_call_arg (stmt, 1), | |
2071 | vec_mask); | |
b731b390 | 2072 | gimple_call_set_lhs (new_stmt, make_ssa_name (vec_dest)); |
5ce9450f JJ |
2073 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2074 | if (i == 0) | |
2075 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2076 | else | |
2077 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2078 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2079 | } | |
2080 | } | |
2081 | ||
3efe2e2c JJ |
2082 | if (!is_store) |
2083 | { | |
2084 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
2085 | from the IL. */ | |
2086 | tree lhs = gimple_call_lhs (stmt); | |
2087 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
2088 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
2089 | set_vinfo_for_stmt (stmt, NULL); | |
2090 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2091 | gsi_replace (gsi, new_stmt, true); | |
2092 | } | |
2093 | ||
5ce9450f JJ |
2094 | return true; |
2095 | } | |
2096 | ||
2097 | ||
ebfd146a IR |
2098 | /* Function vectorizable_call. |
2099 | ||
538dd0b7 | 2100 | Check if GS performs a function call that can be vectorized. |
b8698a0f | 2101 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
2102 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2103 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2104 | ||
2105 | static bool | |
355fe088 | 2106 | vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
190c2236 | 2107 | slp_tree slp_node) |
ebfd146a | 2108 | { |
538dd0b7 | 2109 | gcall *stmt; |
ebfd146a IR |
2110 | tree vec_dest; |
2111 | tree scalar_dest; | |
2112 | tree op, type; | |
2113 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
538dd0b7 | 2114 | stmt_vec_info stmt_info = vinfo_for_stmt (gs), prev_stmt_info; |
ebfd146a IR |
2115 | tree vectype_out, vectype_in; |
2116 | int nunits_in; | |
2117 | int nunits_out; | |
2118 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
190c2236 | 2119 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 2120 | vec_info *vinfo = stmt_info->vinfo; |
81c40241 | 2121 | tree fndecl, new_temp, rhs_type; |
355fe088 | 2122 | gimple *def_stmt; |
0502fb85 UB |
2123 | enum vect_def_type dt[3] |
2124 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
355fe088 | 2125 | gimple *new_stmt = NULL; |
ebfd146a | 2126 | int ncopies, j; |
6e1aa848 | 2127 | vec<tree> vargs = vNULL; |
ebfd146a IR |
2128 | enum { NARROW, NONE, WIDEN } modifier; |
2129 | size_t i, nargs; | |
9d5e7640 | 2130 | tree lhs; |
ebfd146a | 2131 | |
190c2236 | 2132 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2133 | return false; |
2134 | ||
8644a673 | 2135 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
2136 | return false; |
2137 | ||
538dd0b7 DM |
2138 | /* Is GS a vectorizable call? */ |
2139 | stmt = dyn_cast <gcall *> (gs); | |
2140 | if (!stmt) | |
ebfd146a IR |
2141 | return false; |
2142 | ||
5ce9450f JJ |
2143 | if (gimple_call_internal_p (stmt) |
2144 | && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD | |
2145 | || gimple_call_internal_fn (stmt) == IFN_MASK_STORE)) | |
2146 | return vectorizable_mask_load_store (stmt, gsi, vec_stmt, | |
2147 | slp_node); | |
2148 | ||
0136f8f0 AH |
2149 | if (gimple_call_lhs (stmt) == NULL_TREE |
2150 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
ebfd146a IR |
2151 | return false; |
2152 | ||
0136f8f0 | 2153 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); |
5a2c1986 | 2154 | |
b690cc0f RG |
2155 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
2156 | ||
ebfd146a IR |
2157 | /* Process function arguments. */ |
2158 | rhs_type = NULL_TREE; | |
b690cc0f | 2159 | vectype_in = NULL_TREE; |
ebfd146a IR |
2160 | nargs = gimple_call_num_args (stmt); |
2161 | ||
1b1562a5 MM |
2162 | /* Bail out if the function has more than three arguments, we do not have |
2163 | interesting builtin functions to vectorize with more than two arguments | |
2164 | except for fma. No arguments is also not good. */ | |
2165 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
2166 | return false; |
2167 | ||
74bf76ed JJ |
2168 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
2169 | if (gimple_call_internal_p (stmt) | |
2170 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2171 | { | |
2172 | nargs = 0; | |
2173 | rhs_type = unsigned_type_node; | |
2174 | } | |
2175 | ||
ebfd146a IR |
2176 | for (i = 0; i < nargs; i++) |
2177 | { | |
b690cc0f RG |
2178 | tree opvectype; |
2179 | ||
ebfd146a IR |
2180 | op = gimple_call_arg (stmt, i); |
2181 | ||
2182 | /* We can only handle calls with arguments of the same type. */ | |
2183 | if (rhs_type | |
8533c9d8 | 2184 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 2185 | { |
73fbfcad | 2186 | if (dump_enabled_p ()) |
78c60e3d | 2187 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2188 | "argument types differ.\n"); |
ebfd146a IR |
2189 | return false; |
2190 | } | |
b690cc0f RG |
2191 | if (!rhs_type) |
2192 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 2193 | |
81c40241 | 2194 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[i], &opvectype)) |
ebfd146a | 2195 | { |
73fbfcad | 2196 | if (dump_enabled_p ()) |
78c60e3d | 2197 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2198 | "use not simple.\n"); |
ebfd146a IR |
2199 | return false; |
2200 | } | |
ebfd146a | 2201 | |
b690cc0f RG |
2202 | if (!vectype_in) |
2203 | vectype_in = opvectype; | |
2204 | else if (opvectype | |
2205 | && opvectype != vectype_in) | |
2206 | { | |
73fbfcad | 2207 | if (dump_enabled_p ()) |
78c60e3d | 2208 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2209 | "argument vector types differ.\n"); |
b690cc0f RG |
2210 | return false; |
2211 | } | |
2212 | } | |
2213 | /* If all arguments are external or constant defs use a vector type with | |
2214 | the same size as the output vector type. */ | |
ebfd146a | 2215 | if (!vectype_in) |
b690cc0f | 2216 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
2217 | if (vec_stmt) |
2218 | gcc_assert (vectype_in); | |
2219 | if (!vectype_in) | |
2220 | { | |
73fbfcad | 2221 | if (dump_enabled_p ()) |
7d8930a0 | 2222 | { |
78c60e3d SS |
2223 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2224 | "no vectype for scalar type "); | |
2225 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 2226 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
2227 | } |
2228 | ||
2229 | return false; | |
2230 | } | |
ebfd146a IR |
2231 | |
2232 | /* FORNOW */ | |
b690cc0f RG |
2233 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
2234 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
2235 | if (nunits_in == nunits_out / 2) |
2236 | modifier = NARROW; | |
2237 | else if (nunits_out == nunits_in) | |
2238 | modifier = NONE; | |
2239 | else if (nunits_out == nunits_in / 2) | |
2240 | modifier = WIDEN; | |
2241 | else | |
2242 | return false; | |
2243 | ||
2244 | /* For now, we only vectorize functions if a target specific builtin | |
2245 | is available. TODO -- in some cases, it might be profitable to | |
2246 | insert the calls for pieces of the vector, in order to be able | |
2247 | to vectorize other operations in the loop. */ | |
2248 | fndecl = vectorizable_function (stmt, vectype_out, vectype_in); | |
2249 | if (fndecl == NULL_TREE) | |
2250 | { | |
74bf76ed JJ |
2251 | if (gimple_call_internal_p (stmt) |
2252 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE | |
2253 | && !slp_node | |
2254 | && loop_vinfo | |
2255 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2256 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
2257 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2258 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
2259 | { | |
2260 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
2261 | { 0, 1, 2, ... vf - 1 } vector. */ | |
2262 | gcc_assert (nargs == 0); | |
2263 | } | |
2264 | else | |
2265 | { | |
2266 | if (dump_enabled_p ()) | |
2267 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 2268 | "function is not vectorizable.\n"); |
74bf76ed JJ |
2269 | return false; |
2270 | } | |
ebfd146a IR |
2271 | } |
2272 | ||
5006671f | 2273 | gcc_assert (!gimple_vuse (stmt)); |
ebfd146a | 2274 | |
190c2236 JJ |
2275 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
2276 | ncopies = 1; | |
2277 | else if (modifier == NARROW) | |
ebfd146a IR |
2278 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; |
2279 | else | |
2280 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2281 | ||
2282 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
2283 | needs to be generated. */ | |
2284 | gcc_assert (ncopies >= 1); | |
2285 | ||
2286 | if (!vec_stmt) /* transformation not required. */ | |
2287 | { | |
2288 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 2289 | if (dump_enabled_p ()) |
e645e942 TJ |
2290 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ===" |
2291 | "\n"); | |
c3e7ee41 | 2292 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
2293 | return true; |
2294 | } | |
2295 | ||
2296 | /** Transform. **/ | |
2297 | ||
73fbfcad | 2298 | if (dump_enabled_p ()) |
e645e942 | 2299 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
2300 | |
2301 | /* Handle def. */ | |
2302 | scalar_dest = gimple_call_lhs (stmt); | |
2303 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
2304 | ||
2305 | prev_stmt_info = NULL; | |
2306 | switch (modifier) | |
2307 | { | |
2308 | case NONE: | |
2309 | for (j = 0; j < ncopies; ++j) | |
2310 | { | |
2311 | /* Build argument list for the vectorized call. */ | |
2312 | if (j == 0) | |
9771b263 | 2313 | vargs.create (nargs); |
ebfd146a | 2314 | else |
9771b263 | 2315 | vargs.truncate (0); |
ebfd146a | 2316 | |
190c2236 JJ |
2317 | if (slp_node) |
2318 | { | |
ef062b13 | 2319 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2320 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2321 | |
2322 | for (i = 0; i < nargs; i++) | |
9771b263 | 2323 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2324 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2325 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2326 | |
2327 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2328 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
2329 | { |
2330 | size_t k; | |
2331 | for (k = 0; k < nargs; k++) | |
2332 | { | |
37b5ec8f | 2333 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 2334 | vargs[k] = vec_oprndsk[i]; |
190c2236 JJ |
2335 | } |
2336 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2337 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2338 | gimple_call_set_lhs (new_stmt, new_temp); | |
2339 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 2340 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2341 | } |
2342 | ||
2343 | for (i = 0; i < nargs; i++) | |
2344 | { | |
37b5ec8f | 2345 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2346 | vec_oprndsi.release (); |
190c2236 | 2347 | } |
190c2236 JJ |
2348 | continue; |
2349 | } | |
2350 | ||
ebfd146a IR |
2351 | for (i = 0; i < nargs; i++) |
2352 | { | |
2353 | op = gimple_call_arg (stmt, i); | |
2354 | if (j == 0) | |
2355 | vec_oprnd0 | |
81c40241 | 2356 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 2357 | else |
63827fb8 IR |
2358 | { |
2359 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
2360 | vec_oprnd0 | |
2361 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
2362 | } | |
ebfd146a | 2363 | |
9771b263 | 2364 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
2365 | } |
2366 | ||
74bf76ed JJ |
2367 | if (gimple_call_internal_p (stmt) |
2368 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2369 | { | |
2370 | tree *v = XALLOCAVEC (tree, nunits_out); | |
2371 | int k; | |
2372 | for (k = 0; k < nunits_out; ++k) | |
2373 | v[k] = build_int_cst (unsigned_type_node, j * nunits_out + k); | |
2374 | tree cst = build_vector (vectype_out, v); | |
2375 | tree new_var | |
0e22bb5a | 2376 | = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_"); |
355fe088 | 2377 | gimple *init_stmt = gimple_build_assign (new_var, cst); |
74bf76ed | 2378 | vect_init_vector_1 (stmt, init_stmt, NULL); |
b731b390 | 2379 | new_temp = make_ssa_name (vec_dest); |
0e22bb5a | 2380 | new_stmt = gimple_build_assign (new_temp, new_var); |
74bf76ed JJ |
2381 | } |
2382 | else | |
2383 | { | |
2384 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2385 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2386 | gimple_call_set_lhs (new_stmt, new_temp); | |
2387 | } | |
ebfd146a IR |
2388 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2389 | ||
2390 | if (j == 0) | |
2391 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2392 | else | |
2393 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2394 | ||
2395 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2396 | } | |
2397 | ||
2398 | break; | |
2399 | ||
2400 | case NARROW: | |
2401 | for (j = 0; j < ncopies; ++j) | |
2402 | { | |
2403 | /* Build argument list for the vectorized call. */ | |
2404 | if (j == 0) | |
9771b263 | 2405 | vargs.create (nargs * 2); |
ebfd146a | 2406 | else |
9771b263 | 2407 | vargs.truncate (0); |
ebfd146a | 2408 | |
190c2236 JJ |
2409 | if (slp_node) |
2410 | { | |
ef062b13 | 2411 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2412 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2413 | |
2414 | for (i = 0; i < nargs; i++) | |
9771b263 | 2415 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2416 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2417 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2418 | |
2419 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2420 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
2421 | { |
2422 | size_t k; | |
9771b263 | 2423 | vargs.truncate (0); |
190c2236 JJ |
2424 | for (k = 0; k < nargs; k++) |
2425 | { | |
37b5ec8f | 2426 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
2427 | vargs.quick_push (vec_oprndsk[i]); |
2428 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 JJ |
2429 | } |
2430 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2431 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2432 | gimple_call_set_lhs (new_stmt, new_temp); | |
2433 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 2434 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2435 | } |
2436 | ||
2437 | for (i = 0; i < nargs; i++) | |
2438 | { | |
37b5ec8f | 2439 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2440 | vec_oprndsi.release (); |
190c2236 | 2441 | } |
190c2236 JJ |
2442 | continue; |
2443 | } | |
2444 | ||
ebfd146a IR |
2445 | for (i = 0; i < nargs; i++) |
2446 | { | |
2447 | op = gimple_call_arg (stmt, i); | |
2448 | if (j == 0) | |
2449 | { | |
2450 | vec_oprnd0 | |
81c40241 | 2451 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 2452 | vec_oprnd1 |
63827fb8 | 2453 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2454 | } |
2455 | else | |
2456 | { | |
336ecb65 | 2457 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 2458 | vec_oprnd0 |
63827fb8 | 2459 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 2460 | vec_oprnd1 |
63827fb8 | 2461 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2462 | } |
2463 | ||
9771b263 DN |
2464 | vargs.quick_push (vec_oprnd0); |
2465 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
2466 | } |
2467 | ||
2468 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2469 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2470 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
2471 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2472 | ||
2473 | if (j == 0) | |
2474 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2475 | else | |
2476 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2477 | ||
2478 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2479 | } | |
2480 | ||
2481 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
2482 | ||
2483 | break; | |
2484 | ||
2485 | case WIDEN: | |
2486 | /* No current target implements this case. */ | |
2487 | return false; | |
2488 | } | |
2489 | ||
9771b263 | 2490 | vargs.release (); |
ebfd146a | 2491 | |
ebfd146a IR |
2492 | /* The call in STMT might prevent it from being removed in dce. |
2493 | We however cannot remove it here, due to the way the ssa name | |
2494 | it defines is mapped to the new definition. So just replace | |
2495 | rhs of the statement with something harmless. */ | |
2496 | ||
dd34c087 JJ |
2497 | if (slp_node) |
2498 | return true; | |
2499 | ||
ebfd146a | 2500 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
2501 | if (is_pattern_stmt_p (stmt_info)) |
2502 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
2503 | else | |
2504 | lhs = gimple_call_lhs (stmt); | |
3cc2fa2a JJ |
2505 | |
2506 | if (gimple_call_internal_p (stmt) | |
2507 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2508 | { | |
2509 | /* Replace uses of the lhs of GOMP_SIMD_LANE call outside the loop | |
2510 | with vf - 1 rather than 0, that is the last iteration of the | |
2511 | vectorized loop. */ | |
2512 | imm_use_iterator iter; | |
2513 | use_operand_p use_p; | |
355fe088 | 2514 | gimple *use_stmt; |
3cc2fa2a JJ |
2515 | FOR_EACH_IMM_USE_STMT (use_stmt, iter, lhs) |
2516 | { | |
2517 | basic_block use_bb = gimple_bb (use_stmt); | |
2518 | if (use_bb | |
2519 | && !flow_bb_inside_loop_p (LOOP_VINFO_LOOP (loop_vinfo), use_bb)) | |
2520 | { | |
2521 | FOR_EACH_IMM_USE_ON_STMT (use_p, iter) | |
2522 | SET_USE (use_p, build_int_cst (TREE_TYPE (lhs), | |
2523 | ncopies * nunits_out - 1)); | |
2524 | update_stmt (use_stmt); | |
2525 | } | |
2526 | } | |
2527 | } | |
2528 | ||
9d5e7640 | 2529 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); |
ebfd146a | 2530 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 2531 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
2532 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
2533 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
2534 | |
2535 | return true; | |
2536 | } | |
2537 | ||
2538 | ||
0136f8f0 AH |
2539 | struct simd_call_arg_info |
2540 | { | |
2541 | tree vectype; | |
2542 | tree op; | |
2543 | enum vect_def_type dt; | |
2544 | HOST_WIDE_INT linear_step; | |
2545 | unsigned int align; | |
17b658af | 2546 | bool simd_lane_linear; |
0136f8f0 AH |
2547 | }; |
2548 | ||
17b658af JJ |
2549 | /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME, |
2550 | is linear within simd lane (but not within whole loop), note it in | |
2551 | *ARGINFO. */ | |
2552 | ||
2553 | static void | |
2554 | vect_simd_lane_linear (tree op, struct loop *loop, | |
2555 | struct simd_call_arg_info *arginfo) | |
2556 | { | |
355fe088 | 2557 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
17b658af JJ |
2558 | |
2559 | if (!is_gimple_assign (def_stmt) | |
2560 | || gimple_assign_rhs_code (def_stmt) != POINTER_PLUS_EXPR | |
2561 | || !is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))) | |
2562 | return; | |
2563 | ||
2564 | tree base = gimple_assign_rhs1 (def_stmt); | |
2565 | HOST_WIDE_INT linear_step = 0; | |
2566 | tree v = gimple_assign_rhs2 (def_stmt); | |
2567 | while (TREE_CODE (v) == SSA_NAME) | |
2568 | { | |
2569 | tree t; | |
2570 | def_stmt = SSA_NAME_DEF_STMT (v); | |
2571 | if (is_gimple_assign (def_stmt)) | |
2572 | switch (gimple_assign_rhs_code (def_stmt)) | |
2573 | { | |
2574 | case PLUS_EXPR: | |
2575 | t = gimple_assign_rhs2 (def_stmt); | |
2576 | if (linear_step || TREE_CODE (t) != INTEGER_CST) | |
2577 | return; | |
2578 | base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t); | |
2579 | v = gimple_assign_rhs1 (def_stmt); | |
2580 | continue; | |
2581 | case MULT_EXPR: | |
2582 | t = gimple_assign_rhs2 (def_stmt); | |
2583 | if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t)) | |
2584 | return; | |
2585 | linear_step = tree_to_shwi (t); | |
2586 | v = gimple_assign_rhs1 (def_stmt); | |
2587 | continue; | |
2588 | CASE_CONVERT: | |
2589 | t = gimple_assign_rhs1 (def_stmt); | |
2590 | if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE | |
2591 | || (TYPE_PRECISION (TREE_TYPE (v)) | |
2592 | < TYPE_PRECISION (TREE_TYPE (t)))) | |
2593 | return; | |
2594 | if (!linear_step) | |
2595 | linear_step = 1; | |
2596 | v = t; | |
2597 | continue; | |
2598 | default: | |
2599 | return; | |
2600 | } | |
2601 | else if (is_gimple_call (def_stmt) | |
2602 | && gimple_call_internal_p (def_stmt) | |
2603 | && gimple_call_internal_fn (def_stmt) == IFN_GOMP_SIMD_LANE | |
2604 | && loop->simduid | |
2605 | && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME | |
2606 | && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0)) | |
2607 | == loop->simduid)) | |
2608 | { | |
2609 | if (!linear_step) | |
2610 | linear_step = 1; | |
2611 | arginfo->linear_step = linear_step; | |
2612 | arginfo->op = base; | |
2613 | arginfo->simd_lane_linear = true; | |
2614 | return; | |
2615 | } | |
2616 | } | |
2617 | } | |
2618 | ||
0136f8f0 AH |
2619 | /* Function vectorizable_simd_clone_call. |
2620 | ||
2621 | Check if STMT performs a function call that can be vectorized | |
2622 | by calling a simd clone of the function. | |
2623 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2624 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2625 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2626 | ||
2627 | static bool | |
355fe088 TS |
2628 | vectorizable_simd_clone_call (gimple *stmt, gimple_stmt_iterator *gsi, |
2629 | gimple **vec_stmt, slp_tree slp_node) | |
0136f8f0 AH |
2630 | { |
2631 | tree vec_dest; | |
2632 | tree scalar_dest; | |
2633 | tree op, type; | |
2634 | tree vec_oprnd0 = NULL_TREE; | |
2635 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
2636 | tree vectype; | |
2637 | unsigned int nunits; | |
2638 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2639 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
310213d4 | 2640 | vec_info *vinfo = stmt_info->vinfo; |
0136f8f0 | 2641 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
81c40241 | 2642 | tree fndecl, new_temp; |
355fe088 TS |
2643 | gimple *def_stmt; |
2644 | gimple *new_stmt = NULL; | |
0136f8f0 AH |
2645 | int ncopies, j; |
2646 | vec<simd_call_arg_info> arginfo = vNULL; | |
2647 | vec<tree> vargs = vNULL; | |
2648 | size_t i, nargs; | |
2649 | tree lhs, rtype, ratype; | |
2650 | vec<constructor_elt, va_gc> *ret_ctor_elts; | |
2651 | ||
2652 | /* Is STMT a vectorizable call? */ | |
2653 | if (!is_gimple_call (stmt)) | |
2654 | return false; | |
2655 | ||
2656 | fndecl = gimple_call_fndecl (stmt); | |
2657 | if (fndecl == NULL_TREE) | |
2658 | return false; | |
2659 | ||
d52f5295 | 2660 | struct cgraph_node *node = cgraph_node::get (fndecl); |
0136f8f0 AH |
2661 | if (node == NULL || node->simd_clones == NULL) |
2662 | return false; | |
2663 | ||
2664 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
2665 | return false; | |
2666 | ||
2667 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
2668 | return false; | |
2669 | ||
2670 | if (gimple_call_lhs (stmt) | |
2671 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
2672 | return false; | |
2673 | ||
2674 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); | |
2675 | ||
2676 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2677 | ||
2678 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt)) | |
2679 | return false; | |
2680 | ||
2681 | /* FORNOW */ | |
2682 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
2683 | return false; | |
2684 | ||
2685 | /* Process function arguments. */ | |
2686 | nargs = gimple_call_num_args (stmt); | |
2687 | ||
2688 | /* Bail out if the function has zero arguments. */ | |
2689 | if (nargs == 0) | |
2690 | return false; | |
2691 | ||
2692 | arginfo.create (nargs); | |
2693 | ||
2694 | for (i = 0; i < nargs; i++) | |
2695 | { | |
2696 | simd_call_arg_info thisarginfo; | |
2697 | affine_iv iv; | |
2698 | ||
2699 | thisarginfo.linear_step = 0; | |
2700 | thisarginfo.align = 0; | |
2701 | thisarginfo.op = NULL_TREE; | |
17b658af | 2702 | thisarginfo.simd_lane_linear = false; |
0136f8f0 AH |
2703 | |
2704 | op = gimple_call_arg (stmt, i); | |
81c40241 RB |
2705 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &thisarginfo.dt, |
2706 | &thisarginfo.vectype) | |
0136f8f0 AH |
2707 | || thisarginfo.dt == vect_uninitialized_def) |
2708 | { | |
2709 | if (dump_enabled_p ()) | |
2710 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2711 | "use not simple.\n"); | |
2712 | arginfo.release (); | |
2713 | return false; | |
2714 | } | |
2715 | ||
2716 | if (thisarginfo.dt == vect_constant_def | |
2717 | || thisarginfo.dt == vect_external_def) | |
2718 | gcc_assert (thisarginfo.vectype == NULL_TREE); | |
2719 | else | |
2720 | gcc_assert (thisarginfo.vectype != NULL_TREE); | |
2721 | ||
6c9e85fb JJ |
2722 | /* For linear arguments, the analyze phase should have saved |
2723 | the base and step in STMT_VINFO_SIMD_CLONE_INFO. */ | |
17b658af JJ |
2724 | if (i * 3 + 4 <= STMT_VINFO_SIMD_CLONE_INFO (stmt_info).length () |
2725 | && STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]) | |
6c9e85fb JJ |
2726 | { |
2727 | gcc_assert (vec_stmt); | |
2728 | thisarginfo.linear_step | |
17b658af | 2729 | = tree_to_shwi (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]); |
6c9e85fb | 2730 | thisarginfo.op |
17b658af JJ |
2731 | = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 1]; |
2732 | thisarginfo.simd_lane_linear | |
2733 | = (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 3] | |
2734 | == boolean_true_node); | |
6c9e85fb JJ |
2735 | /* If loop has been peeled for alignment, we need to adjust it. */ |
2736 | tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo); | |
2737 | tree n2 = LOOP_VINFO_NITERS (loop_vinfo); | |
17b658af | 2738 | if (n1 != n2 && !thisarginfo.simd_lane_linear) |
6c9e85fb JJ |
2739 | { |
2740 | tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2); | |
17b658af | 2741 | tree step = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]; |
6c9e85fb JJ |
2742 | tree opt = TREE_TYPE (thisarginfo.op); |
2743 | bias = fold_convert (TREE_TYPE (step), bias); | |
2744 | bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step); | |
2745 | thisarginfo.op | |
2746 | = fold_build2 (POINTER_TYPE_P (opt) | |
2747 | ? POINTER_PLUS_EXPR : PLUS_EXPR, opt, | |
2748 | thisarginfo.op, bias); | |
2749 | } | |
2750 | } | |
2751 | else if (!vec_stmt | |
2752 | && thisarginfo.dt != vect_constant_def | |
2753 | && thisarginfo.dt != vect_external_def | |
2754 | && loop_vinfo | |
2755 | && TREE_CODE (op) == SSA_NAME | |
2756 | && simple_iv (loop, loop_containing_stmt (stmt), op, | |
2757 | &iv, false) | |
2758 | && tree_fits_shwi_p (iv.step)) | |
0136f8f0 AH |
2759 | { |
2760 | thisarginfo.linear_step = tree_to_shwi (iv.step); | |
2761 | thisarginfo.op = iv.base; | |
2762 | } | |
2763 | else if ((thisarginfo.dt == vect_constant_def | |
2764 | || thisarginfo.dt == vect_external_def) | |
2765 | && POINTER_TYPE_P (TREE_TYPE (op))) | |
2766 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; | |
17b658af JJ |
2767 | /* Addresses of array elements indexed by GOMP_SIMD_LANE are |
2768 | linear too. */ | |
2769 | if (POINTER_TYPE_P (TREE_TYPE (op)) | |
2770 | && !thisarginfo.linear_step | |
2771 | && !vec_stmt | |
2772 | && thisarginfo.dt != vect_constant_def | |
2773 | && thisarginfo.dt != vect_external_def | |
2774 | && loop_vinfo | |
2775 | && !slp_node | |
2776 | && TREE_CODE (op) == SSA_NAME) | |
2777 | vect_simd_lane_linear (op, loop, &thisarginfo); | |
0136f8f0 AH |
2778 | |
2779 | arginfo.quick_push (thisarginfo); | |
2780 | } | |
2781 | ||
2782 | unsigned int badness = 0; | |
2783 | struct cgraph_node *bestn = NULL; | |
6c9e85fb JJ |
2784 | if (STMT_VINFO_SIMD_CLONE_INFO (stmt_info).exists ()) |
2785 | bestn = cgraph_node::get (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[0]); | |
0136f8f0 AH |
2786 | else |
2787 | for (struct cgraph_node *n = node->simd_clones; n != NULL; | |
2788 | n = n->simdclone->next_clone) | |
2789 | { | |
2790 | unsigned int this_badness = 0; | |
2791 | if (n->simdclone->simdlen | |
2792 | > (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
2793 | || n->simdclone->nargs != nargs) | |
2794 | continue; | |
2795 | if (n->simdclone->simdlen | |
2796 | < (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
2797 | this_badness += (exact_log2 (LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
2798 | - exact_log2 (n->simdclone->simdlen)) * 1024; | |
2799 | if (n->simdclone->inbranch) | |
2800 | this_badness += 2048; | |
2801 | int target_badness = targetm.simd_clone.usable (n); | |
2802 | if (target_badness < 0) | |
2803 | continue; | |
2804 | this_badness += target_badness * 512; | |
2805 | /* FORNOW: Have to add code to add the mask argument. */ | |
2806 | if (n->simdclone->inbranch) | |
2807 | continue; | |
2808 | for (i = 0; i < nargs; i++) | |
2809 | { | |
2810 | switch (n->simdclone->args[i].arg_type) | |
2811 | { | |
2812 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
2813 | if (!useless_type_conversion_p | |
2814 | (n->simdclone->args[i].orig_type, | |
2815 | TREE_TYPE (gimple_call_arg (stmt, i)))) | |
2816 | i = -1; | |
2817 | else if (arginfo[i].dt == vect_constant_def | |
2818 | || arginfo[i].dt == vect_external_def | |
2819 | || arginfo[i].linear_step) | |
2820 | this_badness += 64; | |
2821 | break; | |
2822 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
2823 | if (arginfo[i].dt != vect_constant_def | |
2824 | && arginfo[i].dt != vect_external_def) | |
2825 | i = -1; | |
2826 | break; | |
2827 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
d9a6bd32 | 2828 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
2829 | if (arginfo[i].dt == vect_constant_def |
2830 | || arginfo[i].dt == vect_external_def | |
2831 | || (arginfo[i].linear_step | |
2832 | != n->simdclone->args[i].linear_step)) | |
2833 | i = -1; | |
2834 | break; | |
2835 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
d9a6bd32 JJ |
2836 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
2837 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
e01d41e5 JJ |
2838 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
2839 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
2840 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
2841 | /* FORNOW */ |
2842 | i = -1; | |
2843 | break; | |
2844 | case SIMD_CLONE_ARG_TYPE_MASK: | |
2845 | gcc_unreachable (); | |
2846 | } | |
2847 | if (i == (size_t) -1) | |
2848 | break; | |
2849 | if (n->simdclone->args[i].alignment > arginfo[i].align) | |
2850 | { | |
2851 | i = -1; | |
2852 | break; | |
2853 | } | |
2854 | if (arginfo[i].align) | |
2855 | this_badness += (exact_log2 (arginfo[i].align) | |
2856 | - exact_log2 (n->simdclone->args[i].alignment)); | |
2857 | } | |
2858 | if (i == (size_t) -1) | |
2859 | continue; | |
2860 | if (bestn == NULL || this_badness < badness) | |
2861 | { | |
2862 | bestn = n; | |
2863 | badness = this_badness; | |
2864 | } | |
2865 | } | |
2866 | ||
2867 | if (bestn == NULL) | |
2868 | { | |
2869 | arginfo.release (); | |
2870 | return false; | |
2871 | } | |
2872 | ||
2873 | for (i = 0; i < nargs; i++) | |
2874 | if ((arginfo[i].dt == vect_constant_def | |
2875 | || arginfo[i].dt == vect_external_def) | |
2876 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
2877 | { | |
2878 | arginfo[i].vectype | |
2879 | = get_vectype_for_scalar_type (TREE_TYPE (gimple_call_arg (stmt, | |
2880 | i))); | |
2881 | if (arginfo[i].vectype == NULL | |
2882 | || (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
2883 | > bestn->simdclone->simdlen)) | |
2884 | { | |
2885 | arginfo.release (); | |
2886 | return false; | |
2887 | } | |
2888 | } | |
2889 | ||
2890 | fndecl = bestn->decl; | |
2891 | nunits = bestn->simdclone->simdlen; | |
2892 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
2893 | ||
2894 | /* If the function isn't const, only allow it in simd loops where user | |
2895 | has asserted that at least nunits consecutive iterations can be | |
2896 | performed using SIMD instructions. */ | |
2897 | if ((loop == NULL || (unsigned) loop->safelen < nunits) | |
2898 | && gimple_vuse (stmt)) | |
2899 | { | |
2900 | arginfo.release (); | |
2901 | return false; | |
2902 | } | |
2903 | ||
2904 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
2905 | needs to be generated. */ | |
2906 | gcc_assert (ncopies >= 1); | |
2907 | ||
2908 | if (!vec_stmt) /* transformation not required. */ | |
2909 | { | |
6c9e85fb JJ |
2910 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (bestn->decl); |
2911 | for (i = 0; i < nargs; i++) | |
2912 | if (bestn->simdclone->args[i].arg_type | |
2913 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
2914 | { | |
17b658af | 2915 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_grow_cleared (i * 3 |
6c9e85fb JJ |
2916 | + 1); |
2917 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (arginfo[i].op); | |
2918 | tree lst = POINTER_TYPE_P (TREE_TYPE (arginfo[i].op)) | |
2919 | ? size_type_node : TREE_TYPE (arginfo[i].op); | |
2920 | tree ls = build_int_cst (lst, arginfo[i].linear_step); | |
2921 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (ls); | |
17b658af JJ |
2922 | tree sll = arginfo[i].simd_lane_linear |
2923 | ? boolean_true_node : boolean_false_node; | |
2924 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (sll); | |
6c9e85fb | 2925 | } |
0136f8f0 AH |
2926 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; |
2927 | if (dump_enabled_p ()) | |
2928 | dump_printf_loc (MSG_NOTE, vect_location, | |
2929 | "=== vectorizable_simd_clone_call ===\n"); | |
2930 | /* vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); */ | |
2931 | arginfo.release (); | |
2932 | return true; | |
2933 | } | |
2934 | ||
2935 | /** Transform. **/ | |
2936 | ||
2937 | if (dump_enabled_p ()) | |
2938 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); | |
2939 | ||
2940 | /* Handle def. */ | |
2941 | scalar_dest = gimple_call_lhs (stmt); | |
2942 | vec_dest = NULL_TREE; | |
2943 | rtype = NULL_TREE; | |
2944 | ratype = NULL_TREE; | |
2945 | if (scalar_dest) | |
2946 | { | |
2947 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
2948 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); | |
2949 | if (TREE_CODE (rtype) == ARRAY_TYPE) | |
2950 | { | |
2951 | ratype = rtype; | |
2952 | rtype = TREE_TYPE (ratype); | |
2953 | } | |
2954 | } | |
2955 | ||
2956 | prev_stmt_info = NULL; | |
2957 | for (j = 0; j < ncopies; ++j) | |
2958 | { | |
2959 | /* Build argument list for the vectorized call. */ | |
2960 | if (j == 0) | |
2961 | vargs.create (nargs); | |
2962 | else | |
2963 | vargs.truncate (0); | |
2964 | ||
2965 | for (i = 0; i < nargs; i++) | |
2966 | { | |
2967 | unsigned int k, l, m, o; | |
2968 | tree atype; | |
2969 | op = gimple_call_arg (stmt, i); | |
2970 | switch (bestn->simdclone->args[i].arg_type) | |
2971 | { | |
2972 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
2973 | atype = bestn->simdclone->args[i].vector_type; | |
2974 | o = nunits / TYPE_VECTOR_SUBPARTS (atype); | |
2975 | for (m = j * o; m < (j + 1) * o; m++) | |
2976 | { | |
2977 | if (TYPE_VECTOR_SUBPARTS (atype) | |
2978 | < TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)) | |
2979 | { | |
2980 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); | |
2981 | k = (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
2982 | / TYPE_VECTOR_SUBPARTS (atype)); | |
2983 | gcc_assert ((k & (k - 1)) == 0); | |
2984 | if (m == 0) | |
2985 | vec_oprnd0 | |
81c40241 | 2986 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
2987 | else |
2988 | { | |
2989 | vec_oprnd0 = arginfo[i].op; | |
2990 | if ((m & (k - 1)) == 0) | |
2991 | vec_oprnd0 | |
2992 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
2993 | vec_oprnd0); | |
2994 | } | |
2995 | arginfo[i].op = vec_oprnd0; | |
2996 | vec_oprnd0 | |
2997 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, | |
2998 | size_int (prec), | |
2999 | bitsize_int ((m & (k - 1)) * prec)); | |
3000 | new_stmt | |
b731b390 | 3001 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3002 | vec_oprnd0); |
3003 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3004 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3005 | } | |
3006 | else | |
3007 | { | |
3008 | k = (TYPE_VECTOR_SUBPARTS (atype) | |
3009 | / TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)); | |
3010 | gcc_assert ((k & (k - 1)) == 0); | |
3011 | vec<constructor_elt, va_gc> *ctor_elts; | |
3012 | if (k != 1) | |
3013 | vec_alloc (ctor_elts, k); | |
3014 | else | |
3015 | ctor_elts = NULL; | |
3016 | for (l = 0; l < k; l++) | |
3017 | { | |
3018 | if (m == 0 && l == 0) | |
3019 | vec_oprnd0 | |
81c40241 | 3020 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3021 | else |
3022 | vec_oprnd0 | |
3023 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3024 | arginfo[i].op); | |
3025 | arginfo[i].op = vec_oprnd0; | |
3026 | if (k == 1) | |
3027 | break; | |
3028 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, | |
3029 | vec_oprnd0); | |
3030 | } | |
3031 | if (k == 1) | |
3032 | vargs.safe_push (vec_oprnd0); | |
3033 | else | |
3034 | { | |
3035 | vec_oprnd0 = build_constructor (atype, ctor_elts); | |
3036 | new_stmt | |
b731b390 | 3037 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3038 | vec_oprnd0); |
3039 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3040 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3041 | } | |
3042 | } | |
3043 | } | |
3044 | break; | |
3045 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3046 | vargs.safe_push (op); | |
3047 | break; | |
3048 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
3049 | if (j == 0) | |
3050 | { | |
3051 | gimple_seq stmts; | |
3052 | arginfo[i].op | |
3053 | = force_gimple_operand (arginfo[i].op, &stmts, true, | |
3054 | NULL_TREE); | |
3055 | if (stmts != NULL) | |
3056 | { | |
3057 | basic_block new_bb; | |
3058 | edge pe = loop_preheader_edge (loop); | |
3059 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
3060 | gcc_assert (!new_bb); | |
3061 | } | |
17b658af JJ |
3062 | if (arginfo[i].simd_lane_linear) |
3063 | { | |
3064 | vargs.safe_push (arginfo[i].op); | |
3065 | break; | |
3066 | } | |
b731b390 | 3067 | tree phi_res = copy_ssa_name (op); |
538dd0b7 | 3068 | gphi *new_phi = create_phi_node (phi_res, loop->header); |
0136f8f0 | 3069 | set_vinfo_for_stmt (new_phi, |
310213d4 | 3070 | new_stmt_vec_info (new_phi, loop_vinfo)); |
0136f8f0 AH |
3071 | add_phi_arg (new_phi, arginfo[i].op, |
3072 | loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
3073 | enum tree_code code | |
3074 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3075 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3076 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3077 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3078 | widest_int cst |
3079 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3080 | ncopies * nunits); | |
3081 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3082 | tree phi_arg = copy_ssa_name (op); |
0d0e4a03 JJ |
3083 | new_stmt |
3084 | = gimple_build_assign (phi_arg, code, phi_res, tcst); | |
0136f8f0 AH |
3085 | gimple_stmt_iterator si = gsi_after_labels (loop->header); |
3086 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); | |
3087 | set_vinfo_for_stmt (new_stmt, | |
310213d4 | 3088 | new_stmt_vec_info (new_stmt, loop_vinfo)); |
0136f8f0 AH |
3089 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), |
3090 | UNKNOWN_LOCATION); | |
3091 | arginfo[i].op = phi_res; | |
3092 | vargs.safe_push (phi_res); | |
3093 | } | |
3094 | else | |
3095 | { | |
3096 | enum tree_code code | |
3097 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3098 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3099 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3100 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3101 | widest_int cst |
3102 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3103 | j * nunits); | |
3104 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3105 | new_temp = make_ssa_name (TREE_TYPE (op)); |
0d0e4a03 JJ |
3106 | new_stmt = gimple_build_assign (new_temp, code, |
3107 | arginfo[i].op, tcst); | |
0136f8f0 AH |
3108 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3109 | vargs.safe_push (new_temp); | |
3110 | } | |
3111 | break; | |
3112 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
e01d41e5 JJ |
3113 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3114 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3115 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3116 | default: |
3117 | gcc_unreachable (); | |
3118 | } | |
3119 | } | |
3120 | ||
3121 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
3122 | if (vec_dest) | |
3123 | { | |
3124 | gcc_assert (ratype || TYPE_VECTOR_SUBPARTS (rtype) == nunits); | |
3125 | if (ratype) | |
b731b390 | 3126 | new_temp = create_tmp_var (ratype); |
0136f8f0 AH |
3127 | else if (TYPE_VECTOR_SUBPARTS (vectype) |
3128 | == TYPE_VECTOR_SUBPARTS (rtype)) | |
3129 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3130 | else | |
3131 | new_temp = make_ssa_name (rtype, new_stmt); | |
3132 | gimple_call_set_lhs (new_stmt, new_temp); | |
3133 | } | |
3134 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3135 | ||
3136 | if (vec_dest) | |
3137 | { | |
3138 | if (TYPE_VECTOR_SUBPARTS (vectype) < nunits) | |
3139 | { | |
3140 | unsigned int k, l; | |
3141 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); | |
3142 | k = nunits / TYPE_VECTOR_SUBPARTS (vectype); | |
3143 | gcc_assert ((k & (k - 1)) == 0); | |
3144 | for (l = 0; l < k; l++) | |
3145 | { | |
3146 | tree t; | |
3147 | if (ratype) | |
3148 | { | |
3149 | t = build_fold_addr_expr (new_temp); | |
3150 | t = build2 (MEM_REF, vectype, t, | |
3151 | build_int_cst (TREE_TYPE (t), | |
3152 | l * prec / BITS_PER_UNIT)); | |
3153 | } | |
3154 | else | |
3155 | t = build3 (BIT_FIELD_REF, vectype, new_temp, | |
3156 | size_int (prec), bitsize_int (l * prec)); | |
3157 | new_stmt | |
b731b390 | 3158 | = gimple_build_assign (make_ssa_name (vectype), t); |
0136f8f0 AH |
3159 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3160 | if (j == 0 && l == 0) | |
3161 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3162 | else | |
3163 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3164 | ||
3165 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3166 | } | |
3167 | ||
3168 | if (ratype) | |
3169 | { | |
3170 | tree clobber = build_constructor (ratype, NULL); | |
3171 | TREE_THIS_VOLATILE (clobber) = 1; | |
3172 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3173 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3174 | } | |
3175 | continue; | |
3176 | } | |
3177 | else if (TYPE_VECTOR_SUBPARTS (vectype) > nunits) | |
3178 | { | |
3179 | unsigned int k = (TYPE_VECTOR_SUBPARTS (vectype) | |
3180 | / TYPE_VECTOR_SUBPARTS (rtype)); | |
3181 | gcc_assert ((k & (k - 1)) == 0); | |
3182 | if ((j & (k - 1)) == 0) | |
3183 | vec_alloc (ret_ctor_elts, k); | |
3184 | if (ratype) | |
3185 | { | |
3186 | unsigned int m, o = nunits / TYPE_VECTOR_SUBPARTS (rtype); | |
3187 | for (m = 0; m < o; m++) | |
3188 | { | |
3189 | tree tem = build4 (ARRAY_REF, rtype, new_temp, | |
3190 | size_int (m), NULL_TREE, NULL_TREE); | |
3191 | new_stmt | |
b731b390 | 3192 | = gimple_build_assign (make_ssa_name (rtype), tem); |
0136f8f0 AH |
3193 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3194 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, | |
3195 | gimple_assign_lhs (new_stmt)); | |
3196 | } | |
3197 | tree clobber = build_constructor (ratype, NULL); | |
3198 | TREE_THIS_VOLATILE (clobber) = 1; | |
3199 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3200 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3201 | } | |
3202 | else | |
3203 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); | |
3204 | if ((j & (k - 1)) != k - 1) | |
3205 | continue; | |
3206 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); | |
3207 | new_stmt | |
b731b390 | 3208 | = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0); |
0136f8f0 AH |
3209 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3210 | ||
3211 | if ((unsigned) j == k - 1) | |
3212 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3213 | else | |
3214 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3215 | ||
3216 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3217 | continue; | |
3218 | } | |
3219 | else if (ratype) | |
3220 | { | |
3221 | tree t = build_fold_addr_expr (new_temp); | |
3222 | t = build2 (MEM_REF, vectype, t, | |
3223 | build_int_cst (TREE_TYPE (t), 0)); | |
3224 | new_stmt | |
b731b390 | 3225 | = gimple_build_assign (make_ssa_name (vec_dest), t); |
0136f8f0 AH |
3226 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3227 | tree clobber = build_constructor (ratype, NULL); | |
3228 | TREE_THIS_VOLATILE (clobber) = 1; | |
3229 | vect_finish_stmt_generation (stmt, | |
3230 | gimple_build_assign (new_temp, | |
3231 | clobber), gsi); | |
3232 | } | |
3233 | } | |
3234 | ||
3235 | if (j == 0) | |
3236 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3237 | else | |
3238 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3239 | ||
3240 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3241 | } | |
3242 | ||
3243 | vargs.release (); | |
3244 | ||
3245 | /* The call in STMT might prevent it from being removed in dce. | |
3246 | We however cannot remove it here, due to the way the ssa name | |
3247 | it defines is mapped to the new definition. So just replace | |
3248 | rhs of the statement with something harmless. */ | |
3249 | ||
3250 | if (slp_node) | |
3251 | return true; | |
3252 | ||
3253 | if (scalar_dest) | |
3254 | { | |
3255 | type = TREE_TYPE (scalar_dest); | |
3256 | if (is_pattern_stmt_p (stmt_info)) | |
3257 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
3258 | else | |
3259 | lhs = gimple_call_lhs (stmt); | |
3260 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
3261 | } | |
3262 | else | |
3263 | new_stmt = gimple_build_nop (); | |
3264 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
3265 | set_vinfo_for_stmt (stmt, NULL); | |
3266 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2865f32a | 3267 | gsi_replace (gsi, new_stmt, true); |
0136f8f0 AH |
3268 | unlink_stmt_vdef (stmt); |
3269 | ||
3270 | return true; | |
3271 | } | |
3272 | ||
3273 | ||
ebfd146a IR |
3274 | /* Function vect_gen_widened_results_half |
3275 | ||
3276 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 3277 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 3278 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
3279 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
3280 | needs to be created (DECL is a function-decl of a target-builtin). | |
3281 | STMT is the original scalar stmt that we are vectorizing. */ | |
3282 | ||
355fe088 | 3283 | static gimple * |
ebfd146a IR |
3284 | vect_gen_widened_results_half (enum tree_code code, |
3285 | tree decl, | |
3286 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
3287 | tree vec_dest, gimple_stmt_iterator *gsi, | |
355fe088 | 3288 | gimple *stmt) |
b8698a0f | 3289 | { |
355fe088 | 3290 | gimple *new_stmt; |
b8698a0f L |
3291 | tree new_temp; |
3292 | ||
3293 | /* Generate half of the widened result: */ | |
3294 | if (code == CALL_EXPR) | |
3295 | { | |
3296 | /* Target specific support */ | |
ebfd146a IR |
3297 | if (op_type == binary_op) |
3298 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
3299 | else | |
3300 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
3301 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3302 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
3303 | } |
3304 | else | |
ebfd146a | 3305 | { |
b8698a0f L |
3306 | /* Generic support */ |
3307 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
3308 | if (op_type != binary_op) |
3309 | vec_oprnd1 = NULL; | |
0d0e4a03 | 3310 | new_stmt = gimple_build_assign (vec_dest, code, vec_oprnd0, vec_oprnd1); |
ebfd146a IR |
3311 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3312 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 3313 | } |
ebfd146a IR |
3314 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3315 | ||
ebfd146a IR |
3316 | return new_stmt; |
3317 | } | |
3318 | ||
4a00c761 JJ |
3319 | |
3320 | /* Get vectorized definitions for loop-based vectorization. For the first | |
3321 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
3322 | scalar operand), and for the rest we get a copy with | |
3323 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
3324 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
3325 | The vectors are collected into VEC_OPRNDS. */ | |
3326 | ||
3327 | static void | |
355fe088 | 3328 | vect_get_loop_based_defs (tree *oprnd, gimple *stmt, enum vect_def_type dt, |
9771b263 | 3329 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
3330 | { |
3331 | tree vec_oprnd; | |
3332 | ||
3333 | /* Get first vector operand. */ | |
3334 | /* All the vector operands except the very first one (that is scalar oprnd) | |
3335 | are stmt copies. */ | |
3336 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
81c40241 | 3337 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt); |
4a00c761 JJ |
3338 | else |
3339 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
3340 | ||
9771b263 | 3341 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3342 | |
3343 | /* Get second vector operand. */ | |
3344 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 3345 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3346 | |
3347 | *oprnd = vec_oprnd; | |
3348 | ||
3349 | /* For conversion in multiple steps, continue to get operands | |
3350 | recursively. */ | |
3351 | if (multi_step_cvt) | |
3352 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
3353 | } | |
3354 | ||
3355 | ||
3356 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
3357 | For multi-step conversions store the resulting vectors and call the function | |
3358 | recursively. */ | |
3359 | ||
3360 | static void | |
9771b263 | 3361 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
355fe088 | 3362 | int multi_step_cvt, gimple *stmt, |
9771b263 | 3363 | vec<tree> vec_dsts, |
4a00c761 JJ |
3364 | gimple_stmt_iterator *gsi, |
3365 | slp_tree slp_node, enum tree_code code, | |
3366 | stmt_vec_info *prev_stmt_info) | |
3367 | { | |
3368 | unsigned int i; | |
3369 | tree vop0, vop1, new_tmp, vec_dest; | |
355fe088 | 3370 | gimple *new_stmt; |
4a00c761 JJ |
3371 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
3372 | ||
9771b263 | 3373 | vec_dest = vec_dsts.pop (); |
4a00c761 | 3374 | |
9771b263 | 3375 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
3376 | { |
3377 | /* Create demotion operation. */ | |
9771b263 DN |
3378 | vop0 = (*vec_oprnds)[i]; |
3379 | vop1 = (*vec_oprnds)[i + 1]; | |
0d0e4a03 | 3380 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
4a00c761 JJ |
3381 | new_tmp = make_ssa_name (vec_dest, new_stmt); |
3382 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
3383 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3384 | ||
3385 | if (multi_step_cvt) | |
3386 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 3387 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
3388 | else |
3389 | { | |
3390 | /* This is the last step of the conversion sequence. Store the | |
3391 | vectors in SLP_NODE or in vector info of the scalar statement | |
3392 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
3393 | if (slp_node) | |
9771b263 | 3394 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 3395 | else |
c689ce1e RB |
3396 | { |
3397 | if (!*prev_stmt_info) | |
3398 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3399 | else | |
3400 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
4a00c761 | 3401 | |
c689ce1e RB |
3402 | *prev_stmt_info = vinfo_for_stmt (new_stmt); |
3403 | } | |
4a00c761 JJ |
3404 | } |
3405 | } | |
3406 | ||
3407 | /* For multi-step demotion operations we first generate demotion operations | |
3408 | from the source type to the intermediate types, and then combine the | |
3409 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
3410 | type. */ | |
3411 | if (multi_step_cvt) | |
3412 | { | |
3413 | /* At each level of recursion we have half of the operands we had at the | |
3414 | previous level. */ | |
9771b263 | 3415 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
3416 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
3417 | stmt, vec_dsts, gsi, slp_node, | |
3418 | VEC_PACK_TRUNC_EXPR, | |
3419 | prev_stmt_info); | |
3420 | } | |
3421 | ||
9771b263 | 3422 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3423 | } |
3424 | ||
3425 | ||
3426 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
3427 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
3428 | the resulting vectors and call the function recursively. */ | |
3429 | ||
3430 | static void | |
9771b263 DN |
3431 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
3432 | vec<tree> *vec_oprnds1, | |
355fe088 | 3433 | gimple *stmt, tree vec_dest, |
4a00c761 JJ |
3434 | gimple_stmt_iterator *gsi, |
3435 | enum tree_code code1, | |
3436 | enum tree_code code2, tree decl1, | |
3437 | tree decl2, int op_type) | |
3438 | { | |
3439 | int i; | |
3440 | tree vop0, vop1, new_tmp1, new_tmp2; | |
355fe088 | 3441 | gimple *new_stmt1, *new_stmt2; |
6e1aa848 | 3442 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 3443 | |
9771b263 DN |
3444 | vec_tmp.create (vec_oprnds0->length () * 2); |
3445 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
3446 | { |
3447 | if (op_type == binary_op) | |
9771b263 | 3448 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
3449 | else |
3450 | vop1 = NULL_TREE; | |
3451 | ||
3452 | /* Generate the two halves of promotion operation. */ | |
3453 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
3454 | op_type, vec_dest, gsi, stmt); | |
3455 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
3456 | op_type, vec_dest, gsi, stmt); | |
3457 | if (is_gimple_call (new_stmt1)) | |
3458 | { | |
3459 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
3460 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
3461 | } | |
3462 | else | |
3463 | { | |
3464 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
3465 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
3466 | } | |
3467 | ||
3468 | /* Store the results for the next step. */ | |
9771b263 DN |
3469 | vec_tmp.quick_push (new_tmp1); |
3470 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
3471 | } |
3472 | ||
689eaba3 | 3473 | vec_oprnds0->release (); |
4a00c761 JJ |
3474 | *vec_oprnds0 = vec_tmp; |
3475 | } | |
3476 | ||
3477 | ||
b8698a0f L |
3478 | /* Check if STMT performs a conversion operation, that can be vectorized. |
3479 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 3480 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
3481 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
3482 | ||
3483 | static bool | |
355fe088 TS |
3484 | vectorizable_conversion (gimple *stmt, gimple_stmt_iterator *gsi, |
3485 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
3486 | { |
3487 | tree vec_dest; | |
3488 | tree scalar_dest; | |
4a00c761 | 3489 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
3490 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
3491 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3492 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3493 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 3494 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
3495 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
3496 | tree new_temp; | |
355fe088 | 3497 | gimple *def_stmt; |
ebfd146a | 3498 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
355fe088 | 3499 | gimple *new_stmt = NULL; |
ebfd146a IR |
3500 | stmt_vec_info prev_stmt_info; |
3501 | int nunits_in; | |
3502 | int nunits_out; | |
3503 | tree vectype_out, vectype_in; | |
4a00c761 JJ |
3504 | int ncopies, i, j; |
3505 | tree lhs_type, rhs_type; | |
ebfd146a | 3506 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
3507 | vec<tree> vec_oprnds0 = vNULL; |
3508 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 3509 | tree vop0; |
4a00c761 | 3510 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 3511 | vec_info *vinfo = stmt_info->vinfo; |
4a00c761 | 3512 | int multi_step_cvt = 0; |
6e1aa848 DN |
3513 | vec<tree> vec_dsts = vNULL; |
3514 | vec<tree> interm_types = vNULL; | |
4a00c761 JJ |
3515 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
3516 | int op_type; | |
ef4bddc2 | 3517 | machine_mode rhs_mode; |
4a00c761 | 3518 | unsigned short fltsz; |
ebfd146a IR |
3519 | |
3520 | /* Is STMT a vectorizable conversion? */ | |
3521 | ||
4a00c761 | 3522 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3523 | return false; |
3524 | ||
8644a673 | 3525 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
3526 | return false; |
3527 | ||
3528 | if (!is_gimple_assign (stmt)) | |
3529 | return false; | |
3530 | ||
3531 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3532 | return false; | |
3533 | ||
3534 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
3535 | if (!CONVERT_EXPR_CODE_P (code) |
3536 | && code != FIX_TRUNC_EXPR | |
3537 | && code != FLOAT_EXPR | |
3538 | && code != WIDEN_MULT_EXPR | |
3539 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
3540 | return false; |
3541 | ||
4a00c761 JJ |
3542 | op_type = TREE_CODE_LENGTH (code); |
3543 | ||
ebfd146a | 3544 | /* Check types of lhs and rhs. */ |
b690cc0f | 3545 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 3546 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
3547 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
3548 | ||
ebfd146a IR |
3549 | op0 = gimple_assign_rhs1 (stmt); |
3550 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
3551 | |
3552 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
3553 | && !((INTEGRAL_TYPE_P (lhs_type) | |
3554 | && INTEGRAL_TYPE_P (rhs_type)) | |
3555 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
3556 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
3557 | return false; | |
3558 | ||
3559 | if ((INTEGRAL_TYPE_P (lhs_type) | |
3560 | && (TYPE_PRECISION (lhs_type) | |
3561 | != GET_MODE_PRECISION (TYPE_MODE (lhs_type)))) | |
3562 | || (INTEGRAL_TYPE_P (rhs_type) | |
3563 | && (TYPE_PRECISION (rhs_type) | |
3564 | != GET_MODE_PRECISION (TYPE_MODE (rhs_type))))) | |
3565 | { | |
73fbfcad | 3566 | if (dump_enabled_p ()) |
78c60e3d | 3567 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
3568 | "type conversion to/from bit-precision unsupported." |
3569 | "\n"); | |
4a00c761 JJ |
3570 | return false; |
3571 | } | |
3572 | ||
b690cc0f | 3573 | /* Check the operands of the operation. */ |
81c40241 | 3574 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype_in)) |
b690cc0f | 3575 | { |
73fbfcad | 3576 | if (dump_enabled_p ()) |
78c60e3d | 3577 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3578 | "use not simple.\n"); |
b690cc0f RG |
3579 | return false; |
3580 | } | |
4a00c761 JJ |
3581 | if (op_type == binary_op) |
3582 | { | |
3583 | bool ok; | |
3584 | ||
3585 | op1 = gimple_assign_rhs2 (stmt); | |
3586 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
3587 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
3588 | OP1. */ | |
3589 | if (CONSTANT_CLASS_P (op0)) | |
81c40241 | 3590 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &vectype_in); |
4a00c761 | 3591 | else |
81c40241 | 3592 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1]); |
4a00c761 JJ |
3593 | |
3594 | if (!ok) | |
3595 | { | |
73fbfcad | 3596 | if (dump_enabled_p ()) |
78c60e3d | 3597 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3598 | "use not simple.\n"); |
4a00c761 JJ |
3599 | return false; |
3600 | } | |
3601 | } | |
3602 | ||
b690cc0f RG |
3603 | /* If op0 is an external or constant defs use a vector type of |
3604 | the same size as the output vector type. */ | |
ebfd146a | 3605 | if (!vectype_in) |
b690cc0f | 3606 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
3607 | if (vec_stmt) |
3608 | gcc_assert (vectype_in); | |
3609 | if (!vectype_in) | |
3610 | { | |
73fbfcad | 3611 | if (dump_enabled_p ()) |
4a00c761 | 3612 | { |
78c60e3d SS |
3613 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3614 | "no vectype for scalar type "); | |
3615 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 3616 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 3617 | } |
7d8930a0 IR |
3618 | |
3619 | return false; | |
3620 | } | |
ebfd146a | 3621 | |
b690cc0f RG |
3622 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
3623 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4a00c761 | 3624 | if (nunits_in < nunits_out) |
ebfd146a IR |
3625 | modifier = NARROW; |
3626 | else if (nunits_out == nunits_in) | |
3627 | modifier = NONE; | |
ebfd146a | 3628 | else |
4a00c761 | 3629 | modifier = WIDEN; |
ebfd146a | 3630 | |
ff802fa1 IR |
3631 | /* Multiple types in SLP are handled by creating the appropriate number of |
3632 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3633 | case of SLP. */ | |
437f4a00 | 3634 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
ebfd146a | 3635 | ncopies = 1; |
4a00c761 JJ |
3636 | else if (modifier == NARROW) |
3637 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
3638 | else | |
3639 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
b8698a0f | 3640 | |
ebfd146a IR |
3641 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
3642 | needs to be generated. */ | |
3643 | gcc_assert (ncopies >= 1); | |
3644 | ||
ebfd146a | 3645 | /* Supportable by target? */ |
4a00c761 | 3646 | switch (modifier) |
ebfd146a | 3647 | { |
4a00c761 JJ |
3648 | case NONE: |
3649 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
3650 | return false; | |
3651 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
3652 | &decl1, &code1)) | |
3653 | break; | |
3654 | /* FALLTHRU */ | |
3655 | unsupported: | |
73fbfcad | 3656 | if (dump_enabled_p ()) |
78c60e3d | 3657 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3658 | "conversion not supported by target.\n"); |
ebfd146a | 3659 | return false; |
ebfd146a | 3660 | |
4a00c761 JJ |
3661 | case WIDEN: |
3662 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
3663 | &code1, &code2, &multi_step_cvt, |
3664 | &interm_types)) | |
4a00c761 JJ |
3665 | { |
3666 | /* Binary widening operation can only be supported directly by the | |
3667 | architecture. */ | |
3668 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
3669 | break; | |
3670 | } | |
3671 | ||
3672 | if (code != FLOAT_EXPR | |
3673 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
3674 | <= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
3675 | goto unsupported; | |
3676 | ||
3677 | rhs_mode = TYPE_MODE (rhs_type); | |
3678 | fltsz = GET_MODE_SIZE (TYPE_MODE (lhs_type)); | |
3679 | for (rhs_mode = GET_MODE_2XWIDER_MODE (TYPE_MODE (rhs_type)); | |
3680 | rhs_mode != VOIDmode && GET_MODE_SIZE (rhs_mode) <= fltsz; | |
3681 | rhs_mode = GET_MODE_2XWIDER_MODE (rhs_mode)) | |
3682 | { | |
3683 | cvt_type | |
3684 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
3685 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
3686 | if (cvt_type == NULL_TREE) | |
3687 | goto unsupported; | |
3688 | ||
3689 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
3690 | { | |
3691 | if (!supportable_convert_operation (code, vectype_out, | |
3692 | cvt_type, &decl1, &codecvt1)) | |
3693 | goto unsupported; | |
3694 | } | |
3695 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
3696 | cvt_type, &codecvt1, |
3697 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
3698 | &interm_types)) |
3699 | continue; | |
3700 | else | |
3701 | gcc_assert (multi_step_cvt == 0); | |
3702 | ||
3703 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
3704 | vectype_in, &code1, &code2, |
3705 | &multi_step_cvt, &interm_types)) | |
4a00c761 JJ |
3706 | break; |
3707 | } | |
3708 | ||
3709 | if (rhs_mode == VOIDmode || GET_MODE_SIZE (rhs_mode) > fltsz) | |
3710 | goto unsupported; | |
3711 | ||
3712 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
3713 | codecvt2 = ERROR_MARK; | |
3714 | else | |
3715 | { | |
3716 | multi_step_cvt++; | |
9771b263 | 3717 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
3718 | cvt_type = NULL_TREE; |
3719 | } | |
3720 | break; | |
3721 | ||
3722 | case NARROW: | |
3723 | gcc_assert (op_type == unary_op); | |
3724 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
3725 | &code1, &multi_step_cvt, | |
3726 | &interm_types)) | |
3727 | break; | |
3728 | ||
3729 | if (code != FIX_TRUNC_EXPR | |
3730 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
3731 | >= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
3732 | goto unsupported; | |
3733 | ||
3734 | rhs_mode = TYPE_MODE (rhs_type); | |
3735 | cvt_type | |
3736 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
3737 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
3738 | if (cvt_type == NULL_TREE) | |
3739 | goto unsupported; | |
3740 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
3741 | &decl1, &codecvt1)) | |
3742 | goto unsupported; | |
3743 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
3744 | &code1, &multi_step_cvt, | |
3745 | &interm_types)) | |
3746 | break; | |
3747 | goto unsupported; | |
3748 | ||
3749 | default: | |
3750 | gcc_unreachable (); | |
ebfd146a IR |
3751 | } |
3752 | ||
3753 | if (!vec_stmt) /* transformation not required. */ | |
3754 | { | |
73fbfcad | 3755 | if (dump_enabled_p ()) |
78c60e3d | 3756 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3757 | "=== vectorizable_conversion ===\n"); |
4a00c761 | 3758 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
3759 | { |
3760 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
c3e7ee41 | 3761 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
8bd37302 | 3762 | } |
4a00c761 JJ |
3763 | else if (modifier == NARROW) |
3764 | { | |
3765 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
8bd37302 | 3766 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 JJ |
3767 | } |
3768 | else | |
3769 | { | |
3770 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
8bd37302 | 3771 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 | 3772 | } |
9771b263 | 3773 | interm_types.release (); |
ebfd146a IR |
3774 | return true; |
3775 | } | |
3776 | ||
3777 | /** Transform. **/ | |
73fbfcad | 3778 | if (dump_enabled_p ()) |
78c60e3d | 3779 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3780 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 3781 | |
4a00c761 JJ |
3782 | if (op_type == binary_op) |
3783 | { | |
3784 | if (CONSTANT_CLASS_P (op0)) | |
3785 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
3786 | else if (CONSTANT_CLASS_P (op1)) | |
3787 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
3788 | } | |
3789 | ||
3790 | /* In case of multi-step conversion, we first generate conversion operations | |
3791 | to the intermediate types, and then from that types to the final one. | |
3792 | We create vector destinations for the intermediate type (TYPES) received | |
3793 | from supportable_*_operation, and store them in the correct order | |
3794 | for future use in vect_create_vectorized_*_stmts (). */ | |
9771b263 | 3795 | vec_dsts.create (multi_step_cvt + 1); |
82294ec1 JJ |
3796 | vec_dest = vect_create_destination_var (scalar_dest, |
3797 | (cvt_type && modifier == WIDEN) | |
3798 | ? cvt_type : vectype_out); | |
9771b263 | 3799 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3800 | |
3801 | if (multi_step_cvt) | |
3802 | { | |
9771b263 DN |
3803 | for (i = interm_types.length () - 1; |
3804 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
3805 | { |
3806 | vec_dest = vect_create_destination_var (scalar_dest, | |
3807 | intermediate_type); | |
9771b263 | 3808 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3809 | } |
3810 | } | |
ebfd146a | 3811 | |
4a00c761 | 3812 | if (cvt_type) |
82294ec1 JJ |
3813 | vec_dest = vect_create_destination_var (scalar_dest, |
3814 | modifier == WIDEN | |
3815 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
3816 | |
3817 | if (!slp_node) | |
3818 | { | |
30862efc | 3819 | if (modifier == WIDEN) |
4a00c761 | 3820 | { |
c3284718 | 3821 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 3822 | if (op_type == binary_op) |
9771b263 | 3823 | vec_oprnds1.create (1); |
4a00c761 | 3824 | } |
30862efc | 3825 | else if (modifier == NARROW) |
9771b263 DN |
3826 | vec_oprnds0.create ( |
3827 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
3828 | } |
3829 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 3830 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 3831 | |
4a00c761 | 3832 | last_oprnd = op0; |
ebfd146a IR |
3833 | prev_stmt_info = NULL; |
3834 | switch (modifier) | |
3835 | { | |
3836 | case NONE: | |
3837 | for (j = 0; j < ncopies; j++) | |
3838 | { | |
ebfd146a | 3839 | if (j == 0) |
d092494c IR |
3840 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node, |
3841 | -1); | |
ebfd146a IR |
3842 | else |
3843 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
3844 | ||
9771b263 | 3845 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
3846 | { |
3847 | /* Arguments are ready, create the new vector stmt. */ | |
3848 | if (code1 == CALL_EXPR) | |
3849 | { | |
3850 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
3851 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3852 | gimple_call_set_lhs (new_stmt, new_temp); | |
3853 | } | |
3854 | else | |
3855 | { | |
3856 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
0d0e4a03 | 3857 | new_stmt = gimple_build_assign (vec_dest, code1, vop0); |
4a00c761 JJ |
3858 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3859 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3860 | } | |
3861 | ||
3862 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3863 | if (slp_node) | |
9771b263 | 3864 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
225ce44b RB |
3865 | else |
3866 | { | |
3867 | if (!prev_stmt_info) | |
3868 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3869 | else | |
3870 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3871 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3872 | } | |
4a00c761 | 3873 | } |
ebfd146a IR |
3874 | } |
3875 | break; | |
3876 | ||
3877 | case WIDEN: | |
3878 | /* In case the vectorization factor (VF) is bigger than the number | |
3879 | of elements that we can fit in a vectype (nunits), we have to | |
3880 | generate more than one vector stmt - i.e - we need to "unroll" | |
3881 | the vector stmt by a factor VF/nunits. */ | |
3882 | for (j = 0; j < ncopies; j++) | |
3883 | { | |
4a00c761 | 3884 | /* Handle uses. */ |
ebfd146a | 3885 | if (j == 0) |
4a00c761 JJ |
3886 | { |
3887 | if (slp_node) | |
3888 | { | |
3889 | if (code == WIDEN_LSHIFT_EXPR) | |
3890 | { | |
3891 | unsigned int k; | |
ebfd146a | 3892 | |
4a00c761 JJ |
3893 | vec_oprnd1 = op1; |
3894 | /* Store vec_oprnd1 for every vector stmt to be created | |
3895 | for SLP_NODE. We check during the analysis that all | |
3896 | the shift arguments are the same. */ | |
3897 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 3898 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
3899 | |
3900 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
3901 | slp_node, -1); | |
3902 | } | |
3903 | else | |
3904 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
3905 | &vec_oprnds1, slp_node, -1); | |
3906 | } | |
3907 | else | |
3908 | { | |
81c40241 | 3909 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 3910 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
3911 | if (op_type == binary_op) |
3912 | { | |
3913 | if (code == WIDEN_LSHIFT_EXPR) | |
3914 | vec_oprnd1 = op1; | |
3915 | else | |
81c40241 | 3916 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 3917 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
3918 | } |
3919 | } | |
3920 | } | |
ebfd146a | 3921 | else |
4a00c761 JJ |
3922 | { |
3923 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
3924 | vec_oprnds0.truncate (0); |
3925 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
3926 | if (op_type == binary_op) |
3927 | { | |
3928 | if (code == WIDEN_LSHIFT_EXPR) | |
3929 | vec_oprnd1 = op1; | |
3930 | else | |
3931 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
3932 | vec_oprnd1); | |
9771b263 DN |
3933 | vec_oprnds1.truncate (0); |
3934 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
3935 | } |
3936 | } | |
ebfd146a | 3937 | |
4a00c761 JJ |
3938 | /* Arguments are ready. Create the new vector stmts. */ |
3939 | for (i = multi_step_cvt; i >= 0; i--) | |
3940 | { | |
9771b263 | 3941 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
3942 | enum tree_code c1 = code1, c2 = code2; |
3943 | if (i == 0 && codecvt2 != ERROR_MARK) | |
3944 | { | |
3945 | c1 = codecvt1; | |
3946 | c2 = codecvt2; | |
3947 | } | |
3948 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
3949 | &vec_oprnds1, | |
3950 | stmt, this_dest, gsi, | |
3951 | c1, c2, decl1, decl2, | |
3952 | op_type); | |
3953 | } | |
3954 | ||
9771b263 | 3955 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
3956 | { |
3957 | if (cvt_type) | |
3958 | { | |
3959 | if (codecvt1 == CALL_EXPR) | |
3960 | { | |
3961 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
3962 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3963 | gimple_call_set_lhs (new_stmt, new_temp); | |
3964 | } | |
3965 | else | |
3966 | { | |
3967 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 3968 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
3969 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
3970 | vop0); | |
4a00c761 JJ |
3971 | } |
3972 | ||
3973 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3974 | } | |
3975 | else | |
3976 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
3977 | ||
3978 | if (slp_node) | |
9771b263 | 3979 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 3980 | else |
c689ce1e RB |
3981 | { |
3982 | if (!prev_stmt_info) | |
3983 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3984 | else | |
3985 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3986 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3987 | } | |
4a00c761 | 3988 | } |
ebfd146a | 3989 | } |
4a00c761 JJ |
3990 | |
3991 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
3992 | break; |
3993 | ||
3994 | case NARROW: | |
3995 | /* In case the vectorization factor (VF) is bigger than the number | |
3996 | of elements that we can fit in a vectype (nunits), we have to | |
3997 | generate more than one vector stmt - i.e - we need to "unroll" | |
3998 | the vector stmt by a factor VF/nunits. */ | |
3999 | for (j = 0; j < ncopies; j++) | |
4000 | { | |
4001 | /* Handle uses. */ | |
4a00c761 JJ |
4002 | if (slp_node) |
4003 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
4004 | slp_node, -1); | |
ebfd146a IR |
4005 | else |
4006 | { | |
9771b263 | 4007 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
4008 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
4009 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
4010 | } |
4011 | ||
4a00c761 JJ |
4012 | /* Arguments are ready. Create the new vector stmts. */ |
4013 | if (cvt_type) | |
9771b263 | 4014 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4015 | { |
4016 | if (codecvt1 == CALL_EXPR) | |
4017 | { | |
4018 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4019 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4020 | gimple_call_set_lhs (new_stmt, new_temp); | |
4021 | } | |
4022 | else | |
4023 | { | |
4024 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4025 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4026 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4027 | vop0); | |
4a00c761 | 4028 | } |
ebfd146a | 4029 | |
4a00c761 | 4030 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 4031 | vec_oprnds0[i] = new_temp; |
4a00c761 | 4032 | } |
ebfd146a | 4033 | |
4a00c761 JJ |
4034 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
4035 | stmt, vec_dsts, gsi, | |
4036 | slp_node, code1, | |
4037 | &prev_stmt_info); | |
ebfd146a IR |
4038 | } |
4039 | ||
4040 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 4041 | break; |
ebfd146a IR |
4042 | } |
4043 | ||
9771b263 DN |
4044 | vec_oprnds0.release (); |
4045 | vec_oprnds1.release (); | |
4046 | vec_dsts.release (); | |
4047 | interm_types.release (); | |
ebfd146a IR |
4048 | |
4049 | return true; | |
4050 | } | |
ff802fa1 IR |
4051 | |
4052 | ||
ebfd146a IR |
4053 | /* Function vectorizable_assignment. |
4054 | ||
b8698a0f L |
4055 | Check if STMT performs an assignment (copy) that can be vectorized. |
4056 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4057 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4058 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4059 | ||
4060 | static bool | |
355fe088 TS |
4061 | vectorizable_assignment (gimple *stmt, gimple_stmt_iterator *gsi, |
4062 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
4063 | { |
4064 | tree vec_dest; | |
4065 | tree scalar_dest; | |
4066 | tree op; | |
4067 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a IR |
4068 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4069 | tree new_temp; | |
355fe088 | 4070 | gimple *def_stmt; |
ebfd146a | 4071 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
ebfd146a | 4072 | int ncopies; |
f18b55bd | 4073 | int i, j; |
6e1aa848 | 4074 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 4075 | tree vop; |
a70d6342 | 4076 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4077 | vec_info *vinfo = stmt_info->vinfo; |
355fe088 | 4078 | gimple *new_stmt = NULL; |
f18b55bd | 4079 | stmt_vec_info prev_stmt_info = NULL; |
fde9c428 RG |
4080 | enum tree_code code; |
4081 | tree vectype_in; | |
ebfd146a | 4082 | |
a70d6342 | 4083 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4084 | return false; |
4085 | ||
8644a673 | 4086 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
4087 | return false; |
4088 | ||
4089 | /* Is vectorizable assignment? */ | |
4090 | if (!is_gimple_assign (stmt)) | |
4091 | return false; | |
4092 | ||
4093 | scalar_dest = gimple_assign_lhs (stmt); | |
4094 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4095 | return false; | |
4096 | ||
fde9c428 | 4097 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 4098 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
4099 | || code == PAREN_EXPR |
4100 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
4101 | op = gimple_assign_rhs1 (stmt); |
4102 | else | |
4103 | return false; | |
4104 | ||
7b7ec6c5 RG |
4105 | if (code == VIEW_CONVERT_EXPR) |
4106 | op = TREE_OPERAND (op, 0); | |
4107 | ||
465c8c19 JJ |
4108 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
4109 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4110 | ||
4111 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4112 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4113 | case of SLP. */ | |
4114 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
4115 | ncopies = 1; | |
4116 | else | |
4117 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4118 | ||
4119 | gcc_assert (ncopies >= 1); | |
4120 | ||
81c40241 | 4121 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[0], &vectype_in)) |
ebfd146a | 4122 | { |
73fbfcad | 4123 | if (dump_enabled_p ()) |
78c60e3d | 4124 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4125 | "use not simple.\n"); |
ebfd146a IR |
4126 | return false; |
4127 | } | |
4128 | ||
fde9c428 RG |
4129 | /* We can handle NOP_EXPR conversions that do not change the number |
4130 | of elements or the vector size. */ | |
7b7ec6c5 RG |
4131 | if ((CONVERT_EXPR_CODE_P (code) |
4132 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 RG |
4133 | && (!vectype_in |
4134 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
4135 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
4136 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
4137 | return false; | |
4138 | ||
7b7b1813 RG |
4139 | /* We do not handle bit-precision changes. */ |
4140 | if ((CONVERT_EXPR_CODE_P (code) | |
4141 | || code == VIEW_CONVERT_EXPR) | |
4142 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
4143 | && ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4144 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4145 | || ((TYPE_PRECISION (TREE_TYPE (op)) | |
4146 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))) | |
4147 | /* But a conversion that does not change the bit-pattern is ok. */ | |
4148 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4149 | > TYPE_PRECISION (TREE_TYPE (op))) | |
4150 | && TYPE_UNSIGNED (TREE_TYPE (op)))) | |
4151 | { | |
73fbfcad | 4152 | if (dump_enabled_p ()) |
78c60e3d SS |
4153 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4154 | "type conversion to/from bit-precision " | |
e645e942 | 4155 | "unsupported.\n"); |
7b7b1813 RG |
4156 | return false; |
4157 | } | |
4158 | ||
ebfd146a IR |
4159 | if (!vec_stmt) /* transformation not required. */ |
4160 | { | |
4161 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 4162 | if (dump_enabled_p ()) |
78c60e3d | 4163 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4164 | "=== vectorizable_assignment ===\n"); |
c3e7ee41 | 4165 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
4166 | return true; |
4167 | } | |
4168 | ||
4169 | /** Transform. **/ | |
73fbfcad | 4170 | if (dump_enabled_p ()) |
e645e942 | 4171 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
4172 | |
4173 | /* Handle def. */ | |
4174 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4175 | ||
4176 | /* Handle use. */ | |
f18b55bd | 4177 | for (j = 0; j < ncopies; j++) |
ebfd146a | 4178 | { |
f18b55bd IR |
4179 | /* Handle uses. */ |
4180 | if (j == 0) | |
d092494c | 4181 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node, -1); |
f18b55bd IR |
4182 | else |
4183 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
4184 | ||
4185 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 4186 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 4187 | { |
7b7ec6c5 RG |
4188 | if (CONVERT_EXPR_CODE_P (code) |
4189 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 4190 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
4191 | new_stmt = gimple_build_assign (vec_dest, vop); |
4192 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4193 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4194 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4195 | if (slp_node) | |
9771b263 | 4196 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 4197 | } |
ebfd146a IR |
4198 | |
4199 | if (slp_node) | |
f18b55bd IR |
4200 | continue; |
4201 | ||
4202 | if (j == 0) | |
4203 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4204 | else | |
4205 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4206 | ||
4207 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4208 | } | |
b8698a0f | 4209 | |
9771b263 | 4210 | vec_oprnds.release (); |
ebfd146a IR |
4211 | return true; |
4212 | } | |
4213 | ||
9dc3f7de | 4214 | |
1107f3ae IR |
4215 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
4216 | either as shift by a scalar or by a vector. */ | |
4217 | ||
4218 | bool | |
4219 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
4220 | { | |
4221 | ||
ef4bddc2 | 4222 | machine_mode vec_mode; |
1107f3ae IR |
4223 | optab optab; |
4224 | int icode; | |
4225 | tree vectype; | |
4226 | ||
4227 | vectype = get_vectype_for_scalar_type (scalar_type); | |
4228 | if (!vectype) | |
4229 | return false; | |
4230 | ||
4231 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4232 | if (!optab | |
4233 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
4234 | { | |
4235 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4236 | if (!optab | |
4237 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
4238 | == CODE_FOR_nothing)) | |
4239 | return false; | |
4240 | } | |
4241 | ||
4242 | vec_mode = TYPE_MODE (vectype); | |
4243 | icode = (int) optab_handler (optab, vec_mode); | |
4244 | if (icode == CODE_FOR_nothing) | |
4245 | return false; | |
4246 | ||
4247 | return true; | |
4248 | } | |
4249 | ||
4250 | ||
9dc3f7de IR |
4251 | /* Function vectorizable_shift. |
4252 | ||
4253 | Check if STMT performs a shift operation that can be vectorized. | |
4254 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4255 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
4256 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4257 | ||
4258 | static bool | |
355fe088 TS |
4259 | vectorizable_shift (gimple *stmt, gimple_stmt_iterator *gsi, |
4260 | gimple **vec_stmt, slp_tree slp_node) | |
9dc3f7de IR |
4261 | { |
4262 | tree vec_dest; | |
4263 | tree scalar_dest; | |
4264 | tree op0, op1 = NULL; | |
4265 | tree vec_oprnd1 = NULL_TREE; | |
4266 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4267 | tree vectype; | |
4268 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4269 | enum tree_code code; | |
ef4bddc2 | 4270 | machine_mode vec_mode; |
9dc3f7de IR |
4271 | tree new_temp; |
4272 | optab optab; | |
4273 | int icode; | |
ef4bddc2 | 4274 | machine_mode optab_op2_mode; |
355fe088 | 4275 | gimple *def_stmt; |
9dc3f7de | 4276 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
355fe088 | 4277 | gimple *new_stmt = NULL; |
9dc3f7de IR |
4278 | stmt_vec_info prev_stmt_info; |
4279 | int nunits_in; | |
4280 | int nunits_out; | |
4281 | tree vectype_out; | |
cede2577 | 4282 | tree op1_vectype; |
9dc3f7de IR |
4283 | int ncopies; |
4284 | int j, i; | |
6e1aa848 DN |
4285 | vec<tree> vec_oprnds0 = vNULL; |
4286 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
4287 | tree vop0, vop1; |
4288 | unsigned int k; | |
49eab32e | 4289 | bool scalar_shift_arg = true; |
9dc3f7de | 4290 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4291 | vec_info *vinfo = stmt_info->vinfo; |
9dc3f7de IR |
4292 | int vf; |
4293 | ||
4294 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
4295 | return false; | |
4296 | ||
4297 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
4298 | return false; | |
4299 | ||
4300 | /* Is STMT a vectorizable binary/unary operation? */ | |
4301 | if (!is_gimple_assign (stmt)) | |
4302 | return false; | |
4303 | ||
4304 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4305 | return false; | |
4306 | ||
4307 | code = gimple_assign_rhs_code (stmt); | |
4308 | ||
4309 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
4310 | || code == RROTATE_EXPR)) | |
4311 | return false; | |
4312 | ||
4313 | scalar_dest = gimple_assign_lhs (stmt); | |
4314 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
7b7b1813 RG |
4315 | if (TYPE_PRECISION (TREE_TYPE (scalar_dest)) |
4316 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4317 | { | |
73fbfcad | 4318 | if (dump_enabled_p ()) |
78c60e3d | 4319 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4320 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
4321 | return false; |
4322 | } | |
9dc3f7de IR |
4323 | |
4324 | op0 = gimple_assign_rhs1 (stmt); | |
81c40241 | 4325 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
9dc3f7de | 4326 | { |
73fbfcad | 4327 | if (dump_enabled_p ()) |
78c60e3d | 4328 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4329 | "use not simple.\n"); |
9dc3f7de IR |
4330 | return false; |
4331 | } | |
4332 | /* If op0 is an external or constant def use a vector type with | |
4333 | the same size as the output vector type. */ | |
4334 | if (!vectype) | |
4335 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
4336 | if (vec_stmt) | |
4337 | gcc_assert (vectype); | |
4338 | if (!vectype) | |
4339 | { | |
73fbfcad | 4340 | if (dump_enabled_p ()) |
78c60e3d | 4341 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4342 | "no vectype for scalar type\n"); |
9dc3f7de IR |
4343 | return false; |
4344 | } | |
4345 | ||
4346 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4347 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4348 | if (nunits_out != nunits_in) | |
4349 | return false; | |
4350 | ||
4351 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 4352 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &op1_vectype)) |
9dc3f7de | 4353 | { |
73fbfcad | 4354 | if (dump_enabled_p ()) |
78c60e3d | 4355 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4356 | "use not simple.\n"); |
9dc3f7de IR |
4357 | return false; |
4358 | } | |
4359 | ||
4360 | if (loop_vinfo) | |
4361 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4362 | else | |
4363 | vf = 1; | |
4364 | ||
4365 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4366 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4367 | case of SLP. */ | |
437f4a00 | 4368 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
9dc3f7de IR |
4369 | ncopies = 1; |
4370 | else | |
4371 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
4372 | ||
4373 | gcc_assert (ncopies >= 1); | |
4374 | ||
4375 | /* Determine whether the shift amount is a vector, or scalar. If the | |
4376 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
4377 | ||
dbfa87aa YR |
4378 | if ((dt[1] == vect_internal_def |
4379 | || dt[1] == vect_induction_def) | |
4380 | && !slp_node) | |
49eab32e JJ |
4381 | scalar_shift_arg = false; |
4382 | else if (dt[1] == vect_constant_def | |
4383 | || dt[1] == vect_external_def | |
4384 | || dt[1] == vect_internal_def) | |
4385 | { | |
4386 | /* In SLP, need to check whether the shift count is the same, | |
4387 | in loops if it is a constant or invariant, it is always | |
4388 | a scalar shift. */ | |
4389 | if (slp_node) | |
4390 | { | |
355fe088 TS |
4391 | vec<gimple *> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
4392 | gimple *slpstmt; | |
49eab32e | 4393 | |
9771b263 | 4394 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
4395 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
4396 | scalar_shift_arg = false; | |
4397 | } | |
4398 | } | |
4399 | else | |
4400 | { | |
73fbfcad | 4401 | if (dump_enabled_p ()) |
78c60e3d | 4402 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4403 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
4404 | return false; |
4405 | } | |
4406 | ||
9dc3f7de | 4407 | /* Vector shifted by vector. */ |
49eab32e | 4408 | if (!scalar_shift_arg) |
9dc3f7de IR |
4409 | { |
4410 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 4411 | if (dump_enabled_p ()) |
78c60e3d | 4412 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4413 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 4414 | |
aa948027 JJ |
4415 | if (!op1_vectype) |
4416 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
4417 | if (op1_vectype == NULL_TREE | |
4418 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 4419 | { |
73fbfcad | 4420 | if (dump_enabled_p ()) |
78c60e3d SS |
4421 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4422 | "unusable type for last operand in" | |
e645e942 | 4423 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
4424 | return false; |
4425 | } | |
9dc3f7de IR |
4426 | } |
4427 | /* See if the machine has a vector shifted by scalar insn and if not | |
4428 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 4429 | else |
9dc3f7de IR |
4430 | { |
4431 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4432 | if (optab | |
4433 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
4434 | { | |
73fbfcad | 4435 | if (dump_enabled_p ()) |
78c60e3d | 4436 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4437 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
4438 | } |
4439 | else | |
4440 | { | |
4441 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4442 | if (optab | |
4443 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
4444 | != CODE_FOR_nothing)) | |
4445 | { | |
49eab32e JJ |
4446 | scalar_shift_arg = false; |
4447 | ||
73fbfcad | 4448 | if (dump_enabled_p ()) |
78c60e3d | 4449 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4450 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
4451 | |
4452 | /* Unlike the other binary operators, shifts/rotates have | |
4453 | the rhs being int, instead of the same type as the lhs, | |
4454 | so make sure the scalar is the right type if we are | |
aa948027 | 4455 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
4456 | if (dt[1] == vect_constant_def) |
4457 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
4458 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
4459 | TREE_TYPE (op1))) | |
4460 | { | |
4461 | if (slp_node | |
4462 | && TYPE_MODE (TREE_TYPE (vectype)) | |
4463 | != TYPE_MODE (TREE_TYPE (op1))) | |
4464 | { | |
73fbfcad | 4465 | if (dump_enabled_p ()) |
78c60e3d SS |
4466 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4467 | "unusable type for last operand in" | |
e645e942 | 4468 | " vector/vector shift/rotate.\n"); |
21c0a521 | 4469 | return false; |
aa948027 JJ |
4470 | } |
4471 | if (vec_stmt && !slp_node) | |
4472 | { | |
4473 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
4474 | op1 = vect_init_vector (stmt, op1, | |
4475 | TREE_TYPE (vectype), NULL); | |
4476 | } | |
4477 | } | |
9dc3f7de IR |
4478 | } |
4479 | } | |
4480 | } | |
9dc3f7de IR |
4481 | |
4482 | /* Supportable by target? */ | |
4483 | if (!optab) | |
4484 | { | |
73fbfcad | 4485 | if (dump_enabled_p ()) |
78c60e3d | 4486 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4487 | "no optab.\n"); |
9dc3f7de IR |
4488 | return false; |
4489 | } | |
4490 | vec_mode = TYPE_MODE (vectype); | |
4491 | icode = (int) optab_handler (optab, vec_mode); | |
4492 | if (icode == CODE_FOR_nothing) | |
4493 | { | |
73fbfcad | 4494 | if (dump_enabled_p ()) |
78c60e3d | 4495 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4496 | "op not supported by target.\n"); |
9dc3f7de IR |
4497 | /* Check only during analysis. */ |
4498 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
4499 | || (vf < vect_min_worthwhile_factor (code) | |
4500 | && !vec_stmt)) | |
4501 | return false; | |
73fbfcad | 4502 | if (dump_enabled_p ()) |
e645e942 TJ |
4503 | dump_printf_loc (MSG_NOTE, vect_location, |
4504 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
4505 | } |
4506 | ||
4507 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
4508 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
4509 | && vf < vect_min_worthwhile_factor (code) | |
4510 | && !vec_stmt) | |
4511 | { | |
73fbfcad | 4512 | if (dump_enabled_p ()) |
78c60e3d | 4513 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4514 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
4515 | return false; |
4516 | } | |
4517 | ||
4518 | if (!vec_stmt) /* transformation not required. */ | |
4519 | { | |
4520 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 4521 | if (dump_enabled_p ()) |
e645e942 TJ |
4522 | dump_printf_loc (MSG_NOTE, vect_location, |
4523 | "=== vectorizable_shift ===\n"); | |
c3e7ee41 | 4524 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
9dc3f7de IR |
4525 | return true; |
4526 | } | |
4527 | ||
4528 | /** Transform. **/ | |
4529 | ||
73fbfcad | 4530 | if (dump_enabled_p ()) |
78c60e3d | 4531 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4532 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
4533 | |
4534 | /* Handle def. */ | |
4535 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4536 | ||
9dc3f7de IR |
4537 | prev_stmt_info = NULL; |
4538 | for (j = 0; j < ncopies; j++) | |
4539 | { | |
4540 | /* Handle uses. */ | |
4541 | if (j == 0) | |
4542 | { | |
4543 | if (scalar_shift_arg) | |
4544 | { | |
4545 | /* Vector shl and shr insn patterns can be defined with scalar | |
4546 | operand 2 (shift operand). In this case, use constant or loop | |
4547 | invariant op1 directly, without extending it to vector mode | |
4548 | first. */ | |
4549 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
4550 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
4551 | { | |
73fbfcad | 4552 | if (dump_enabled_p ()) |
78c60e3d | 4553 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4554 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 4555 | vec_oprnd1 = op1; |
8930f723 | 4556 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 4557 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
4558 | if (slp_node) |
4559 | { | |
4560 | /* Store vec_oprnd1 for every vector stmt to be created | |
4561 | for SLP_NODE. We check during the analysis that all | |
4562 | the shift arguments are the same. | |
4563 | TODO: Allow different constants for different vector | |
4564 | stmts generated for an SLP instance. */ | |
4565 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4566 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
4567 | } |
4568 | } | |
4569 | } | |
4570 | ||
4571 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
4572 | (a special case for certain kind of vector shifts); otherwise, | |
4573 | operand 1 should be of a vector type (the usual case). */ | |
4574 | if (vec_oprnd1) | |
4575 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
d092494c | 4576 | slp_node, -1); |
9dc3f7de IR |
4577 | else |
4578 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
d092494c | 4579 | slp_node, -1); |
9dc3f7de IR |
4580 | } |
4581 | else | |
4582 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
4583 | ||
4584 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 4585 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 4586 | { |
9771b263 | 4587 | vop1 = vec_oprnds1[i]; |
0d0e4a03 | 4588 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
9dc3f7de IR |
4589 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4590 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4591 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4592 | if (slp_node) | |
9771b263 | 4593 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
4594 | } |
4595 | ||
4596 | if (slp_node) | |
4597 | continue; | |
4598 | ||
4599 | if (j == 0) | |
4600 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4601 | else | |
4602 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4603 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4604 | } | |
4605 | ||
9771b263 DN |
4606 | vec_oprnds0.release (); |
4607 | vec_oprnds1.release (); | |
9dc3f7de IR |
4608 | |
4609 | return true; | |
4610 | } | |
4611 | ||
4612 | ||
ebfd146a IR |
4613 | /* Function vectorizable_operation. |
4614 | ||
16949072 RG |
4615 | Check if STMT performs a binary, unary or ternary operation that can |
4616 | be vectorized. | |
b8698a0f | 4617 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
4618 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4619 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4620 | ||
4621 | static bool | |
355fe088 TS |
4622 | vectorizable_operation (gimple *stmt, gimple_stmt_iterator *gsi, |
4623 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a | 4624 | { |
00f07b86 | 4625 | tree vec_dest; |
ebfd146a | 4626 | tree scalar_dest; |
16949072 | 4627 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 4628 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 4629 | tree vectype; |
ebfd146a IR |
4630 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4631 | enum tree_code code; | |
ef4bddc2 | 4632 | machine_mode vec_mode; |
ebfd146a IR |
4633 | tree new_temp; |
4634 | int op_type; | |
00f07b86 | 4635 | optab optab; |
523ba738 | 4636 | bool target_support_p; |
355fe088 | 4637 | gimple *def_stmt; |
16949072 RG |
4638 | enum vect_def_type dt[3] |
4639 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
355fe088 | 4640 | gimple *new_stmt = NULL; |
ebfd146a | 4641 | stmt_vec_info prev_stmt_info; |
b690cc0f | 4642 | int nunits_in; |
ebfd146a IR |
4643 | int nunits_out; |
4644 | tree vectype_out; | |
4645 | int ncopies; | |
4646 | int j, i; | |
6e1aa848 DN |
4647 | vec<tree> vec_oprnds0 = vNULL; |
4648 | vec<tree> vec_oprnds1 = vNULL; | |
4649 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 4650 | tree vop0, vop1, vop2; |
a70d6342 | 4651 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4652 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 IR |
4653 | int vf; |
4654 | ||
a70d6342 | 4655 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4656 | return false; |
4657 | ||
8644a673 | 4658 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
4659 | return false; |
4660 | ||
4661 | /* Is STMT a vectorizable binary/unary operation? */ | |
4662 | if (!is_gimple_assign (stmt)) | |
4663 | return false; | |
4664 | ||
4665 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4666 | return false; | |
4667 | ||
ebfd146a IR |
4668 | code = gimple_assign_rhs_code (stmt); |
4669 | ||
4670 | /* For pointer addition, we should use the normal plus for | |
4671 | the vector addition. */ | |
4672 | if (code == POINTER_PLUS_EXPR) | |
4673 | code = PLUS_EXPR; | |
4674 | ||
4675 | /* Support only unary or binary operations. */ | |
4676 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 4677 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 4678 | { |
73fbfcad | 4679 | if (dump_enabled_p ()) |
78c60e3d | 4680 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4681 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 4682 | op_type); |
ebfd146a IR |
4683 | return false; |
4684 | } | |
4685 | ||
b690cc0f RG |
4686 | scalar_dest = gimple_assign_lhs (stmt); |
4687 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
4688 | ||
7b7b1813 RG |
4689 | /* Most operations cannot handle bit-precision types without extra |
4690 | truncations. */ | |
4691 | if ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4692 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4693 | /* Exception are bitwise binary operations. */ | |
4694 | && code != BIT_IOR_EXPR | |
4695 | && code != BIT_XOR_EXPR | |
4696 | && code != BIT_AND_EXPR) | |
4697 | { | |
73fbfcad | 4698 | if (dump_enabled_p ()) |
78c60e3d | 4699 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4700 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
4701 | return false; |
4702 | } | |
4703 | ||
ebfd146a | 4704 | op0 = gimple_assign_rhs1 (stmt); |
81c40241 | 4705 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
ebfd146a | 4706 | { |
73fbfcad | 4707 | if (dump_enabled_p ()) |
78c60e3d | 4708 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4709 | "use not simple.\n"); |
ebfd146a IR |
4710 | return false; |
4711 | } | |
b690cc0f RG |
4712 | /* If op0 is an external or constant def use a vector type with |
4713 | the same size as the output vector type. */ | |
4714 | if (!vectype) | |
b036c6c5 IE |
4715 | { |
4716 | /* For boolean type we cannot determine vectype by | |
4717 | invariant value (don't know whether it is a vector | |
4718 | of booleans or vector of integers). We use output | |
4719 | vectype because operations on boolean don't change | |
4720 | type. */ | |
4721 | if (TREE_CODE (TREE_TYPE (op0)) == BOOLEAN_TYPE) | |
4722 | { | |
4723 | if (TREE_CODE (TREE_TYPE (scalar_dest)) != BOOLEAN_TYPE) | |
4724 | { | |
4725 | if (dump_enabled_p ()) | |
4726 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
4727 | "not supported operation on bool value.\n"); | |
4728 | return false; | |
4729 | } | |
4730 | vectype = vectype_out; | |
4731 | } | |
4732 | else | |
4733 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
4734 | } | |
7d8930a0 IR |
4735 | if (vec_stmt) |
4736 | gcc_assert (vectype); | |
4737 | if (!vectype) | |
4738 | { | |
73fbfcad | 4739 | if (dump_enabled_p ()) |
7d8930a0 | 4740 | { |
78c60e3d SS |
4741 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4742 | "no vectype for scalar type "); | |
4743 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
4744 | TREE_TYPE (op0)); | |
e645e942 | 4745 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
4746 | } |
4747 | ||
4748 | return false; | |
4749 | } | |
b690cc0f RG |
4750 | |
4751 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4752 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4753 | if (nunits_out != nunits_in) | |
4754 | return false; | |
ebfd146a | 4755 | |
16949072 | 4756 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
4757 | { |
4758 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 4759 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1])) |
ebfd146a | 4760 | { |
73fbfcad | 4761 | if (dump_enabled_p ()) |
78c60e3d | 4762 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4763 | "use not simple.\n"); |
ebfd146a IR |
4764 | return false; |
4765 | } | |
4766 | } | |
16949072 RG |
4767 | if (op_type == ternary_op) |
4768 | { | |
4769 | op2 = gimple_assign_rhs3 (stmt); | |
81c40241 | 4770 | if (!vect_is_simple_use (op2, vinfo, &def_stmt, &dt[2])) |
16949072 | 4771 | { |
73fbfcad | 4772 | if (dump_enabled_p ()) |
78c60e3d | 4773 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4774 | "use not simple.\n"); |
16949072 RG |
4775 | return false; |
4776 | } | |
4777 | } | |
ebfd146a | 4778 | |
b690cc0f RG |
4779 | if (loop_vinfo) |
4780 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4781 | else | |
4782 | vf = 1; | |
4783 | ||
4784 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 4785 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 4786 | case of SLP. */ |
437f4a00 | 4787 | if (slp_node || PURE_SLP_STMT (stmt_info)) |
b690cc0f RG |
4788 | ncopies = 1; |
4789 | else | |
4790 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
4791 | ||
4792 | gcc_assert (ncopies >= 1); | |
4793 | ||
9dc3f7de | 4794 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
4795 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
4796 | || code == RROTATE_EXPR) | |
9dc3f7de | 4797 | return false; |
ebfd146a | 4798 | |
ebfd146a | 4799 | /* Supportable by target? */ |
00f07b86 RH |
4800 | |
4801 | vec_mode = TYPE_MODE (vectype); | |
4802 | if (code == MULT_HIGHPART_EXPR) | |
523ba738 | 4803 | target_support_p = can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype)); |
00f07b86 RH |
4804 | else |
4805 | { | |
4806 | optab = optab_for_tree_code (code, vectype, optab_default); | |
4807 | if (!optab) | |
5deb57cb | 4808 | { |
73fbfcad | 4809 | if (dump_enabled_p ()) |
78c60e3d | 4810 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4811 | "no optab.\n"); |
00f07b86 | 4812 | return false; |
5deb57cb | 4813 | } |
523ba738 RS |
4814 | target_support_p = (optab_handler (optab, vec_mode) |
4815 | != CODE_FOR_nothing); | |
5deb57cb JJ |
4816 | } |
4817 | ||
523ba738 | 4818 | if (!target_support_p) |
ebfd146a | 4819 | { |
73fbfcad | 4820 | if (dump_enabled_p ()) |
78c60e3d | 4821 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4822 | "op not supported by target.\n"); |
ebfd146a IR |
4823 | /* Check only during analysis. */ |
4824 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
5deb57cb | 4825 | || (!vec_stmt && vf < vect_min_worthwhile_factor (code))) |
ebfd146a | 4826 | return false; |
73fbfcad | 4827 | if (dump_enabled_p ()) |
e645e942 TJ |
4828 | dump_printf_loc (MSG_NOTE, vect_location, |
4829 | "proceeding using word mode.\n"); | |
383d9c83 IR |
4830 | } |
4831 | ||
4a00c761 | 4832 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
4833 | if (!VECTOR_MODE_P (vec_mode) |
4834 | && !vec_stmt | |
4835 | && vf < vect_min_worthwhile_factor (code)) | |
7d8930a0 | 4836 | { |
73fbfcad | 4837 | if (dump_enabled_p ()) |
78c60e3d | 4838 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4839 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 4840 | return false; |
7d8930a0 | 4841 | } |
ebfd146a | 4842 | |
ebfd146a IR |
4843 | if (!vec_stmt) /* transformation not required. */ |
4844 | { | |
4a00c761 | 4845 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 4846 | if (dump_enabled_p ()) |
78c60e3d | 4847 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4848 | "=== vectorizable_operation ===\n"); |
c3e7ee41 | 4849 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
4850 | return true; |
4851 | } | |
4852 | ||
4853 | /** Transform. **/ | |
4854 | ||
73fbfcad | 4855 | if (dump_enabled_p ()) |
78c60e3d | 4856 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4857 | "transform binary/unary operation.\n"); |
383d9c83 | 4858 | |
ebfd146a | 4859 | /* Handle def. */ |
00f07b86 | 4860 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 4861 | |
ebfd146a IR |
4862 | /* In case the vectorization factor (VF) is bigger than the number |
4863 | of elements that we can fit in a vectype (nunits), we have to generate | |
4864 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
4865 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
4866 | from one copy of the vector stmt to the next, in the field | |
4867 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
4868 | stages to find the correct vector defs to be used when vectorizing | |
4869 | stmts that use the defs of the current stmt. The example below | |
4870 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
4871 | we need to create 4 vectorized stmts): | |
4872 | ||
4873 | before vectorization: | |
4874 | RELATED_STMT VEC_STMT | |
4875 | S1: x = memref - - | |
4876 | S2: z = x + 1 - - | |
4877 | ||
4878 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
4879 | there): | |
4880 | RELATED_STMT VEC_STMT | |
4881 | VS1_0: vx0 = memref0 VS1_1 - | |
4882 | VS1_1: vx1 = memref1 VS1_2 - | |
4883 | VS1_2: vx2 = memref2 VS1_3 - | |
4884 | VS1_3: vx3 = memref3 - - | |
4885 | S1: x = load - VS1_0 | |
4886 | S2: z = x + 1 - - | |
4887 | ||
4888 | step2: vectorize stmt S2 (done here): | |
4889 | To vectorize stmt S2 we first need to find the relevant vector | |
4890 | def for the first operand 'x'. This is, as usual, obtained from | |
4891 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
4892 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
4893 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
4894 | the vector stmt VS2_0, and as usual, record it in the | |
4895 | STMT_VINFO_VEC_STMT of stmt S2. | |
4896 | When creating the second copy (VS2_1), we obtain the relevant vector | |
4897 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
4898 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
4899 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
4900 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
4901 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
4902 | chain of stmts and pointers: | |
4903 | RELATED_STMT VEC_STMT | |
4904 | VS1_0: vx0 = memref0 VS1_1 - | |
4905 | VS1_1: vx1 = memref1 VS1_2 - | |
4906 | VS1_2: vx2 = memref2 VS1_3 - | |
4907 | VS1_3: vx3 = memref3 - - | |
4908 | S1: x = load - VS1_0 | |
4909 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
4910 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
4911 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
4912 | VS2_3: vz3 = vx3 + v1 - - | |
4913 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
4914 | |
4915 | prev_stmt_info = NULL; | |
4916 | for (j = 0; j < ncopies; j++) | |
4917 | { | |
4918 | /* Handle uses. */ | |
4919 | if (j == 0) | |
4a00c761 JJ |
4920 | { |
4921 | if (op_type == binary_op || op_type == ternary_op) | |
4922 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
4923 | slp_node, -1); | |
4924 | else | |
4925 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
4926 | slp_node, -1); | |
4927 | if (op_type == ternary_op) | |
36ba4aae | 4928 | { |
9771b263 DN |
4929 | vec_oprnds2.create (1); |
4930 | vec_oprnds2.quick_push (vect_get_vec_def_for_operand (op2, | |
81c40241 | 4931 | stmt)); |
36ba4aae | 4932 | } |
4a00c761 | 4933 | } |
ebfd146a | 4934 | else |
4a00c761 JJ |
4935 | { |
4936 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
4937 | if (op_type == ternary_op) | |
4938 | { | |
9771b263 DN |
4939 | tree vec_oprnd = vec_oprnds2.pop (); |
4940 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
4941 | vec_oprnd)); | |
4a00c761 JJ |
4942 | } |
4943 | } | |
4944 | ||
4945 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 4946 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 4947 | { |
4a00c761 | 4948 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 4949 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 4950 | vop2 = ((op_type == ternary_op) |
9771b263 | 4951 | ? vec_oprnds2[i] : NULL_TREE); |
0d0e4a03 | 4952 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1, vop2); |
4a00c761 JJ |
4953 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4954 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4955 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4956 | if (slp_node) | |
9771b263 | 4957 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
4958 | } |
4959 | ||
4a00c761 JJ |
4960 | if (slp_node) |
4961 | continue; | |
4962 | ||
4963 | if (j == 0) | |
4964 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4965 | else | |
4966 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4967 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
4968 | } |
4969 | ||
9771b263 DN |
4970 | vec_oprnds0.release (); |
4971 | vec_oprnds1.release (); | |
4972 | vec_oprnds2.release (); | |
ebfd146a | 4973 | |
ebfd146a IR |
4974 | return true; |
4975 | } | |
4976 | ||
c716e67f XDL |
4977 | /* A helper function to ensure data reference DR's base alignment |
4978 | for STMT_INFO. */ | |
4979 | ||
4980 | static void | |
4981 | ensure_base_align (stmt_vec_info stmt_info, struct data_reference *dr) | |
4982 | { | |
4983 | if (!dr->aux) | |
4984 | return; | |
4985 | ||
52639a61 | 4986 | if (DR_VECT_AUX (dr)->base_misaligned) |
c716e67f XDL |
4987 | { |
4988 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
52639a61 | 4989 | tree base_decl = DR_VECT_AUX (dr)->base_decl; |
c716e67f | 4990 | |
428f0c67 JH |
4991 | if (decl_in_symtab_p (base_decl)) |
4992 | symtab_node::get (base_decl)->increase_alignment (TYPE_ALIGN (vectype)); | |
4993 | else | |
4994 | { | |
4995 | DECL_ALIGN (base_decl) = TYPE_ALIGN (vectype); | |
4996 | DECL_USER_ALIGN (base_decl) = 1; | |
4997 | } | |
52639a61 | 4998 | DR_VECT_AUX (dr)->base_misaligned = false; |
c716e67f XDL |
4999 | } |
5000 | } | |
5001 | ||
ebfd146a | 5002 | |
09dfa495 BM |
5003 | /* Given a vector type VECTYPE returns the VECTOR_CST mask that implements |
5004 | reversal of the vector elements. If that is impossible to do, | |
5005 | returns NULL. */ | |
5006 | ||
5007 | static tree | |
5008 | perm_mask_for_reverse (tree vectype) | |
5009 | { | |
5010 | int i, nunits; | |
5011 | unsigned char *sel; | |
5012 | ||
5013 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
5014 | sel = XALLOCAVEC (unsigned char, nunits); | |
5015 | ||
5016 | for (i = 0; i < nunits; ++i) | |
5017 | sel[i] = nunits - 1 - i; | |
5018 | ||
557be5a8 AL |
5019 | if (!can_vec_perm_p (TYPE_MODE (vectype), false, sel)) |
5020 | return NULL_TREE; | |
5021 | return vect_gen_perm_mask_checked (vectype, sel); | |
09dfa495 BM |
5022 | } |
5023 | ||
ebfd146a IR |
5024 | /* Function vectorizable_store. |
5025 | ||
b8698a0f L |
5026 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
5027 | can be vectorized. | |
5028 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
5029 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5030 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5031 | ||
5032 | static bool | |
355fe088 | 5033 | vectorizable_store (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 5034 | slp_tree slp_node) |
ebfd146a IR |
5035 | { |
5036 | tree scalar_dest; | |
5037 | tree data_ref; | |
5038 | tree op; | |
5039 | tree vec_oprnd = NULL_TREE; | |
5040 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5041 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
272c6793 | 5042 | tree elem_type; |
ebfd146a | 5043 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 5044 | struct loop *loop = NULL; |
ef4bddc2 | 5045 | machine_mode vec_mode; |
ebfd146a IR |
5046 | tree dummy; |
5047 | enum dr_alignment_support alignment_support_scheme; | |
355fe088 | 5048 | gimple *def_stmt; |
ebfd146a IR |
5049 | enum vect_def_type dt; |
5050 | stmt_vec_info prev_stmt_info = NULL; | |
5051 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 5052 | tree dataref_offset = NULL_TREE; |
355fe088 | 5053 | gimple *ptr_incr = NULL; |
ebfd146a IR |
5054 | int ncopies; |
5055 | int j; | |
355fe088 | 5056 | gimple *next_stmt, *first_stmt = NULL; |
0d0293ac | 5057 | bool grouped_store = false; |
272c6793 | 5058 | bool store_lanes_p = false; |
ebfd146a | 5059 | unsigned int group_size, i; |
6e1aa848 DN |
5060 | vec<tree> dr_chain = vNULL; |
5061 | vec<tree> oprnds = vNULL; | |
5062 | vec<tree> result_chain = vNULL; | |
ebfd146a | 5063 | bool inv_p; |
09dfa495 BM |
5064 | bool negative = false; |
5065 | tree offset = NULL_TREE; | |
6e1aa848 | 5066 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 5067 | bool slp = (slp_node != NULL); |
ebfd146a | 5068 | unsigned int vec_num; |
a70d6342 | 5069 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5070 | vec_info *vinfo = stmt_info->vinfo; |
272c6793 | 5071 | tree aggr_type; |
3bab6342 AT |
5072 | tree scatter_base = NULL_TREE, scatter_off = NULL_TREE; |
5073 | tree scatter_off_vectype = NULL_TREE, scatter_decl = NULL_TREE; | |
5074 | int scatter_scale = 1; | |
5075 | enum vect_def_type scatter_idx_dt = vect_unknown_def_type; | |
5076 | enum vect_def_type scatter_src_dt = vect_unknown_def_type; | |
355fe088 | 5077 | gimple *new_stmt; |
a70d6342 | 5078 | |
a70d6342 | 5079 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5080 | return false; |
5081 | ||
8644a673 | 5082 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) |
ebfd146a IR |
5083 | return false; |
5084 | ||
5085 | /* Is vectorizable store? */ | |
5086 | ||
5087 | if (!is_gimple_assign (stmt)) | |
5088 | return false; | |
5089 | ||
5090 | scalar_dest = gimple_assign_lhs (stmt); | |
ab0ef706 JJ |
5091 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
5092 | && is_pattern_stmt_p (stmt_info)) | |
5093 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
ebfd146a | 5094 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
38000232 | 5095 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF |
ebfd146a | 5096 | && TREE_CODE (scalar_dest) != INDIRECT_REF |
e9dbe7bb IR |
5097 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
5098 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
5099 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
5100 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
5101 | return false; |
5102 | ||
5103 | gcc_assert (gimple_assign_single_p (stmt)); | |
465c8c19 JJ |
5104 | |
5105 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
5106 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
5107 | ||
5108 | if (loop_vinfo) | |
5109 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
5110 | ||
5111 | /* Multiple types in SLP are handled by creating the appropriate number of | |
5112 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5113 | case of SLP. */ | |
5114 | if (slp || PURE_SLP_STMT (stmt_info)) | |
5115 | ncopies = 1; | |
5116 | else | |
5117 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
5118 | ||
5119 | gcc_assert (ncopies >= 1); | |
5120 | ||
5121 | /* FORNOW. This restriction should be relaxed. */ | |
5122 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) | |
5123 | { | |
5124 | if (dump_enabled_p ()) | |
5125 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5126 | "multiple types in nested loop.\n"); | |
5127 | return false; | |
5128 | } | |
5129 | ||
ebfd146a | 5130 | op = gimple_assign_rhs1 (stmt); |
81c40241 | 5131 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt)) |
ebfd146a | 5132 | { |
73fbfcad | 5133 | if (dump_enabled_p ()) |
78c60e3d | 5134 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5135 | "use not simple.\n"); |
ebfd146a IR |
5136 | return false; |
5137 | } | |
5138 | ||
272c6793 | 5139 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 5140 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 5141 | |
ebfd146a IR |
5142 | /* FORNOW. In some cases can vectorize even if data-type not supported |
5143 | (e.g. - array initialization with 0). */ | |
947131ba | 5144 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
5145 | return false; |
5146 | ||
5147 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
5148 | return false; | |
5149 | ||
f2e2a985 | 5150 | if (!STMT_VINFO_STRIDED_P (stmt_info)) |
09dfa495 | 5151 | { |
f2e2a985 MM |
5152 | negative = |
5153 | tree_int_cst_compare (loop && nested_in_vect_loop_p (loop, stmt) | |
5154 | ? STMT_VINFO_DR_STEP (stmt_info) : DR_STEP (dr), | |
5155 | size_zero_node) < 0; | |
5156 | if (negative && ncopies > 1) | |
09dfa495 BM |
5157 | { |
5158 | if (dump_enabled_p ()) | |
5159 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
f2e2a985 | 5160 | "multiple types with negative step.\n"); |
09dfa495 BM |
5161 | return false; |
5162 | } | |
f2e2a985 | 5163 | if (negative) |
09dfa495 | 5164 | { |
f2e2a985 MM |
5165 | gcc_assert (!grouped_store); |
5166 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
5167 | if (alignment_support_scheme != dr_aligned | |
5168 | && alignment_support_scheme != dr_unaligned_supported) | |
5169 | { | |
5170 | if (dump_enabled_p ()) | |
5171 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5172 | "negative step but alignment required.\n"); | |
5173 | return false; | |
5174 | } | |
5175 | if (dt != vect_constant_def | |
5176 | && dt != vect_external_def | |
5177 | && !perm_mask_for_reverse (vectype)) | |
5178 | { | |
5179 | if (dump_enabled_p ()) | |
5180 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5181 | "negative step and reversing not supported.\n"); | |
5182 | return false; | |
5183 | } | |
09dfa495 BM |
5184 | } |
5185 | } | |
5186 | ||
0d0293ac | 5187 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 5188 | { |
0d0293ac | 5189 | grouped_store = true; |
e14c1050 | 5190 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
cee62fee MM |
5191 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
5192 | if (!slp | |
5193 | && !PURE_SLP_STMT (stmt_info) | |
5194 | && !STMT_VINFO_STRIDED_P (stmt_info)) | |
b602d918 | 5195 | { |
272c6793 RS |
5196 | if (vect_store_lanes_supported (vectype, group_size)) |
5197 | store_lanes_p = true; | |
0d0293ac | 5198 | else if (!vect_grouped_store_supported (vectype, group_size)) |
b602d918 RS |
5199 | return false; |
5200 | } | |
b8698a0f | 5201 | |
cee62fee MM |
5202 | if (STMT_VINFO_STRIDED_P (stmt_info) |
5203 | && (slp || PURE_SLP_STMT (stmt_info)) | |
5204 | && (group_size > nunits | |
5205 | || nunits % group_size != 0)) | |
5206 | { | |
5207 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5208 | "unhandled strided group store\n"); | |
5209 | return false; | |
5210 | } | |
5211 | ||
ebfd146a IR |
5212 | if (first_stmt == stmt) |
5213 | { | |
5214 | /* STMT is the leader of the group. Check the operands of all the | |
5215 | stmts of the group. */ | |
e14c1050 | 5216 | next_stmt = GROUP_NEXT_ELEMENT (stmt_info); |
ebfd146a IR |
5217 | while (next_stmt) |
5218 | { | |
5219 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
5220 | op = gimple_assign_rhs1 (next_stmt); | |
81c40241 | 5221 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt)) |
ebfd146a | 5222 | { |
73fbfcad | 5223 | if (dump_enabled_p ()) |
78c60e3d | 5224 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5225 | "use not simple.\n"); |
ebfd146a IR |
5226 | return false; |
5227 | } | |
e14c1050 | 5228 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
5229 | } |
5230 | } | |
5231 | } | |
5232 | ||
3bab6342 AT |
5233 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5234 | { | |
355fe088 | 5235 | gimple *def_stmt; |
3bab6342 AT |
5236 | scatter_decl = vect_check_gather_scatter (stmt, loop_vinfo, &scatter_base, |
5237 | &scatter_off, &scatter_scale); | |
5238 | gcc_assert (scatter_decl); | |
81c40241 RB |
5239 | if (!vect_is_simple_use (scatter_off, vinfo, &def_stmt, &scatter_idx_dt, |
5240 | &scatter_off_vectype)) | |
3bab6342 AT |
5241 | { |
5242 | if (dump_enabled_p ()) | |
5243 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5244 | "scatter index use not simple."); | |
5245 | return false; | |
5246 | } | |
5247 | } | |
5248 | ||
ebfd146a IR |
5249 | if (!vec_stmt) /* transformation not required. */ |
5250 | { | |
5251 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
2e8ab70c RB |
5252 | /* The SLP costs are calculated during SLP analysis. */ |
5253 | if (!PURE_SLP_STMT (stmt_info)) | |
5254 | vect_model_store_cost (stmt_info, ncopies, store_lanes_p, dt, | |
5255 | NULL, NULL, NULL); | |
ebfd146a IR |
5256 | return true; |
5257 | } | |
5258 | ||
5259 | /** Transform. **/ | |
5260 | ||
c716e67f XDL |
5261 | ensure_base_align (stmt_info, dr); |
5262 | ||
3bab6342 AT |
5263 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5264 | { | |
5265 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE, op, src; | |
5266 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (scatter_decl)); | |
5267 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
5268 | tree ptr, mask, var, scale, perm_mask = NULL_TREE; | |
5269 | edge pe = loop_preheader_edge (loop); | |
5270 | gimple_seq seq; | |
5271 | basic_block new_bb; | |
5272 | enum { NARROW, NONE, WIDEN } modifier; | |
5273 | int scatter_off_nunits = TYPE_VECTOR_SUBPARTS (scatter_off_vectype); | |
5274 | ||
5275 | if (nunits == (unsigned int) scatter_off_nunits) | |
5276 | modifier = NONE; | |
5277 | else if (nunits == (unsigned int) scatter_off_nunits / 2) | |
5278 | { | |
5279 | unsigned char *sel = XALLOCAVEC (unsigned char, scatter_off_nunits); | |
5280 | modifier = WIDEN; | |
5281 | ||
5282 | for (i = 0; i < (unsigned int) scatter_off_nunits; ++i) | |
5283 | sel[i] = i | nunits; | |
5284 | ||
5285 | perm_mask = vect_gen_perm_mask_checked (scatter_off_vectype, sel); | |
5286 | gcc_assert (perm_mask != NULL_TREE); | |
5287 | } | |
5288 | else if (nunits == (unsigned int) scatter_off_nunits * 2) | |
5289 | { | |
5290 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
5291 | modifier = NARROW; | |
5292 | ||
5293 | for (i = 0; i < (unsigned int) nunits; ++i) | |
5294 | sel[i] = i | scatter_off_nunits; | |
5295 | ||
5296 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); | |
5297 | gcc_assert (perm_mask != NULL_TREE); | |
5298 | ncopies *= 2; | |
5299 | } | |
5300 | else | |
5301 | gcc_unreachable (); | |
5302 | ||
5303 | rettype = TREE_TYPE (TREE_TYPE (scatter_decl)); | |
5304 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5305 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5306 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5307 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5308 | scaletype = TREE_VALUE (arglist); | |
5309 | ||
5310 | gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE | |
5311 | && TREE_CODE (rettype) == VOID_TYPE); | |
5312 | ||
5313 | ptr = fold_convert (ptrtype, scatter_base); | |
5314 | if (!is_gimple_min_invariant (ptr)) | |
5315 | { | |
5316 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
5317 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
5318 | gcc_assert (!new_bb); | |
5319 | } | |
5320 | ||
5321 | /* Currently we support only unconditional scatter stores, | |
5322 | so mask should be all ones. */ | |
5323 | mask = build_int_cst (masktype, -1); | |
5324 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
5325 | ||
5326 | scale = build_int_cst (scaletype, scatter_scale); | |
5327 | ||
5328 | prev_stmt_info = NULL; | |
5329 | for (j = 0; j < ncopies; ++j) | |
5330 | { | |
5331 | if (j == 0) | |
5332 | { | |
5333 | src = vec_oprnd1 | |
81c40241 | 5334 | = vect_get_vec_def_for_operand (gimple_assign_rhs1 (stmt), stmt); |
3bab6342 | 5335 | op = vec_oprnd0 |
81c40241 | 5336 | = vect_get_vec_def_for_operand (scatter_off, stmt); |
3bab6342 AT |
5337 | } |
5338 | else if (modifier != NONE && (j & 1)) | |
5339 | { | |
5340 | if (modifier == WIDEN) | |
5341 | { | |
5342 | src = vec_oprnd1 | |
5343 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5344 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, perm_mask, | |
5345 | stmt, gsi); | |
5346 | } | |
5347 | else if (modifier == NARROW) | |
5348 | { | |
5349 | src = permute_vec_elements (vec_oprnd1, vec_oprnd1, perm_mask, | |
5350 | stmt, gsi); | |
5351 | op = vec_oprnd0 | |
5352 | = vect_get_vec_def_for_stmt_copy (scatter_idx_dt, vec_oprnd0); | |
5353 | } | |
5354 | else | |
5355 | gcc_unreachable (); | |
5356 | } | |
5357 | else | |
5358 | { | |
5359 | src = vec_oprnd1 | |
5360 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5361 | op = vec_oprnd0 | |
5362 | = vect_get_vec_def_for_stmt_copy (scatter_idx_dt, vec_oprnd0); | |
5363 | } | |
5364 | ||
5365 | if (!useless_type_conversion_p (srctype, TREE_TYPE (src))) | |
5366 | { | |
5367 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)) | |
5368 | == TYPE_VECTOR_SUBPARTS (srctype)); | |
0e22bb5a | 5369 | var = vect_get_new_ssa_name (srctype, vect_simple_var); |
3bab6342 AT |
5370 | src = build1 (VIEW_CONVERT_EXPR, srctype, src); |
5371 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, src); | |
5372 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5373 | src = var; | |
5374 | } | |
5375 | ||
5376 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
5377 | { | |
5378 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
5379 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 5380 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
3bab6342 AT |
5381 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
5382 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
5383 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5384 | op = var; | |
5385 | } | |
5386 | ||
5387 | new_stmt | |
5388 | = gimple_build_call (scatter_decl, 5, ptr, mask, op, src, scale); | |
5389 | ||
5390 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5391 | ||
5392 | if (prev_stmt_info == NULL) | |
5393 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5394 | else | |
5395 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5396 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5397 | } | |
5398 | return true; | |
5399 | } | |
5400 | ||
0d0293ac | 5401 | if (grouped_store) |
ebfd146a IR |
5402 | { |
5403 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 5404 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 5405 | |
e14c1050 | 5406 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
5407 | |
5408 | /* FORNOW */ | |
a70d6342 | 5409 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
5410 | |
5411 | /* We vectorize all the stmts of the interleaving group when we | |
5412 | reach the last stmt in the group. */ | |
e14c1050 IR |
5413 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
5414 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
5415 | && !slp) |
5416 | { | |
5417 | *vec_stmt = NULL; | |
5418 | return true; | |
5419 | } | |
5420 | ||
5421 | if (slp) | |
4b5caab7 | 5422 | { |
0d0293ac | 5423 | grouped_store = false; |
4b5caab7 IR |
5424 | /* VEC_NUM is the number of vect stmts to be created for this |
5425 | group. */ | |
5426 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 5427 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
4b5caab7 | 5428 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
d092494c | 5429 | op = gimple_assign_rhs1 (first_stmt); |
4b5caab7 | 5430 | } |
ebfd146a | 5431 | else |
4b5caab7 IR |
5432 | /* VEC_NUM is the number of vect stmts to be created for this |
5433 | group. */ | |
ebfd146a IR |
5434 | vec_num = group_size; |
5435 | } | |
b8698a0f | 5436 | else |
ebfd146a IR |
5437 | { |
5438 | first_stmt = stmt; | |
5439 | first_dr = dr; | |
5440 | group_size = vec_num = 1; | |
ebfd146a | 5441 | } |
b8698a0f | 5442 | |
73fbfcad | 5443 | if (dump_enabled_p ()) |
78c60e3d | 5444 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5445 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 5446 | |
f2e2a985 MM |
5447 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
5448 | { | |
5449 | gimple_stmt_iterator incr_gsi; | |
5450 | bool insert_after; | |
355fe088 | 5451 | gimple *incr; |
f2e2a985 MM |
5452 | tree offvar; |
5453 | tree ivstep; | |
5454 | tree running_off; | |
5455 | gimple_seq stmts = NULL; | |
5456 | tree stride_base, stride_step, alias_off; | |
5457 | tree vec_oprnd; | |
f502d50e | 5458 | unsigned int g; |
f2e2a985 MM |
5459 | |
5460 | gcc_assert (!nested_in_vect_loop_p (loop, stmt)); | |
5461 | ||
5462 | stride_base | |
5463 | = fold_build_pointer_plus | |
f502d50e | 5464 | (unshare_expr (DR_BASE_ADDRESS (first_dr)), |
f2e2a985 | 5465 | size_binop (PLUS_EXPR, |
f502d50e MM |
5466 | convert_to_ptrofftype (unshare_expr (DR_OFFSET (first_dr))), |
5467 | convert_to_ptrofftype (DR_INIT(first_dr)))); | |
5468 | stride_step = fold_convert (sizetype, unshare_expr (DR_STEP (first_dr))); | |
f2e2a985 MM |
5469 | |
5470 | /* For a store with loop-invariant (but other than power-of-2) | |
5471 | stride (i.e. not a grouped access) like so: | |
5472 | ||
5473 | for (i = 0; i < n; i += stride) | |
5474 | array[i] = ...; | |
5475 | ||
5476 | we generate a new induction variable and new stores from | |
5477 | the components of the (vectorized) rhs: | |
5478 | ||
5479 | for (j = 0; ; j += VF*stride) | |
5480 | vectemp = ...; | |
5481 | tmp1 = vectemp[0]; | |
5482 | array[j] = tmp1; | |
5483 | tmp2 = vectemp[1]; | |
5484 | array[j + stride] = tmp2; | |
5485 | ... | |
5486 | */ | |
5487 | ||
cee62fee MM |
5488 | unsigned nstores = nunits; |
5489 | tree ltype = elem_type; | |
5490 | if (slp) | |
5491 | { | |
5492 | nstores = nunits / group_size; | |
5493 | if (group_size < nunits) | |
5494 | ltype = build_vector_type (elem_type, group_size); | |
5495 | else | |
5496 | ltype = vectype; | |
5497 | ltype = build_aligned_type (ltype, TYPE_ALIGN (elem_type)); | |
5498 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
f502d50e | 5499 | group_size = 1; |
cee62fee MM |
5500 | } |
5501 | ||
f2e2a985 MM |
5502 | ivstep = stride_step; |
5503 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
5504 | build_int_cst (TREE_TYPE (ivstep), | |
cee62fee | 5505 | ncopies * nstores)); |
f2e2a985 MM |
5506 | |
5507 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
5508 | ||
5509 | create_iv (stride_base, ivstep, NULL, | |
5510 | loop, &incr_gsi, insert_after, | |
5511 | &offvar, NULL); | |
5512 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 5513 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
f2e2a985 MM |
5514 | |
5515 | stride_step = force_gimple_operand (stride_step, &stmts, true, NULL_TREE); | |
5516 | if (stmts) | |
5517 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
5518 | ||
5519 | prev_stmt_info = NULL; | |
f502d50e MM |
5520 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (first_dr)), 0); |
5521 | next_stmt = first_stmt; | |
5522 | for (g = 0; g < group_size; g++) | |
f2e2a985 | 5523 | { |
f502d50e MM |
5524 | running_off = offvar; |
5525 | if (g) | |
f2e2a985 | 5526 | { |
f502d50e MM |
5527 | tree size = TYPE_SIZE_UNIT (ltype); |
5528 | tree pos = fold_build2 (MULT_EXPR, sizetype, size_int (g), | |
f2e2a985 | 5529 | size); |
f502d50e | 5530 | tree newoff = copy_ssa_name (running_off, NULL); |
f2e2a985 | 5531 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
f502d50e | 5532 | running_off, pos); |
f2e2a985 | 5533 | vect_finish_stmt_generation (stmt, incr, gsi); |
f2e2a985 | 5534 | running_off = newoff; |
f502d50e MM |
5535 | } |
5536 | for (j = 0; j < ncopies; j++) | |
5537 | { | |
5538 | /* We've set op and dt above, from gimple_assign_rhs1(stmt), | |
5539 | and first_stmt == stmt. */ | |
5540 | if (j == 0) | |
5541 | { | |
5542 | if (slp) | |
5543 | { | |
5544 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, NULL, | |
5545 | slp_node, -1); | |
5546 | vec_oprnd = vec_oprnds[0]; | |
5547 | } | |
5548 | else | |
5549 | { | |
5550 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
5551 | op = gimple_assign_rhs1 (next_stmt); | |
81c40241 | 5552 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
f502d50e MM |
5553 | } |
5554 | } | |
f2e2a985 | 5555 | else |
f502d50e MM |
5556 | { |
5557 | if (slp) | |
5558 | vec_oprnd = vec_oprnds[j]; | |
5559 | else | |
c079cbac | 5560 | { |
81c40241 | 5561 | vect_is_simple_use (vec_oprnd, vinfo, &def_stmt, &dt); |
c079cbac RB |
5562 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); |
5563 | } | |
f502d50e MM |
5564 | } |
5565 | ||
5566 | for (i = 0; i < nstores; i++) | |
5567 | { | |
5568 | tree newref, newoff; | |
355fe088 | 5569 | gimple *incr, *assign; |
f502d50e MM |
5570 | tree size = TYPE_SIZE (ltype); |
5571 | /* Extract the i'th component. */ | |
5572 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, | |
5573 | bitsize_int (i), size); | |
5574 | tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd, | |
5575 | size, pos); | |
5576 | ||
5577 | elem = force_gimple_operand_gsi (gsi, elem, true, | |
5578 | NULL_TREE, true, | |
5579 | GSI_SAME_STMT); | |
5580 | ||
5581 | newref = build2 (MEM_REF, ltype, | |
5582 | running_off, alias_off); | |
5583 | ||
5584 | /* And store it to *running_off. */ | |
5585 | assign = gimple_build_assign (newref, elem); | |
5586 | vect_finish_stmt_generation (stmt, assign, gsi); | |
5587 | ||
5588 | newoff = copy_ssa_name (running_off, NULL); | |
5589 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
5590 | running_off, stride_step); | |
5591 | vect_finish_stmt_generation (stmt, incr, gsi); | |
5592 | ||
5593 | running_off = newoff; | |
225ce44b RB |
5594 | if (g == group_size - 1 |
5595 | && !slp) | |
f502d50e MM |
5596 | { |
5597 | if (j == 0 && i == 0) | |
225ce44b RB |
5598 | STMT_VINFO_VEC_STMT (stmt_info) |
5599 | = *vec_stmt = assign; | |
f502d50e MM |
5600 | else |
5601 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = assign; | |
5602 | prev_stmt_info = vinfo_for_stmt (assign); | |
5603 | } | |
5604 | } | |
f2e2a985 | 5605 | } |
f502d50e | 5606 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
f2e2a985 MM |
5607 | } |
5608 | return true; | |
5609 | } | |
5610 | ||
9771b263 DN |
5611 | dr_chain.create (group_size); |
5612 | oprnds.create (group_size); | |
ebfd146a | 5613 | |
720f5239 | 5614 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 5615 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
5616 | /* Targets with store-lane instructions must not require explicit |
5617 | realignment. */ | |
5618 | gcc_assert (!store_lanes_p | |
5619 | || alignment_support_scheme == dr_aligned | |
5620 | || alignment_support_scheme == dr_unaligned_supported); | |
5621 | ||
09dfa495 BM |
5622 | if (negative) |
5623 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
5624 | ||
272c6793 RS |
5625 | if (store_lanes_p) |
5626 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
5627 | else | |
5628 | aggr_type = vectype; | |
ebfd146a IR |
5629 | |
5630 | /* In case the vectorization factor (VF) is bigger than the number | |
5631 | of elements that we can fit in a vectype (nunits), we have to generate | |
5632 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 5633 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
5634 | vect_get_vec_def_for_copy_stmt. */ |
5635 | ||
0d0293ac | 5636 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
5637 | |
5638 | S1: &base + 2 = x2 | |
5639 | S2: &base = x0 | |
5640 | S3: &base + 1 = x1 | |
5641 | S4: &base + 3 = x3 | |
5642 | ||
5643 | We create vectorized stores starting from base address (the access of the | |
5644 | first stmt in the chain (S2 in the above example), when the last store stmt | |
5645 | of the chain (S4) is reached: | |
5646 | ||
5647 | VS1: &base = vx2 | |
5648 | VS2: &base + vec_size*1 = vx0 | |
5649 | VS3: &base + vec_size*2 = vx1 | |
5650 | VS4: &base + vec_size*3 = vx3 | |
5651 | ||
5652 | Then permutation statements are generated: | |
5653 | ||
3fcc1b55 JJ |
5654 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
5655 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 5656 | ... |
b8698a0f | 5657 | |
ebfd146a IR |
5658 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
5659 | (the order of the data-refs in the output of vect_permute_store_chain | |
5660 | corresponds to the order of scalar stmts in the interleaving chain - see | |
5661 | the documentation of vect_permute_store_chain()). | |
5662 | ||
5663 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 5664 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 5665 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 5666 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
5667 | */ |
5668 | ||
5669 | prev_stmt_info = NULL; | |
5670 | for (j = 0; j < ncopies; j++) | |
5671 | { | |
ebfd146a IR |
5672 | |
5673 | if (j == 0) | |
5674 | { | |
5675 | if (slp) | |
5676 | { | |
5677 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c IR |
5678 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
5679 | NULL, slp_node, -1); | |
ebfd146a | 5680 | |
9771b263 | 5681 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
5682 | } |
5683 | else | |
5684 | { | |
b8698a0f L |
5685 | /* For interleaved stores we collect vectorized defs for all the |
5686 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
5687 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
5688 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
5689 | ||
0d0293ac | 5690 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 5691 | OPRNDS are of size 1. */ |
b8698a0f | 5692 | next_stmt = first_stmt; |
ebfd146a IR |
5693 | for (i = 0; i < group_size; i++) |
5694 | { | |
b8698a0f L |
5695 | /* Since gaps are not supported for interleaved stores, |
5696 | GROUP_SIZE is the exact number of stmts in the chain. | |
5697 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
5698 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
5699 | iteration of the loop will be executed. */ |
5700 | gcc_assert (next_stmt | |
5701 | && gimple_assign_single_p (next_stmt)); | |
5702 | op = gimple_assign_rhs1 (next_stmt); | |
5703 | ||
81c40241 | 5704 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
9771b263 DN |
5705 | dr_chain.quick_push (vec_oprnd); |
5706 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 5707 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
5708 | } |
5709 | } | |
5710 | ||
5711 | /* We should have catched mismatched types earlier. */ | |
5712 | gcc_assert (useless_type_conversion_p (vectype, | |
5713 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
5714 | bool simd_lane_access_p |
5715 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
5716 | if (simd_lane_access_p | |
5717 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
5718 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
5719 | && integer_zerop (DR_OFFSET (first_dr)) | |
5720 | && integer_zerop (DR_INIT (first_dr)) | |
5721 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
5722 | get_alias_set (DR_REF (first_dr)))) | |
5723 | { | |
5724 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
5725 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
5726 | (DR_REF (first_dr)), 0); | |
8928eff3 | 5727 | inv_p = false; |
74bf76ed JJ |
5728 | } |
5729 | else | |
5730 | dataref_ptr | |
5731 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
5732 | simd_lane_access_p ? loop : NULL, | |
09dfa495 | 5733 | offset, &dummy, gsi, &ptr_incr, |
74bf76ed | 5734 | simd_lane_access_p, &inv_p); |
a70d6342 | 5735 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 5736 | } |
b8698a0f | 5737 | else |
ebfd146a | 5738 | { |
b8698a0f L |
5739 | /* For interleaved stores we created vectorized defs for all the |
5740 | defs stored in OPRNDS in the previous iteration (previous copy). | |
5741 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
5742 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
5743 | next copy. | |
0d0293ac | 5744 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
5745 | OPRNDS are of size 1. */ |
5746 | for (i = 0; i < group_size; i++) | |
5747 | { | |
9771b263 | 5748 | op = oprnds[i]; |
81c40241 | 5749 | vect_is_simple_use (op, vinfo, &def_stmt, &dt); |
b8698a0f | 5750 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
9771b263 DN |
5751 | dr_chain[i] = vec_oprnd; |
5752 | oprnds[i] = vec_oprnd; | |
ebfd146a | 5753 | } |
74bf76ed JJ |
5754 | if (dataref_offset) |
5755 | dataref_offset | |
5756 | = int_const_binop (PLUS_EXPR, dataref_offset, | |
5757 | TYPE_SIZE_UNIT (aggr_type)); | |
5758 | else | |
5759 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
5760 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
5761 | } |
5762 | ||
272c6793 | 5763 | if (store_lanes_p) |
ebfd146a | 5764 | { |
272c6793 | 5765 | tree vec_array; |
267d3070 | 5766 | |
272c6793 RS |
5767 | /* Combine all the vectors into an array. */ |
5768 | vec_array = create_vector_array (vectype, vec_num); | |
5769 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 5770 | { |
9771b263 | 5771 | vec_oprnd = dr_chain[i]; |
272c6793 | 5772 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 5773 | } |
b8698a0f | 5774 | |
272c6793 RS |
5775 | /* Emit: |
5776 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
5777 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
5778 | new_stmt = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array); | |
5779 | gimple_call_set_lhs (new_stmt, data_ref); | |
267d3070 | 5780 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
5781 | } |
5782 | else | |
5783 | { | |
5784 | new_stmt = NULL; | |
0d0293ac | 5785 | if (grouped_store) |
272c6793 | 5786 | { |
b6b9227d JJ |
5787 | if (j == 0) |
5788 | result_chain.create (group_size); | |
272c6793 RS |
5789 | /* Permute. */ |
5790 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
5791 | &result_chain); | |
5792 | } | |
c2d7ab2a | 5793 | |
272c6793 RS |
5794 | next_stmt = first_stmt; |
5795 | for (i = 0; i < vec_num; i++) | |
5796 | { | |
644ffefd | 5797 | unsigned align, misalign; |
272c6793 RS |
5798 | |
5799 | if (i > 0) | |
5800 | /* Bump the vector pointer. */ | |
5801 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
5802 | stmt, NULL_TREE); | |
5803 | ||
5804 | if (slp) | |
9771b263 | 5805 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
5806 | else if (grouped_store) |
5807 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 5808 | vect_permute_store_chain(). */ |
9771b263 | 5809 | vec_oprnd = result_chain[i]; |
272c6793 | 5810 | |
aed93b23 RB |
5811 | data_ref = fold_build2 (MEM_REF, TREE_TYPE (vec_oprnd), |
5812 | dataref_ptr, | |
5813 | dataref_offset | |
5814 | ? dataref_offset | |
5815 | : build_int_cst (reference_alias_ptr_type | |
5816 | (DR_REF (first_dr)), 0)); | |
644ffefd | 5817 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 | 5818 | if (aligned_access_p (first_dr)) |
644ffefd | 5819 | misalign = 0; |
272c6793 RS |
5820 | else if (DR_MISALIGNMENT (first_dr) == -1) |
5821 | { | |
52639a61 RB |
5822 | if (DR_VECT_AUX (first_dr)->base_element_aligned) |
5823 | align = TYPE_ALIGN_UNIT (elem_type); | |
5824 | else | |
5825 | align = get_object_alignment (DR_REF (first_dr)) | |
5826 | / BITS_PER_UNIT; | |
5827 | misalign = 0; | |
272c6793 RS |
5828 | TREE_TYPE (data_ref) |
5829 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 5830 | align * BITS_PER_UNIT); |
272c6793 RS |
5831 | } |
5832 | else | |
5833 | { | |
5834 | TREE_TYPE (data_ref) | |
5835 | = build_aligned_type (TREE_TYPE (data_ref), | |
5836 | TYPE_ALIGN (elem_type)); | |
644ffefd | 5837 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 5838 | } |
aed93b23 RB |
5839 | if (dataref_offset == NULL_TREE |
5840 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
5841 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
5842 | misalign); | |
c2d7ab2a | 5843 | |
f234d260 BM |
5844 | if (negative |
5845 | && dt != vect_constant_def | |
5846 | && dt != vect_external_def) | |
09dfa495 BM |
5847 | { |
5848 | tree perm_mask = perm_mask_for_reverse (vectype); | |
5849 | tree perm_dest | |
5850 | = vect_create_destination_var (gimple_assign_rhs1 (stmt), | |
5851 | vectype); | |
b731b390 | 5852 | tree new_temp = make_ssa_name (perm_dest); |
09dfa495 BM |
5853 | |
5854 | /* Generate the permute statement. */ | |
355fe088 | 5855 | gimple *perm_stmt |
0d0e4a03 JJ |
5856 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
5857 | vec_oprnd, perm_mask); | |
09dfa495 BM |
5858 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
5859 | ||
5860 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); | |
5861 | vec_oprnd = new_temp; | |
5862 | } | |
5863 | ||
272c6793 RS |
5864 | /* Arguments are ready. Create the new vector stmt. */ |
5865 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
5866 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
5867 | |
5868 | if (slp) | |
5869 | continue; | |
5870 | ||
e14c1050 | 5871 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
5872 | if (!next_stmt) |
5873 | break; | |
5874 | } | |
ebfd146a | 5875 | } |
1da0876c RS |
5876 | if (!slp) |
5877 | { | |
5878 | if (j == 0) | |
5879 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5880 | else | |
5881 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5882 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5883 | } | |
ebfd146a IR |
5884 | } |
5885 | ||
9771b263 DN |
5886 | dr_chain.release (); |
5887 | oprnds.release (); | |
5888 | result_chain.release (); | |
5889 | vec_oprnds.release (); | |
ebfd146a IR |
5890 | |
5891 | return true; | |
5892 | } | |
5893 | ||
557be5a8 AL |
5894 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
5895 | VECTOR_CST mask. No checks are made that the target platform supports the | |
5896 | mask, so callers may wish to test can_vec_perm_p separately, or use | |
5897 | vect_gen_perm_mask_checked. */ | |
a1e53f3f | 5898 | |
3fcc1b55 | 5899 | tree |
557be5a8 | 5900 | vect_gen_perm_mask_any (tree vectype, const unsigned char *sel) |
a1e53f3f | 5901 | { |
d2a12ae7 | 5902 | tree mask_elt_type, mask_type, mask_vec, *mask_elts; |
2635892a | 5903 | int i, nunits; |
a1e53f3f | 5904 | |
22e4dee7 | 5905 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
22e4dee7 | 5906 | |
96f9265a RG |
5907 | mask_elt_type = lang_hooks.types.type_for_mode |
5908 | (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype))), 1); | |
22e4dee7 | 5909 | mask_type = get_vectype_for_scalar_type (mask_elt_type); |
a1e53f3f | 5910 | |
d2a12ae7 | 5911 | mask_elts = XALLOCAVEC (tree, nunits); |
aec7ae7d | 5912 | for (i = nunits - 1; i >= 0; i--) |
d2a12ae7 RG |
5913 | mask_elts[i] = build_int_cst (mask_elt_type, sel[i]); |
5914 | mask_vec = build_vector (mask_type, mask_elts); | |
a1e53f3f | 5915 | |
2635892a | 5916 | return mask_vec; |
a1e53f3f L |
5917 | } |
5918 | ||
cf7aa6a3 AL |
5919 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_p, |
5920 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ | |
557be5a8 AL |
5921 | |
5922 | tree | |
5923 | vect_gen_perm_mask_checked (tree vectype, const unsigned char *sel) | |
5924 | { | |
5925 | gcc_assert (can_vec_perm_p (TYPE_MODE (vectype), false, sel)); | |
5926 | return vect_gen_perm_mask_any (vectype, sel); | |
5927 | } | |
5928 | ||
aec7ae7d JJ |
5929 | /* Given a vector variable X and Y, that was generated for the scalar |
5930 | STMT, generate instructions to permute the vector elements of X and Y | |
5931 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
5932 | permuted vector variable. */ | |
a1e53f3f L |
5933 | |
5934 | static tree | |
355fe088 | 5935 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple *stmt, |
aec7ae7d | 5936 | gimple_stmt_iterator *gsi) |
a1e53f3f L |
5937 | { |
5938 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 5939 | tree perm_dest, data_ref; |
355fe088 | 5940 | gimple *perm_stmt; |
a1e53f3f | 5941 | |
acdcd61b | 5942 | perm_dest = vect_create_destination_var (gimple_get_lhs (stmt), vectype); |
b731b390 | 5943 | data_ref = make_ssa_name (perm_dest); |
a1e53f3f L |
5944 | |
5945 | /* Generate the permute statement. */ | |
0d0e4a03 | 5946 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
a1e53f3f L |
5947 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
5948 | ||
5949 | return data_ref; | |
5950 | } | |
5951 | ||
6b916b36 RB |
5952 | /* Hoist the definitions of all SSA uses on STMT out of the loop LOOP, |
5953 | inserting them on the loops preheader edge. Returns true if we | |
5954 | were successful in doing so (and thus STMT can be moved then), | |
5955 | otherwise returns false. */ | |
5956 | ||
5957 | static bool | |
355fe088 | 5958 | hoist_defs_of_uses (gimple *stmt, struct loop *loop) |
6b916b36 RB |
5959 | { |
5960 | ssa_op_iter i; | |
5961 | tree op; | |
5962 | bool any = false; | |
5963 | ||
5964 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
5965 | { | |
355fe088 | 5966 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
5967 | if (!gimple_nop_p (def_stmt) |
5968 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
5969 | { | |
5970 | /* Make sure we don't need to recurse. While we could do | |
5971 | so in simple cases when there are more complex use webs | |
5972 | we don't have an easy way to preserve stmt order to fulfil | |
5973 | dependencies within them. */ | |
5974 | tree op2; | |
5975 | ssa_op_iter i2; | |
d1417442 JJ |
5976 | if (gimple_code (def_stmt) == GIMPLE_PHI) |
5977 | return false; | |
6b916b36 RB |
5978 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
5979 | { | |
355fe088 | 5980 | gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2); |
6b916b36 RB |
5981 | if (!gimple_nop_p (def_stmt2) |
5982 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2))) | |
5983 | return false; | |
5984 | } | |
5985 | any = true; | |
5986 | } | |
5987 | } | |
5988 | ||
5989 | if (!any) | |
5990 | return true; | |
5991 | ||
5992 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
5993 | { | |
355fe088 | 5994 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
5995 | if (!gimple_nop_p (def_stmt) |
5996 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
5997 | { | |
5998 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); | |
5999 | gsi_remove (&gsi, false); | |
6000 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); | |
6001 | } | |
6002 | } | |
6003 | ||
6004 | return true; | |
6005 | } | |
6006 | ||
ebfd146a IR |
6007 | /* vectorizable_load. |
6008 | ||
b8698a0f L |
6009 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
6010 | can be vectorized. | |
6011 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
6012 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
6013 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
6014 | ||
6015 | static bool | |
355fe088 | 6016 | vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 6017 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a IR |
6018 | { |
6019 | tree scalar_dest; | |
6020 | tree vec_dest = NULL; | |
6021 | tree data_ref = NULL; | |
6022 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 6023 | stmt_vec_info prev_stmt_info; |
ebfd146a | 6024 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 6025 | struct loop *loop = NULL; |
ebfd146a | 6026 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 6027 | bool nested_in_vect_loop = false; |
c716e67f | 6028 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
272c6793 | 6029 | tree elem_type; |
ebfd146a | 6030 | tree new_temp; |
ef4bddc2 | 6031 | machine_mode mode; |
355fe088 | 6032 | gimple *new_stmt = NULL; |
ebfd146a IR |
6033 | tree dummy; |
6034 | enum dr_alignment_support alignment_support_scheme; | |
6035 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 6036 | tree dataref_offset = NULL_TREE; |
355fe088 | 6037 | gimple *ptr_incr = NULL; |
ebfd146a | 6038 | int ncopies; |
9b999e8c | 6039 | int i, j, group_size = -1, group_gap_adj; |
ebfd146a IR |
6040 | tree msq = NULL_TREE, lsq; |
6041 | tree offset = NULL_TREE; | |
356bbc4c | 6042 | tree byte_offset = NULL_TREE; |
ebfd146a | 6043 | tree realignment_token = NULL_TREE; |
538dd0b7 | 6044 | gphi *phi = NULL; |
6e1aa848 | 6045 | vec<tree> dr_chain = vNULL; |
0d0293ac | 6046 | bool grouped_load = false; |
272c6793 | 6047 | bool load_lanes_p = false; |
355fe088 | 6048 | gimple *first_stmt; |
ebfd146a | 6049 | bool inv_p; |
319e6439 | 6050 | bool negative = false; |
ebfd146a IR |
6051 | bool compute_in_loop = false; |
6052 | struct loop *at_loop; | |
6053 | int vec_num; | |
6054 | bool slp = (slp_node != NULL); | |
6055 | bool slp_perm = false; | |
6056 | enum tree_code code; | |
a70d6342 IR |
6057 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
6058 | int vf; | |
272c6793 | 6059 | tree aggr_type; |
aec7ae7d JJ |
6060 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; |
6061 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
6062 | int gather_scale = 1; | |
6063 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
310213d4 | 6064 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 6065 | |
465c8c19 JJ |
6066 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
6067 | return false; | |
6068 | ||
6069 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) | |
6070 | return false; | |
6071 | ||
6072 | /* Is vectorizable load? */ | |
6073 | if (!is_gimple_assign (stmt)) | |
6074 | return false; | |
6075 | ||
6076 | scalar_dest = gimple_assign_lhs (stmt); | |
6077 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
6078 | return false; | |
6079 | ||
6080 | code = gimple_assign_rhs_code (stmt); | |
6081 | if (code != ARRAY_REF | |
6082 | && code != BIT_FIELD_REF | |
6083 | && code != INDIRECT_REF | |
6084 | && code != COMPONENT_REF | |
6085 | && code != IMAGPART_EXPR | |
6086 | && code != REALPART_EXPR | |
6087 | && code != MEM_REF | |
6088 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
6089 | return false; | |
6090 | ||
6091 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
6092 | return false; | |
6093 | ||
6094 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
6095 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
6096 | ||
a70d6342 IR |
6097 | if (loop_vinfo) |
6098 | { | |
6099 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
6100 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
6101 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
6102 | } | |
6103 | else | |
3533e503 | 6104 | vf = 1; |
ebfd146a IR |
6105 | |
6106 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 6107 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 6108 | case of SLP. */ |
437f4a00 | 6109 | if (slp || PURE_SLP_STMT (stmt_info)) |
ebfd146a IR |
6110 | ncopies = 1; |
6111 | else | |
6112 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
6113 | ||
6114 | gcc_assert (ncopies >= 1); | |
6115 | ||
6116 | /* FORNOW. This restriction should be relaxed. */ | |
6117 | if (nested_in_vect_loop && ncopies > 1) | |
6118 | { | |
73fbfcad | 6119 | if (dump_enabled_p ()) |
78c60e3d | 6120 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6121 | "multiple types in nested loop.\n"); |
ebfd146a IR |
6122 | return false; |
6123 | } | |
6124 | ||
f2556b68 RB |
6125 | /* Invalidate assumptions made by dependence analysis when vectorization |
6126 | on the unrolled body effectively re-orders stmts. */ | |
6127 | if (ncopies > 1 | |
6128 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
6129 | && ((unsigned)LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
6130 | > STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
6131 | { | |
6132 | if (dump_enabled_p ()) | |
6133 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6134 | "cannot perform implicit CSE when unrolling " | |
6135 | "with negative dependence distance\n"); | |
6136 | return false; | |
6137 | } | |
6138 | ||
7b7b1813 | 6139 | elem_type = TREE_TYPE (vectype); |
947131ba | 6140 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
6141 | |
6142 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
6143 | (e.g. - data copies). */ | |
947131ba | 6144 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 6145 | { |
73fbfcad | 6146 | if (dump_enabled_p ()) |
78c60e3d | 6147 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6148 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
6149 | return false; |
6150 | } | |
6151 | ||
ebfd146a | 6152 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 6153 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 6154 | { |
0d0293ac | 6155 | grouped_load = true; |
ebfd146a | 6156 | /* FORNOW */ |
3bab6342 | 6157 | gcc_assert (!nested_in_vect_loop && !STMT_VINFO_GATHER_SCATTER_P (stmt_info)); |
ebfd146a | 6158 | |
e14c1050 | 6159 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
d5f035ea RB |
6160 | |
6161 | /* If this is single-element interleaving with an element distance | |
6162 | that leaves unused vector loads around punt - we at least create | |
6163 | very sub-optimal code in that case (and blow up memory, | |
6164 | see PR65518). */ | |
6165 | if (first_stmt == stmt | |
6166 | && !GROUP_NEXT_ELEMENT (stmt_info) | |
6167 | && GROUP_SIZE (stmt_info) > TYPE_VECTOR_SUBPARTS (vectype)) | |
6168 | { | |
6169 | if (dump_enabled_p ()) | |
6170 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6171 | "single-element interleaving not supported " | |
6172 | "for not adjacent vector loads\n"); | |
6173 | return false; | |
6174 | } | |
6175 | ||
b1af7da6 RB |
6176 | if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
6177 | slp_perm = true; | |
6178 | ||
7b5fc413 RB |
6179 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
6180 | if (!slp | |
6181 | && !PURE_SLP_STMT (stmt_info) | |
f2e2a985 | 6182 | && !STMT_VINFO_STRIDED_P (stmt_info)) |
b602d918 | 6183 | { |
272c6793 RS |
6184 | if (vect_load_lanes_supported (vectype, group_size)) |
6185 | load_lanes_p = true; | |
0d0293ac | 6186 | else if (!vect_grouped_load_supported (vectype, group_size)) |
b602d918 RS |
6187 | return false; |
6188 | } | |
f2556b68 RB |
6189 | |
6190 | /* Invalidate assumptions made by dependence analysis when vectorization | |
6191 | on the unrolled body effectively re-orders stmts. */ | |
6192 | if (!PURE_SLP_STMT (stmt_info) | |
6193 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
6194 | && ((unsigned)LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
6195 | > STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
6196 | { | |
6197 | if (dump_enabled_p ()) | |
6198 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6199 | "cannot perform implicit CSE when performing " | |
6200 | "group loads with negative dependence distance\n"); | |
6201 | return false; | |
6202 | } | |
96bb56b2 RB |
6203 | |
6204 | /* Similarly when the stmt is a load that is both part of a SLP | |
6205 | instance and a loop vectorized stmt via the same-dr mechanism | |
6206 | we have to give up. */ | |
6207 | if (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info) | |
6208 | && (STMT_SLP_TYPE (stmt_info) | |
6209 | != STMT_SLP_TYPE (vinfo_for_stmt | |
6210 | (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info))))) | |
6211 | { | |
6212 | if (dump_enabled_p ()) | |
6213 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6214 | "conflicting SLP types for CSEd load\n"); | |
6215 | return false; | |
6216 | } | |
ebfd146a IR |
6217 | } |
6218 | ||
a1e53f3f | 6219 | |
3bab6342 | 6220 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
aec7ae7d | 6221 | { |
355fe088 | 6222 | gimple *def_stmt; |
3bab6342 AT |
6223 | gather_decl = vect_check_gather_scatter (stmt, loop_vinfo, &gather_base, |
6224 | &gather_off, &gather_scale); | |
aec7ae7d | 6225 | gcc_assert (gather_decl); |
81c40241 RB |
6226 | if (!vect_is_simple_use (gather_off, vinfo, &def_stmt, &gather_dt, |
6227 | &gather_off_vectype)) | |
aec7ae7d | 6228 | { |
73fbfcad | 6229 | if (dump_enabled_p ()) |
78c60e3d | 6230 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6231 | "gather index use not simple.\n"); |
aec7ae7d JJ |
6232 | return false; |
6233 | } | |
6234 | } | |
f2e2a985 | 6235 | else if (STMT_VINFO_STRIDED_P (stmt_info)) |
7b5fc413 RB |
6236 | { |
6237 | if ((grouped_load | |
6238 | && (slp || PURE_SLP_STMT (stmt_info))) | |
6239 | && (group_size > nunits | |
b266b968 | 6240 | || nunits % group_size != 0)) |
7b5fc413 RB |
6241 | { |
6242 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6243 | "unhandled strided group load\n"); | |
6244 | return false; | |
6245 | } | |
6246 | } | |
319e6439 RG |
6247 | else |
6248 | { | |
6249 | negative = tree_int_cst_compare (nested_in_vect_loop | |
6250 | ? STMT_VINFO_DR_STEP (stmt_info) | |
6251 | : DR_STEP (dr), | |
6252 | size_zero_node) < 0; | |
6253 | if (negative && ncopies > 1) | |
6254 | { | |
73fbfcad | 6255 | if (dump_enabled_p ()) |
78c60e3d | 6256 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6257 | "multiple types with negative step.\n"); |
319e6439 RG |
6258 | return false; |
6259 | } | |
6260 | ||
6261 | if (negative) | |
6262 | { | |
08940f33 RB |
6263 | if (grouped_load) |
6264 | { | |
6265 | if (dump_enabled_p ()) | |
6266 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 TJ |
6267 | "negative step for group load not supported" |
6268 | "\n"); | |
08940f33 RB |
6269 | return false; |
6270 | } | |
319e6439 RG |
6271 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
6272 | if (alignment_support_scheme != dr_aligned | |
6273 | && alignment_support_scheme != dr_unaligned_supported) | |
6274 | { | |
73fbfcad | 6275 | if (dump_enabled_p ()) |
78c60e3d | 6276 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6277 | "negative step but alignment required.\n"); |
319e6439 RG |
6278 | return false; |
6279 | } | |
6280 | if (!perm_mask_for_reverse (vectype)) | |
6281 | { | |
73fbfcad | 6282 | if (dump_enabled_p ()) |
78c60e3d | 6283 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
6284 | "negative step and reversing not supported." |
6285 | "\n"); | |
319e6439 RG |
6286 | return false; |
6287 | } | |
6288 | } | |
7d75abc8 | 6289 | } |
aec7ae7d | 6290 | |
ebfd146a IR |
6291 | if (!vec_stmt) /* transformation not required. */ |
6292 | { | |
6293 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
2e8ab70c RB |
6294 | /* The SLP costs are calculated during SLP analysis. */ |
6295 | if (!PURE_SLP_STMT (stmt_info)) | |
6296 | vect_model_load_cost (stmt_info, ncopies, load_lanes_p, | |
6297 | NULL, NULL, NULL); | |
ebfd146a IR |
6298 | return true; |
6299 | } | |
6300 | ||
73fbfcad | 6301 | if (dump_enabled_p ()) |
78c60e3d | 6302 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6303 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a IR |
6304 | |
6305 | /** Transform. **/ | |
6306 | ||
c716e67f XDL |
6307 | ensure_base_align (stmt_info, dr); |
6308 | ||
3bab6342 | 6309 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
aec7ae7d JJ |
6310 | { |
6311 | tree vec_oprnd0 = NULL_TREE, op; | |
6312 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
6313 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
d3c2fee0 | 6314 | tree ptr, mask, var, scale, merge, perm_mask = NULL_TREE, prev_res = NULL_TREE; |
aec7ae7d JJ |
6315 | edge pe = loop_preheader_edge (loop); |
6316 | gimple_seq seq; | |
6317 | basic_block new_bb; | |
6318 | enum { NARROW, NONE, WIDEN } modifier; | |
6319 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
6320 | ||
6321 | if (nunits == gather_off_nunits) | |
6322 | modifier = NONE; | |
6323 | else if (nunits == gather_off_nunits / 2) | |
6324 | { | |
6325 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
6326 | modifier = WIDEN; | |
6327 | ||
6328 | for (i = 0; i < gather_off_nunits; ++i) | |
6329 | sel[i] = i | nunits; | |
6330 | ||
557be5a8 | 6331 | perm_mask = vect_gen_perm_mask_checked (gather_off_vectype, sel); |
aec7ae7d JJ |
6332 | } |
6333 | else if (nunits == gather_off_nunits * 2) | |
6334 | { | |
6335 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
6336 | modifier = NARROW; | |
6337 | ||
6338 | for (i = 0; i < nunits; ++i) | |
6339 | sel[i] = i < gather_off_nunits | |
6340 | ? i : i + nunits - gather_off_nunits; | |
6341 | ||
557be5a8 | 6342 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); |
aec7ae7d JJ |
6343 | ncopies *= 2; |
6344 | } | |
6345 | else | |
6346 | gcc_unreachable (); | |
6347 | ||
6348 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); | |
6349 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6350 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6351 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6352 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6353 | scaletype = TREE_VALUE (arglist); | |
d3c2fee0 | 6354 | gcc_checking_assert (types_compatible_p (srctype, rettype)); |
aec7ae7d JJ |
6355 | |
6356 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
6357 | ||
6358 | ptr = fold_convert (ptrtype, gather_base); | |
6359 | if (!is_gimple_min_invariant (ptr)) | |
6360 | { | |
6361 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
6362 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
6363 | gcc_assert (!new_bb); | |
6364 | } | |
6365 | ||
6366 | /* Currently we support only unconditional gather loads, | |
6367 | so mask should be all ones. */ | |
d3c2fee0 AI |
6368 | if (TREE_CODE (masktype) == INTEGER_TYPE) |
6369 | mask = build_int_cst (masktype, -1); | |
6370 | else if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
6371 | { | |
6372 | mask = build_int_cst (TREE_TYPE (masktype), -1); | |
6373 | mask = build_vector_from_val (masktype, mask); | |
03b9e8e4 | 6374 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
d3c2fee0 | 6375 | } |
aec7ae7d JJ |
6376 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) |
6377 | { | |
6378 | REAL_VALUE_TYPE r; | |
6379 | long tmp[6]; | |
6380 | for (j = 0; j < 6; ++j) | |
6381 | tmp[j] = -1; | |
6382 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
6383 | mask = build_real (TREE_TYPE (masktype), r); | |
d3c2fee0 | 6384 | mask = build_vector_from_val (masktype, mask); |
03b9e8e4 | 6385 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
aec7ae7d JJ |
6386 | } |
6387 | else | |
6388 | gcc_unreachable (); | |
aec7ae7d JJ |
6389 | |
6390 | scale = build_int_cst (scaletype, gather_scale); | |
6391 | ||
d3c2fee0 AI |
6392 | if (TREE_CODE (TREE_TYPE (rettype)) == INTEGER_TYPE) |
6393 | merge = build_int_cst (TREE_TYPE (rettype), 0); | |
6394 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (rettype))) | |
6395 | { | |
6396 | REAL_VALUE_TYPE r; | |
6397 | long tmp[6]; | |
6398 | for (j = 0; j < 6; ++j) | |
6399 | tmp[j] = 0; | |
6400 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (rettype))); | |
6401 | merge = build_real (TREE_TYPE (rettype), r); | |
6402 | } | |
6403 | else | |
6404 | gcc_unreachable (); | |
6405 | merge = build_vector_from_val (rettype, merge); | |
6406 | merge = vect_init_vector (stmt, merge, rettype, NULL); | |
6407 | ||
aec7ae7d JJ |
6408 | prev_stmt_info = NULL; |
6409 | for (j = 0; j < ncopies; ++j) | |
6410 | { | |
6411 | if (modifier == WIDEN && (j & 1)) | |
6412 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
6413 | perm_mask, stmt, gsi); | |
6414 | else if (j == 0) | |
6415 | op = vec_oprnd0 | |
81c40241 | 6416 | = vect_get_vec_def_for_operand (gather_off, stmt); |
aec7ae7d JJ |
6417 | else |
6418 | op = vec_oprnd0 | |
6419 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
6420 | ||
6421 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
6422 | { | |
6423 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
6424 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 6425 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
aec7ae7d JJ |
6426 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
6427 | new_stmt | |
0d0e4a03 | 6428 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
6429 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6430 | op = var; | |
6431 | } | |
6432 | ||
6433 | new_stmt | |
d3c2fee0 | 6434 | = gimple_build_call (gather_decl, 5, merge, ptr, op, mask, scale); |
aec7ae7d JJ |
6435 | |
6436 | if (!useless_type_conversion_p (vectype, rettype)) | |
6437 | { | |
6438 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
6439 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 6440 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
aec7ae7d JJ |
6441 | gimple_call_set_lhs (new_stmt, op); |
6442 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 6443 | var = make_ssa_name (vec_dest); |
aec7ae7d JJ |
6444 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
6445 | new_stmt | |
0d0e4a03 | 6446 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
6447 | } |
6448 | else | |
6449 | { | |
6450 | var = make_ssa_name (vec_dest, new_stmt); | |
6451 | gimple_call_set_lhs (new_stmt, var); | |
6452 | } | |
6453 | ||
6454 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6455 | ||
6456 | if (modifier == NARROW) | |
6457 | { | |
6458 | if ((j & 1) == 0) | |
6459 | { | |
6460 | prev_res = var; | |
6461 | continue; | |
6462 | } | |
6463 | var = permute_vec_elements (prev_res, var, | |
6464 | perm_mask, stmt, gsi); | |
6465 | new_stmt = SSA_NAME_DEF_STMT (var); | |
6466 | } | |
6467 | ||
6468 | if (prev_stmt_info == NULL) | |
6469 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6470 | else | |
6471 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6472 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6473 | } | |
6474 | return true; | |
6475 | } | |
f2e2a985 | 6476 | else if (STMT_VINFO_STRIDED_P (stmt_info)) |
7d75abc8 MM |
6477 | { |
6478 | gimple_stmt_iterator incr_gsi; | |
6479 | bool insert_after; | |
355fe088 | 6480 | gimple *incr; |
7d75abc8 | 6481 | tree offvar; |
7d75abc8 MM |
6482 | tree ivstep; |
6483 | tree running_off; | |
9771b263 | 6484 | vec<constructor_elt, va_gc> *v = NULL; |
7d75abc8 | 6485 | gimple_seq stmts = NULL; |
14ac6aa2 RB |
6486 | tree stride_base, stride_step, alias_off; |
6487 | ||
6488 | gcc_assert (!nested_in_vect_loop); | |
7d75abc8 | 6489 | |
f502d50e | 6490 | if (slp && grouped_load) |
ab313a8c RB |
6491 | first_dr = STMT_VINFO_DATA_REF |
6492 | (vinfo_for_stmt (GROUP_FIRST_ELEMENT (stmt_info))); | |
6493 | else | |
6494 | first_dr = dr; | |
6495 | ||
14ac6aa2 RB |
6496 | stride_base |
6497 | = fold_build_pointer_plus | |
ab313a8c | 6498 | (DR_BASE_ADDRESS (first_dr), |
14ac6aa2 | 6499 | size_binop (PLUS_EXPR, |
ab313a8c RB |
6500 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
6501 | convert_to_ptrofftype (DR_INIT (first_dr)))); | |
6502 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); | |
7d75abc8 MM |
6503 | |
6504 | /* For a load with loop-invariant (but other than power-of-2) | |
6505 | stride (i.e. not a grouped access) like so: | |
6506 | ||
6507 | for (i = 0; i < n; i += stride) | |
6508 | ... = array[i]; | |
6509 | ||
6510 | we generate a new induction variable and new accesses to | |
6511 | form a new vector (or vectors, depending on ncopies): | |
6512 | ||
6513 | for (j = 0; ; j += VF*stride) | |
6514 | tmp1 = array[j]; | |
6515 | tmp2 = array[j + stride]; | |
6516 | ... | |
6517 | vectemp = {tmp1, tmp2, ...} | |
6518 | */ | |
6519 | ||
ab313a8c RB |
6520 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step, |
6521 | build_int_cst (TREE_TYPE (stride_step), vf)); | |
7d75abc8 MM |
6522 | |
6523 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
6524 | ||
ab313a8c | 6525 | create_iv (unshare_expr (stride_base), unshare_expr (ivstep), NULL, |
7d75abc8 MM |
6526 | loop, &incr_gsi, insert_after, |
6527 | &offvar, NULL); | |
6528 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 6529 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
7d75abc8 | 6530 | |
ab313a8c RB |
6531 | stride_step = force_gimple_operand (unshare_expr (stride_step), |
6532 | &stmts, true, NULL_TREE); | |
7d75abc8 MM |
6533 | if (stmts) |
6534 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
6535 | ||
6536 | prev_stmt_info = NULL; | |
6537 | running_off = offvar; | |
ab313a8c | 6538 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (first_dr)), 0); |
7b5fc413 RB |
6539 | int nloads = nunits; |
6540 | tree ltype = TREE_TYPE (vectype); | |
b266b968 | 6541 | auto_vec<tree> dr_chain; |
7b5fc413 RB |
6542 | if (slp) |
6543 | { | |
6544 | nloads = nunits / group_size; | |
6545 | if (group_size < nunits) | |
6546 | ltype = build_vector_type (TREE_TYPE (vectype), group_size); | |
6547 | else | |
6548 | ltype = vectype; | |
6549 | ltype = build_aligned_type (ltype, TYPE_ALIGN (TREE_TYPE (vectype))); | |
6550 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
b266b968 RB |
6551 | if (slp_perm) |
6552 | dr_chain.create (ncopies); | |
7b5fc413 | 6553 | } |
7d75abc8 MM |
6554 | for (j = 0; j < ncopies; j++) |
6555 | { | |
6556 | tree vec_inv; | |
6557 | ||
7b5fc413 RB |
6558 | if (nloads > 1) |
6559 | { | |
6560 | vec_alloc (v, nloads); | |
6561 | for (i = 0; i < nloads; i++) | |
6562 | { | |
6563 | tree newref, newoff; | |
355fe088 | 6564 | gimple *incr; |
7b5fc413 RB |
6565 | newref = build2 (MEM_REF, ltype, running_off, alias_off); |
6566 | ||
6567 | newref = force_gimple_operand_gsi (gsi, newref, true, | |
6568 | NULL_TREE, true, | |
6569 | GSI_SAME_STMT); | |
6570 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, newref); | |
6571 | newoff = copy_ssa_name (running_off); | |
6572 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
6573 | running_off, stride_step); | |
6574 | vect_finish_stmt_generation (stmt, incr, gsi); | |
6575 | ||
6576 | running_off = newoff; | |
6577 | } | |
6578 | ||
6579 | vec_inv = build_constructor (vectype, v); | |
6580 | new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi); | |
6581 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
6582 | } | |
6583 | else | |
7d75abc8 | 6584 | { |
7b5fc413 RB |
6585 | new_stmt = gimple_build_assign (make_ssa_name (ltype), |
6586 | build2 (MEM_REF, ltype, | |
6587 | running_off, alias_off)); | |
6588 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6589 | ||
6590 | tree newoff = copy_ssa_name (running_off); | |
355fe088 | 6591 | gimple *incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
0d0e4a03 | 6592 | running_off, stride_step); |
7d75abc8 MM |
6593 | vect_finish_stmt_generation (stmt, incr, gsi); |
6594 | ||
6595 | running_off = newoff; | |
6596 | } | |
6597 | ||
7b5fc413 | 6598 | if (slp) |
b266b968 RB |
6599 | { |
6600 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
6601 | if (slp_perm) | |
6602 | dr_chain.quick_push (gimple_assign_lhs (new_stmt)); | |
6603 | } | |
7d75abc8 | 6604 | else |
225ce44b RB |
6605 | { |
6606 | if (j == 0) | |
6607 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6608 | else | |
6609 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6610 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6611 | } | |
7d75abc8 | 6612 | } |
b266b968 RB |
6613 | if (slp_perm) |
6614 | vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, | |
6615 | slp_node_instance, false); | |
7d75abc8 MM |
6616 | return true; |
6617 | } | |
aec7ae7d | 6618 | |
0d0293ac | 6619 | if (grouped_load) |
ebfd146a | 6620 | { |
e14c1050 | 6621 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
6aa904c4 | 6622 | if (slp |
01d8bf07 | 6623 | && !SLP_TREE_LOAD_PERMUTATION (slp_node).exists () |
9771b263 DN |
6624 | && first_stmt != SLP_TREE_SCALAR_STMTS (slp_node)[0]) |
6625 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 6626 | |
ebfd146a | 6627 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
6628 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
6629 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
6630 | ??? But we can only do so if there is exactly one | |
6631 | as we have no way to get at the rest. Leave the CSE | |
6632 | opportunity alone. | |
6633 | ??? With the group load eventually participating | |
6634 | in multiple different permutations (having multiple | |
6635 | slp nodes which refer to the same group) the CSE | |
6636 | is even wrong code. See PR56270. */ | |
6637 | && !slp) | |
ebfd146a IR |
6638 | { |
6639 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
6640 | return true; | |
6641 | } | |
6642 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 6643 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
9b999e8c | 6644 | group_gap_adj = 0; |
ebfd146a IR |
6645 | |
6646 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
6647 | if (slp) | |
6648 | { | |
0d0293ac | 6649 | grouped_load = false; |
91ff1504 RB |
6650 | /* For SLP permutation support we need to load the whole group, |
6651 | not only the number of vector stmts the permutation result | |
6652 | fits in. */ | |
6653 | if (slp_perm) | |
6654 | vec_num = (group_size * vf + nunits - 1) / nunits; | |
6655 | else | |
6656 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9b999e8c | 6657 | group_gap_adj = vf * group_size - nunits * vec_num; |
a70d6342 | 6658 | } |
ebfd146a | 6659 | else |
9b999e8c | 6660 | vec_num = group_size; |
ebfd146a IR |
6661 | } |
6662 | else | |
6663 | { | |
6664 | first_stmt = stmt; | |
6665 | first_dr = dr; | |
6666 | group_size = vec_num = 1; | |
9b999e8c | 6667 | group_gap_adj = 0; |
ebfd146a IR |
6668 | } |
6669 | ||
720f5239 | 6670 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 6671 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
6672 | /* Targets with load-lane instructions must not require explicit |
6673 | realignment. */ | |
6674 | gcc_assert (!load_lanes_p | |
6675 | || alignment_support_scheme == dr_aligned | |
6676 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
6677 | |
6678 | /* In case the vectorization factor (VF) is bigger than the number | |
6679 | of elements that we can fit in a vectype (nunits), we have to generate | |
6680 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 6681 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 6682 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 6683 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 6684 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
6685 | stmts that use the defs of the current stmt. The example below |
6686 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
6687 | need to create 4 vectorized stmts): | |
ebfd146a IR |
6688 | |
6689 | before vectorization: | |
6690 | RELATED_STMT VEC_STMT | |
6691 | S1: x = memref - - | |
6692 | S2: z = x + 1 - - | |
6693 | ||
6694 | step 1: vectorize stmt S1: | |
6695 | We first create the vector stmt VS1_0, and, as usual, record a | |
6696 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
6697 | Next, we create the vector stmt VS1_1, and record a pointer to | |
6698 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 6699 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
6700 | stmts and pointers: |
6701 | RELATED_STMT VEC_STMT | |
6702 | VS1_0: vx0 = memref0 VS1_1 - | |
6703 | VS1_1: vx1 = memref1 VS1_2 - | |
6704 | VS1_2: vx2 = memref2 VS1_3 - | |
6705 | VS1_3: vx3 = memref3 - - | |
6706 | S1: x = load - VS1_0 | |
6707 | S2: z = x + 1 - - | |
6708 | ||
b8698a0f L |
6709 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
6710 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
6711 | stmt S2. */ |
6712 | ||
0d0293ac | 6713 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
6714 | |
6715 | S1: x2 = &base + 2 | |
6716 | S2: x0 = &base | |
6717 | S3: x1 = &base + 1 | |
6718 | S4: x3 = &base + 3 | |
6719 | ||
b8698a0f | 6720 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
6721 | starting from the access of the first stmt of the chain: |
6722 | ||
6723 | VS1: vx0 = &base | |
6724 | VS2: vx1 = &base + vec_size*1 | |
6725 | VS3: vx3 = &base + vec_size*2 | |
6726 | VS4: vx4 = &base + vec_size*3 | |
6727 | ||
6728 | Then permutation statements are generated: | |
6729 | ||
e2c83630 RH |
6730 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
6731 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
6732 | ... |
6733 | ||
6734 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
6735 | (the order of the data-refs in the output of vect_permute_load_chain | |
6736 | corresponds to the order of scalar stmts in the interleaving chain - see | |
6737 | the documentation of vect_permute_load_chain()). | |
6738 | The generation of permutation stmts and recording them in | |
0d0293ac | 6739 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 6740 | |
b8698a0f | 6741 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
6742 | permutation stmts above are created for every copy. The result vector |
6743 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
6744 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
6745 | |
6746 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
6747 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
6748 | we generate the following code: | |
6749 | p = initial_addr; | |
6750 | indx = 0; | |
6751 | loop { | |
6752 | p = p + indx * vectype_size; | |
6753 | vec_dest = *(p); | |
6754 | indx = indx + 1; | |
6755 | } | |
6756 | ||
6757 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 6758 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
6759 | then generate the following code, in which the data in each iteration is |
6760 | obtained by two vector loads, one from the previous iteration, and one | |
6761 | from the current iteration: | |
6762 | p1 = initial_addr; | |
6763 | msq_init = *(floor(p1)) | |
6764 | p2 = initial_addr + VS - 1; | |
6765 | realignment_token = call target_builtin; | |
6766 | indx = 0; | |
6767 | loop { | |
6768 | p2 = p2 + indx * vectype_size | |
6769 | lsq = *(floor(p2)) | |
6770 | vec_dest = realign_load (msq, lsq, realignment_token) | |
6771 | indx = indx + 1; | |
6772 | msq = lsq; | |
6773 | } */ | |
6774 | ||
6775 | /* If the misalignment remains the same throughout the execution of the | |
6776 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 6777 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
6778 | This can only occur when vectorizing memory accesses in the inner-loop |
6779 | nested within an outer-loop that is being vectorized. */ | |
6780 | ||
d1e4b493 | 6781 | if (nested_in_vect_loop |
211bea38 | 6782 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
ebfd146a IR |
6783 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) |
6784 | { | |
6785 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
6786 | compute_in_loop = true; | |
6787 | } | |
6788 | ||
6789 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
6790 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 6791 | && !compute_in_loop) |
ebfd146a IR |
6792 | { |
6793 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
6794 | alignment_support_scheme, NULL_TREE, | |
6795 | &at_loop); | |
6796 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
6797 | { | |
538dd0b7 | 6798 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (msq)); |
356bbc4c JJ |
6799 | byte_offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
6800 | size_one_node); | |
ebfd146a IR |
6801 | } |
6802 | } | |
6803 | else | |
6804 | at_loop = loop; | |
6805 | ||
a1e53f3f L |
6806 | if (negative) |
6807 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
6808 | ||
272c6793 RS |
6809 | if (load_lanes_p) |
6810 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
6811 | else | |
6812 | aggr_type = vectype; | |
6813 | ||
ebfd146a IR |
6814 | prev_stmt_info = NULL; |
6815 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 6816 | { |
272c6793 | 6817 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 6818 | if (j == 0) |
74bf76ed JJ |
6819 | { |
6820 | bool simd_lane_access_p | |
6821 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
6822 | if (simd_lane_access_p | |
6823 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
6824 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
6825 | && integer_zerop (DR_OFFSET (first_dr)) | |
6826 | && integer_zerop (DR_INIT (first_dr)) | |
6827 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
6828 | get_alias_set (DR_REF (first_dr))) | |
6829 | && (alignment_support_scheme == dr_aligned | |
6830 | || alignment_support_scheme == dr_unaligned_supported)) | |
6831 | { | |
6832 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
6833 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
6834 | (DR_REF (first_dr)), 0); | |
8928eff3 | 6835 | inv_p = false; |
74bf76ed JJ |
6836 | } |
6837 | else | |
6838 | dataref_ptr | |
6839 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
6840 | offset, &dummy, gsi, &ptr_incr, | |
356bbc4c JJ |
6841 | simd_lane_access_p, &inv_p, |
6842 | byte_offset); | |
74bf76ed JJ |
6843 | } |
6844 | else if (dataref_offset) | |
6845 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
6846 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 6847 | else |
272c6793 RS |
6848 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
6849 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 6850 | |
0d0293ac | 6851 | if (grouped_load || slp_perm) |
9771b263 | 6852 | dr_chain.create (vec_num); |
5ce1ee7f | 6853 | |
272c6793 | 6854 | if (load_lanes_p) |
ebfd146a | 6855 | { |
272c6793 RS |
6856 | tree vec_array; |
6857 | ||
6858 | vec_array = create_vector_array (vectype, vec_num); | |
6859 | ||
6860 | /* Emit: | |
6861 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
6862 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
6863 | new_stmt = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
6864 | gimple_call_set_lhs (new_stmt, vec_array); | |
6865 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a | 6866 | |
272c6793 RS |
6867 | /* Extract each vector into an SSA_NAME. */ |
6868 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 6869 | { |
272c6793 RS |
6870 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
6871 | vec_array, i); | |
9771b263 | 6872 | dr_chain.quick_push (new_temp); |
272c6793 RS |
6873 | } |
6874 | ||
6875 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 6876 | vect_record_grouped_load_vectors (stmt, dr_chain); |
272c6793 RS |
6877 | } |
6878 | else | |
6879 | { | |
6880 | for (i = 0; i < vec_num; i++) | |
6881 | { | |
6882 | if (i > 0) | |
6883 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
6884 | stmt, NULL_TREE); | |
6885 | ||
6886 | /* 2. Create the vector-load in the loop. */ | |
6887 | switch (alignment_support_scheme) | |
6888 | { | |
6889 | case dr_aligned: | |
6890 | case dr_unaligned_supported: | |
be1ac4ec | 6891 | { |
644ffefd MJ |
6892 | unsigned int align, misalign; |
6893 | ||
272c6793 | 6894 | data_ref |
aed93b23 RB |
6895 | = fold_build2 (MEM_REF, vectype, dataref_ptr, |
6896 | dataref_offset | |
6897 | ? dataref_offset | |
6898 | : build_int_cst (reference_alias_ptr_type | |
6899 | (DR_REF (first_dr)), 0)); | |
644ffefd | 6900 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 RS |
6901 | if (alignment_support_scheme == dr_aligned) |
6902 | { | |
6903 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 6904 | misalign = 0; |
272c6793 RS |
6905 | } |
6906 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
6907 | { | |
52639a61 RB |
6908 | if (DR_VECT_AUX (first_dr)->base_element_aligned) |
6909 | align = TYPE_ALIGN_UNIT (elem_type); | |
6910 | else | |
6911 | align = (get_object_alignment (DR_REF (first_dr)) | |
6912 | / BITS_PER_UNIT); | |
6913 | misalign = 0; | |
272c6793 RS |
6914 | TREE_TYPE (data_ref) |
6915 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 6916 | align * BITS_PER_UNIT); |
272c6793 RS |
6917 | } |
6918 | else | |
6919 | { | |
6920 | TREE_TYPE (data_ref) | |
6921 | = build_aligned_type (TREE_TYPE (data_ref), | |
6922 | TYPE_ALIGN (elem_type)); | |
644ffefd | 6923 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 6924 | } |
aed93b23 RB |
6925 | if (dataref_offset == NULL_TREE |
6926 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
6927 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), |
6928 | align, misalign); | |
272c6793 | 6929 | break; |
be1ac4ec | 6930 | } |
272c6793 | 6931 | case dr_explicit_realign: |
267d3070 | 6932 | { |
272c6793 | 6933 | tree ptr, bump; |
272c6793 | 6934 | |
d88981fc | 6935 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
272c6793 RS |
6936 | |
6937 | if (compute_in_loop) | |
6938 | msq = vect_setup_realignment (first_stmt, gsi, | |
6939 | &realignment_token, | |
6940 | dr_explicit_realign, | |
6941 | dataref_ptr, NULL); | |
6942 | ||
aed93b23 RB |
6943 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
6944 | ptr = copy_ssa_name (dataref_ptr); | |
6945 | else | |
6946 | ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
0d0e4a03 JJ |
6947 | new_stmt = gimple_build_assign |
6948 | (ptr, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
6949 | build_int_cst |
6950 | (TREE_TYPE (dataref_ptr), | |
6951 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
6952 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6953 | data_ref | |
6954 | = build2 (MEM_REF, vectype, ptr, | |
6955 | build_int_cst (reference_alias_ptr_type | |
6956 | (DR_REF (first_dr)), 0)); | |
6957 | vec_dest = vect_create_destination_var (scalar_dest, | |
6958 | vectype); | |
6959 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
6960 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
6961 | gimple_assign_set_lhs (new_stmt, new_temp); | |
6962 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
6963 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
6964 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6965 | msq = new_temp; | |
6966 | ||
d88981fc | 6967 | bump = size_binop (MULT_EXPR, vs, |
7b7b1813 | 6968 | TYPE_SIZE_UNIT (elem_type)); |
d88981fc | 6969 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
272c6793 | 6970 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
0d0e4a03 JJ |
6971 | new_stmt = gimple_build_assign |
6972 | (NULL_TREE, BIT_AND_EXPR, ptr, | |
272c6793 RS |
6973 | build_int_cst |
6974 | (TREE_TYPE (ptr), | |
6975 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
aed93b23 | 6976 | ptr = copy_ssa_name (ptr, new_stmt); |
272c6793 RS |
6977 | gimple_assign_set_lhs (new_stmt, ptr); |
6978 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6979 | data_ref | |
6980 | = build2 (MEM_REF, vectype, ptr, | |
6981 | build_int_cst (reference_alias_ptr_type | |
6982 | (DR_REF (first_dr)), 0)); | |
6983 | break; | |
267d3070 | 6984 | } |
272c6793 | 6985 | case dr_explicit_realign_optimized: |
aed93b23 RB |
6986 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
6987 | new_temp = copy_ssa_name (dataref_ptr); | |
6988 | else | |
6989 | new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
0d0e4a03 JJ |
6990 | new_stmt = gimple_build_assign |
6991 | (new_temp, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
6992 | build_int_cst |
6993 | (TREE_TYPE (dataref_ptr), | |
6994 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
6995 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6996 | data_ref | |
6997 | = build2 (MEM_REF, vectype, new_temp, | |
6998 | build_int_cst (reference_alias_ptr_type | |
6999 | (DR_REF (first_dr)), 0)); | |
7000 | break; | |
7001 | default: | |
7002 | gcc_unreachable (); | |
7003 | } | |
ebfd146a | 7004 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 7005 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
7006 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7007 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7008 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7009 | ||
272c6793 RS |
7010 | /* 3. Handle explicit realignment if necessary/supported. |
7011 | Create in loop: | |
7012 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
7013 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
7014 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 7015 | { |
272c6793 RS |
7016 | lsq = gimple_assign_lhs (new_stmt); |
7017 | if (!realignment_token) | |
7018 | realignment_token = dataref_ptr; | |
7019 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
0d0e4a03 JJ |
7020 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, |
7021 | msq, lsq, realignment_token); | |
272c6793 RS |
7022 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7023 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7024 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7025 | ||
7026 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7027 | { | |
7028 | gcc_assert (phi); | |
7029 | if (i == vec_num - 1 && j == ncopies - 1) | |
7030 | add_phi_arg (phi, lsq, | |
7031 | loop_latch_edge (containing_loop), | |
9e227d60 | 7032 | UNKNOWN_LOCATION); |
272c6793 RS |
7033 | msq = lsq; |
7034 | } | |
ebfd146a | 7035 | } |
ebfd146a | 7036 | |
59fd17e3 RB |
7037 | /* 4. Handle invariant-load. */ |
7038 | if (inv_p && !bb_vinfo) | |
7039 | { | |
59fd17e3 | 7040 | gcc_assert (!grouped_load); |
d1417442 JJ |
7041 | /* If we have versioned for aliasing or the loop doesn't |
7042 | have any data dependencies that would preclude this, | |
7043 | then we are sure this is a loop invariant load and | |
7044 | thus we can insert it on the preheader edge. */ | |
7045 | if (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) | |
7046 | && !nested_in_vect_loop | |
6b916b36 | 7047 | && hoist_defs_of_uses (stmt, loop)) |
a0e35eb0 RB |
7048 | { |
7049 | if (dump_enabled_p ()) | |
7050 | { | |
7051 | dump_printf_loc (MSG_NOTE, vect_location, | |
7052 | "hoisting out of the vectorized " | |
7053 | "loop: "); | |
7054 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
a0e35eb0 | 7055 | } |
b731b390 | 7056 | tree tem = copy_ssa_name (scalar_dest); |
a0e35eb0 RB |
7057 | gsi_insert_on_edge_immediate |
7058 | (loop_preheader_edge (loop), | |
7059 | gimple_build_assign (tem, | |
7060 | unshare_expr | |
7061 | (gimple_assign_rhs1 (stmt)))); | |
7062 | new_temp = vect_init_vector (stmt, tem, vectype, NULL); | |
7063 | } | |
7064 | else | |
7065 | { | |
7066 | gimple_stmt_iterator gsi2 = *gsi; | |
7067 | gsi_next (&gsi2); | |
7068 | new_temp = vect_init_vector (stmt, scalar_dest, | |
7069 | vectype, &gsi2); | |
7070 | } | |
59fd17e3 | 7071 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
a0e35eb0 | 7072 | set_vinfo_for_stmt (new_stmt, |
310213d4 | 7073 | new_stmt_vec_info (new_stmt, vinfo)); |
59fd17e3 RB |
7074 | } |
7075 | ||
272c6793 RS |
7076 | if (negative) |
7077 | { | |
aec7ae7d JJ |
7078 | tree perm_mask = perm_mask_for_reverse (vectype); |
7079 | new_temp = permute_vec_elements (new_temp, new_temp, | |
7080 | perm_mask, stmt, gsi); | |
ebfd146a IR |
7081 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
7082 | } | |
267d3070 | 7083 | |
272c6793 | 7084 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
7085 | vect_transform_grouped_load (). */ |
7086 | if (grouped_load || slp_perm) | |
9771b263 | 7087 | dr_chain.quick_push (new_temp); |
267d3070 | 7088 | |
272c6793 RS |
7089 | /* Store vector loads in the corresponding SLP_NODE. */ |
7090 | if (slp && !slp_perm) | |
9771b263 | 7091 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
272c6793 | 7092 | } |
9b999e8c RB |
7093 | /* Bump the vector pointer to account for a gap or for excess |
7094 | elements loaded for a permuted SLP load. */ | |
7095 | if (group_gap_adj != 0) | |
a64b9c26 | 7096 | { |
9b999e8c RB |
7097 | bool ovf; |
7098 | tree bump | |
7099 | = wide_int_to_tree (sizetype, | |
7100 | wi::smul (TYPE_SIZE_UNIT (elem_type), | |
7101 | group_gap_adj, &ovf)); | |
a64b9c26 RB |
7102 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
7103 | stmt, bump); | |
7104 | } | |
ebfd146a IR |
7105 | } |
7106 | ||
7107 | if (slp && !slp_perm) | |
7108 | continue; | |
7109 | ||
7110 | if (slp_perm) | |
7111 | { | |
01d8bf07 | 7112 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
ebfd146a IR |
7113 | slp_node_instance, false)) |
7114 | { | |
9771b263 | 7115 | dr_chain.release (); |
ebfd146a IR |
7116 | return false; |
7117 | } | |
7118 | } | |
7119 | else | |
7120 | { | |
0d0293ac | 7121 | if (grouped_load) |
ebfd146a | 7122 | { |
272c6793 | 7123 | if (!load_lanes_p) |
0d0293ac | 7124 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 7125 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
7126 | } |
7127 | else | |
7128 | { | |
7129 | if (j == 0) | |
7130 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7131 | else | |
7132 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7133 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7134 | } | |
7135 | } | |
9771b263 | 7136 | dr_chain.release (); |
ebfd146a IR |
7137 | } |
7138 | ||
ebfd146a IR |
7139 | return true; |
7140 | } | |
7141 | ||
7142 | /* Function vect_is_simple_cond. | |
b8698a0f | 7143 | |
ebfd146a IR |
7144 | Input: |
7145 | LOOP - the loop that is being vectorized. | |
7146 | COND - Condition that is checked for simple use. | |
7147 | ||
e9e1d143 RG |
7148 | Output: |
7149 | *COMP_VECTYPE - the vector type for the comparison. | |
7150 | ||
ebfd146a IR |
7151 | Returns whether a COND can be vectorized. Checks whether |
7152 | condition operands are supportable using vec_is_simple_use. */ | |
7153 | ||
87aab9b2 | 7154 | static bool |
81c40241 | 7155 | vect_is_simple_cond (tree cond, vec_info *vinfo, tree *comp_vectype) |
ebfd146a IR |
7156 | { |
7157 | tree lhs, rhs; | |
ebfd146a | 7158 | enum vect_def_type dt; |
e9e1d143 | 7159 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a IR |
7160 | |
7161 | if (!COMPARISON_CLASS_P (cond)) | |
7162 | return false; | |
7163 | ||
7164 | lhs = TREE_OPERAND (cond, 0); | |
7165 | rhs = TREE_OPERAND (cond, 1); | |
7166 | ||
7167 | if (TREE_CODE (lhs) == SSA_NAME) | |
7168 | { | |
355fe088 | 7169 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); |
81c40241 | 7170 | if (!vect_is_simple_use (lhs, vinfo, &lhs_def_stmt, &dt, &vectype1)) |
ebfd146a IR |
7171 | return false; |
7172 | } | |
7173 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
7174 | && TREE_CODE (lhs) != FIXED_CST) | |
7175 | return false; | |
7176 | ||
7177 | if (TREE_CODE (rhs) == SSA_NAME) | |
7178 | { | |
355fe088 | 7179 | gimple *rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); |
81c40241 | 7180 | if (!vect_is_simple_use (rhs, vinfo, &rhs_def_stmt, &dt, &vectype2)) |
ebfd146a IR |
7181 | return false; |
7182 | } | |
f7e531cf | 7183 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST |
ebfd146a IR |
7184 | && TREE_CODE (rhs) != FIXED_CST) |
7185 | return false; | |
7186 | ||
e9e1d143 | 7187 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
ebfd146a IR |
7188 | return true; |
7189 | } | |
7190 | ||
7191 | /* vectorizable_condition. | |
7192 | ||
b8698a0f L |
7193 | Check if STMT is conditional modify expression that can be vectorized. |
7194 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
7195 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
7196 | at GSI. |
7197 | ||
7198 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
7199 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
0ad23163 | 7200 | else clause if it is 2). |
ebfd146a IR |
7201 | |
7202 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7203 | ||
4bbe8262 | 7204 | bool |
355fe088 TS |
7205 | vectorizable_condition (gimple *stmt, gimple_stmt_iterator *gsi, |
7206 | gimple **vec_stmt, tree reduc_def, int reduc_index, | |
f7e531cf | 7207 | slp_tree slp_node) |
ebfd146a IR |
7208 | { |
7209 | tree scalar_dest = NULL_TREE; | |
7210 | tree vec_dest = NULL_TREE; | |
ebfd146a IR |
7211 | tree cond_expr, then_clause, else_clause; |
7212 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
df11cc78 | 7213 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
7214 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
7215 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
ebfd146a IR |
7216 | tree vec_compare, vec_cond_expr; |
7217 | tree new_temp; | |
7218 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a855b1b1 | 7219 | enum vect_def_type dt, dts[4]; |
f7e531cf | 7220 | int ncopies; |
ebfd146a | 7221 | enum tree_code code; |
a855b1b1 | 7222 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
7223 | int i, j; |
7224 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
7225 | vec<tree> vec_oprnds0 = vNULL; |
7226 | vec<tree> vec_oprnds1 = vNULL; | |
7227 | vec<tree> vec_oprnds2 = vNULL; | |
7228 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 7229 | tree vec_cmp_type; |
b8698a0f | 7230 | |
f7e531cf IR |
7231 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
7232 | return false; | |
7233 | ||
af29617a AH |
7234 | if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) == TREE_CODE_REDUCTION) |
7235 | { | |
7236 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
7237 | return false; | |
ebfd146a | 7238 | |
af29617a AH |
7239 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7240 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
7241 | && reduc_def)) | |
7242 | return false; | |
ebfd146a | 7243 | |
af29617a AH |
7244 | /* FORNOW: not yet supported. */ |
7245 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
7246 | { | |
7247 | if (dump_enabled_p ()) | |
7248 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7249 | "value used after loop.\n"); | |
7250 | return false; | |
7251 | } | |
ebfd146a IR |
7252 | } |
7253 | ||
7254 | /* Is vectorizable conditional operation? */ | |
7255 | if (!is_gimple_assign (stmt)) | |
7256 | return false; | |
7257 | ||
7258 | code = gimple_assign_rhs_code (stmt); | |
7259 | ||
7260 | if (code != COND_EXPR) | |
7261 | return false; | |
7262 | ||
465c8c19 JJ |
7263 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
7264 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
7265 | ||
7266 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
7267 | ncopies = 1; | |
7268 | else | |
7269 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
7270 | ||
7271 | gcc_assert (ncopies >= 1); | |
7272 | if (reduc_index && ncopies > 1) | |
7273 | return false; /* FORNOW */ | |
7274 | ||
4e71066d RG |
7275 | cond_expr = gimple_assign_rhs1 (stmt); |
7276 | then_clause = gimple_assign_rhs2 (stmt); | |
7277 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 7278 | |
81c40241 | 7279 | if (!vect_is_simple_cond (cond_expr, stmt_info->vinfo, &comp_vectype) |
e9e1d143 | 7280 | || !comp_vectype) |
ebfd146a IR |
7281 | return false; |
7282 | ||
81c40241 RB |
7283 | gimple *def_stmt; |
7284 | if (!vect_is_simple_use (then_clause, stmt_info->vinfo, &def_stmt, &dt)) | |
ebfd146a | 7285 | return false; |
81c40241 | 7286 | if (!vect_is_simple_use (else_clause, stmt_info->vinfo, &def_stmt, &dt)) |
ebfd146a IR |
7287 | return false; |
7288 | ||
9f47c7e5 | 7289 | vec_cmp_type = build_same_sized_truth_vector_type (comp_vectype); |
74946978 MP |
7290 | if (vec_cmp_type == NULL_TREE) |
7291 | return false; | |
784fb9b3 | 7292 | |
b8698a0f | 7293 | if (!vec_stmt) |
ebfd146a IR |
7294 | { |
7295 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
e9e1d143 | 7296 | return expand_vec_cond_expr_p (vectype, comp_vectype); |
ebfd146a IR |
7297 | } |
7298 | ||
f7e531cf IR |
7299 | /* Transform. */ |
7300 | ||
7301 | if (!slp_node) | |
7302 | { | |
9771b263 DN |
7303 | vec_oprnds0.create (1); |
7304 | vec_oprnds1.create (1); | |
7305 | vec_oprnds2.create (1); | |
7306 | vec_oprnds3.create (1); | |
f7e531cf | 7307 | } |
ebfd146a IR |
7308 | |
7309 | /* Handle def. */ | |
7310 | scalar_dest = gimple_assign_lhs (stmt); | |
7311 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
7312 | ||
7313 | /* Handle cond expr. */ | |
a855b1b1 MM |
7314 | for (j = 0; j < ncopies; j++) |
7315 | { | |
538dd0b7 | 7316 | gassign *new_stmt = NULL; |
a855b1b1 MM |
7317 | if (j == 0) |
7318 | { | |
f7e531cf IR |
7319 | if (slp_node) |
7320 | { | |
00f96dc9 TS |
7321 | auto_vec<tree, 4> ops; |
7322 | auto_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 7323 | |
9771b263 DN |
7324 | ops.safe_push (TREE_OPERAND (cond_expr, 0)); |
7325 | ops.safe_push (TREE_OPERAND (cond_expr, 1)); | |
7326 | ops.safe_push (then_clause); | |
7327 | ops.safe_push (else_clause); | |
f7e531cf | 7328 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); |
37b5ec8f JJ |
7329 | vec_oprnds3 = vec_defs.pop (); |
7330 | vec_oprnds2 = vec_defs.pop (); | |
7331 | vec_oprnds1 = vec_defs.pop (); | |
7332 | vec_oprnds0 = vec_defs.pop (); | |
f7e531cf | 7333 | |
9771b263 DN |
7334 | ops.release (); |
7335 | vec_defs.release (); | |
f7e531cf IR |
7336 | } |
7337 | else | |
7338 | { | |
355fe088 | 7339 | gimple *gtemp; |
f7e531cf | 7340 | vec_cond_lhs = |
42fd8198 IE |
7341 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), |
7342 | stmt, comp_vectype); | |
81c40241 RB |
7343 | vect_is_simple_use (TREE_OPERAND (cond_expr, 0), |
7344 | loop_vinfo, >emp, &dts[0]); | |
f7e531cf IR |
7345 | |
7346 | vec_cond_rhs = | |
7347 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), | |
42fd8198 | 7348 | stmt, comp_vectype); |
81c40241 RB |
7349 | vect_is_simple_use (TREE_OPERAND (cond_expr, 1), |
7350 | loop_vinfo, >emp, &dts[1]); | |
f7e531cf IR |
7351 | if (reduc_index == 1) |
7352 | vec_then_clause = reduc_def; | |
7353 | else | |
7354 | { | |
7355 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
81c40241 RB |
7356 | stmt); |
7357 | vect_is_simple_use (then_clause, loop_vinfo, | |
7358 | >emp, &dts[2]); | |
f7e531cf IR |
7359 | } |
7360 | if (reduc_index == 2) | |
7361 | vec_else_clause = reduc_def; | |
7362 | else | |
7363 | { | |
7364 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
81c40241 RB |
7365 | stmt); |
7366 | vect_is_simple_use (else_clause, loop_vinfo, >emp, &dts[3]); | |
f7e531cf | 7367 | } |
a855b1b1 MM |
7368 | } |
7369 | } | |
7370 | else | |
7371 | { | |
f7e531cf | 7372 | vec_cond_lhs = vect_get_vec_def_for_stmt_copy (dts[0], |
9771b263 | 7373 | vec_oprnds0.pop ()); |
f7e531cf | 7374 | vec_cond_rhs = vect_get_vec_def_for_stmt_copy (dts[1], |
9771b263 | 7375 | vec_oprnds1.pop ()); |
a855b1b1 | 7376 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 7377 | vec_oprnds2.pop ()); |
a855b1b1 | 7378 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 7379 | vec_oprnds3.pop ()); |
f7e531cf IR |
7380 | } |
7381 | ||
7382 | if (!slp_node) | |
7383 | { | |
9771b263 DN |
7384 | vec_oprnds0.quick_push (vec_cond_lhs); |
7385 | vec_oprnds1.quick_push (vec_cond_rhs); | |
7386 | vec_oprnds2.quick_push (vec_then_clause); | |
7387 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
7388 | } |
7389 | ||
9dc3f7de | 7390 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 7391 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 7392 | { |
9771b263 DN |
7393 | vec_cond_rhs = vec_oprnds1[i]; |
7394 | vec_then_clause = vec_oprnds2[i]; | |
7395 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 7396 | |
784fb9b3 JJ |
7397 | vec_compare = build2 (TREE_CODE (cond_expr), vec_cmp_type, |
7398 | vec_cond_lhs, vec_cond_rhs); | |
f7e531cf IR |
7399 | vec_cond_expr = build3 (VEC_COND_EXPR, vectype, |
7400 | vec_compare, vec_then_clause, vec_else_clause); | |
a855b1b1 | 7401 | |
f7e531cf IR |
7402 | new_stmt = gimple_build_assign (vec_dest, vec_cond_expr); |
7403 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
7404 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7405 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7406 | if (slp_node) | |
9771b263 | 7407 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
7408 | } |
7409 | ||
7410 | if (slp_node) | |
7411 | continue; | |
7412 | ||
7413 | if (j == 0) | |
7414 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7415 | else | |
7416 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7417 | ||
7418 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 7419 | } |
b8698a0f | 7420 | |
9771b263 DN |
7421 | vec_oprnds0.release (); |
7422 | vec_oprnds1.release (); | |
7423 | vec_oprnds2.release (); | |
7424 | vec_oprnds3.release (); | |
f7e531cf | 7425 | |
ebfd146a IR |
7426 | return true; |
7427 | } | |
7428 | ||
42fd8198 IE |
7429 | /* vectorizable_comparison. |
7430 | ||
7431 | Check if STMT is comparison expression that can be vectorized. | |
7432 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
7433 | comparison, put it in VEC_STMT, and insert it at GSI. | |
7434 | ||
7435 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7436 | ||
7437 | bool | |
7438 | vectorizable_comparison (gimple *stmt, gimple_stmt_iterator *gsi, | |
7439 | gimple **vec_stmt, tree reduc_def, | |
7440 | slp_tree slp_node) | |
7441 | { | |
7442 | tree lhs, rhs1, rhs2; | |
7443 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
7444 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; | |
7445 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
7446 | tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE; | |
7447 | tree new_temp; | |
7448 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
7449 | enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
7450 | unsigned nunits; | |
7451 | int ncopies; | |
7452 | enum tree_code code; | |
7453 | stmt_vec_info prev_stmt_info = NULL; | |
7454 | int i, j; | |
7455 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
7456 | vec<tree> vec_oprnds0 = vNULL; | |
7457 | vec<tree> vec_oprnds1 = vNULL; | |
7458 | gimple *def_stmt; | |
7459 | tree mask_type; | |
7460 | tree mask; | |
7461 | ||
7462 | if (!VECTOR_BOOLEAN_TYPE_P (vectype)) | |
7463 | return false; | |
7464 | ||
7465 | mask_type = vectype; | |
7466 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
7467 | ||
7468 | if (slp_node || PURE_SLP_STMT (stmt_info)) | |
7469 | ncopies = 1; | |
7470 | else | |
7471 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
7472 | ||
7473 | gcc_assert (ncopies >= 1); | |
7474 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
7475 | return false; | |
7476 | ||
7477 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def | |
7478 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
7479 | && reduc_def)) | |
7480 | return false; | |
7481 | ||
7482 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
7483 | { | |
7484 | if (dump_enabled_p ()) | |
7485 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7486 | "value used after loop.\n"); | |
7487 | return false; | |
7488 | } | |
7489 | ||
7490 | if (!is_gimple_assign (stmt)) | |
7491 | return false; | |
7492 | ||
7493 | code = gimple_assign_rhs_code (stmt); | |
7494 | ||
7495 | if (TREE_CODE_CLASS (code) != tcc_comparison) | |
7496 | return false; | |
7497 | ||
7498 | rhs1 = gimple_assign_rhs1 (stmt); | |
7499 | rhs2 = gimple_assign_rhs2 (stmt); | |
7500 | ||
7501 | if (!vect_is_simple_use (rhs1, stmt_info->vinfo, &def_stmt, | |
7502 | &dts[0], &vectype1)) | |
7503 | return false; | |
7504 | ||
7505 | if (!vect_is_simple_use (rhs2, stmt_info->vinfo, &def_stmt, | |
7506 | &dts[1], &vectype2)) | |
7507 | return false; | |
7508 | ||
7509 | if (vectype1 && vectype2 | |
7510 | && TYPE_VECTOR_SUBPARTS (vectype1) != TYPE_VECTOR_SUBPARTS (vectype2)) | |
7511 | return false; | |
7512 | ||
7513 | vectype = vectype1 ? vectype1 : vectype2; | |
7514 | ||
7515 | /* Invariant comparison. */ | |
7516 | if (!vectype) | |
7517 | { | |
7518 | vectype = build_vector_type (TREE_TYPE (rhs1), nunits); | |
7519 | if (tree_to_shwi (TYPE_SIZE_UNIT (vectype)) != current_vector_size) | |
7520 | return false; | |
7521 | } | |
7522 | else if (nunits != TYPE_VECTOR_SUBPARTS (vectype)) | |
7523 | return false; | |
7524 | ||
7525 | if (!vec_stmt) | |
7526 | { | |
7527 | STMT_VINFO_TYPE (stmt_info) = comparison_vec_info_type; | |
7528 | vect_model_simple_cost (stmt_info, ncopies, dts, NULL, NULL); | |
7529 | return expand_vec_cmp_expr_p (vectype, mask_type); | |
7530 | } | |
7531 | ||
7532 | /* Transform. */ | |
7533 | if (!slp_node) | |
7534 | { | |
7535 | vec_oprnds0.create (1); | |
7536 | vec_oprnds1.create (1); | |
7537 | } | |
7538 | ||
7539 | /* Handle def. */ | |
7540 | lhs = gimple_assign_lhs (stmt); | |
7541 | mask = vect_create_destination_var (lhs, mask_type); | |
7542 | ||
7543 | /* Handle cmp expr. */ | |
7544 | for (j = 0; j < ncopies; j++) | |
7545 | { | |
7546 | gassign *new_stmt = NULL; | |
7547 | if (j == 0) | |
7548 | { | |
7549 | if (slp_node) | |
7550 | { | |
7551 | auto_vec<tree, 2> ops; | |
7552 | auto_vec<vec<tree>, 2> vec_defs; | |
7553 | ||
7554 | ops.safe_push (rhs1); | |
7555 | ops.safe_push (rhs2); | |
7556 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); | |
7557 | vec_oprnds1 = vec_defs.pop (); | |
7558 | vec_oprnds0 = vec_defs.pop (); | |
7559 | } | |
7560 | else | |
7561 | { | |
7562 | vec_rhs1 = vect_get_vec_def_for_operand (rhs1, stmt, NULL); | |
7563 | vec_rhs2 = vect_get_vec_def_for_operand (rhs2, stmt, NULL); | |
7564 | } | |
7565 | } | |
7566 | else | |
7567 | { | |
7568 | vec_rhs1 = vect_get_vec_def_for_stmt_copy (dts[0], | |
7569 | vec_oprnds0.pop ()); | |
7570 | vec_rhs2 = vect_get_vec_def_for_stmt_copy (dts[1], | |
7571 | vec_oprnds1.pop ()); | |
7572 | } | |
7573 | ||
7574 | if (!slp_node) | |
7575 | { | |
7576 | vec_oprnds0.quick_push (vec_rhs1); | |
7577 | vec_oprnds1.quick_push (vec_rhs2); | |
7578 | } | |
7579 | ||
7580 | /* Arguments are ready. Create the new vector stmt. */ | |
7581 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1) | |
7582 | { | |
7583 | vec_rhs2 = vec_oprnds1[i]; | |
7584 | ||
7585 | new_temp = make_ssa_name (mask); | |
7586 | new_stmt = gimple_build_assign (new_temp, code, vec_rhs1, vec_rhs2); | |
7587 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7588 | if (slp_node) | |
7589 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
7590 | } | |
7591 | ||
7592 | if (slp_node) | |
7593 | continue; | |
7594 | ||
7595 | if (j == 0) | |
7596 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7597 | else | |
7598 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7599 | ||
7600 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7601 | } | |
7602 | ||
7603 | vec_oprnds0.release (); | |
7604 | vec_oprnds1.release (); | |
7605 | ||
7606 | return true; | |
7607 | } | |
ebfd146a | 7608 | |
8644a673 | 7609 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
7610 | |
7611 | bool | |
355fe088 | 7612 | vect_analyze_stmt (gimple *stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 7613 | { |
8644a673 | 7614 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 7615 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 7616 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 7617 | bool ok; |
a70d6342 | 7618 | tree scalar_type, vectype; |
355fe088 | 7619 | gimple *pattern_stmt; |
363477c0 | 7620 | gimple_seq pattern_def_seq; |
ebfd146a | 7621 | |
73fbfcad | 7622 | if (dump_enabled_p ()) |
ebfd146a | 7623 | { |
78c60e3d SS |
7624 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
7625 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
8644a673 | 7626 | } |
ebfd146a | 7627 | |
1825a1f3 | 7628 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 7629 | { |
73fbfcad | 7630 | if (dump_enabled_p ()) |
78c60e3d | 7631 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7632 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
7633 | |
7634 | return false; | |
7635 | } | |
b8698a0f L |
7636 | |
7637 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
7638 | to include: |
7639 | - the COND_EXPR which is the loop exit condition | |
7640 | - any LABEL_EXPRs in the loop | |
b8698a0f | 7641 | - computations that are used only for array indexing or loop control. |
8644a673 | 7642 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 7643 | instance, therefore, all the statements are relevant. |
ebfd146a | 7644 | |
d092494c | 7645 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 7646 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
7647 | statements. In basic blocks we are called from some SLP instance |
7648 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
7649 | already will be part of SLP instance. */ | |
83197f37 IR |
7650 | |
7651 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 7652 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 7653 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 7654 | { |
9d5e7640 | 7655 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 7656 | && pattern_stmt |
9d5e7640 IR |
7657 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
7658 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
7659 | { | |
83197f37 | 7660 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
7661 | stmt = pattern_stmt; |
7662 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 7663 | if (dump_enabled_p ()) |
9d5e7640 | 7664 | { |
78c60e3d SS |
7665 | dump_printf_loc (MSG_NOTE, vect_location, |
7666 | "==> examining pattern statement: "); | |
7667 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
9d5e7640 IR |
7668 | } |
7669 | } | |
7670 | else | |
7671 | { | |
73fbfcad | 7672 | if (dump_enabled_p ()) |
e645e942 | 7673 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 7674 | |
9d5e7640 IR |
7675 | return true; |
7676 | } | |
8644a673 | 7677 | } |
83197f37 | 7678 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 7679 | && node == NULL |
83197f37 IR |
7680 | && pattern_stmt |
7681 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
7682 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
7683 | { | |
7684 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 7685 | if (dump_enabled_p ()) |
83197f37 | 7686 | { |
78c60e3d SS |
7687 | dump_printf_loc (MSG_NOTE, vect_location, |
7688 | "==> examining pattern statement: "); | |
7689 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
83197f37 IR |
7690 | } |
7691 | ||
7692 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node)) | |
7693 | return false; | |
7694 | } | |
ebfd146a | 7695 | |
1107f3ae | 7696 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 7697 | && node == NULL |
363477c0 | 7698 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 7699 | { |
363477c0 | 7700 | gimple_stmt_iterator si; |
1107f3ae | 7701 | |
363477c0 JJ |
7702 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
7703 | { | |
355fe088 | 7704 | gimple *pattern_def_stmt = gsi_stmt (si); |
363477c0 JJ |
7705 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) |
7706 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
7707 | { | |
7708 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 7709 | if (dump_enabled_p ()) |
363477c0 | 7710 | { |
78c60e3d SS |
7711 | dump_printf_loc (MSG_NOTE, vect_location, |
7712 | "==> examining pattern def statement: "); | |
7713 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
363477c0 | 7714 | } |
1107f3ae | 7715 | |
363477c0 JJ |
7716 | if (!vect_analyze_stmt (pattern_def_stmt, |
7717 | need_to_vectorize, node)) | |
7718 | return false; | |
7719 | } | |
7720 | } | |
7721 | } | |
1107f3ae | 7722 | |
8644a673 IR |
7723 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
7724 | { | |
7725 | case vect_internal_def: | |
7726 | break; | |
ebfd146a | 7727 | |
8644a673 | 7728 | case vect_reduction_def: |
7c5222ff | 7729 | case vect_nested_cycle: |
14a61437 RB |
7730 | gcc_assert (!bb_vinfo |
7731 | && (relevance == vect_used_in_outer | |
7732 | || relevance == vect_used_in_outer_by_reduction | |
7733 | || relevance == vect_used_by_reduction | |
7734 | || relevance == vect_unused_in_scope)); | |
8644a673 IR |
7735 | break; |
7736 | ||
7737 | case vect_induction_def: | |
7738 | case vect_constant_def: | |
7739 | case vect_external_def: | |
7740 | case vect_unknown_def_type: | |
7741 | default: | |
7742 | gcc_unreachable (); | |
7743 | } | |
ebfd146a | 7744 | |
a70d6342 IR |
7745 | if (bb_vinfo) |
7746 | { | |
7747 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
7748 | ||
b690cc0f | 7749 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
73fbfcad | 7750 | if (dump_enabled_p ()) |
a70d6342 | 7751 | { |
78c60e3d SS |
7752 | dump_printf_loc (MSG_NOTE, vect_location, |
7753 | "get vectype for scalar type: "); | |
7754 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
e645e942 | 7755 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
7756 | } |
7757 | ||
7758 | vectype = get_vectype_for_scalar_type (scalar_type); | |
7759 | if (!vectype) | |
7760 | { | |
73fbfcad | 7761 | if (dump_enabled_p ()) |
a70d6342 | 7762 | { |
78c60e3d SS |
7763 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
7764 | "not SLPed: unsupported data-type "); | |
7765 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
7766 | scalar_type); | |
e645e942 | 7767 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
a70d6342 IR |
7768 | } |
7769 | return false; | |
7770 | } | |
7771 | ||
73fbfcad | 7772 | if (dump_enabled_p ()) |
a70d6342 | 7773 | { |
78c60e3d SS |
7774 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); |
7775 | dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); | |
e645e942 | 7776 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
7777 | } |
7778 | ||
7779 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
7780 | } | |
7781 | ||
8644a673 | 7782 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 7783 | { |
8644a673 | 7784 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
0136f8f0 AH |
7785 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
7786 | || (is_gimple_call (stmt) | |
7787 | && gimple_call_lhs (stmt) == NULL_TREE)); | |
8644a673 | 7788 | *need_to_vectorize = true; |
ebfd146a IR |
7789 | } |
7790 | ||
b1af7da6 RB |
7791 | if (PURE_SLP_STMT (stmt_info) && !node) |
7792 | { | |
7793 | dump_printf_loc (MSG_NOTE, vect_location, | |
7794 | "handled only by SLP analysis\n"); | |
7795 | return true; | |
7796 | } | |
7797 | ||
7798 | ok = true; | |
7799 | if (!bb_vinfo | |
7800 | && (STMT_VINFO_RELEVANT_P (stmt_info) | |
7801 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
7802 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
7803 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
7804 | || vectorizable_shift (stmt, NULL, NULL, node) | |
7805 | || vectorizable_operation (stmt, NULL, NULL, node) | |
7806 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
7807 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
7808 | || vectorizable_call (stmt, NULL, NULL, node) | |
7809 | || vectorizable_store (stmt, NULL, NULL, node) | |
7810 | || vectorizable_reduction (stmt, NULL, NULL, node) | |
42fd8198 IE |
7811 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
7812 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 RB |
7813 | else |
7814 | { | |
7815 | if (bb_vinfo) | |
7816 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
7817 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
7818 | || vectorizable_shift (stmt, NULL, NULL, node) | |
7819 | || vectorizable_operation (stmt, NULL, NULL, node) | |
7820 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
7821 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
7822 | || vectorizable_call (stmt, NULL, NULL, node) | |
7823 | || vectorizable_store (stmt, NULL, NULL, node) | |
42fd8198 IE |
7824 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
7825 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 | 7826 | } |
8644a673 IR |
7827 | |
7828 | if (!ok) | |
ebfd146a | 7829 | { |
73fbfcad | 7830 | if (dump_enabled_p ()) |
8644a673 | 7831 | { |
78c60e3d SS |
7832 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
7833 | "not vectorized: relevant stmt not "); | |
7834 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
7835 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 7836 | } |
b8698a0f | 7837 | |
ebfd146a IR |
7838 | return false; |
7839 | } | |
7840 | ||
a70d6342 IR |
7841 | if (bb_vinfo) |
7842 | return true; | |
7843 | ||
8644a673 IR |
7844 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
7845 | need extra handling, except for vectorizable reductions. */ | |
7846 | if (STMT_VINFO_LIVE_P (stmt_info) | |
7847 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
7848 | ok = vectorizable_live_operation (stmt, NULL, NULL); | |
ebfd146a | 7849 | |
8644a673 | 7850 | if (!ok) |
ebfd146a | 7851 | { |
73fbfcad | 7852 | if (dump_enabled_p ()) |
8644a673 | 7853 | { |
78c60e3d SS |
7854 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
7855 | "not vectorized: live stmt not "); | |
7856 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
7857 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 7858 | } |
b8698a0f | 7859 | |
8644a673 | 7860 | return false; |
ebfd146a IR |
7861 | } |
7862 | ||
ebfd146a IR |
7863 | return true; |
7864 | } | |
7865 | ||
7866 | ||
7867 | /* Function vect_transform_stmt. | |
7868 | ||
7869 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
7870 | ||
7871 | bool | |
355fe088 | 7872 | vect_transform_stmt (gimple *stmt, gimple_stmt_iterator *gsi, |
0d0293ac | 7873 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
7874 | slp_instance slp_node_instance) |
7875 | { | |
7876 | bool is_store = false; | |
355fe088 | 7877 | gimple *vec_stmt = NULL; |
ebfd146a | 7878 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
ebfd146a | 7879 | bool done; |
ebfd146a | 7880 | |
355fe088 | 7881 | gimple *old_vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
225ce44b | 7882 | |
ebfd146a IR |
7883 | switch (STMT_VINFO_TYPE (stmt_info)) |
7884 | { | |
7885 | case type_demotion_vec_info_type: | |
ebfd146a | 7886 | case type_promotion_vec_info_type: |
ebfd146a IR |
7887 | case type_conversion_vec_info_type: |
7888 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
7889 | gcc_assert (done); | |
7890 | break; | |
7891 | ||
7892 | case induc_vec_info_type: | |
7893 | gcc_assert (!slp_node); | |
7894 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
7895 | gcc_assert (done); | |
7896 | break; | |
7897 | ||
9dc3f7de IR |
7898 | case shift_vec_info_type: |
7899 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
7900 | gcc_assert (done); | |
7901 | break; | |
7902 | ||
ebfd146a IR |
7903 | case op_vec_info_type: |
7904 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
7905 | gcc_assert (done); | |
7906 | break; | |
7907 | ||
7908 | case assignment_vec_info_type: | |
7909 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
7910 | gcc_assert (done); | |
7911 | break; | |
7912 | ||
7913 | case load_vec_info_type: | |
b8698a0f | 7914 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
7915 | slp_node_instance); |
7916 | gcc_assert (done); | |
7917 | break; | |
7918 | ||
7919 | case store_vec_info_type: | |
7920 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
7921 | gcc_assert (done); | |
0d0293ac | 7922 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
7923 | { |
7924 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 7925 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
7926 | one are skipped, and there vec_stmt_info shouldn't be freed |
7927 | meanwhile. */ | |
0d0293ac | 7928 | *grouped_store = true; |
ebfd146a IR |
7929 | if (STMT_VINFO_VEC_STMT (stmt_info)) |
7930 | is_store = true; | |
7931 | } | |
7932 | else | |
7933 | is_store = true; | |
7934 | break; | |
7935 | ||
7936 | case condition_vec_info_type: | |
f7e531cf | 7937 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
7938 | gcc_assert (done); |
7939 | break; | |
7940 | ||
42fd8198 IE |
7941 | case comparison_vec_info_type: |
7942 | done = vectorizable_comparison (stmt, gsi, &vec_stmt, NULL, slp_node); | |
7943 | gcc_assert (done); | |
7944 | break; | |
7945 | ||
ebfd146a | 7946 | case call_vec_info_type: |
190c2236 | 7947 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 7948 | stmt = gsi_stmt (*gsi); |
5ce9450f JJ |
7949 | if (is_gimple_call (stmt) |
7950 | && gimple_call_internal_p (stmt) | |
7951 | && gimple_call_internal_fn (stmt) == IFN_MASK_STORE) | |
7952 | is_store = true; | |
ebfd146a IR |
7953 | break; |
7954 | ||
0136f8f0 AH |
7955 | case call_simd_clone_vec_info_type: |
7956 | done = vectorizable_simd_clone_call (stmt, gsi, &vec_stmt, slp_node); | |
7957 | stmt = gsi_stmt (*gsi); | |
7958 | break; | |
7959 | ||
ebfd146a | 7960 | case reduc_vec_info_type: |
b5aeb3bb | 7961 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
7962 | gcc_assert (done); |
7963 | break; | |
7964 | ||
7965 | default: | |
7966 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
7967 | { | |
73fbfcad | 7968 | if (dump_enabled_p ()) |
78c60e3d | 7969 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7970 | "stmt not supported.\n"); |
ebfd146a IR |
7971 | gcc_unreachable (); |
7972 | } | |
7973 | } | |
7974 | ||
225ce44b RB |
7975 | /* Verify SLP vectorization doesn't mess with STMT_VINFO_VEC_STMT. |
7976 | This would break hybrid SLP vectorization. */ | |
7977 | if (slp_node) | |
d90f8440 RB |
7978 | gcc_assert (!vec_stmt |
7979 | && STMT_VINFO_VEC_STMT (stmt_info) == old_vec_stmt); | |
225ce44b | 7980 | |
ebfd146a IR |
7981 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that |
7982 | is being vectorized, but outside the immediately enclosing loop. */ | |
7983 | if (vec_stmt | |
a70d6342 IR |
7984 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
7985 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
7986 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
7987 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
7988 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 7989 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 7990 | vect_used_in_outer_by_reduction)) |
ebfd146a | 7991 | { |
a70d6342 IR |
7992 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
7993 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
7994 | imm_use_iterator imm_iter; |
7995 | use_operand_p use_p; | |
7996 | tree scalar_dest; | |
355fe088 | 7997 | gimple *exit_phi; |
ebfd146a | 7998 | |
73fbfcad | 7999 | if (dump_enabled_p ()) |
78c60e3d | 8000 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 8001 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
8002 | |
8003 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
8004 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
8005 | STMT). */ | |
8006 | if (gimple_code (stmt) == GIMPLE_PHI) | |
8007 | scalar_dest = PHI_RESULT (stmt); | |
8008 | else | |
8009 | scalar_dest = gimple_assign_lhs (stmt); | |
8010 | ||
8011 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
8012 | { | |
8013 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
8014 | { | |
8015 | exit_phi = USE_STMT (use_p); | |
8016 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
8017 | } | |
8018 | } | |
8019 | } | |
8020 | ||
8021 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
8022 | being vectorized. */ | |
8023 | if (STMT_VINFO_LIVE_P (stmt_info) | |
8024 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
8025 | { | |
8026 | done = vectorizable_live_operation (stmt, gsi, &vec_stmt); | |
8027 | gcc_assert (done); | |
8028 | } | |
8029 | ||
8030 | if (vec_stmt) | |
83197f37 | 8031 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 8032 | |
b8698a0f | 8033 | return is_store; |
ebfd146a IR |
8034 | } |
8035 | ||
8036 | ||
b8698a0f | 8037 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
8038 | stmt_vec_info. */ |
8039 | ||
8040 | void | |
355fe088 | 8041 | vect_remove_stores (gimple *first_stmt) |
ebfd146a | 8042 | { |
355fe088 TS |
8043 | gimple *next = first_stmt; |
8044 | gimple *tmp; | |
ebfd146a IR |
8045 | gimple_stmt_iterator next_si; |
8046 | ||
8047 | while (next) | |
8048 | { | |
78048b1c JJ |
8049 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
8050 | ||
8051 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
8052 | if (is_pattern_stmt_p (stmt_info)) | |
8053 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
8054 | /* Free the attached stmt_vec_info and remove the stmt. */ |
8055 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 8056 | unlink_stmt_vdef (next); |
ebfd146a | 8057 | gsi_remove (&next_si, true); |
3d3f2249 | 8058 | release_defs (next); |
ebfd146a IR |
8059 | free_stmt_vec_info (next); |
8060 | next = tmp; | |
8061 | } | |
8062 | } | |
8063 | ||
8064 | ||
8065 | /* Function new_stmt_vec_info. | |
8066 | ||
8067 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
8068 | ||
8069 | stmt_vec_info | |
310213d4 | 8070 | new_stmt_vec_info (gimple *stmt, vec_info *vinfo) |
ebfd146a IR |
8071 | { |
8072 | stmt_vec_info res; | |
8073 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
8074 | ||
8075 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
8076 | STMT_VINFO_STMT (res) = stmt; | |
310213d4 | 8077 | res->vinfo = vinfo; |
8644a673 | 8078 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
8079 | STMT_VINFO_LIVE_P (res) = false; |
8080 | STMT_VINFO_VECTYPE (res) = NULL; | |
8081 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 8082 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
8083 | STMT_VINFO_IN_PATTERN_P (res) = false; |
8084 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 8085 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a | 8086 | STMT_VINFO_DATA_REF (res) = NULL; |
af29617a | 8087 | STMT_VINFO_VEC_REDUCTION_TYPE (res) = TREE_CODE_REDUCTION; |
ebfd146a IR |
8088 | |
8089 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
8090 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
8091 | STMT_VINFO_DR_INIT (res) = NULL; | |
8092 | STMT_VINFO_DR_STEP (res) = NULL; | |
8093 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
8094 | ||
8095 | if (gimple_code (stmt) == GIMPLE_PHI | |
8096 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
8097 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
8098 | else | |
8644a673 IR |
8099 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
8100 | ||
9771b263 | 8101 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 8102 | STMT_SLP_TYPE (res) = loop_vect; |
e14c1050 IR |
8103 | GROUP_FIRST_ELEMENT (res) = NULL; |
8104 | GROUP_NEXT_ELEMENT (res) = NULL; | |
8105 | GROUP_SIZE (res) = 0; | |
8106 | GROUP_STORE_COUNT (res) = 0; | |
8107 | GROUP_GAP (res) = 0; | |
8108 | GROUP_SAME_DR_STMT (res) = NULL; | |
ebfd146a IR |
8109 | |
8110 | return res; | |
8111 | } | |
8112 | ||
8113 | ||
8114 | /* Create a hash table for stmt_vec_info. */ | |
8115 | ||
8116 | void | |
8117 | init_stmt_vec_info_vec (void) | |
8118 | { | |
9771b263 DN |
8119 | gcc_assert (!stmt_vec_info_vec.exists ()); |
8120 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
8121 | } |
8122 | ||
8123 | ||
8124 | /* Free hash table for stmt_vec_info. */ | |
8125 | ||
8126 | void | |
8127 | free_stmt_vec_info_vec (void) | |
8128 | { | |
93675444 | 8129 | unsigned int i; |
3161455c | 8130 | stmt_vec_info info; |
93675444 JJ |
8131 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) |
8132 | if (info != NULL) | |
3161455c | 8133 | free_stmt_vec_info (STMT_VINFO_STMT (info)); |
9771b263 DN |
8134 | gcc_assert (stmt_vec_info_vec.exists ()); |
8135 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
8136 | } |
8137 | ||
8138 | ||
8139 | /* Free stmt vectorization related info. */ | |
8140 | ||
8141 | void | |
355fe088 | 8142 | free_stmt_vec_info (gimple *stmt) |
ebfd146a IR |
8143 | { |
8144 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8145 | ||
8146 | if (!stmt_info) | |
8147 | return; | |
8148 | ||
78048b1c JJ |
8149 | /* Check if this statement has a related "pattern stmt" |
8150 | (introduced by the vectorizer during the pattern recognition | |
8151 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
8152 | too. */ | |
8153 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
8154 | { | |
8155 | stmt_vec_info patt_info | |
8156 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
8157 | if (patt_info) | |
8158 | { | |
363477c0 | 8159 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
355fe088 | 8160 | gimple *patt_stmt = STMT_VINFO_STMT (patt_info); |
f0281fde RB |
8161 | gimple_set_bb (patt_stmt, NULL); |
8162 | tree lhs = gimple_get_lhs (patt_stmt); | |
8163 | if (TREE_CODE (lhs) == SSA_NAME) | |
8164 | release_ssa_name (lhs); | |
363477c0 JJ |
8165 | if (seq) |
8166 | { | |
8167 | gimple_stmt_iterator si; | |
8168 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
f0281fde | 8169 | { |
355fe088 | 8170 | gimple *seq_stmt = gsi_stmt (si); |
f0281fde | 8171 | gimple_set_bb (seq_stmt, NULL); |
7532abf2 | 8172 | lhs = gimple_get_lhs (seq_stmt); |
f0281fde RB |
8173 | if (TREE_CODE (lhs) == SSA_NAME) |
8174 | release_ssa_name (lhs); | |
8175 | free_stmt_vec_info (seq_stmt); | |
8176 | } | |
363477c0 | 8177 | } |
f0281fde | 8178 | free_stmt_vec_info (patt_stmt); |
78048b1c JJ |
8179 | } |
8180 | } | |
8181 | ||
9771b263 | 8182 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
6c9e85fb | 8183 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).release (); |
ebfd146a IR |
8184 | set_vinfo_for_stmt (stmt, NULL); |
8185 | free (stmt_info); | |
8186 | } | |
8187 | ||
8188 | ||
bb67d9c7 | 8189 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 8190 | |
bb67d9c7 | 8191 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
8192 | by the target. */ |
8193 | ||
bb67d9c7 RG |
8194 | static tree |
8195 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a | 8196 | { |
ef4bddc2 RS |
8197 | machine_mode inner_mode = TYPE_MODE (scalar_type); |
8198 | machine_mode simd_mode; | |
2f816591 | 8199 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
8200 | int nunits; |
8201 | tree vectype; | |
8202 | ||
cc4b5170 | 8203 | if (nbytes == 0) |
ebfd146a IR |
8204 | return NULL_TREE; |
8205 | ||
48f2e373 RB |
8206 | if (GET_MODE_CLASS (inner_mode) != MODE_INT |
8207 | && GET_MODE_CLASS (inner_mode) != MODE_FLOAT) | |
8208 | return NULL_TREE; | |
8209 | ||
7b7b1813 RG |
8210 | /* For vector types of elements whose mode precision doesn't |
8211 | match their types precision we use a element type of mode | |
8212 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
8213 | they support the proper result truncation/extension. |
8214 | We also make sure to build vector types with INTEGER_TYPE | |
8215 | component type only. */ | |
6d7971b8 | 8216 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
8217 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
8218 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
8219 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
8220 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 8221 | |
ccbf5bb4 RG |
8222 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
8223 | When the component mode passes the above test simply use a type | |
8224 | corresponding to that mode. The theory is that any use that | |
8225 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 8226 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 8227 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
8228 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
8229 | ||
8230 | /* We can't build a vector type of elements with alignment bigger than | |
8231 | their size. */ | |
dfc2e2ac | 8232 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
8233 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
8234 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 8235 | |
dfc2e2ac RB |
8236 | /* If we felt back to using the mode fail if there was |
8237 | no scalar type for it. */ | |
8238 | if (scalar_type == NULL_TREE) | |
8239 | return NULL_TREE; | |
8240 | ||
bb67d9c7 RG |
8241 | /* If no size was supplied use the mode the target prefers. Otherwise |
8242 | lookup a vector mode of the specified size. */ | |
8243 | if (size == 0) | |
8244 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
8245 | else | |
8246 | simd_mode = mode_for_vector (inner_mode, size / nbytes); | |
cc4b5170 RG |
8247 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
8248 | if (nunits <= 1) | |
8249 | return NULL_TREE; | |
ebfd146a IR |
8250 | |
8251 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
8252 | |
8253 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
8254 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 8255 | return NULL_TREE; |
ebfd146a IR |
8256 | |
8257 | return vectype; | |
8258 | } | |
8259 | ||
bb67d9c7 RG |
8260 | unsigned int current_vector_size; |
8261 | ||
8262 | /* Function get_vectype_for_scalar_type. | |
8263 | ||
8264 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
8265 | by the target. */ | |
8266 | ||
8267 | tree | |
8268 | get_vectype_for_scalar_type (tree scalar_type) | |
8269 | { | |
8270 | tree vectype; | |
8271 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
8272 | current_vector_size); | |
8273 | if (vectype | |
8274 | && current_vector_size == 0) | |
8275 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
8276 | return vectype; | |
8277 | } | |
8278 | ||
42fd8198 IE |
8279 | /* Function get_mask_type_for_scalar_type. |
8280 | ||
8281 | Returns the mask type corresponding to a result of comparison | |
8282 | of vectors of specified SCALAR_TYPE as supported by target. */ | |
8283 | ||
8284 | tree | |
8285 | get_mask_type_for_scalar_type (tree scalar_type) | |
8286 | { | |
8287 | tree vectype = get_vectype_for_scalar_type (scalar_type); | |
8288 | ||
8289 | if (!vectype) | |
8290 | return NULL; | |
8291 | ||
8292 | return build_truth_vector_type (TYPE_VECTOR_SUBPARTS (vectype), | |
8293 | current_vector_size); | |
8294 | } | |
8295 | ||
b690cc0f RG |
8296 | /* Function get_same_sized_vectype |
8297 | ||
8298 | Returns a vector type corresponding to SCALAR_TYPE of size | |
8299 | VECTOR_TYPE if supported by the target. */ | |
8300 | ||
8301 | tree | |
bb67d9c7 | 8302 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 8303 | { |
9f47c7e5 IE |
8304 | if (TREE_CODE (scalar_type) == BOOLEAN_TYPE) |
8305 | return build_same_sized_truth_vector_type (vector_type); | |
8306 | ||
bb67d9c7 RG |
8307 | return get_vectype_for_scalar_type_and_size |
8308 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
8309 | } |
8310 | ||
ebfd146a IR |
8311 | /* Function vect_is_simple_use. |
8312 | ||
8313 | Input: | |
81c40241 RB |
8314 | VINFO - the vect info of the loop or basic block that is being vectorized. |
8315 | OPERAND - operand in the loop or bb. | |
8316 | Output: | |
8317 | DEF_STMT - the defining stmt in case OPERAND is an SSA_NAME. | |
8318 | DT - the type of definition | |
ebfd146a IR |
8319 | |
8320 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 8321 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 8322 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 8323 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
8324 | is the case in reduction/induction computations). |
8325 | For basic blocks, supportable operands are constants and bb invariants. | |
8326 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
8327 | |
8328 | bool | |
81c40241 RB |
8329 | vect_is_simple_use (tree operand, vec_info *vinfo, |
8330 | gimple **def_stmt, enum vect_def_type *dt) | |
b8698a0f | 8331 | { |
ebfd146a | 8332 | *def_stmt = NULL; |
3fc356dc | 8333 | *dt = vect_unknown_def_type; |
b8698a0f | 8334 | |
73fbfcad | 8335 | if (dump_enabled_p ()) |
ebfd146a | 8336 | { |
78c60e3d SS |
8337 | dump_printf_loc (MSG_NOTE, vect_location, |
8338 | "vect_is_simple_use: operand "); | |
8339 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
e645e942 | 8340 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 8341 | } |
b8698a0f | 8342 | |
b758f602 | 8343 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
8344 | { |
8345 | *dt = vect_constant_def; | |
8346 | return true; | |
8347 | } | |
b8698a0f | 8348 | |
ebfd146a IR |
8349 | if (is_gimple_min_invariant (operand)) |
8350 | { | |
8644a673 | 8351 | *dt = vect_external_def; |
ebfd146a IR |
8352 | return true; |
8353 | } | |
8354 | ||
ebfd146a IR |
8355 | if (TREE_CODE (operand) != SSA_NAME) |
8356 | { | |
73fbfcad | 8357 | if (dump_enabled_p ()) |
af29617a AH |
8358 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8359 | "not ssa-name.\n"); | |
ebfd146a IR |
8360 | return false; |
8361 | } | |
b8698a0f | 8362 | |
3fc356dc | 8363 | if (SSA_NAME_IS_DEFAULT_DEF (operand)) |
ebfd146a | 8364 | { |
3fc356dc RB |
8365 | *dt = vect_external_def; |
8366 | return true; | |
ebfd146a IR |
8367 | } |
8368 | ||
3fc356dc | 8369 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
73fbfcad | 8370 | if (dump_enabled_p ()) |
ebfd146a | 8371 | { |
78c60e3d SS |
8372 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
8373 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
ebfd146a IR |
8374 | } |
8375 | ||
61d371eb | 8376 | if (! vect_stmt_in_region_p (vinfo, *def_stmt)) |
8644a673 | 8377 | *dt = vect_external_def; |
ebfd146a IR |
8378 | else |
8379 | { | |
3fc356dc | 8380 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (*def_stmt); |
310213d4 | 8381 | if (is_a <bb_vec_info> (vinfo) && !STMT_VINFO_VECTORIZABLE (stmt_vinfo)) |
90dd6e3d RB |
8382 | *dt = vect_external_def; |
8383 | else | |
8384 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); | |
ebfd146a IR |
8385 | } |
8386 | ||
2e8ab70c RB |
8387 | if (dump_enabled_p ()) |
8388 | { | |
8389 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: "); | |
8390 | switch (*dt) | |
8391 | { | |
8392 | case vect_uninitialized_def: | |
8393 | dump_printf (MSG_NOTE, "uninitialized\n"); | |
8394 | break; | |
8395 | case vect_constant_def: | |
8396 | dump_printf (MSG_NOTE, "constant\n"); | |
8397 | break; | |
8398 | case vect_external_def: | |
8399 | dump_printf (MSG_NOTE, "external\n"); | |
8400 | break; | |
8401 | case vect_internal_def: | |
8402 | dump_printf (MSG_NOTE, "internal\n"); | |
8403 | break; | |
8404 | case vect_induction_def: | |
8405 | dump_printf (MSG_NOTE, "induction\n"); | |
8406 | break; | |
8407 | case vect_reduction_def: | |
8408 | dump_printf (MSG_NOTE, "reduction\n"); | |
8409 | break; | |
8410 | case vect_double_reduction_def: | |
8411 | dump_printf (MSG_NOTE, "double reduction\n"); | |
8412 | break; | |
8413 | case vect_nested_cycle: | |
8414 | dump_printf (MSG_NOTE, "nested cycle\n"); | |
8415 | break; | |
8416 | case vect_unknown_def_type: | |
8417 | dump_printf (MSG_NOTE, "unknown\n"); | |
8418 | break; | |
8419 | } | |
8420 | } | |
8421 | ||
81c40241 | 8422 | if (*dt == vect_unknown_def_type) |
ebfd146a | 8423 | { |
73fbfcad | 8424 | if (dump_enabled_p ()) |
78c60e3d | 8425 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8426 | "Unsupported pattern.\n"); |
ebfd146a IR |
8427 | return false; |
8428 | } | |
8429 | ||
ebfd146a IR |
8430 | switch (gimple_code (*def_stmt)) |
8431 | { | |
8432 | case GIMPLE_PHI: | |
ebfd146a | 8433 | case GIMPLE_ASSIGN: |
ebfd146a | 8434 | case GIMPLE_CALL: |
81c40241 | 8435 | break; |
ebfd146a | 8436 | default: |
73fbfcad | 8437 | if (dump_enabled_p ()) |
78c60e3d | 8438 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8439 | "unsupported defining stmt:\n"); |
ebfd146a IR |
8440 | return false; |
8441 | } | |
8442 | ||
8443 | return true; | |
8444 | } | |
8445 | ||
81c40241 | 8446 | /* Function vect_is_simple_use. |
b690cc0f | 8447 | |
81c40241 | 8448 | Same as vect_is_simple_use but also determines the vector operand |
b690cc0f RG |
8449 | type of OPERAND and stores it to *VECTYPE. If the definition of |
8450 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
8451 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
8452 | is responsible to compute the best suited vector type for the | |
8453 | scalar operand. */ | |
8454 | ||
8455 | bool | |
81c40241 RB |
8456 | vect_is_simple_use (tree operand, vec_info *vinfo, |
8457 | gimple **def_stmt, enum vect_def_type *dt, tree *vectype) | |
b690cc0f | 8458 | { |
81c40241 | 8459 | if (!vect_is_simple_use (operand, vinfo, def_stmt, dt)) |
b690cc0f RG |
8460 | return false; |
8461 | ||
8462 | /* Now get a vector type if the def is internal, otherwise supply | |
8463 | NULL_TREE and leave it up to the caller to figure out a proper | |
8464 | type for the use stmt. */ | |
8465 | if (*dt == vect_internal_def | |
8466 | || *dt == vect_induction_def | |
8467 | || *dt == vect_reduction_def | |
8468 | || *dt == vect_double_reduction_def | |
8469 | || *dt == vect_nested_cycle) | |
8470 | { | |
8471 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
8472 | |
8473 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
8474 | && !STMT_VINFO_RELEVANT (stmt_info) | |
8475 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 8476 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 8477 | |
b690cc0f RG |
8478 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
8479 | gcc_assert (*vectype != NULL_TREE); | |
8480 | } | |
8481 | else if (*dt == vect_uninitialized_def | |
8482 | || *dt == vect_constant_def | |
8483 | || *dt == vect_external_def) | |
8484 | *vectype = NULL_TREE; | |
8485 | else | |
8486 | gcc_unreachable (); | |
8487 | ||
8488 | return true; | |
8489 | } | |
8490 | ||
ebfd146a IR |
8491 | |
8492 | /* Function supportable_widening_operation | |
8493 | ||
b8698a0f L |
8494 | Check whether an operation represented by the code CODE is a |
8495 | widening operation that is supported by the target platform in | |
b690cc0f RG |
8496 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
8497 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 8498 | |
ebfd146a IR |
8499 | Widening operations we currently support are NOP (CONVERT), FLOAT |
8500 | and WIDEN_MULT. This function checks if these operations are supported | |
8501 | by the target platform either directly (via vector tree-codes), or via | |
8502 | target builtins. | |
8503 | ||
8504 | Output: | |
b8698a0f L |
8505 | - CODE1 and CODE2 are codes of vector operations to be used when |
8506 | vectorizing the operation, if available. | |
ebfd146a IR |
8507 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
8508 | case of multi-step conversion (like char->short->int - in that case | |
8509 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
8510 | - INTERM_TYPES contains the intermediate type required to perform the |
8511 | widening operation (short in the above example). */ | |
ebfd146a IR |
8512 | |
8513 | bool | |
355fe088 | 8514 | supportable_widening_operation (enum tree_code code, gimple *stmt, |
b690cc0f | 8515 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
8516 | enum tree_code *code1, enum tree_code *code2, |
8517 | int *multi_step_cvt, | |
9771b263 | 8518 | vec<tree> *interm_types) |
ebfd146a IR |
8519 | { |
8520 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8521 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 8522 | struct loop *vect_loop = NULL; |
ef4bddc2 | 8523 | machine_mode vec_mode; |
81f40b79 | 8524 | enum insn_code icode1, icode2; |
ebfd146a | 8525 | optab optab1, optab2; |
b690cc0f RG |
8526 | tree vectype = vectype_in; |
8527 | tree wide_vectype = vectype_out; | |
ebfd146a | 8528 | enum tree_code c1, c2; |
4a00c761 JJ |
8529 | int i; |
8530 | tree prev_type, intermediate_type; | |
ef4bddc2 | 8531 | machine_mode intermediate_mode, prev_mode; |
4a00c761 | 8532 | optab optab3, optab4; |
ebfd146a | 8533 | |
4a00c761 | 8534 | *multi_step_cvt = 0; |
4ef69dfc IR |
8535 | if (loop_info) |
8536 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
8537 | ||
ebfd146a IR |
8538 | switch (code) |
8539 | { | |
8540 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
8541 | /* The result of a vectorized widening operation usually requires |
8542 | two vectors (because the widened results do not fit into one vector). | |
8543 | The generated vector results would normally be expected to be | |
8544 | generated in the same order as in the original scalar computation, | |
8545 | i.e. if 8 results are generated in each vector iteration, they are | |
8546 | to be organized as follows: | |
8547 | vect1: [res1,res2,res3,res4], | |
8548 | vect2: [res5,res6,res7,res8]. | |
8549 | ||
8550 | However, in the special case that the result of the widening | |
8551 | operation is used in a reduction computation only, the order doesn't | |
8552 | matter (because when vectorizing a reduction we change the order of | |
8553 | the computation). Some targets can take advantage of this and | |
8554 | generate more efficient code. For example, targets like Altivec, | |
8555 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
8556 | generate the following vectors: | |
8557 | vect1: [res1,res3,res5,res7], | |
8558 | vect2: [res2,res4,res6,res8]. | |
8559 | ||
8560 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
8561 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
8562 | iterations in parallel). We therefore don't allow to change the | |
8563 | order of the computation in the inner-loop during outer-loop | |
8564 | vectorization. */ | |
8565 | /* TODO: Another case in which order doesn't *really* matter is when we | |
8566 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
8567 | Normally, pack_trunc performs an even/odd permute, whereas the | |
8568 | repack from an even/odd expansion would be an interleave, which | |
8569 | would be significantly simpler for e.g. AVX2. */ | |
8570 | /* In any case, in order to avoid duplicating the code below, recurse | |
8571 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
8572 | are properly set up for the caller. If we fail, we'll continue with | |
8573 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
8574 | if (vect_loop | |
8575 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
8576 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
8577 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
8578 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
8579 | code1, code2, multi_step_cvt, |
8580 | interm_types)) | |
ebc047a2 CH |
8581 | { |
8582 | /* Elements in a vector with vect_used_by_reduction property cannot | |
8583 | be reordered if the use chain with this property does not have the | |
8584 | same operation. One such an example is s += a * b, where elements | |
8585 | in a and b cannot be reordered. Here we check if the vector defined | |
8586 | by STMT is only directly used in the reduction statement. */ | |
8587 | tree lhs = gimple_assign_lhs (stmt); | |
8588 | use_operand_p dummy; | |
355fe088 | 8589 | gimple *use_stmt; |
ebc047a2 CH |
8590 | stmt_vec_info use_stmt_info = NULL; |
8591 | if (single_imm_use (lhs, &dummy, &use_stmt) | |
8592 | && (use_stmt_info = vinfo_for_stmt (use_stmt)) | |
8593 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) | |
8594 | return true; | |
8595 | } | |
4a00c761 JJ |
8596 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
8597 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
8598 | break; |
8599 | ||
81c40241 RB |
8600 | case DOT_PROD_EXPR: |
8601 | c1 = DOT_PROD_EXPR; | |
8602 | c2 = DOT_PROD_EXPR; | |
8603 | break; | |
8604 | ||
8605 | case SAD_EXPR: | |
8606 | c1 = SAD_EXPR; | |
8607 | c2 = SAD_EXPR; | |
8608 | break; | |
8609 | ||
6ae6116f RH |
8610 | case VEC_WIDEN_MULT_EVEN_EXPR: |
8611 | /* Support the recursion induced just above. */ | |
8612 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
8613 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
8614 | break; | |
8615 | ||
36ba4aae | 8616 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
8617 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
8618 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
8619 | break; |
8620 | ||
ebfd146a | 8621 | CASE_CONVERT: |
4a00c761 JJ |
8622 | c1 = VEC_UNPACK_LO_EXPR; |
8623 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
8624 | break; |
8625 | ||
8626 | case FLOAT_EXPR: | |
4a00c761 JJ |
8627 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
8628 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
8629 | break; |
8630 | ||
8631 | case FIX_TRUNC_EXPR: | |
8632 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
8633 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
8634 | computing the operation. */ | |
8635 | return false; | |
8636 | ||
8637 | default: | |
8638 | gcc_unreachable (); | |
8639 | } | |
8640 | ||
6ae6116f | 8641 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
6b4db501 | 8642 | std::swap (c1, c2); |
4a00c761 | 8643 | |
ebfd146a IR |
8644 | if (code == FIX_TRUNC_EXPR) |
8645 | { | |
8646 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
8647 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
8648 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
8649 | } |
8650 | else | |
8651 | { | |
8652 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
8653 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
8654 | } | |
8655 | ||
8656 | if (!optab1 || !optab2) | |
8657 | return false; | |
8658 | ||
8659 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
8660 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
8661 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
8662 | return false; |
8663 | ||
4a00c761 JJ |
8664 | *code1 = c1; |
8665 | *code2 = c2; | |
8666 | ||
8667 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
8668 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
8669 | return true; | |
8670 | ||
b8698a0f | 8671 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 8672 | types. */ |
ebfd146a | 8673 | |
4a00c761 JJ |
8674 | prev_type = vectype; |
8675 | prev_mode = vec_mode; | |
b8698a0f | 8676 | |
4a00c761 JJ |
8677 | if (!CONVERT_EXPR_CODE_P (code)) |
8678 | return false; | |
b8698a0f | 8679 | |
4a00c761 JJ |
8680 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
8681 | intermediate steps in promotion sequence. We try | |
8682 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
8683 | not. */ | |
9771b263 | 8684 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
8685 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
8686 | { | |
8687 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
8688 | intermediate_type | |
8689 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
8690 | TYPE_UNSIGNED (prev_type)); | |
8691 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); | |
8692 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
8693 | ||
8694 | if (!optab3 || !optab4 | |
8695 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
8696 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
8697 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
8698 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
8699 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
8700 | == CODE_FOR_nothing) | |
8701 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
8702 | == CODE_FOR_nothing)) | |
8703 | break; | |
ebfd146a | 8704 | |
9771b263 | 8705 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
8706 | (*multi_step_cvt)++; |
8707 | ||
8708 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
8709 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
8710 | return true; | |
8711 | ||
8712 | prev_type = intermediate_type; | |
8713 | prev_mode = intermediate_mode; | |
ebfd146a IR |
8714 | } |
8715 | ||
9771b263 | 8716 | interm_types->release (); |
4a00c761 | 8717 | return false; |
ebfd146a IR |
8718 | } |
8719 | ||
8720 | ||
8721 | /* Function supportable_narrowing_operation | |
8722 | ||
b8698a0f L |
8723 | Check whether an operation represented by the code CODE is a |
8724 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
8725 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
8726 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 8727 | |
ebfd146a | 8728 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 8729 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
8730 | the target platform directly via vector tree-codes. |
8731 | ||
8732 | Output: | |
b8698a0f L |
8733 | - CODE1 is the code of a vector operation to be used when |
8734 | vectorizing the operation, if available. | |
ebfd146a IR |
8735 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
8736 | case of multi-step conversion (like int->short->char - in that case | |
8737 | MULTI_STEP_CVT will be 1). | |
8738 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 8739 | narrowing operation (short in the above example). */ |
ebfd146a IR |
8740 | |
8741 | bool | |
8742 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 8743 | tree vectype_out, tree vectype_in, |
ebfd146a | 8744 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 8745 | vec<tree> *interm_types) |
ebfd146a | 8746 | { |
ef4bddc2 | 8747 | machine_mode vec_mode; |
ebfd146a IR |
8748 | enum insn_code icode1; |
8749 | optab optab1, interm_optab; | |
b690cc0f RG |
8750 | tree vectype = vectype_in; |
8751 | tree narrow_vectype = vectype_out; | |
ebfd146a | 8752 | enum tree_code c1; |
4a00c761 | 8753 | tree intermediate_type; |
ef4bddc2 | 8754 | machine_mode intermediate_mode, prev_mode; |
ebfd146a | 8755 | int i; |
4a00c761 | 8756 | bool uns; |
ebfd146a | 8757 | |
4a00c761 | 8758 | *multi_step_cvt = 0; |
ebfd146a IR |
8759 | switch (code) |
8760 | { | |
8761 | CASE_CONVERT: | |
8762 | c1 = VEC_PACK_TRUNC_EXPR; | |
8763 | break; | |
8764 | ||
8765 | case FIX_TRUNC_EXPR: | |
8766 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
8767 | break; | |
8768 | ||
8769 | case FLOAT_EXPR: | |
8770 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
8771 | tree code and optabs used for computing the operation. */ | |
8772 | return false; | |
8773 | ||
8774 | default: | |
8775 | gcc_unreachable (); | |
8776 | } | |
8777 | ||
8778 | if (code == FIX_TRUNC_EXPR) | |
8779 | /* The signedness is determined from output operand. */ | |
b690cc0f | 8780 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
8781 | else |
8782 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
8783 | ||
8784 | if (!optab1) | |
8785 | return false; | |
8786 | ||
8787 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 8788 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
8789 | return false; |
8790 | ||
4a00c761 JJ |
8791 | *code1 = c1; |
8792 | ||
8793 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
8794 | return true; | |
8795 | ||
ebfd146a IR |
8796 | /* Check if it's a multi-step conversion that can be done using intermediate |
8797 | types. */ | |
4a00c761 JJ |
8798 | prev_mode = vec_mode; |
8799 | if (code == FIX_TRUNC_EXPR) | |
8800 | uns = TYPE_UNSIGNED (vectype_out); | |
8801 | else | |
8802 | uns = TYPE_UNSIGNED (vectype); | |
8803 | ||
8804 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
8805 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
8806 | costly than signed. */ | |
8807 | if (code == FIX_TRUNC_EXPR && uns) | |
8808 | { | |
8809 | enum insn_code icode2; | |
8810 | ||
8811 | intermediate_type | |
8812 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
8813 | interm_optab | |
8814 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 8815 | if (interm_optab != unknown_optab |
4a00c761 JJ |
8816 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
8817 | && insn_data[icode1].operand[0].mode | |
8818 | == insn_data[icode2].operand[0].mode) | |
8819 | { | |
8820 | uns = false; | |
8821 | optab1 = interm_optab; | |
8822 | icode1 = icode2; | |
8823 | } | |
8824 | } | |
ebfd146a | 8825 | |
4a00c761 JJ |
8826 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
8827 | intermediate steps in promotion sequence. We try | |
8828 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 8829 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
8830 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
8831 | { | |
8832 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
8833 | intermediate_type | |
8834 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
8835 | interm_optab | |
8836 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
8837 | optab_default); | |
8838 | if (!interm_optab | |
8839 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
8840 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
8841 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
8842 | == CODE_FOR_nothing)) | |
8843 | break; | |
8844 | ||
9771b263 | 8845 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
8846 | (*multi_step_cvt)++; |
8847 | ||
8848 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
8849 | return true; | |
8850 | ||
8851 | prev_mode = intermediate_mode; | |
8852 | optab1 = interm_optab; | |
ebfd146a IR |
8853 | } |
8854 | ||
9771b263 | 8855 | interm_types->release (); |
4a00c761 | 8856 | return false; |
ebfd146a | 8857 | } |