]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
cbe34bb5 | 2 | Copyright (C) 2003-2017 Free Software Foundation, Inc. |
b8698a0f | 3 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
4 | and Ira Rosen <irar@il.ibm.com> |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 3, or (at your option) any later | |
11 | version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
c7131fb2 | 25 | #include "backend.h" |
957060b5 AM |
26 | #include "target.h" |
27 | #include "rtl.h" | |
ebfd146a | 28 | #include "tree.h" |
c7131fb2 | 29 | #include "gimple.h" |
c7131fb2 | 30 | #include "ssa.h" |
957060b5 AM |
31 | #include "optabs-tree.h" |
32 | #include "insn-config.h" | |
33 | #include "recog.h" /* FIXME: for insn_data */ | |
34 | #include "cgraph.h" | |
957060b5 | 35 | #include "dumpfile.h" |
c7131fb2 | 36 | #include "alias.h" |
40e23961 | 37 | #include "fold-const.h" |
d8a2d370 | 38 | #include "stor-layout.h" |
2fb9a547 | 39 | #include "tree-eh.h" |
45b0be94 | 40 | #include "gimplify.h" |
5be5c238 | 41 | #include "gimple-iterator.h" |
18f429e2 | 42 | #include "gimplify-me.h" |
442b4905 | 43 | #include "tree-cfg.h" |
e28030cf | 44 | #include "tree-ssa-loop-manip.h" |
ebfd146a | 45 | #include "cfgloop.h" |
0136f8f0 AH |
46 | #include "tree-ssa-loop.h" |
47 | #include "tree-scalar-evolution.h" | |
ebfd146a | 48 | #include "tree-vectorizer.h" |
9b2b7279 | 49 | #include "builtins.h" |
70439f0d | 50 | #include "internal-fn.h" |
ebfd146a | 51 | |
7ee2468b SB |
52 | /* For lang_hooks.types.type_for_mode. */ |
53 | #include "langhooks.h" | |
ebfd146a | 54 | |
2de001ee RS |
55 | /* Says whether a statement is a load, a store of a vectorized statement |
56 | result, or a store of an invariant value. */ | |
57 | enum vec_load_store_type { | |
58 | VLS_LOAD, | |
59 | VLS_STORE, | |
60 | VLS_STORE_INVARIANT | |
61 | }; | |
62 | ||
c3e7ee41 BS |
63 | /* Return the vectorized type for the given statement. */ |
64 | ||
65 | tree | |
66 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
67 | { | |
68 | return STMT_VINFO_VECTYPE (stmt_info); | |
69 | } | |
70 | ||
71 | /* Return TRUE iff the given statement is in an inner loop relative to | |
72 | the loop being vectorized. */ | |
73 | bool | |
74 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
75 | { | |
355fe088 | 76 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
c3e7ee41 BS |
77 | basic_block bb = gimple_bb (stmt); |
78 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
79 | struct loop* loop; | |
80 | ||
81 | if (!loop_vinfo) | |
82 | return false; | |
83 | ||
84 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
85 | ||
86 | return (bb->loop_father == loop->inner); | |
87 | } | |
88 | ||
89 | /* Record the cost of a statement, either by directly informing the | |
90 | target model or by saving it in a vector for later processing. | |
91 | Return a preliminary estimate of the statement's cost. */ | |
92 | ||
93 | unsigned | |
92345349 | 94 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 95 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 96 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 97 | { |
92345349 | 98 | if (body_cost_vec) |
c3e7ee41 | 99 | { |
92345349 | 100 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
ddf56386 RB |
101 | stmt_info_for_cost si = { count, kind, |
102 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
103 | misalign }; | |
104 | body_cost_vec->safe_push (si); | |
c3e7ee41 | 105 | return (unsigned) |
92345349 | 106 | (builtin_vectorization_cost (kind, vectype, misalign) * count); |
c3e7ee41 BS |
107 | } |
108 | else | |
310213d4 RB |
109 | return add_stmt_cost (stmt_info->vinfo->target_cost_data, |
110 | count, kind, stmt_info, misalign, where); | |
c3e7ee41 BS |
111 | } |
112 | ||
272c6793 RS |
113 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
114 | ||
115 | static tree | |
116 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
117 | { | |
118 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
119 | "vect_array"); | |
120 | } | |
121 | ||
122 | /* ARRAY is an array of vectors created by create_vector_array. | |
123 | Return an SSA_NAME for the vector in index N. The reference | |
124 | is part of the vectorization of STMT and the vector is associated | |
125 | with scalar destination SCALAR_DEST. */ | |
126 | ||
127 | static tree | |
355fe088 | 128 | read_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree scalar_dest, |
272c6793 RS |
129 | tree array, unsigned HOST_WIDE_INT n) |
130 | { | |
131 | tree vect_type, vect, vect_name, array_ref; | |
355fe088 | 132 | gimple *new_stmt; |
272c6793 RS |
133 | |
134 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
135 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
136 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
137 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
138 | build_int_cst (size_type_node, n), | |
139 | NULL_TREE, NULL_TREE); | |
140 | ||
141 | new_stmt = gimple_build_assign (vect, array_ref); | |
142 | vect_name = make_ssa_name (vect, new_stmt); | |
143 | gimple_assign_set_lhs (new_stmt, vect_name); | |
144 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
145 | |
146 | return vect_name; | |
147 | } | |
148 | ||
149 | /* ARRAY is an array of vectors created by create_vector_array. | |
150 | Emit code to store SSA_NAME VECT in index N of the array. | |
151 | The store is part of the vectorization of STMT. */ | |
152 | ||
153 | static void | |
355fe088 | 154 | write_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree vect, |
272c6793 RS |
155 | tree array, unsigned HOST_WIDE_INT n) |
156 | { | |
157 | tree array_ref; | |
355fe088 | 158 | gimple *new_stmt; |
272c6793 RS |
159 | |
160 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
161 | build_int_cst (size_type_node, n), | |
162 | NULL_TREE, NULL_TREE); | |
163 | ||
164 | new_stmt = gimple_build_assign (array_ref, vect); | |
165 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
166 | } |
167 | ||
168 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
169 | of *PTR. The memory reference replaces those in FIRST_DR | |
170 | (and its group). */ | |
171 | ||
172 | static tree | |
44fc7854 | 173 | create_array_ref (tree type, tree ptr, tree alias_ptr_type) |
272c6793 | 174 | { |
44fc7854 | 175 | tree mem_ref; |
272c6793 | 176 | |
272c6793 RS |
177 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); |
178 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 179 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
180 | return mem_ref; |
181 | } | |
182 | ||
ebfd146a IR |
183 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
184 | ||
185 | /* Function vect_mark_relevant. | |
186 | ||
187 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
188 | ||
189 | static void | |
355fe088 | 190 | vect_mark_relevant (vec<gimple *> *worklist, gimple *stmt, |
97ecdb46 | 191 | enum vect_relevant relevant, bool live_p) |
ebfd146a IR |
192 | { |
193 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
194 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
195 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
355fe088 | 196 | gimple *pattern_stmt; |
ebfd146a | 197 | |
73fbfcad | 198 | if (dump_enabled_p ()) |
66c16fd9 RB |
199 | { |
200 | dump_printf_loc (MSG_NOTE, vect_location, | |
201 | "mark relevant %d, live %d: ", relevant, live_p); | |
202 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
203 | } | |
ebfd146a | 204 | |
83197f37 IR |
205 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
206 | related pattern stmt instead of the original stmt. However, such stmts | |
207 | may have their own uses that are not in any pattern, in such cases the | |
208 | stmt itself should be marked. */ | |
ebfd146a IR |
209 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
210 | { | |
97ecdb46 JJ |
211 | /* This is the last stmt in a sequence that was detected as a |
212 | pattern that can potentially be vectorized. Don't mark the stmt | |
213 | as relevant/live because it's not going to be vectorized. | |
214 | Instead mark the pattern-stmt that replaces it. */ | |
83197f37 | 215 | |
97ecdb46 JJ |
216 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); |
217 | ||
218 | if (dump_enabled_p ()) | |
219 | dump_printf_loc (MSG_NOTE, vect_location, | |
220 | "last stmt in pattern. don't mark" | |
221 | " relevant/live.\n"); | |
222 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
223 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
224 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
225 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
226 | stmt = pattern_stmt; | |
ebfd146a IR |
227 | } |
228 | ||
229 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
230 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
231 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
232 | ||
233 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
234 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
235 | { | |
73fbfcad | 236 | if (dump_enabled_p ()) |
78c60e3d | 237 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 238 | "already marked relevant/live.\n"); |
ebfd146a IR |
239 | return; |
240 | } | |
241 | ||
9771b263 | 242 | worklist->safe_push (stmt); |
ebfd146a IR |
243 | } |
244 | ||
245 | ||
b28ead45 AH |
246 | /* Function is_simple_and_all_uses_invariant |
247 | ||
248 | Return true if STMT is simple and all uses of it are invariant. */ | |
249 | ||
250 | bool | |
251 | is_simple_and_all_uses_invariant (gimple *stmt, loop_vec_info loop_vinfo) | |
252 | { | |
253 | tree op; | |
254 | gimple *def_stmt; | |
255 | ssa_op_iter iter; | |
256 | ||
257 | if (!is_gimple_assign (stmt)) | |
258 | return false; | |
259 | ||
260 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) | |
261 | { | |
262 | enum vect_def_type dt = vect_uninitialized_def; | |
263 | ||
264 | if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt)) | |
265 | { | |
266 | if (dump_enabled_p ()) | |
267 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
268 | "use not simple.\n"); | |
269 | return false; | |
270 | } | |
271 | ||
272 | if (dt != vect_external_def && dt != vect_constant_def) | |
273 | return false; | |
274 | } | |
275 | return true; | |
276 | } | |
277 | ||
ebfd146a IR |
278 | /* Function vect_stmt_relevant_p. |
279 | ||
280 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
281 | "relevant for vectorization". | |
282 | ||
283 | A stmt is considered "relevant for vectorization" if: | |
284 | - it has uses outside the loop. | |
285 | - it has vdefs (it alters memory). | |
286 | - control stmts in the loop (except for the exit condition). | |
287 | ||
288 | CHECKME: what other side effects would the vectorizer allow? */ | |
289 | ||
290 | static bool | |
355fe088 | 291 | vect_stmt_relevant_p (gimple *stmt, loop_vec_info loop_vinfo, |
ebfd146a IR |
292 | enum vect_relevant *relevant, bool *live_p) |
293 | { | |
294 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
295 | ssa_op_iter op_iter; | |
296 | imm_use_iterator imm_iter; | |
297 | use_operand_p use_p; | |
298 | def_operand_p def_p; | |
299 | ||
8644a673 | 300 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
301 | *live_p = false; |
302 | ||
303 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
304 | if (is_ctrl_stmt (stmt) |
305 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
306 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 307 | *relevant = vect_used_in_scope; |
ebfd146a IR |
308 | |
309 | /* changing memory. */ | |
310 | if (gimple_code (stmt) != GIMPLE_PHI) | |
ac6aeab4 RB |
311 | if (gimple_vdef (stmt) |
312 | && !gimple_clobber_p (stmt)) | |
ebfd146a | 313 | { |
73fbfcad | 314 | if (dump_enabled_p ()) |
78c60e3d | 315 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 316 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 317 | *relevant = vect_used_in_scope; |
ebfd146a IR |
318 | } |
319 | ||
320 | /* uses outside the loop. */ | |
321 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
322 | { | |
323 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
324 | { | |
325 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
326 | if (!flow_bb_inside_loop_p (loop, bb)) | |
327 | { | |
73fbfcad | 328 | if (dump_enabled_p ()) |
78c60e3d | 329 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 330 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 331 | |
3157b0c2 AO |
332 | if (is_gimple_debug (USE_STMT (use_p))) |
333 | continue; | |
334 | ||
ebfd146a IR |
335 | /* We expect all such uses to be in the loop exit phis |
336 | (because of loop closed form) */ | |
337 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
338 | gcc_assert (bb == single_exit (loop)->dest); | |
339 | ||
340 | *live_p = true; | |
341 | } | |
342 | } | |
343 | } | |
344 | ||
3a2edf4c AH |
345 | if (*live_p && *relevant == vect_unused_in_scope |
346 | && !is_simple_and_all_uses_invariant (stmt, loop_vinfo)) | |
b28ead45 AH |
347 | { |
348 | if (dump_enabled_p ()) | |
349 | dump_printf_loc (MSG_NOTE, vect_location, | |
350 | "vec_stmt_relevant_p: stmt live but not relevant.\n"); | |
351 | *relevant = vect_used_only_live; | |
352 | } | |
353 | ||
ebfd146a IR |
354 | return (*live_p || *relevant); |
355 | } | |
356 | ||
357 | ||
b8698a0f | 358 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 359 | |
ff802fa1 | 360 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
361 | used in STMT for anything other than indexing an array. */ |
362 | ||
363 | static bool | |
355fe088 | 364 | exist_non_indexing_operands_for_use_p (tree use, gimple *stmt) |
ebfd146a IR |
365 | { |
366 | tree operand; | |
367 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 368 | |
ff802fa1 | 369 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
370 | reference in STMT, then any operand that corresponds to USE |
371 | is not indexing an array. */ | |
372 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
373 | return true; | |
59a05b0c | 374 | |
ebfd146a IR |
375 | /* STMT has a data_ref. FORNOW this means that its of one of |
376 | the following forms: | |
377 | -1- ARRAY_REF = var | |
378 | -2- var = ARRAY_REF | |
379 | (This should have been verified in analyze_data_refs). | |
380 | ||
381 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 382 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
383 | for array indexing. |
384 | ||
385 | Therefore, all we need to check is if STMT falls into the | |
386 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
387 | |
388 | if (!gimple_assign_copy_p (stmt)) | |
5ce9450f JJ |
389 | { |
390 | if (is_gimple_call (stmt) | |
391 | && gimple_call_internal_p (stmt)) | |
392 | switch (gimple_call_internal_fn (stmt)) | |
393 | { | |
394 | case IFN_MASK_STORE: | |
395 | operand = gimple_call_arg (stmt, 3); | |
396 | if (operand == use) | |
397 | return true; | |
398 | /* FALLTHRU */ | |
399 | case IFN_MASK_LOAD: | |
400 | operand = gimple_call_arg (stmt, 2); | |
401 | if (operand == use) | |
402 | return true; | |
403 | break; | |
404 | default: | |
405 | break; | |
406 | } | |
407 | return false; | |
408 | } | |
409 | ||
59a05b0c EB |
410 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
411 | return false; | |
ebfd146a | 412 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
413 | if (TREE_CODE (operand) != SSA_NAME) |
414 | return false; | |
415 | ||
416 | if (operand == use) | |
417 | return true; | |
418 | ||
419 | return false; | |
420 | } | |
421 | ||
422 | ||
b8698a0f | 423 | /* |
ebfd146a IR |
424 | Function process_use. |
425 | ||
426 | Inputs: | |
427 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b28ead45 | 428 | - RELEVANT - enum value to be set in the STMT_VINFO of the stmt |
ff802fa1 | 429 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 430 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
431 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
432 | be performed. | |
ebfd146a IR |
433 | |
434 | Outputs: | |
435 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
436 | relevance info of the DEF_STMT of this USE: | |
437 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
438 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
439 | Exceptions: | |
440 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 441 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 442 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
443 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
444 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
445 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
446 | be modified accordingly. | |
447 | ||
448 | Return true if everything is as expected. Return false otherwise. */ | |
449 | ||
450 | static bool | |
b28ead45 | 451 | process_use (gimple *stmt, tree use, loop_vec_info loop_vinfo, |
355fe088 | 452 | enum vect_relevant relevant, vec<gimple *> *worklist, |
aec7ae7d | 453 | bool force) |
ebfd146a IR |
454 | { |
455 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
456 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
457 | stmt_vec_info dstmt_vinfo; | |
458 | basic_block bb, def_bb; | |
355fe088 | 459 | gimple *def_stmt; |
ebfd146a IR |
460 | enum vect_def_type dt; |
461 | ||
b8698a0f | 462 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 463 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 464 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
465 | return true; |
466 | ||
81c40241 | 467 | if (!vect_is_simple_use (use, loop_vinfo, &def_stmt, &dt)) |
b8698a0f | 468 | { |
73fbfcad | 469 | if (dump_enabled_p ()) |
78c60e3d | 470 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 471 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
472 | return false; |
473 | } | |
474 | ||
475 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
476 | return true; | |
477 | ||
478 | def_bb = gimple_bb (def_stmt); | |
479 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
480 | { | |
73fbfcad | 481 | if (dump_enabled_p ()) |
e645e942 | 482 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n"); |
ebfd146a IR |
483 | return true; |
484 | } | |
485 | ||
b8698a0f L |
486 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
487 | DEF_STMT must have already been processed, because this should be the | |
488 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
489 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
490 | check that everything is as expected, and we are done. */ |
491 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
492 | bb = gimple_bb (stmt); | |
493 | if (gimple_code (stmt) == GIMPLE_PHI | |
494 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
495 | && gimple_code (def_stmt) != GIMPLE_PHI | |
496 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
497 | && bb->loop_father == def_bb->loop_father) | |
498 | { | |
73fbfcad | 499 | if (dump_enabled_p ()) |
78c60e3d | 500 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 501 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a IR |
502 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
503 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
504 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 505 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 506 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
507 | return true; |
508 | } | |
509 | ||
510 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
511 | outer-loop-header-bb: | |
512 | d = def_stmt | |
513 | inner-loop: | |
514 | stmt # use (d) | |
515 | outer-loop-tail-bb: | |
516 | ... */ | |
517 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
518 | { | |
73fbfcad | 519 | if (dump_enabled_p ()) |
78c60e3d | 520 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 521 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 522 | |
ebfd146a IR |
523 | switch (relevant) |
524 | { | |
8644a673 | 525 | case vect_unused_in_scope: |
7c5222ff IR |
526 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
527 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 528 | break; |
7c5222ff | 529 | |
ebfd146a | 530 | case vect_used_in_outer_by_reduction: |
7c5222ff | 531 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
532 | relevant = vect_used_by_reduction; |
533 | break; | |
7c5222ff | 534 | |
ebfd146a | 535 | case vect_used_in_outer: |
7c5222ff | 536 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 537 | relevant = vect_used_in_scope; |
ebfd146a | 538 | break; |
7c5222ff | 539 | |
8644a673 | 540 | case vect_used_in_scope: |
ebfd146a IR |
541 | break; |
542 | ||
543 | default: | |
544 | gcc_unreachable (); | |
b8698a0f | 545 | } |
ebfd146a IR |
546 | } |
547 | ||
548 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
549 | outer-loop-header-bb: | |
550 | ... | |
551 | inner-loop: | |
552 | d = def_stmt | |
06066f92 | 553 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
554 | stmt # use (d) */ |
555 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
556 | { | |
73fbfcad | 557 | if (dump_enabled_p ()) |
78c60e3d | 558 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 559 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 560 | |
ebfd146a IR |
561 | switch (relevant) |
562 | { | |
8644a673 | 563 | case vect_unused_in_scope: |
b8698a0f | 564 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 565 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 566 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
567 | break; |
568 | ||
ebfd146a | 569 | case vect_used_by_reduction: |
b28ead45 | 570 | case vect_used_only_live: |
ebfd146a IR |
571 | relevant = vect_used_in_outer_by_reduction; |
572 | break; | |
573 | ||
8644a673 | 574 | case vect_used_in_scope: |
ebfd146a IR |
575 | relevant = vect_used_in_outer; |
576 | break; | |
577 | ||
578 | default: | |
579 | gcc_unreachable (); | |
580 | } | |
581 | } | |
643a9684 RB |
582 | /* We are also not interested in uses on loop PHI backedges that are |
583 | inductions. Otherwise we'll needlessly vectorize the IV increment | |
e294f495 RB |
584 | and cause hybrid SLP for SLP inductions. Unless the PHI is live |
585 | of course. */ | |
643a9684 RB |
586 | else if (gimple_code (stmt) == GIMPLE_PHI |
587 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_induction_def | |
e294f495 | 588 | && ! STMT_VINFO_LIVE_P (stmt_vinfo) |
643a9684 RB |
589 | && (PHI_ARG_DEF_FROM_EDGE (stmt, loop_latch_edge (bb->loop_father)) |
590 | == use)) | |
591 | { | |
592 | if (dump_enabled_p ()) | |
593 | dump_printf_loc (MSG_NOTE, vect_location, | |
594 | "induction value on backedge.\n"); | |
595 | return true; | |
596 | } | |
597 | ||
ebfd146a | 598 | |
b28ead45 | 599 | vect_mark_relevant (worklist, def_stmt, relevant, false); |
ebfd146a IR |
600 | return true; |
601 | } | |
602 | ||
603 | ||
604 | /* Function vect_mark_stmts_to_be_vectorized. | |
605 | ||
606 | Not all stmts in the loop need to be vectorized. For example: | |
607 | ||
608 | for i... | |
609 | for j... | |
610 | 1. T0 = i + j | |
611 | 2. T1 = a[T0] | |
612 | ||
613 | 3. j = j + 1 | |
614 | ||
615 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
616 | addressing of vectorized data-refs are handled differently. | |
617 | ||
618 | This pass detects such stmts. */ | |
619 | ||
620 | bool | |
621 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
622 | { | |
ebfd146a IR |
623 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
624 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
625 | unsigned int nbbs = loop->num_nodes; | |
626 | gimple_stmt_iterator si; | |
355fe088 | 627 | gimple *stmt; |
ebfd146a IR |
628 | unsigned int i; |
629 | stmt_vec_info stmt_vinfo; | |
630 | basic_block bb; | |
355fe088 | 631 | gimple *phi; |
ebfd146a | 632 | bool live_p; |
b28ead45 | 633 | enum vect_relevant relevant; |
ebfd146a | 634 | |
73fbfcad | 635 | if (dump_enabled_p ()) |
78c60e3d | 636 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 637 | "=== vect_mark_stmts_to_be_vectorized ===\n"); |
ebfd146a | 638 | |
355fe088 | 639 | auto_vec<gimple *, 64> worklist; |
ebfd146a IR |
640 | |
641 | /* 1. Init worklist. */ | |
642 | for (i = 0; i < nbbs; i++) | |
643 | { | |
644 | bb = bbs[i]; | |
645 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 646 | { |
ebfd146a | 647 | phi = gsi_stmt (si); |
73fbfcad | 648 | if (dump_enabled_p ()) |
ebfd146a | 649 | { |
78c60e3d SS |
650 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
651 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
652 | } |
653 | ||
654 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 655 | vect_mark_relevant (&worklist, phi, relevant, live_p); |
ebfd146a IR |
656 | } |
657 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
658 | { | |
659 | stmt = gsi_stmt (si); | |
73fbfcad | 660 | if (dump_enabled_p ()) |
ebfd146a | 661 | { |
78c60e3d SS |
662 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
663 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
b8698a0f | 664 | } |
ebfd146a IR |
665 | |
666 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 667 | vect_mark_relevant (&worklist, stmt, relevant, live_p); |
ebfd146a IR |
668 | } |
669 | } | |
670 | ||
671 | /* 2. Process_worklist */ | |
9771b263 | 672 | while (worklist.length () > 0) |
ebfd146a IR |
673 | { |
674 | use_operand_p use_p; | |
675 | ssa_op_iter iter; | |
676 | ||
9771b263 | 677 | stmt = worklist.pop (); |
73fbfcad | 678 | if (dump_enabled_p ()) |
ebfd146a | 679 | { |
78c60e3d SS |
680 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
681 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
ebfd146a IR |
682 | } |
683 | ||
b8698a0f | 684 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
b28ead45 AH |
685 | (DEF_STMT) as relevant/irrelevant according to the relevance property |
686 | of STMT. */ | |
ebfd146a IR |
687 | stmt_vinfo = vinfo_for_stmt (stmt); |
688 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
ebfd146a | 689 | |
b28ead45 AH |
690 | /* Generally, the relevance property of STMT (in STMT_VINFO_RELEVANT) is |
691 | propagated as is to the DEF_STMTs of its USEs. | |
ebfd146a IR |
692 | |
693 | One exception is when STMT has been identified as defining a reduction | |
b28ead45 | 694 | variable; in this case we set the relevance to vect_used_by_reduction. |
ebfd146a | 695 | This is because we distinguish between two kinds of relevant stmts - |
b8698a0f | 696 | those that are used by a reduction computation, and those that are |
ff802fa1 | 697 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 698 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 699 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 700 | |
b28ead45 | 701 | switch (STMT_VINFO_DEF_TYPE (stmt_vinfo)) |
ebfd146a | 702 | { |
06066f92 | 703 | case vect_reduction_def: |
b28ead45 AH |
704 | gcc_assert (relevant != vect_unused_in_scope); |
705 | if (relevant != vect_unused_in_scope | |
706 | && relevant != vect_used_in_scope | |
707 | && relevant != vect_used_by_reduction | |
708 | && relevant != vect_used_only_live) | |
06066f92 | 709 | { |
b28ead45 AH |
710 | if (dump_enabled_p ()) |
711 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
712 | "unsupported use of reduction.\n"); | |
713 | return false; | |
06066f92 | 714 | } |
06066f92 | 715 | break; |
b8698a0f | 716 | |
06066f92 | 717 | case vect_nested_cycle: |
b28ead45 AH |
718 | if (relevant != vect_unused_in_scope |
719 | && relevant != vect_used_in_outer_by_reduction | |
720 | && relevant != vect_used_in_outer) | |
06066f92 | 721 | { |
73fbfcad | 722 | if (dump_enabled_p ()) |
78c60e3d | 723 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 724 | "unsupported use of nested cycle.\n"); |
7c5222ff | 725 | |
06066f92 IR |
726 | return false; |
727 | } | |
b8698a0f L |
728 | break; |
729 | ||
06066f92 | 730 | case vect_double_reduction_def: |
b28ead45 AH |
731 | if (relevant != vect_unused_in_scope |
732 | && relevant != vect_used_by_reduction | |
733 | && relevant != vect_used_only_live) | |
06066f92 | 734 | { |
73fbfcad | 735 | if (dump_enabled_p ()) |
78c60e3d | 736 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 737 | "unsupported use of double reduction.\n"); |
7c5222ff | 738 | |
7c5222ff | 739 | return false; |
06066f92 | 740 | } |
b8698a0f | 741 | break; |
7c5222ff | 742 | |
06066f92 IR |
743 | default: |
744 | break; | |
7c5222ff | 745 | } |
b8698a0f | 746 | |
aec7ae7d | 747 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
748 | { |
749 | /* Pattern statements are not inserted into the code, so | |
750 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
751 | have to scan the RHS or function arguments instead. */ | |
752 | if (is_gimple_assign (stmt)) | |
753 | { | |
69d2aade JJ |
754 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
755 | tree op = gimple_assign_rhs1 (stmt); | |
756 | ||
757 | i = 1; | |
758 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
759 | { | |
760 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
b28ead45 | 761 | relevant, &worklist, false) |
69d2aade | 762 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
b28ead45 | 763 | relevant, &worklist, false)) |
566d377a | 764 | return false; |
69d2aade JJ |
765 | i = 2; |
766 | } | |
767 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 768 | { |
69d2aade | 769 | op = gimple_op (stmt, i); |
afbe6325 | 770 | if (TREE_CODE (op) == SSA_NAME |
b28ead45 | 771 | && !process_use (stmt, op, loop_vinfo, relevant, |
afbe6325 | 772 | &worklist, false)) |
07687835 | 773 | return false; |
9d5e7640 IR |
774 | } |
775 | } | |
776 | else if (is_gimple_call (stmt)) | |
777 | { | |
778 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
779 | { | |
780 | tree arg = gimple_call_arg (stmt, i); | |
b28ead45 | 781 | if (!process_use (stmt, arg, loop_vinfo, relevant, |
aec7ae7d | 782 | &worklist, false)) |
07687835 | 783 | return false; |
9d5e7640 IR |
784 | } |
785 | } | |
786 | } | |
787 | else | |
788 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
789 | { | |
790 | tree op = USE_FROM_PTR (use_p); | |
b28ead45 | 791 | if (!process_use (stmt, op, loop_vinfo, relevant, |
aec7ae7d | 792 | &worklist, false)) |
07687835 | 793 | return false; |
9d5e7640 | 794 | } |
aec7ae7d | 795 | |
3bab6342 | 796 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo)) |
aec7ae7d | 797 | { |
134c85ca RS |
798 | gather_scatter_info gs_info; |
799 | if (!vect_check_gather_scatter (stmt, loop_vinfo, &gs_info)) | |
800 | gcc_unreachable (); | |
801 | if (!process_use (stmt, gs_info.offset, loop_vinfo, relevant, | |
802 | &worklist, true)) | |
566d377a | 803 | return false; |
aec7ae7d | 804 | } |
ebfd146a IR |
805 | } /* while worklist */ |
806 | ||
ebfd146a IR |
807 | return true; |
808 | } | |
809 | ||
810 | ||
b8698a0f | 811 | /* Function vect_model_simple_cost. |
ebfd146a | 812 | |
b8698a0f | 813 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
814 | single op. Right now, this does not account for multiple insns that could |
815 | be generated for the single vector op. We will handle that shortly. */ | |
816 | ||
817 | void | |
b8698a0f | 818 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 | 819 | enum vect_def_type *dt, |
4fc5ebf1 | 820 | int ndts, |
92345349 BS |
821 | stmt_vector_for_cost *prologue_cost_vec, |
822 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
823 | { |
824 | int i; | |
92345349 | 825 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
826 | |
827 | /* The SLP costs were already calculated during SLP tree build. */ | |
828 | if (PURE_SLP_STMT (stmt_info)) | |
829 | return; | |
830 | ||
4fc5ebf1 JG |
831 | /* Cost the "broadcast" of a scalar operand in to a vector operand. |
832 | Use scalar_to_vec to cost the broadcast, as elsewhere in the vector | |
833 | cost model. */ | |
834 | for (i = 0; i < ndts; i++) | |
92345349 | 835 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
4fc5ebf1 | 836 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
92345349 | 837 | stmt_info, 0, vect_prologue); |
c3e7ee41 BS |
838 | |
839 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
92345349 BS |
840 | inside_cost = record_stmt_cost (body_cost_vec, ncopies, vector_stmt, |
841 | stmt_info, 0, vect_body); | |
c3e7ee41 | 842 | |
73fbfcad | 843 | if (dump_enabled_p ()) |
78c60e3d SS |
844 | dump_printf_loc (MSG_NOTE, vect_location, |
845 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 846 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
847 | } |
848 | ||
849 | ||
8bd37302 BS |
850 | /* Model cost for type demotion and promotion operations. PWR is normally |
851 | zero for single-step promotions and demotions. It will be one if | |
852 | two-step promotion/demotion is required, and so on. Each additional | |
853 | step doubles the number of instructions required. */ | |
854 | ||
855 | static void | |
856 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
857 | enum vect_def_type *dt, int pwr) | |
858 | { | |
859 | int i, tmp; | |
92345349 | 860 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 BS |
861 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
862 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
863 | void *target_cost_data; | |
8bd37302 BS |
864 | |
865 | /* The SLP costs were already calculated during SLP tree build. */ | |
866 | if (PURE_SLP_STMT (stmt_info)) | |
867 | return; | |
868 | ||
c3e7ee41 BS |
869 | if (loop_vinfo) |
870 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
871 | else | |
872 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
873 | ||
8bd37302 BS |
874 | for (i = 0; i < pwr + 1; i++) |
875 | { | |
876 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
877 | (i + 1) : i; | |
c3e7ee41 | 878 | inside_cost += add_stmt_cost (target_cost_data, vect_pow2 (tmp), |
92345349 BS |
879 | vec_promote_demote, stmt_info, 0, |
880 | vect_body); | |
8bd37302 BS |
881 | } |
882 | ||
883 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
884 | for (i = 0; i < 2; i++) | |
92345349 BS |
885 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
886 | prologue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt, | |
887 | stmt_info, 0, vect_prologue); | |
8bd37302 | 888 | |
73fbfcad | 889 | if (dump_enabled_p ()) |
78c60e3d SS |
890 | dump_printf_loc (MSG_NOTE, vect_location, |
891 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 892 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
893 | } |
894 | ||
ebfd146a IR |
895 | /* Function vect_model_store_cost |
896 | ||
0d0293ac MM |
897 | Models cost for stores. In the case of grouped accesses, one access |
898 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a IR |
899 | |
900 | void | |
b8698a0f | 901 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
2de001ee RS |
902 | vect_memory_access_type memory_access_type, |
903 | enum vect_def_type dt, slp_tree slp_node, | |
92345349 BS |
904 | stmt_vector_for_cost *prologue_cost_vec, |
905 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a | 906 | { |
92345349 | 907 | unsigned int inside_cost = 0, prologue_cost = 0; |
892a981f RS |
908 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
909 | gimple *first_stmt = STMT_VINFO_STMT (stmt_info); | |
910 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); | |
ebfd146a | 911 | |
8644a673 | 912 | if (dt == vect_constant_def || dt == vect_external_def) |
92345349 BS |
913 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
914 | stmt_info, 0, vect_prologue); | |
ebfd146a | 915 | |
892a981f RS |
916 | /* Grouped stores update all elements in the group at once, |
917 | so we want the DR for the first statement. */ | |
918 | if (!slp_node && grouped_access_p) | |
720f5239 | 919 | { |
892a981f RS |
920 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
921 | dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
720f5239 | 922 | } |
ebfd146a | 923 | |
892a981f RS |
924 | /* True if we should include any once-per-group costs as well as |
925 | the cost of the statement itself. For SLP we only get called | |
926 | once per group anyhow. */ | |
927 | bool first_stmt_p = (first_stmt == STMT_VINFO_STMT (stmt_info)); | |
928 | ||
272c6793 | 929 | /* We assume that the cost of a single store-lanes instruction is |
0d0293ac | 930 | equivalent to the cost of GROUP_SIZE separate stores. If a grouped |
272c6793 | 931 | access is instead being provided by a permute-and-store operation, |
2de001ee RS |
932 | include the cost of the permutes. */ |
933 | if (first_stmt_p | |
934 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 935 | { |
e1377713 ES |
936 | /* Uses a high and low interleave or shuffle operations for each |
937 | needed permute. */ | |
892a981f | 938 | int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
e1377713 | 939 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
92345349 BS |
940 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
941 | stmt_info, 0, vect_body); | |
ebfd146a | 942 | |
73fbfcad | 943 | if (dump_enabled_p ()) |
78c60e3d | 944 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 945 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 946 | group_size); |
ebfd146a IR |
947 | } |
948 | ||
cee62fee | 949 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a | 950 | /* Costs of the stores. */ |
067bc855 RB |
951 | if (memory_access_type == VMAT_ELEMENTWISE |
952 | || memory_access_type == VMAT_GATHER_SCATTER) | |
2de001ee RS |
953 | /* N scalar stores plus extracting the elements. */ |
954 | inside_cost += record_stmt_cost (body_cost_vec, | |
955 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), | |
956 | scalar_store, stmt_info, 0, vect_body); | |
f2e2a985 | 957 | else |
892a981f | 958 | vect_get_store_cost (dr, ncopies, &inside_cost, body_cost_vec); |
ebfd146a | 959 | |
2de001ee RS |
960 | if (memory_access_type == VMAT_ELEMENTWISE |
961 | || memory_access_type == VMAT_STRIDED_SLP) | |
cee62fee MM |
962 | inside_cost += record_stmt_cost (body_cost_vec, |
963 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), | |
964 | vec_to_scalar, stmt_info, 0, vect_body); | |
965 | ||
73fbfcad | 966 | if (dump_enabled_p ()) |
78c60e3d SS |
967 | dump_printf_loc (MSG_NOTE, vect_location, |
968 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 969 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
970 | } |
971 | ||
972 | ||
720f5239 IR |
973 | /* Calculate cost of DR's memory access. */ |
974 | void | |
975 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 976 | unsigned int *inside_cost, |
92345349 | 977 | stmt_vector_for_cost *body_cost_vec) |
720f5239 IR |
978 | { |
979 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 980 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 981 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
982 | |
983 | switch (alignment_support_scheme) | |
984 | { | |
985 | case dr_aligned: | |
986 | { | |
92345349 BS |
987 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
988 | vector_store, stmt_info, 0, | |
989 | vect_body); | |
720f5239 | 990 | |
73fbfcad | 991 | if (dump_enabled_p ()) |
78c60e3d | 992 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 993 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
994 | break; |
995 | } | |
996 | ||
997 | case dr_unaligned_supported: | |
998 | { | |
720f5239 | 999 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1000 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1001 | unaligned_store, stmt_info, |
92345349 | 1002 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1003 | if (dump_enabled_p ()) |
78c60e3d SS |
1004 | dump_printf_loc (MSG_NOTE, vect_location, |
1005 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1006 | "hardware.\n"); |
720f5239 IR |
1007 | break; |
1008 | } | |
1009 | ||
38eec4c6 UW |
1010 | case dr_unaligned_unsupported: |
1011 | { | |
1012 | *inside_cost = VECT_MAX_COST; | |
1013 | ||
73fbfcad | 1014 | if (dump_enabled_p ()) |
78c60e3d | 1015 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1016 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1017 | break; |
1018 | } | |
1019 | ||
720f5239 IR |
1020 | default: |
1021 | gcc_unreachable (); | |
1022 | } | |
1023 | } | |
1024 | ||
1025 | ||
ebfd146a IR |
1026 | /* Function vect_model_load_cost |
1027 | ||
892a981f RS |
1028 | Models cost for loads. In the case of grouped accesses, one access has |
1029 | the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1030 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1031 | access scheme chosen. */ |
1032 | ||
1033 | void | |
92345349 | 1034 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, |
2de001ee RS |
1035 | vect_memory_access_type memory_access_type, |
1036 | slp_tree slp_node, | |
92345349 BS |
1037 | stmt_vector_for_cost *prologue_cost_vec, |
1038 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a | 1039 | { |
892a981f RS |
1040 | gimple *first_stmt = STMT_VINFO_STMT (stmt_info); |
1041 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
92345349 | 1042 | unsigned int inside_cost = 0, prologue_cost = 0; |
892a981f | 1043 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); |
ebfd146a | 1044 | |
892a981f RS |
1045 | /* Grouped loads read all elements in the group at once, |
1046 | so we want the DR for the first statement. */ | |
1047 | if (!slp_node && grouped_access_p) | |
ebfd146a | 1048 | { |
892a981f RS |
1049 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
1050 | dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
ebfd146a IR |
1051 | } |
1052 | ||
892a981f RS |
1053 | /* True if we should include any once-per-group costs as well as |
1054 | the cost of the statement itself. For SLP we only get called | |
1055 | once per group anyhow. */ | |
1056 | bool first_stmt_p = (first_stmt == STMT_VINFO_STMT (stmt_info)); | |
1057 | ||
272c6793 | 1058 | /* We assume that the cost of a single load-lanes instruction is |
0d0293ac | 1059 | equivalent to the cost of GROUP_SIZE separate loads. If a grouped |
272c6793 | 1060 | access is instead being provided by a load-and-permute operation, |
2de001ee RS |
1061 | include the cost of the permutes. */ |
1062 | if (first_stmt_p | |
1063 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 1064 | { |
2c23db6d ES |
1065 | /* Uses an even and odd extract operations or shuffle operations |
1066 | for each needed permute. */ | |
892a981f | 1067 | int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
2c23db6d ES |
1068 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
1069 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, | |
1070 | stmt_info, 0, vect_body); | |
ebfd146a | 1071 | |
73fbfcad | 1072 | if (dump_enabled_p ()) |
e645e942 TJ |
1073 | dump_printf_loc (MSG_NOTE, vect_location, |
1074 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1075 | group_size); |
ebfd146a IR |
1076 | } |
1077 | ||
1078 | /* The loads themselves. */ | |
067bc855 RB |
1079 | if (memory_access_type == VMAT_ELEMENTWISE |
1080 | || memory_access_type == VMAT_GATHER_SCATTER) | |
a82960aa | 1081 | { |
a21892ad BS |
1082 | /* N scalar loads plus gathering them into a vector. */ |
1083 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
92345349 | 1084 | inside_cost += record_stmt_cost (body_cost_vec, |
c3e7ee41 | 1085 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), |
92345349 | 1086 | scalar_load, stmt_info, 0, vect_body); |
a82960aa RG |
1087 | } |
1088 | else | |
892a981f | 1089 | vect_get_load_cost (dr, ncopies, first_stmt_p, |
92345349 BS |
1090 | &inside_cost, &prologue_cost, |
1091 | prologue_cost_vec, body_cost_vec, true); | |
2de001ee RS |
1092 | if (memory_access_type == VMAT_ELEMENTWISE |
1093 | || memory_access_type == VMAT_STRIDED_SLP) | |
892a981f RS |
1094 | inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct, |
1095 | stmt_info, 0, vect_body); | |
720f5239 | 1096 | |
73fbfcad | 1097 | if (dump_enabled_p ()) |
78c60e3d SS |
1098 | dump_printf_loc (MSG_NOTE, vect_location, |
1099 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1100 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1101 | } |
1102 | ||
1103 | ||
1104 | /* Calculate cost of DR's memory access. */ | |
1105 | void | |
1106 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1107 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1108 | unsigned int *prologue_cost, |
1109 | stmt_vector_for_cost *prologue_cost_vec, | |
1110 | stmt_vector_for_cost *body_cost_vec, | |
1111 | bool record_prologue_costs) | |
720f5239 IR |
1112 | { |
1113 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 1114 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 1115 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
1116 | |
1117 | switch (alignment_support_scheme) | |
ebfd146a IR |
1118 | { |
1119 | case dr_aligned: | |
1120 | { | |
92345349 BS |
1121 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1122 | stmt_info, 0, vect_body); | |
ebfd146a | 1123 | |
73fbfcad | 1124 | if (dump_enabled_p ()) |
78c60e3d | 1125 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1126 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1127 | |
1128 | break; | |
1129 | } | |
1130 | case dr_unaligned_supported: | |
1131 | { | |
720f5239 | 1132 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1133 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1134 | unaligned_load, stmt_info, |
92345349 | 1135 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1136 | |
73fbfcad | 1137 | if (dump_enabled_p ()) |
78c60e3d SS |
1138 | dump_printf_loc (MSG_NOTE, vect_location, |
1139 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1140 | "hardware.\n"); |
ebfd146a IR |
1141 | |
1142 | break; | |
1143 | } | |
1144 | case dr_explicit_realign: | |
1145 | { | |
92345349 BS |
1146 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1147 | vector_load, stmt_info, 0, vect_body); | |
1148 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1149 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1150 | |
1151 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1152 | the containing loop, the following cost should be added to the | |
92345349 | 1153 | prologue costs. */ |
ebfd146a | 1154 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1155 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1156 | stmt_info, 0, vect_body); | |
ebfd146a | 1157 | |
73fbfcad | 1158 | if (dump_enabled_p ()) |
e645e942 TJ |
1159 | dump_printf_loc (MSG_NOTE, vect_location, |
1160 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1161 | |
ebfd146a IR |
1162 | break; |
1163 | } | |
1164 | case dr_explicit_realign_optimized: | |
1165 | { | |
73fbfcad | 1166 | if (dump_enabled_p ()) |
e645e942 | 1167 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1168 | "vect_model_load_cost: unaligned software " |
e645e942 | 1169 | "pipelined.\n"); |
ebfd146a IR |
1170 | |
1171 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1172 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1173 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1174 | access, then the above cost should only be considered for one |
ff802fa1 | 1175 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1176 | and a realignment op. */ |
1177 | ||
92345349 | 1178 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1179 | { |
92345349 BS |
1180 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1181 | vector_stmt, stmt_info, | |
1182 | 0, vect_prologue); | |
ebfd146a | 1183 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1184 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1185 | vector_stmt, stmt_info, | |
1186 | 0, vect_prologue); | |
ebfd146a IR |
1187 | } |
1188 | ||
92345349 BS |
1189 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1190 | stmt_info, 0, vect_body); | |
1191 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1192 | stmt_info, 0, vect_body); | |
8bd37302 | 1193 | |
73fbfcad | 1194 | if (dump_enabled_p ()) |
78c60e3d | 1195 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1196 | "vect_model_load_cost: explicit realign optimized" |
1197 | "\n"); | |
8bd37302 | 1198 | |
ebfd146a IR |
1199 | break; |
1200 | } | |
1201 | ||
38eec4c6 UW |
1202 | case dr_unaligned_unsupported: |
1203 | { | |
1204 | *inside_cost = VECT_MAX_COST; | |
1205 | ||
73fbfcad | 1206 | if (dump_enabled_p ()) |
78c60e3d | 1207 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1208 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1209 | break; |
1210 | } | |
1211 | ||
ebfd146a IR |
1212 | default: |
1213 | gcc_unreachable (); | |
1214 | } | |
ebfd146a IR |
1215 | } |
1216 | ||
418b7df3 RG |
1217 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1218 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1219 | |
418b7df3 | 1220 | static void |
355fe088 | 1221 | vect_init_vector_1 (gimple *stmt, gimple *new_stmt, gimple_stmt_iterator *gsi) |
ebfd146a | 1222 | { |
ebfd146a | 1223 | if (gsi) |
418b7df3 | 1224 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1225 | else |
1226 | { | |
418b7df3 | 1227 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1228 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1229 | |
a70d6342 IR |
1230 | if (loop_vinfo) |
1231 | { | |
1232 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1233 | basic_block new_bb; |
1234 | edge pe; | |
a70d6342 IR |
1235 | |
1236 | if (nested_in_vect_loop_p (loop, stmt)) | |
1237 | loop = loop->inner; | |
b8698a0f | 1238 | |
a70d6342 | 1239 | pe = loop_preheader_edge (loop); |
418b7df3 | 1240 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1241 | gcc_assert (!new_bb); |
1242 | } | |
1243 | else | |
1244 | { | |
1245 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1246 | basic_block bb; | |
1247 | gimple_stmt_iterator gsi_bb_start; | |
1248 | ||
1249 | gcc_assert (bb_vinfo); | |
1250 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1251 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1252 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1253 | } |
ebfd146a IR |
1254 | } |
1255 | ||
73fbfcad | 1256 | if (dump_enabled_p ()) |
ebfd146a | 1257 | { |
78c60e3d SS |
1258 | dump_printf_loc (MSG_NOTE, vect_location, |
1259 | "created new init_stmt: "); | |
1260 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
ebfd146a | 1261 | } |
418b7df3 RG |
1262 | } |
1263 | ||
1264 | /* Function vect_init_vector. | |
ebfd146a | 1265 | |
5467ee52 RG |
1266 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1267 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1268 | vector type a vector with all elements equal to VAL is created first. | |
1269 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1270 | initialization at the loop preheader. | |
418b7df3 RG |
1271 | Return the DEF of INIT_STMT. |
1272 | It will be used in the vectorization of STMT. */ | |
1273 | ||
1274 | tree | |
355fe088 | 1275 | vect_init_vector (gimple *stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 | 1276 | { |
355fe088 | 1277 | gimple *init_stmt; |
418b7df3 RG |
1278 | tree new_temp; |
1279 | ||
e412ece4 RB |
1280 | /* We abuse this function to push sth to a SSA name with initial 'val'. */ |
1281 | if (! useless_type_conversion_p (type, TREE_TYPE (val))) | |
418b7df3 | 1282 | { |
e412ece4 RB |
1283 | gcc_assert (TREE_CODE (type) == VECTOR_TYPE); |
1284 | if (! types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) | |
418b7df3 | 1285 | { |
5a308cf1 IE |
1286 | /* Scalar boolean value should be transformed into |
1287 | all zeros or all ones value before building a vector. */ | |
1288 | if (VECTOR_BOOLEAN_TYPE_P (type)) | |
1289 | { | |
b3d51f23 IE |
1290 | tree true_val = build_all_ones_cst (TREE_TYPE (type)); |
1291 | tree false_val = build_zero_cst (TREE_TYPE (type)); | |
5a308cf1 IE |
1292 | |
1293 | if (CONSTANT_CLASS_P (val)) | |
1294 | val = integer_zerop (val) ? false_val : true_val; | |
1295 | else | |
1296 | { | |
1297 | new_temp = make_ssa_name (TREE_TYPE (type)); | |
1298 | init_stmt = gimple_build_assign (new_temp, COND_EXPR, | |
1299 | val, true_val, false_val); | |
1300 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
1301 | val = new_temp; | |
1302 | } | |
1303 | } | |
1304 | else if (CONSTANT_CLASS_P (val)) | |
42fd8198 | 1305 | val = fold_convert (TREE_TYPE (type), val); |
418b7df3 RG |
1306 | else |
1307 | { | |
b731b390 | 1308 | new_temp = make_ssa_name (TREE_TYPE (type)); |
e412ece4 RB |
1309 | if (! INTEGRAL_TYPE_P (TREE_TYPE (val))) |
1310 | init_stmt = gimple_build_assign (new_temp, | |
1311 | fold_build1 (VIEW_CONVERT_EXPR, | |
1312 | TREE_TYPE (type), | |
1313 | val)); | |
1314 | else | |
1315 | init_stmt = gimple_build_assign (new_temp, NOP_EXPR, val); | |
418b7df3 | 1316 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1317 | val = new_temp; |
418b7df3 RG |
1318 | } |
1319 | } | |
5467ee52 | 1320 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1321 | } |
1322 | ||
0e22bb5a RB |
1323 | new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_"); |
1324 | init_stmt = gimple_build_assign (new_temp, val); | |
418b7df3 | 1325 | vect_init_vector_1 (stmt, init_stmt, gsi); |
0e22bb5a | 1326 | return new_temp; |
ebfd146a IR |
1327 | } |
1328 | ||
c83a894c | 1329 | /* Function vect_get_vec_def_for_operand_1. |
a70d6342 | 1330 | |
c83a894c AH |
1331 | For a defining stmt DEF_STMT of a scalar stmt, return a vector def with type |
1332 | DT that will be used in the vectorized stmt. */ | |
ebfd146a IR |
1333 | |
1334 | tree | |
c83a894c | 1335 | vect_get_vec_def_for_operand_1 (gimple *def_stmt, enum vect_def_type dt) |
ebfd146a IR |
1336 | { |
1337 | tree vec_oprnd; | |
355fe088 | 1338 | gimple *vec_stmt; |
ebfd146a | 1339 | stmt_vec_info def_stmt_info = NULL; |
ebfd146a IR |
1340 | |
1341 | switch (dt) | |
1342 | { | |
81c40241 | 1343 | /* operand is a constant or a loop invariant. */ |
ebfd146a | 1344 | case vect_constant_def: |
81c40241 | 1345 | case vect_external_def: |
c83a894c AH |
1346 | /* Code should use vect_get_vec_def_for_operand. */ |
1347 | gcc_unreachable (); | |
ebfd146a | 1348 | |
81c40241 | 1349 | /* operand is defined inside the loop. */ |
8644a673 | 1350 | case vect_internal_def: |
ebfd146a | 1351 | { |
ebfd146a IR |
1352 | /* Get the def from the vectorized stmt. */ |
1353 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1354 | |
ebfd146a | 1355 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1356 | /* Get vectorized pattern statement. */ |
1357 | if (!vec_stmt | |
1358 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1359 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1360 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1361 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1362 | gcc_assert (vec_stmt); |
1363 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1364 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1365 | else if (is_gimple_call (vec_stmt)) | |
1366 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1367 | else | |
1368 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1369 | return vec_oprnd; | |
1370 | } | |
1371 | ||
c78e3652 | 1372 | /* operand is defined by a loop header phi. */ |
ebfd146a | 1373 | case vect_reduction_def: |
06066f92 | 1374 | case vect_double_reduction_def: |
7c5222ff | 1375 | case vect_nested_cycle: |
ebfd146a IR |
1376 | case vect_induction_def: |
1377 | { | |
1378 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1379 | ||
1380 | /* Get the def from the vectorized stmt. */ | |
1381 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1382 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1383 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1384 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1385 | else | |
1386 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1387 | return vec_oprnd; |
1388 | } | |
1389 | ||
1390 | default: | |
1391 | gcc_unreachable (); | |
1392 | } | |
1393 | } | |
1394 | ||
1395 | ||
c83a894c AH |
1396 | /* Function vect_get_vec_def_for_operand. |
1397 | ||
1398 | OP is an operand in STMT. This function returns a (vector) def that will be | |
1399 | used in the vectorized stmt for STMT. | |
1400 | ||
1401 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1402 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1403 | ||
1404 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1405 | needs to be introduced. VECTYPE may be used to specify a required type for | |
1406 | vector invariant. */ | |
1407 | ||
1408 | tree | |
1409 | vect_get_vec_def_for_operand (tree op, gimple *stmt, tree vectype) | |
1410 | { | |
1411 | gimple *def_stmt; | |
1412 | enum vect_def_type dt; | |
1413 | bool is_simple_use; | |
1414 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
1415 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
1416 | ||
1417 | if (dump_enabled_p ()) | |
1418 | { | |
1419 | dump_printf_loc (MSG_NOTE, vect_location, | |
1420 | "vect_get_vec_def_for_operand: "); | |
1421 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
1422 | dump_printf (MSG_NOTE, "\n"); | |
1423 | } | |
1424 | ||
1425 | is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt); | |
1426 | gcc_assert (is_simple_use); | |
1427 | if (def_stmt && dump_enabled_p ()) | |
1428 | { | |
1429 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1430 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
1431 | } | |
1432 | ||
1433 | if (dt == vect_constant_def || dt == vect_external_def) | |
1434 | { | |
1435 | tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo); | |
1436 | tree vector_type; | |
1437 | ||
1438 | if (vectype) | |
1439 | vector_type = vectype; | |
2568d8a1 | 1440 | else if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op)) |
c83a894c AH |
1441 | && VECTOR_BOOLEAN_TYPE_P (stmt_vectype)) |
1442 | vector_type = build_same_sized_truth_vector_type (stmt_vectype); | |
1443 | else | |
1444 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
1445 | ||
1446 | gcc_assert (vector_type); | |
1447 | return vect_init_vector (stmt, op, vector_type, NULL); | |
1448 | } | |
1449 | else | |
1450 | return vect_get_vec_def_for_operand_1 (def_stmt, dt); | |
1451 | } | |
1452 | ||
1453 | ||
ebfd146a IR |
1454 | /* Function vect_get_vec_def_for_stmt_copy |
1455 | ||
ff802fa1 | 1456 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1457 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1458 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1459 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1460 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1461 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1462 | DT is the type of the vector def VEC_OPRND. |
1463 | ||
1464 | Context: | |
1465 | In case the vectorization factor (VF) is bigger than the number | |
1466 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1467 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1468 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1469 | smallest data-type determines the VF, and as a result, when vectorizing |
1470 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1471 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1472 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1473 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1474 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1475 | ||
1476 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1477 | |
ebfd146a IR |
1478 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1479 | VS1.1: vx.1 = memref1 VS1.2 | |
1480 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1481 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1482 | |
1483 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1484 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1485 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1486 | VSnew.3: vz3 = vx.3 + ... | |
1487 | ||
1488 | The vectorization of S1 is explained in vectorizable_load. | |
1489 | The vectorization of S2: | |
b8698a0f L |
1490 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1491 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1492 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1493 | returns the vector-def 'vx.0'. |
1494 | ||
b8698a0f L |
1495 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1496 | function is called to get the relevant vector-def for each operand. It is | |
1497 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1498 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1499 | ||
b8698a0f L |
1500 | For example, to obtain the vector-def 'vx.1' in order to create the |
1501 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1502 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1503 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1504 | and return its def ('vx.1'). | |
1505 | Overall, to create the above sequence this function will be called 3 times: | |
1506 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1507 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1508 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1509 | ||
1510 | tree | |
1511 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1512 | { | |
355fe088 | 1513 | gimple *vec_stmt_for_operand; |
ebfd146a IR |
1514 | stmt_vec_info def_stmt_info; |
1515 | ||
1516 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1517 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1518 | return vec_oprnd; |
1519 | ||
1520 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1521 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1522 | gcc_assert (def_stmt_info); | |
1523 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1524 | gcc_assert (vec_stmt_for_operand); | |
ebfd146a IR |
1525 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) |
1526 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1527 | else | |
1528 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1529 | return vec_oprnd; | |
1530 | } | |
1531 | ||
1532 | ||
1533 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1534 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a | 1535 | |
c78e3652 | 1536 | void |
b8698a0f | 1537 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1538 | vec<tree> *vec_oprnds0, |
1539 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1540 | { |
9771b263 | 1541 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1542 | |
1543 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1544 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1545 | |
9771b263 | 1546 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1547 | { |
9771b263 | 1548 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1549 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1550 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1551 | } |
1552 | } | |
1553 | ||
1554 | ||
c78e3652 | 1555 | /* Get vectorized definitions for OP0 and OP1. */ |
ebfd146a | 1556 | |
c78e3652 | 1557 | void |
355fe088 | 1558 | vect_get_vec_defs (tree op0, tree op1, gimple *stmt, |
9771b263 DN |
1559 | vec<tree> *vec_oprnds0, |
1560 | vec<tree> *vec_oprnds1, | |
306b0c92 | 1561 | slp_tree slp_node) |
ebfd146a IR |
1562 | { |
1563 | if (slp_node) | |
d092494c IR |
1564 | { |
1565 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1566 | auto_vec<tree> ops (nops); |
1567 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1568 | |
9771b263 | 1569 | ops.quick_push (op0); |
d092494c | 1570 | if (op1) |
9771b263 | 1571 | ops.quick_push (op1); |
d092494c | 1572 | |
306b0c92 | 1573 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
d092494c | 1574 | |
37b5ec8f | 1575 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1576 | if (op1) |
37b5ec8f | 1577 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1578 | } |
ebfd146a IR |
1579 | else |
1580 | { | |
1581 | tree vec_oprnd; | |
1582 | ||
9771b263 | 1583 | vec_oprnds0->create (1); |
81c40241 | 1584 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 1585 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1586 | |
1587 | if (op1) | |
1588 | { | |
9771b263 | 1589 | vec_oprnds1->create (1); |
81c40241 | 1590 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 1591 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1592 | } |
1593 | } | |
1594 | } | |
1595 | ||
1596 | ||
1597 | /* Function vect_finish_stmt_generation. | |
1598 | ||
1599 | Insert a new stmt. */ | |
1600 | ||
1601 | void | |
355fe088 | 1602 | vect_finish_stmt_generation (gimple *stmt, gimple *vec_stmt, |
ebfd146a IR |
1603 | gimple_stmt_iterator *gsi) |
1604 | { | |
1605 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
310213d4 | 1606 | vec_info *vinfo = stmt_info->vinfo; |
ebfd146a IR |
1607 | |
1608 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1609 | ||
54e8e2c3 RG |
1610 | if (!gsi_end_p (*gsi) |
1611 | && gimple_has_mem_ops (vec_stmt)) | |
1612 | { | |
355fe088 | 1613 | gimple *at_stmt = gsi_stmt (*gsi); |
54e8e2c3 RG |
1614 | tree vuse = gimple_vuse (at_stmt); |
1615 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1616 | { | |
1617 | tree vdef = gimple_vdef (at_stmt); | |
1618 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1619 | /* If we have an SSA vuse and insert a store, update virtual | |
1620 | SSA form to avoid triggering the renamer. Do so only | |
1621 | if we can easily see all uses - which is what almost always | |
1622 | happens with the way vectorized stmts are inserted. */ | |
1623 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1624 | && ((is_gimple_assign (vec_stmt) | |
1625 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1626 | || (is_gimple_call (vec_stmt) | |
1627 | && !(gimple_call_flags (vec_stmt) | |
1628 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1629 | { | |
1630 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1631 | gimple_set_vdef (vec_stmt, new_vdef); | |
1632 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1633 | } | |
1634 | } | |
1635 | } | |
ebfd146a IR |
1636 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1637 | ||
310213d4 | 1638 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, vinfo)); |
ebfd146a | 1639 | |
73fbfcad | 1640 | if (dump_enabled_p ()) |
ebfd146a | 1641 | { |
78c60e3d SS |
1642 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); |
1643 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
ebfd146a IR |
1644 | } |
1645 | ||
ad885386 | 1646 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
8e91d222 JJ |
1647 | |
1648 | /* While EH edges will generally prevent vectorization, stmt might | |
1649 | e.g. be in a must-not-throw region. Ensure newly created stmts | |
1650 | that could throw are part of the same region. */ | |
1651 | int lp_nr = lookup_stmt_eh_lp (stmt); | |
1652 | if (lp_nr != 0 && stmt_could_throw_p (vec_stmt)) | |
1653 | add_stmt_to_eh_lp (vec_stmt, lp_nr); | |
ebfd146a IR |
1654 | } |
1655 | ||
70439f0d RS |
1656 | /* We want to vectorize a call to combined function CFN with function |
1657 | decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN | |
1658 | as the types of all inputs. Check whether this is possible using | |
1659 | an internal function, returning its code if so or IFN_LAST if not. */ | |
ebfd146a | 1660 | |
70439f0d RS |
1661 | static internal_fn |
1662 | vectorizable_internal_function (combined_fn cfn, tree fndecl, | |
1663 | tree vectype_out, tree vectype_in) | |
ebfd146a | 1664 | { |
70439f0d RS |
1665 | internal_fn ifn; |
1666 | if (internal_fn_p (cfn)) | |
1667 | ifn = as_internal_fn (cfn); | |
1668 | else | |
1669 | ifn = associated_internal_fn (fndecl); | |
1670 | if (ifn != IFN_LAST && direct_internal_fn_p (ifn)) | |
1671 | { | |
1672 | const direct_internal_fn_info &info = direct_internal_fn (ifn); | |
1673 | if (info.vectorizable) | |
1674 | { | |
1675 | tree type0 = (info.type0 < 0 ? vectype_out : vectype_in); | |
1676 | tree type1 = (info.type1 < 0 ? vectype_out : vectype_in); | |
d95ab70a RS |
1677 | if (direct_internal_fn_supported_p (ifn, tree_pair (type0, type1), |
1678 | OPTIMIZE_FOR_SPEED)) | |
70439f0d RS |
1679 | return ifn; |
1680 | } | |
1681 | } | |
1682 | return IFN_LAST; | |
ebfd146a IR |
1683 | } |
1684 | ||
5ce9450f | 1685 | |
355fe088 | 1686 | static tree permute_vec_elements (tree, tree, tree, gimple *, |
5ce9450f JJ |
1687 | gimple_stmt_iterator *); |
1688 | ||
62da9e14 RS |
1689 | /* STMT is a non-strided load or store, meaning that it accesses |
1690 | elements with a known constant step. Return -1 if that step | |
1691 | is negative, 0 if it is zero, and 1 if it is greater than zero. */ | |
1692 | ||
1693 | static int | |
1694 | compare_step_with_zero (gimple *stmt) | |
1695 | { | |
1696 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3f5e8a76 RS |
1697 | data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
1698 | return tree_int_cst_compare (vect_dr_behavior (dr)->step, | |
1699 | size_zero_node); | |
62da9e14 RS |
1700 | } |
1701 | ||
1702 | /* If the target supports a permute mask that reverses the elements in | |
1703 | a vector of type VECTYPE, return that mask, otherwise return null. */ | |
1704 | ||
1705 | static tree | |
1706 | perm_mask_for_reverse (tree vectype) | |
1707 | { | |
1708 | int i, nunits; | |
62da9e14 RS |
1709 | |
1710 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
62da9e14 | 1711 | |
908a1a16 | 1712 | auto_vec_perm_indices sel (nunits); |
62da9e14 | 1713 | for (i = 0; i < nunits; ++i) |
908a1a16 | 1714 | sel.quick_push (nunits - 1 - i); |
62da9e14 | 1715 | |
908a1a16 | 1716 | if (!can_vec_perm_p (TYPE_MODE (vectype), false, &sel)) |
62da9e14 RS |
1717 | return NULL_TREE; |
1718 | return vect_gen_perm_mask_checked (vectype, sel); | |
1719 | } | |
5ce9450f | 1720 | |
2de001ee RS |
1721 | /* A subroutine of get_load_store_type, with a subset of the same |
1722 | arguments. Handle the case where STMT is part of a grouped load | |
1723 | or store. | |
1724 | ||
1725 | For stores, the statements in the group are all consecutive | |
1726 | and there is no gap at the end. For loads, the statements in the | |
1727 | group might not be consecutive; there can be gaps between statements | |
1728 | as well as at the end. */ | |
1729 | ||
1730 | static bool | |
1731 | get_group_load_store_type (gimple *stmt, tree vectype, bool slp, | |
1732 | vec_load_store_type vls_type, | |
1733 | vect_memory_access_type *memory_access_type) | |
1734 | { | |
1735 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1736 | vec_info *vinfo = stmt_info->vinfo; | |
1737 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1738 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; | |
1739 | gimple *first_stmt = GROUP_FIRST_ELEMENT (stmt_info); | |
1740 | unsigned int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
1741 | bool single_element_p = (stmt == first_stmt | |
1742 | && !GROUP_NEXT_ELEMENT (stmt_info)); | |
1743 | unsigned HOST_WIDE_INT gap = GROUP_GAP (vinfo_for_stmt (first_stmt)); | |
522fcdd7 | 1744 | unsigned nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2de001ee RS |
1745 | |
1746 | /* True if the vectorized statements would access beyond the last | |
1747 | statement in the group. */ | |
1748 | bool overrun_p = false; | |
1749 | ||
1750 | /* True if we can cope with such overrun by peeling for gaps, so that | |
1751 | there is at least one final scalar iteration after the vector loop. */ | |
1752 | bool can_overrun_p = (vls_type == VLS_LOAD && loop_vinfo && !loop->inner); | |
1753 | ||
1754 | /* There can only be a gap at the end of the group if the stride is | |
1755 | known at compile time. */ | |
1756 | gcc_assert (!STMT_VINFO_STRIDED_P (stmt_info) || gap == 0); | |
1757 | ||
1758 | /* Stores can't yet have gaps. */ | |
1759 | gcc_assert (slp || vls_type == VLS_LOAD || gap == 0); | |
1760 | ||
1761 | if (slp) | |
1762 | { | |
1763 | if (STMT_VINFO_STRIDED_P (stmt_info)) | |
1764 | { | |
1765 | /* Try to use consecutive accesses of GROUP_SIZE elements, | |
1766 | separated by the stride, until we have a complete vector. | |
1767 | Fall back to scalar accesses if that isn't possible. */ | |
1768 | if (nunits % group_size == 0) | |
1769 | *memory_access_type = VMAT_STRIDED_SLP; | |
1770 | else | |
1771 | *memory_access_type = VMAT_ELEMENTWISE; | |
1772 | } | |
1773 | else | |
1774 | { | |
1775 | overrun_p = loop_vinfo && gap != 0; | |
1776 | if (overrun_p && vls_type != VLS_LOAD) | |
1777 | { | |
1778 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1779 | "Grouped store with gaps requires" | |
1780 | " non-consecutive accesses\n"); | |
1781 | return false; | |
1782 | } | |
f9ef2c76 RB |
1783 | /* If the access is aligned an overrun is fine. */ |
1784 | if (overrun_p | |
1785 | && aligned_access_p | |
1786 | (STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)))) | |
1787 | overrun_p = false; | |
2de001ee RS |
1788 | if (overrun_p && !can_overrun_p) |
1789 | { | |
1790 | if (dump_enabled_p ()) | |
1791 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1792 | "Peeling for outer loop is not supported\n"); | |
1793 | return false; | |
1794 | } | |
1795 | *memory_access_type = VMAT_CONTIGUOUS; | |
1796 | } | |
1797 | } | |
1798 | else | |
1799 | { | |
1800 | /* We can always handle this case using elementwise accesses, | |
1801 | but see if something more efficient is available. */ | |
1802 | *memory_access_type = VMAT_ELEMENTWISE; | |
1803 | ||
1804 | /* If there is a gap at the end of the group then these optimizations | |
1805 | would access excess elements in the last iteration. */ | |
1806 | bool would_overrun_p = (gap != 0); | |
522fcdd7 RB |
1807 | /* If the access is aligned an overrun is fine, but only if the |
1808 | overrun is not inside an unused vector (if the gap is as large | |
1809 | or larger than a vector). */ | |
f9ef2c76 | 1810 | if (would_overrun_p |
522fcdd7 RB |
1811 | && gap < nunits |
1812 | && aligned_access_p | |
1813 | (STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)))) | |
f9ef2c76 | 1814 | would_overrun_p = false; |
2de001ee | 1815 | if (!STMT_VINFO_STRIDED_P (stmt_info) |
62da9e14 RS |
1816 | && (can_overrun_p || !would_overrun_p) |
1817 | && compare_step_with_zero (stmt) > 0) | |
2de001ee RS |
1818 | { |
1819 | /* First try using LOAD/STORE_LANES. */ | |
1820 | if (vls_type == VLS_LOAD | |
1821 | ? vect_load_lanes_supported (vectype, group_size) | |
1822 | : vect_store_lanes_supported (vectype, group_size)) | |
1823 | { | |
1824 | *memory_access_type = VMAT_LOAD_STORE_LANES; | |
1825 | overrun_p = would_overrun_p; | |
1826 | } | |
1827 | ||
1828 | /* If that fails, try using permuting loads. */ | |
1829 | if (*memory_access_type == VMAT_ELEMENTWISE | |
1830 | && (vls_type == VLS_LOAD | |
1831 | ? vect_grouped_load_supported (vectype, single_element_p, | |
1832 | group_size) | |
1833 | : vect_grouped_store_supported (vectype, group_size))) | |
1834 | { | |
1835 | *memory_access_type = VMAT_CONTIGUOUS_PERMUTE; | |
1836 | overrun_p = would_overrun_p; | |
1837 | } | |
1838 | } | |
1839 | } | |
1840 | ||
1841 | if (vls_type != VLS_LOAD && first_stmt == stmt) | |
1842 | { | |
1843 | /* STMT is the leader of the group. Check the operands of all the | |
1844 | stmts of the group. */ | |
1845 | gimple *next_stmt = GROUP_NEXT_ELEMENT (stmt_info); | |
1846 | while (next_stmt) | |
1847 | { | |
1848 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
1849 | tree op = gimple_assign_rhs1 (next_stmt); | |
1850 | gimple *def_stmt; | |
1851 | enum vect_def_type dt; | |
1852 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt)) | |
1853 | { | |
1854 | if (dump_enabled_p ()) | |
1855 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1856 | "use not simple.\n"); | |
1857 | return false; | |
1858 | } | |
1859 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); | |
1860 | } | |
1861 | } | |
1862 | ||
1863 | if (overrun_p) | |
1864 | { | |
1865 | gcc_assert (can_overrun_p); | |
1866 | if (dump_enabled_p ()) | |
1867 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1868 | "Data access with gaps requires scalar " | |
1869 | "epilogue loop\n"); | |
1870 | LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true; | |
1871 | } | |
1872 | ||
1873 | return true; | |
1874 | } | |
1875 | ||
62da9e14 RS |
1876 | /* A subroutine of get_load_store_type, with a subset of the same |
1877 | arguments. Handle the case where STMT is a load or store that | |
1878 | accesses consecutive elements with a negative step. */ | |
1879 | ||
1880 | static vect_memory_access_type | |
1881 | get_negative_load_store_type (gimple *stmt, tree vectype, | |
1882 | vec_load_store_type vls_type, | |
1883 | unsigned int ncopies) | |
1884 | { | |
1885 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1886 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
1887 | dr_alignment_support alignment_support_scheme; | |
1888 | ||
1889 | if (ncopies > 1) | |
1890 | { | |
1891 | if (dump_enabled_p ()) | |
1892 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1893 | "multiple types with negative step.\n"); | |
1894 | return VMAT_ELEMENTWISE; | |
1895 | } | |
1896 | ||
1897 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
1898 | if (alignment_support_scheme != dr_aligned | |
1899 | && alignment_support_scheme != dr_unaligned_supported) | |
1900 | { | |
1901 | if (dump_enabled_p ()) | |
1902 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1903 | "negative step but alignment required.\n"); | |
1904 | return VMAT_ELEMENTWISE; | |
1905 | } | |
1906 | ||
1907 | if (vls_type == VLS_STORE_INVARIANT) | |
1908 | { | |
1909 | if (dump_enabled_p ()) | |
1910 | dump_printf_loc (MSG_NOTE, vect_location, | |
1911 | "negative step with invariant source;" | |
1912 | " no permute needed.\n"); | |
1913 | return VMAT_CONTIGUOUS_DOWN; | |
1914 | } | |
1915 | ||
1916 | if (!perm_mask_for_reverse (vectype)) | |
1917 | { | |
1918 | if (dump_enabled_p ()) | |
1919 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1920 | "negative step and reversing not supported.\n"); | |
1921 | return VMAT_ELEMENTWISE; | |
1922 | } | |
1923 | ||
1924 | return VMAT_CONTIGUOUS_REVERSE; | |
1925 | } | |
1926 | ||
2de001ee RS |
1927 | /* Analyze load or store statement STMT of type VLS_TYPE. Return true |
1928 | if there is a memory access type that the vectorized form can use, | |
1929 | storing it in *MEMORY_ACCESS_TYPE if so. If we decide to use gathers | |
1930 | or scatters, fill in GS_INFO accordingly. | |
1931 | ||
1932 | SLP says whether we're performing SLP rather than loop vectorization. | |
62da9e14 RS |
1933 | VECTYPE is the vector type that the vectorized statements will use. |
1934 | NCOPIES is the number of vector statements that will be needed. */ | |
2de001ee RS |
1935 | |
1936 | static bool | |
1937 | get_load_store_type (gimple *stmt, tree vectype, bool slp, | |
62da9e14 | 1938 | vec_load_store_type vls_type, unsigned int ncopies, |
2de001ee RS |
1939 | vect_memory_access_type *memory_access_type, |
1940 | gather_scatter_info *gs_info) | |
1941 | { | |
1942 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1943 | vec_info *vinfo = stmt_info->vinfo; | |
1944 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1945 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
1946 | { | |
1947 | *memory_access_type = VMAT_GATHER_SCATTER; | |
1948 | gimple *def_stmt; | |
1949 | if (!vect_check_gather_scatter (stmt, loop_vinfo, gs_info)) | |
1950 | gcc_unreachable (); | |
1951 | else if (!vect_is_simple_use (gs_info->offset, vinfo, &def_stmt, | |
1952 | &gs_info->offset_dt, | |
1953 | &gs_info->offset_vectype)) | |
1954 | { | |
1955 | if (dump_enabled_p ()) | |
1956 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1957 | "%s index use not simple.\n", | |
1958 | vls_type == VLS_LOAD ? "gather" : "scatter"); | |
1959 | return false; | |
1960 | } | |
1961 | } | |
1962 | else if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1963 | { | |
1964 | if (!get_group_load_store_type (stmt, vectype, slp, vls_type, | |
1965 | memory_access_type)) | |
1966 | return false; | |
1967 | } | |
1968 | else if (STMT_VINFO_STRIDED_P (stmt_info)) | |
1969 | { | |
1970 | gcc_assert (!slp); | |
1971 | *memory_access_type = VMAT_ELEMENTWISE; | |
1972 | } | |
1973 | else | |
62da9e14 RS |
1974 | { |
1975 | int cmp = compare_step_with_zero (stmt); | |
1976 | if (cmp < 0) | |
1977 | *memory_access_type = get_negative_load_store_type | |
1978 | (stmt, vectype, vls_type, ncopies); | |
1979 | else if (cmp == 0) | |
1980 | { | |
1981 | gcc_assert (vls_type == VLS_LOAD); | |
1982 | *memory_access_type = VMAT_INVARIANT; | |
1983 | } | |
1984 | else | |
1985 | *memory_access_type = VMAT_CONTIGUOUS; | |
1986 | } | |
2de001ee RS |
1987 | |
1988 | /* FIXME: At the moment the cost model seems to underestimate the | |
1989 | cost of using elementwise accesses. This check preserves the | |
1990 | traditional behavior until that can be fixed. */ | |
1991 | if (*memory_access_type == VMAT_ELEMENTWISE | |
1992 | && !STMT_VINFO_STRIDED_P (stmt_info)) | |
1993 | { | |
1994 | if (dump_enabled_p ()) | |
1995 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1996 | "not falling back to elementwise accesses\n"); | |
1997 | return false; | |
1998 | } | |
1999 | return true; | |
2000 | } | |
2001 | ||
5ce9450f JJ |
2002 | /* Function vectorizable_mask_load_store. |
2003 | ||
2004 | Check if STMT performs a conditional load or store that can be vectorized. | |
2005 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2006 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. | |
2007 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2008 | ||
2009 | static bool | |
355fe088 TS |
2010 | vectorizable_mask_load_store (gimple *stmt, gimple_stmt_iterator *gsi, |
2011 | gimple **vec_stmt, slp_tree slp_node) | |
5ce9450f JJ |
2012 | { |
2013 | tree vec_dest = NULL; | |
2014 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2015 | stmt_vec_info prev_stmt_info; | |
2016 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2017 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2018 | bool nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
2019 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
2020 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
57e2f6ad | 2021 | tree rhs_vectype = NULL_TREE; |
045c1278 | 2022 | tree mask_vectype; |
5ce9450f | 2023 | tree elem_type; |
355fe088 | 2024 | gimple *new_stmt; |
5ce9450f JJ |
2025 | tree dummy; |
2026 | tree dataref_ptr = NULL_TREE; | |
355fe088 | 2027 | gimple *ptr_incr; |
5ce9450f JJ |
2028 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2029 | int ncopies; | |
2030 | int i, j; | |
2031 | bool inv_p; | |
134c85ca | 2032 | gather_scatter_info gs_info; |
2de001ee | 2033 | vec_load_store_type vls_type; |
5ce9450f | 2034 | tree mask; |
355fe088 | 2035 | gimple *def_stmt; |
5ce9450f JJ |
2036 | enum vect_def_type dt; |
2037 | ||
2038 | if (slp_node != NULL) | |
2039 | return false; | |
2040 | ||
2041 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
2042 | gcc_assert (ncopies >= 1); | |
2043 | ||
5ce9450f | 2044 | mask = gimple_call_arg (stmt, 2); |
045c1278 | 2045 | |
2568d8a1 | 2046 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (mask))) |
5ce9450f JJ |
2047 | return false; |
2048 | ||
2049 | /* FORNOW. This restriction should be relaxed. */ | |
2050 | if (nested_in_vect_loop && ncopies > 1) | |
2051 | { | |
2052 | if (dump_enabled_p ()) | |
2053 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2054 | "multiple types in nested loop."); | |
2055 | return false; | |
2056 | } | |
2057 | ||
2058 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) | |
2059 | return false; | |
2060 | ||
66c16fd9 RB |
2061 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
2062 | && ! vec_stmt) | |
5ce9450f JJ |
2063 | return false; |
2064 | ||
2065 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
2066 | return false; | |
2067 | ||
2068 | elem_type = TREE_TYPE (vectype); | |
2069 | ||
045c1278 IE |
2070 | if (TREE_CODE (mask) != SSA_NAME) |
2071 | return false; | |
2072 | ||
2073 | if (!vect_is_simple_use (mask, loop_vinfo, &def_stmt, &dt, &mask_vectype)) | |
2074 | return false; | |
2075 | ||
2076 | if (!mask_vectype) | |
2077 | mask_vectype = get_mask_type_for_scalar_type (TREE_TYPE (vectype)); | |
2078 | ||
dc6a3147 IE |
2079 | if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype) |
2080 | || TYPE_VECTOR_SUBPARTS (mask_vectype) != TYPE_VECTOR_SUBPARTS (vectype)) | |
045c1278 IE |
2081 | return false; |
2082 | ||
2de001ee | 2083 | if (gimple_call_internal_fn (stmt) == IFN_MASK_STORE) |
57e2f6ad IE |
2084 | { |
2085 | tree rhs = gimple_call_arg (stmt, 3); | |
2086 | if (!vect_is_simple_use (rhs, loop_vinfo, &def_stmt, &dt, &rhs_vectype)) | |
2087 | return false; | |
2de001ee RS |
2088 | if (dt == vect_constant_def || dt == vect_external_def) |
2089 | vls_type = VLS_STORE_INVARIANT; | |
2090 | else | |
2091 | vls_type = VLS_STORE; | |
57e2f6ad | 2092 | } |
2de001ee RS |
2093 | else |
2094 | vls_type = VLS_LOAD; | |
57e2f6ad | 2095 | |
2de001ee | 2096 | vect_memory_access_type memory_access_type; |
62da9e14 | 2097 | if (!get_load_store_type (stmt, vectype, false, vls_type, ncopies, |
2de001ee RS |
2098 | &memory_access_type, &gs_info)) |
2099 | return false; | |
03b9e8e4 | 2100 | |
2de001ee RS |
2101 | if (memory_access_type == VMAT_GATHER_SCATTER) |
2102 | { | |
134c85ca | 2103 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); |
03b9e8e4 JJ |
2104 | tree masktype |
2105 | = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (TREE_CHAIN (arglist)))); | |
2106 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
2107 | { | |
2108 | if (dump_enabled_p ()) | |
2109 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2110 | "masked gather with integer mask not supported."); | |
2111 | return false; | |
2112 | } | |
5ce9450f | 2113 | } |
2de001ee RS |
2114 | else if (memory_access_type != VMAT_CONTIGUOUS) |
2115 | { | |
2116 | if (dump_enabled_p ()) | |
2117 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2118 | "unsupported access type for masked %s.\n", | |
2119 | vls_type == VLS_LOAD ? "load" : "store"); | |
2120 | return false; | |
2121 | } | |
5ce9450f | 2122 | else if (!VECTOR_MODE_P (TYPE_MODE (vectype)) |
045c1278 IE |
2123 | || !can_vec_mask_load_store_p (TYPE_MODE (vectype), |
2124 | TYPE_MODE (mask_vectype), | |
2de001ee | 2125 | vls_type == VLS_LOAD) |
57e2f6ad IE |
2126 | || (rhs_vectype |
2127 | && !useless_type_conversion_p (vectype, rhs_vectype))) | |
5ce9450f JJ |
2128 | return false; |
2129 | ||
5ce9450f JJ |
2130 | if (!vec_stmt) /* transformation not required. */ |
2131 | { | |
2de001ee | 2132 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; |
5ce9450f | 2133 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; |
2de001ee RS |
2134 | if (vls_type == VLS_LOAD) |
2135 | vect_model_load_cost (stmt_info, ncopies, memory_access_type, | |
2136 | NULL, NULL, NULL); | |
5ce9450f | 2137 | else |
2de001ee RS |
2138 | vect_model_store_cost (stmt_info, ncopies, memory_access_type, |
2139 | dt, NULL, NULL, NULL); | |
5ce9450f JJ |
2140 | return true; |
2141 | } | |
2de001ee | 2142 | gcc_assert (memory_access_type == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); |
5ce9450f | 2143 | |
67b8dbac | 2144 | /* Transform. */ |
5ce9450f | 2145 | |
2de001ee | 2146 | if (memory_access_type == VMAT_GATHER_SCATTER) |
5ce9450f JJ |
2147 | { |
2148 | tree vec_oprnd0 = NULL_TREE, op; | |
134c85ca | 2149 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); |
5ce9450f | 2150 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; |
acdcd61b | 2151 | tree ptr, vec_mask = NULL_TREE, mask_op = NULL_TREE, var, scale; |
5ce9450f | 2152 | tree perm_mask = NULL_TREE, prev_res = NULL_TREE; |
acdcd61b | 2153 | tree mask_perm_mask = NULL_TREE; |
5ce9450f JJ |
2154 | edge pe = loop_preheader_edge (loop); |
2155 | gimple_seq seq; | |
2156 | basic_block new_bb; | |
2157 | enum { NARROW, NONE, WIDEN } modifier; | |
134c85ca | 2158 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gs_info.offset_vectype); |
5ce9450f | 2159 | |
134c85ca | 2160 | rettype = TREE_TYPE (TREE_TYPE (gs_info.decl)); |
acdcd61b JJ |
2161 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
2162 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2163 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2164 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2165 | scaletype = TREE_VALUE (arglist); | |
2166 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
2167 | && types_compatible_p (srctype, masktype)); | |
2168 | ||
5ce9450f JJ |
2169 | if (nunits == gather_off_nunits) |
2170 | modifier = NONE; | |
2171 | else if (nunits == gather_off_nunits / 2) | |
2172 | { | |
5ce9450f JJ |
2173 | modifier = WIDEN; |
2174 | ||
908a1a16 | 2175 | auto_vec_perm_indices sel (gather_off_nunits); |
5ce9450f | 2176 | for (i = 0; i < gather_off_nunits; ++i) |
908a1a16 | 2177 | sel.quick_push (i | nunits); |
5ce9450f | 2178 | |
134c85ca | 2179 | perm_mask = vect_gen_perm_mask_checked (gs_info.offset_vectype, sel); |
5ce9450f JJ |
2180 | } |
2181 | else if (nunits == gather_off_nunits * 2) | |
2182 | { | |
5ce9450f JJ |
2183 | modifier = NARROW; |
2184 | ||
908a1a16 RS |
2185 | auto_vec_perm_indices sel (nunits); |
2186 | sel.quick_grow (nunits); | |
5ce9450f JJ |
2187 | for (i = 0; i < nunits; ++i) |
2188 | sel[i] = i < gather_off_nunits | |
2189 | ? i : i + nunits - gather_off_nunits; | |
2190 | ||
557be5a8 | 2191 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); |
5ce9450f | 2192 | ncopies *= 2; |
acdcd61b JJ |
2193 | for (i = 0; i < nunits; ++i) |
2194 | sel[i] = i | gather_off_nunits; | |
557be5a8 | 2195 | mask_perm_mask = vect_gen_perm_mask_checked (masktype, sel); |
5ce9450f JJ |
2196 | } |
2197 | else | |
2198 | gcc_unreachable (); | |
2199 | ||
5ce9450f JJ |
2200 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); |
2201 | ||
134c85ca | 2202 | ptr = fold_convert (ptrtype, gs_info.base); |
5ce9450f JJ |
2203 | if (!is_gimple_min_invariant (ptr)) |
2204 | { | |
2205 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
2206 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
2207 | gcc_assert (!new_bb); | |
2208 | } | |
2209 | ||
134c85ca | 2210 | scale = build_int_cst (scaletype, gs_info.scale); |
5ce9450f JJ |
2211 | |
2212 | prev_stmt_info = NULL; | |
2213 | for (j = 0; j < ncopies; ++j) | |
2214 | { | |
2215 | if (modifier == WIDEN && (j & 1)) | |
2216 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
2217 | perm_mask, stmt, gsi); | |
2218 | else if (j == 0) | |
2219 | op = vec_oprnd0 | |
134c85ca | 2220 | = vect_get_vec_def_for_operand (gs_info.offset, stmt); |
5ce9450f JJ |
2221 | else |
2222 | op = vec_oprnd0 | |
134c85ca | 2223 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, vec_oprnd0); |
5ce9450f JJ |
2224 | |
2225 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
2226 | { | |
2227 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
2228 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 2229 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
5ce9450f JJ |
2230 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
2231 | new_stmt | |
0d0e4a03 | 2232 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
2233 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2234 | op = var; | |
2235 | } | |
2236 | ||
acdcd61b JJ |
2237 | if (mask_perm_mask && (j & 1)) |
2238 | mask_op = permute_vec_elements (mask_op, mask_op, | |
2239 | mask_perm_mask, stmt, gsi); | |
5ce9450f JJ |
2240 | else |
2241 | { | |
acdcd61b | 2242 | if (j == 0) |
81c40241 | 2243 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); |
acdcd61b JJ |
2244 | else |
2245 | { | |
81c40241 | 2246 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
acdcd61b JJ |
2247 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2248 | } | |
5ce9450f | 2249 | |
acdcd61b JJ |
2250 | mask_op = vec_mask; |
2251 | if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask))) | |
2252 | { | |
2253 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)) | |
2254 | == TYPE_VECTOR_SUBPARTS (masktype)); | |
0e22bb5a | 2255 | var = vect_get_new_ssa_name (masktype, vect_simple_var); |
acdcd61b JJ |
2256 | mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op); |
2257 | new_stmt | |
0d0e4a03 | 2258 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, mask_op); |
acdcd61b JJ |
2259 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2260 | mask_op = var; | |
2261 | } | |
5ce9450f JJ |
2262 | } |
2263 | ||
2264 | new_stmt | |
134c85ca | 2265 | = gimple_build_call (gs_info.decl, 5, mask_op, ptr, op, mask_op, |
5ce9450f JJ |
2266 | scale); |
2267 | ||
2268 | if (!useless_type_conversion_p (vectype, rettype)) | |
2269 | { | |
2270 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
2271 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 2272 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
5ce9450f JJ |
2273 | gimple_call_set_lhs (new_stmt, op); |
2274 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 2275 | var = make_ssa_name (vec_dest); |
5ce9450f | 2276 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
0d0e4a03 | 2277 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
2278 | } |
2279 | else | |
2280 | { | |
2281 | var = make_ssa_name (vec_dest, new_stmt); | |
2282 | gimple_call_set_lhs (new_stmt, var); | |
2283 | } | |
2284 | ||
2285 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2286 | ||
2287 | if (modifier == NARROW) | |
2288 | { | |
2289 | if ((j & 1) == 0) | |
2290 | { | |
2291 | prev_res = var; | |
2292 | continue; | |
2293 | } | |
2294 | var = permute_vec_elements (prev_res, var, | |
2295 | perm_mask, stmt, gsi); | |
2296 | new_stmt = SSA_NAME_DEF_STMT (var); | |
2297 | } | |
2298 | ||
2299 | if (prev_stmt_info == NULL) | |
2300 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2301 | else | |
2302 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2303 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2304 | } | |
3efe2e2c JJ |
2305 | |
2306 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
2307 | from the IL. */ | |
e6f5c25d IE |
2308 | if (STMT_VINFO_RELATED_STMT (stmt_info)) |
2309 | { | |
2310 | stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
2311 | stmt_info = vinfo_for_stmt (stmt); | |
2312 | } | |
3efe2e2c JJ |
2313 | tree lhs = gimple_call_lhs (stmt); |
2314 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
2315 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
2316 | set_vinfo_for_stmt (stmt, NULL); | |
2317 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2318 | gsi_replace (gsi, new_stmt, true); | |
5ce9450f JJ |
2319 | return true; |
2320 | } | |
2de001ee | 2321 | else if (vls_type != VLS_LOAD) |
5ce9450f JJ |
2322 | { |
2323 | tree vec_rhs = NULL_TREE, vec_mask = NULL_TREE; | |
2324 | prev_stmt_info = NULL; | |
2d4dc223 | 2325 | LOOP_VINFO_HAS_MASK_STORE (loop_vinfo) = true; |
5ce9450f JJ |
2326 | for (i = 0; i < ncopies; i++) |
2327 | { | |
2328 | unsigned align, misalign; | |
2329 | ||
2330 | if (i == 0) | |
2331 | { | |
2332 | tree rhs = gimple_call_arg (stmt, 3); | |
81c40241 RB |
2333 | vec_rhs = vect_get_vec_def_for_operand (rhs, stmt); |
2334 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); | |
5ce9450f JJ |
2335 | /* We should have catched mismatched types earlier. */ |
2336 | gcc_assert (useless_type_conversion_p (vectype, | |
2337 | TREE_TYPE (vec_rhs))); | |
2338 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, | |
2339 | NULL_TREE, &dummy, gsi, | |
2340 | &ptr_incr, false, &inv_p); | |
2341 | gcc_assert (!inv_p); | |
2342 | } | |
2343 | else | |
2344 | { | |
81c40241 | 2345 | vect_is_simple_use (vec_rhs, loop_vinfo, &def_stmt, &dt); |
5ce9450f | 2346 | vec_rhs = vect_get_vec_def_for_stmt_copy (dt, vec_rhs); |
81c40241 | 2347 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2348 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2349 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2350 | TYPE_SIZE_UNIT (vectype)); | |
2351 | } | |
2352 | ||
2353 | align = TYPE_ALIGN_UNIT (vectype); | |
2354 | if (aligned_access_p (dr)) | |
2355 | misalign = 0; | |
2356 | else if (DR_MISALIGNMENT (dr) == -1) | |
2357 | { | |
2358 | align = TYPE_ALIGN_UNIT (elem_type); | |
2359 | misalign = 0; | |
2360 | } | |
2361 | else | |
2362 | misalign = DR_MISALIGNMENT (dr); | |
2363 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2364 | misalign); | |
08554c26 | 2365 | tree ptr = build_int_cst (TREE_TYPE (gimple_call_arg (stmt, 1)), |
146ec50f | 2366 | misalign ? least_bit_hwi (misalign) : align); |
a844293d | 2367 | gcall *call |
5ce9450f | 2368 | = gimple_build_call_internal (IFN_MASK_STORE, 4, dataref_ptr, |
08554c26 | 2369 | ptr, vec_mask, vec_rhs); |
a844293d RS |
2370 | gimple_call_set_nothrow (call, true); |
2371 | new_stmt = call; | |
5ce9450f JJ |
2372 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2373 | if (i == 0) | |
2374 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2375 | else | |
2376 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2377 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2378 | } | |
2379 | } | |
2380 | else | |
2381 | { | |
2382 | tree vec_mask = NULL_TREE; | |
2383 | prev_stmt_info = NULL; | |
2384 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); | |
2385 | for (i = 0; i < ncopies; i++) | |
2386 | { | |
2387 | unsigned align, misalign; | |
2388 | ||
2389 | if (i == 0) | |
2390 | { | |
81c40241 | 2391 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); |
5ce9450f JJ |
2392 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, |
2393 | NULL_TREE, &dummy, gsi, | |
2394 | &ptr_incr, false, &inv_p); | |
2395 | gcc_assert (!inv_p); | |
2396 | } | |
2397 | else | |
2398 | { | |
81c40241 | 2399 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2400 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2401 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2402 | TYPE_SIZE_UNIT (vectype)); | |
2403 | } | |
2404 | ||
2405 | align = TYPE_ALIGN_UNIT (vectype); | |
2406 | if (aligned_access_p (dr)) | |
2407 | misalign = 0; | |
2408 | else if (DR_MISALIGNMENT (dr) == -1) | |
2409 | { | |
2410 | align = TYPE_ALIGN_UNIT (elem_type); | |
2411 | misalign = 0; | |
2412 | } | |
2413 | else | |
2414 | misalign = DR_MISALIGNMENT (dr); | |
2415 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2416 | misalign); | |
08554c26 | 2417 | tree ptr = build_int_cst (TREE_TYPE (gimple_call_arg (stmt, 1)), |
146ec50f | 2418 | misalign ? least_bit_hwi (misalign) : align); |
a844293d | 2419 | gcall *call |
5ce9450f | 2420 | = gimple_build_call_internal (IFN_MASK_LOAD, 3, dataref_ptr, |
08554c26 | 2421 | ptr, vec_mask); |
a844293d RS |
2422 | gimple_call_set_lhs (call, make_ssa_name (vec_dest)); |
2423 | gimple_call_set_nothrow (call, true); | |
2424 | vect_finish_stmt_generation (stmt, call, gsi); | |
5ce9450f | 2425 | if (i == 0) |
a844293d | 2426 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = call; |
5ce9450f | 2427 | else |
a844293d RS |
2428 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = call; |
2429 | prev_stmt_info = vinfo_for_stmt (call); | |
5ce9450f JJ |
2430 | } |
2431 | } | |
2432 | ||
2de001ee | 2433 | if (vls_type == VLS_LOAD) |
3efe2e2c JJ |
2434 | { |
2435 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
2436 | from the IL. */ | |
e6f5c25d IE |
2437 | if (STMT_VINFO_RELATED_STMT (stmt_info)) |
2438 | { | |
2439 | stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
2440 | stmt_info = vinfo_for_stmt (stmt); | |
2441 | } | |
3efe2e2c JJ |
2442 | tree lhs = gimple_call_lhs (stmt); |
2443 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
2444 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
2445 | set_vinfo_for_stmt (stmt, NULL); | |
2446 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2447 | gsi_replace (gsi, new_stmt, true); | |
2448 | } | |
2449 | ||
5ce9450f JJ |
2450 | return true; |
2451 | } | |
2452 | ||
37b14185 RB |
2453 | /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64}. */ |
2454 | ||
2455 | static bool | |
2456 | vectorizable_bswap (gimple *stmt, gimple_stmt_iterator *gsi, | |
2457 | gimple **vec_stmt, slp_tree slp_node, | |
2458 | tree vectype_in, enum vect_def_type *dt) | |
2459 | { | |
2460 | tree op, vectype; | |
2461 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2462 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2463 | unsigned ncopies, nunits; | |
2464 | ||
2465 | op = gimple_call_arg (stmt, 0); | |
2466 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2467 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
2468 | ||
2469 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2470 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2471 | case of SLP. */ | |
2472 | if (slp_node) | |
2473 | ncopies = 1; | |
2474 | else | |
2475 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
2476 | ||
2477 | gcc_assert (ncopies >= 1); | |
2478 | ||
2479 | tree char_vectype = get_same_sized_vectype (char_type_node, vectype_in); | |
2480 | if (! char_vectype) | |
2481 | return false; | |
2482 | ||
794e3180 | 2483 | unsigned int num_bytes = TYPE_VECTOR_SUBPARTS (char_vectype); |
794e3180 | 2484 | unsigned word_bytes = num_bytes / nunits; |
908a1a16 RS |
2485 | |
2486 | auto_vec_perm_indices elts (num_bytes); | |
37b14185 RB |
2487 | for (unsigned i = 0; i < nunits; ++i) |
2488 | for (unsigned j = 0; j < word_bytes; ++j) | |
908a1a16 | 2489 | elts.quick_push ((i + 1) * word_bytes - j - 1); |
37b14185 | 2490 | |
908a1a16 | 2491 | if (! can_vec_perm_p (TYPE_MODE (char_vectype), false, &elts)) |
37b14185 RB |
2492 | return false; |
2493 | ||
2494 | if (! vec_stmt) | |
2495 | { | |
2496 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
2497 | if (dump_enabled_p ()) | |
2498 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_bswap ===" | |
2499 | "\n"); | |
2500 | if (! PURE_SLP_STMT (stmt_info)) | |
2501 | { | |
2502 | add_stmt_cost (stmt_info->vinfo->target_cost_data, | |
2503 | 1, vector_stmt, stmt_info, 0, vect_prologue); | |
2504 | add_stmt_cost (stmt_info->vinfo->target_cost_data, | |
2505 | ncopies, vec_perm, stmt_info, 0, vect_body); | |
2506 | } | |
2507 | return true; | |
2508 | } | |
2509 | ||
794e3180 RS |
2510 | auto_vec<tree, 32> telts (num_bytes); |
2511 | for (unsigned i = 0; i < num_bytes; ++i) | |
2512 | telts.quick_push (build_int_cst (char_type_node, elts[i])); | |
37b14185 RB |
2513 | tree bswap_vconst = build_vector (char_vectype, telts); |
2514 | ||
2515 | /* Transform. */ | |
2516 | vec<tree> vec_oprnds = vNULL; | |
2517 | gimple *new_stmt = NULL; | |
2518 | stmt_vec_info prev_stmt_info = NULL; | |
2519 | for (unsigned j = 0; j < ncopies; j++) | |
2520 | { | |
2521 | /* Handle uses. */ | |
2522 | if (j == 0) | |
306b0c92 | 2523 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); |
37b14185 RB |
2524 | else |
2525 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
2526 | ||
2527 | /* Arguments are ready. create the new vector stmt. */ | |
2528 | unsigned i; | |
2529 | tree vop; | |
2530 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) | |
2531 | { | |
2532 | tree tem = make_ssa_name (char_vectype); | |
2533 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
2534 | char_vectype, vop)); | |
2535 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2536 | tree tem2 = make_ssa_name (char_vectype); | |
2537 | new_stmt = gimple_build_assign (tem2, VEC_PERM_EXPR, | |
2538 | tem, tem, bswap_vconst); | |
2539 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2540 | tem = make_ssa_name (vectype); | |
2541 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
2542 | vectype, tem2)); | |
2543 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2544 | if (slp_node) | |
2545 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
2546 | } | |
2547 | ||
2548 | if (slp_node) | |
2549 | continue; | |
2550 | ||
2551 | if (j == 0) | |
2552 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2553 | else | |
2554 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2555 | ||
2556 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2557 | } | |
2558 | ||
2559 | vec_oprnds.release (); | |
2560 | return true; | |
2561 | } | |
2562 | ||
b1b6836e RS |
2563 | /* Return true if vector types VECTYPE_IN and VECTYPE_OUT have |
2564 | integer elements and if we can narrow VECTYPE_IN to VECTYPE_OUT | |
2565 | in a single step. On success, store the binary pack code in | |
2566 | *CONVERT_CODE. */ | |
2567 | ||
2568 | static bool | |
2569 | simple_integer_narrowing (tree vectype_out, tree vectype_in, | |
2570 | tree_code *convert_code) | |
2571 | { | |
2572 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_out)) | |
2573 | || !INTEGRAL_TYPE_P (TREE_TYPE (vectype_in))) | |
2574 | return false; | |
2575 | ||
2576 | tree_code code; | |
2577 | int multi_step_cvt = 0; | |
2578 | auto_vec <tree, 8> interm_types; | |
2579 | if (!supportable_narrowing_operation (NOP_EXPR, vectype_out, vectype_in, | |
2580 | &code, &multi_step_cvt, | |
2581 | &interm_types) | |
2582 | || multi_step_cvt) | |
2583 | return false; | |
2584 | ||
2585 | *convert_code = code; | |
2586 | return true; | |
2587 | } | |
5ce9450f | 2588 | |
ebfd146a IR |
2589 | /* Function vectorizable_call. |
2590 | ||
538dd0b7 | 2591 | Check if GS performs a function call that can be vectorized. |
b8698a0f | 2592 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
2593 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2594 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2595 | ||
2596 | static bool | |
355fe088 | 2597 | vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
190c2236 | 2598 | slp_tree slp_node) |
ebfd146a | 2599 | { |
538dd0b7 | 2600 | gcall *stmt; |
ebfd146a IR |
2601 | tree vec_dest; |
2602 | tree scalar_dest; | |
2603 | tree op, type; | |
2604 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
538dd0b7 | 2605 | stmt_vec_info stmt_info = vinfo_for_stmt (gs), prev_stmt_info; |
ebfd146a IR |
2606 | tree vectype_out, vectype_in; |
2607 | int nunits_in; | |
2608 | int nunits_out; | |
2609 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
190c2236 | 2610 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 2611 | vec_info *vinfo = stmt_info->vinfo; |
81c40241 | 2612 | tree fndecl, new_temp, rhs_type; |
355fe088 | 2613 | gimple *def_stmt; |
0502fb85 UB |
2614 | enum vect_def_type dt[3] |
2615 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 2616 | int ndts = 3; |
355fe088 | 2617 | gimple *new_stmt = NULL; |
ebfd146a | 2618 | int ncopies, j; |
6e1aa848 | 2619 | vec<tree> vargs = vNULL; |
ebfd146a IR |
2620 | enum { NARROW, NONE, WIDEN } modifier; |
2621 | size_t i, nargs; | |
9d5e7640 | 2622 | tree lhs; |
ebfd146a | 2623 | |
190c2236 | 2624 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2625 | return false; |
2626 | ||
66c16fd9 RB |
2627 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
2628 | && ! vec_stmt) | |
ebfd146a IR |
2629 | return false; |
2630 | ||
538dd0b7 DM |
2631 | /* Is GS a vectorizable call? */ |
2632 | stmt = dyn_cast <gcall *> (gs); | |
2633 | if (!stmt) | |
ebfd146a IR |
2634 | return false; |
2635 | ||
5ce9450f JJ |
2636 | if (gimple_call_internal_p (stmt) |
2637 | && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD | |
2638 | || gimple_call_internal_fn (stmt) == IFN_MASK_STORE)) | |
2639 | return vectorizable_mask_load_store (stmt, gsi, vec_stmt, | |
2640 | slp_node); | |
2641 | ||
0136f8f0 AH |
2642 | if (gimple_call_lhs (stmt) == NULL_TREE |
2643 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
ebfd146a IR |
2644 | return false; |
2645 | ||
0136f8f0 | 2646 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); |
5a2c1986 | 2647 | |
b690cc0f RG |
2648 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
2649 | ||
ebfd146a IR |
2650 | /* Process function arguments. */ |
2651 | rhs_type = NULL_TREE; | |
b690cc0f | 2652 | vectype_in = NULL_TREE; |
ebfd146a IR |
2653 | nargs = gimple_call_num_args (stmt); |
2654 | ||
1b1562a5 MM |
2655 | /* Bail out if the function has more than three arguments, we do not have |
2656 | interesting builtin functions to vectorize with more than two arguments | |
2657 | except for fma. No arguments is also not good. */ | |
2658 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
2659 | return false; |
2660 | ||
74bf76ed JJ |
2661 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
2662 | if (gimple_call_internal_p (stmt) | |
2663 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2664 | { | |
2665 | nargs = 0; | |
2666 | rhs_type = unsigned_type_node; | |
2667 | } | |
2668 | ||
ebfd146a IR |
2669 | for (i = 0; i < nargs; i++) |
2670 | { | |
b690cc0f RG |
2671 | tree opvectype; |
2672 | ||
ebfd146a IR |
2673 | op = gimple_call_arg (stmt, i); |
2674 | ||
2675 | /* We can only handle calls with arguments of the same type. */ | |
2676 | if (rhs_type | |
8533c9d8 | 2677 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 2678 | { |
73fbfcad | 2679 | if (dump_enabled_p ()) |
78c60e3d | 2680 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2681 | "argument types differ.\n"); |
ebfd146a IR |
2682 | return false; |
2683 | } | |
b690cc0f RG |
2684 | if (!rhs_type) |
2685 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 2686 | |
81c40241 | 2687 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[i], &opvectype)) |
ebfd146a | 2688 | { |
73fbfcad | 2689 | if (dump_enabled_p ()) |
78c60e3d | 2690 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2691 | "use not simple.\n"); |
ebfd146a IR |
2692 | return false; |
2693 | } | |
ebfd146a | 2694 | |
b690cc0f RG |
2695 | if (!vectype_in) |
2696 | vectype_in = opvectype; | |
2697 | else if (opvectype | |
2698 | && opvectype != vectype_in) | |
2699 | { | |
73fbfcad | 2700 | if (dump_enabled_p ()) |
78c60e3d | 2701 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2702 | "argument vector types differ.\n"); |
b690cc0f RG |
2703 | return false; |
2704 | } | |
2705 | } | |
2706 | /* If all arguments are external or constant defs use a vector type with | |
2707 | the same size as the output vector type. */ | |
ebfd146a | 2708 | if (!vectype_in) |
b690cc0f | 2709 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
2710 | if (vec_stmt) |
2711 | gcc_assert (vectype_in); | |
2712 | if (!vectype_in) | |
2713 | { | |
73fbfcad | 2714 | if (dump_enabled_p ()) |
7d8930a0 | 2715 | { |
78c60e3d SS |
2716 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2717 | "no vectype for scalar type "); | |
2718 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 2719 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
2720 | } |
2721 | ||
2722 | return false; | |
2723 | } | |
ebfd146a IR |
2724 | |
2725 | /* FORNOW */ | |
b690cc0f RG |
2726 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
2727 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
2728 | if (nunits_in == nunits_out / 2) |
2729 | modifier = NARROW; | |
2730 | else if (nunits_out == nunits_in) | |
2731 | modifier = NONE; | |
2732 | else if (nunits_out == nunits_in / 2) | |
2733 | modifier = WIDEN; | |
2734 | else | |
2735 | return false; | |
2736 | ||
70439f0d RS |
2737 | /* We only handle functions that do not read or clobber memory. */ |
2738 | if (gimple_vuse (stmt)) | |
2739 | { | |
2740 | if (dump_enabled_p ()) | |
2741 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2742 | "function reads from or writes to memory.\n"); | |
2743 | return false; | |
2744 | } | |
2745 | ||
ebfd146a IR |
2746 | /* For now, we only vectorize functions if a target specific builtin |
2747 | is available. TODO -- in some cases, it might be profitable to | |
2748 | insert the calls for pieces of the vector, in order to be able | |
2749 | to vectorize other operations in the loop. */ | |
70439f0d RS |
2750 | fndecl = NULL_TREE; |
2751 | internal_fn ifn = IFN_LAST; | |
2752 | combined_fn cfn = gimple_call_combined_fn (stmt); | |
2753 | tree callee = gimple_call_fndecl (stmt); | |
2754 | ||
2755 | /* First try using an internal function. */ | |
b1b6836e RS |
2756 | tree_code convert_code = ERROR_MARK; |
2757 | if (cfn != CFN_LAST | |
2758 | && (modifier == NONE | |
2759 | || (modifier == NARROW | |
2760 | && simple_integer_narrowing (vectype_out, vectype_in, | |
2761 | &convert_code)))) | |
70439f0d RS |
2762 | ifn = vectorizable_internal_function (cfn, callee, vectype_out, |
2763 | vectype_in); | |
2764 | ||
2765 | /* If that fails, try asking for a target-specific built-in function. */ | |
2766 | if (ifn == IFN_LAST) | |
2767 | { | |
2768 | if (cfn != CFN_LAST) | |
2769 | fndecl = targetm.vectorize.builtin_vectorized_function | |
2770 | (cfn, vectype_out, vectype_in); | |
2771 | else | |
2772 | fndecl = targetm.vectorize.builtin_md_vectorized_function | |
2773 | (callee, vectype_out, vectype_in); | |
2774 | } | |
2775 | ||
2776 | if (ifn == IFN_LAST && !fndecl) | |
ebfd146a | 2777 | { |
70439f0d | 2778 | if (cfn == CFN_GOMP_SIMD_LANE |
74bf76ed JJ |
2779 | && !slp_node |
2780 | && loop_vinfo | |
2781 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2782 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
2783 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2784 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
2785 | { | |
2786 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
2787 | { 0, 1, 2, ... vf - 1 } vector. */ | |
2788 | gcc_assert (nargs == 0); | |
2789 | } | |
37b14185 RB |
2790 | else if (modifier == NONE |
2791 | && (gimple_call_builtin_p (stmt, BUILT_IN_BSWAP16) | |
2792 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP32) | |
2793 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP64))) | |
2794 | return vectorizable_bswap (stmt, gsi, vec_stmt, slp_node, | |
2795 | vectype_in, dt); | |
74bf76ed JJ |
2796 | else |
2797 | { | |
2798 | if (dump_enabled_p ()) | |
2799 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 2800 | "function is not vectorizable.\n"); |
74bf76ed JJ |
2801 | return false; |
2802 | } | |
ebfd146a IR |
2803 | } |
2804 | ||
fce57248 | 2805 | if (slp_node) |
190c2236 | 2806 | ncopies = 1; |
b1b6836e | 2807 | else if (modifier == NARROW && ifn == IFN_LAST) |
ebfd146a IR |
2808 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; |
2809 | else | |
2810 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2811 | ||
2812 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
2813 | needs to be generated. */ | |
2814 | gcc_assert (ncopies >= 1); | |
2815 | ||
2816 | if (!vec_stmt) /* transformation not required. */ | |
2817 | { | |
2818 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 2819 | if (dump_enabled_p ()) |
e645e942 TJ |
2820 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ===" |
2821 | "\n"); | |
4fc5ebf1 | 2822 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); |
b1b6836e RS |
2823 | if (ifn != IFN_LAST && modifier == NARROW && !slp_node) |
2824 | add_stmt_cost (stmt_info->vinfo->target_cost_data, ncopies / 2, | |
2825 | vec_promote_demote, stmt_info, 0, vect_body); | |
2826 | ||
ebfd146a IR |
2827 | return true; |
2828 | } | |
2829 | ||
67b8dbac | 2830 | /* Transform. */ |
ebfd146a | 2831 | |
73fbfcad | 2832 | if (dump_enabled_p ()) |
e645e942 | 2833 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
2834 | |
2835 | /* Handle def. */ | |
2836 | scalar_dest = gimple_call_lhs (stmt); | |
2837 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
2838 | ||
2839 | prev_stmt_info = NULL; | |
b1b6836e | 2840 | if (modifier == NONE || ifn != IFN_LAST) |
ebfd146a | 2841 | { |
b1b6836e | 2842 | tree prev_res = NULL_TREE; |
ebfd146a IR |
2843 | for (j = 0; j < ncopies; ++j) |
2844 | { | |
2845 | /* Build argument list for the vectorized call. */ | |
2846 | if (j == 0) | |
9771b263 | 2847 | vargs.create (nargs); |
ebfd146a | 2848 | else |
9771b263 | 2849 | vargs.truncate (0); |
ebfd146a | 2850 | |
190c2236 JJ |
2851 | if (slp_node) |
2852 | { | |
ef062b13 | 2853 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2854 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2855 | |
2856 | for (i = 0; i < nargs; i++) | |
9771b263 | 2857 | vargs.quick_push (gimple_call_arg (stmt, i)); |
306b0c92 | 2858 | vect_get_slp_defs (vargs, slp_node, &vec_defs); |
37b5ec8f | 2859 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2860 | |
2861 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2862 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
2863 | { |
2864 | size_t k; | |
2865 | for (k = 0; k < nargs; k++) | |
2866 | { | |
37b5ec8f | 2867 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 2868 | vargs[k] = vec_oprndsk[i]; |
190c2236 | 2869 | } |
b1b6836e RS |
2870 | if (modifier == NARROW) |
2871 | { | |
2872 | tree half_res = make_ssa_name (vectype_in); | |
a844293d RS |
2873 | gcall *call |
2874 | = gimple_build_call_internal_vec (ifn, vargs); | |
2875 | gimple_call_set_lhs (call, half_res); | |
2876 | gimple_call_set_nothrow (call, true); | |
2877 | new_stmt = call; | |
b1b6836e RS |
2878 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2879 | if ((i & 1) == 0) | |
2880 | { | |
2881 | prev_res = half_res; | |
2882 | continue; | |
2883 | } | |
2884 | new_temp = make_ssa_name (vec_dest); | |
2885 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
2886 | prev_res, half_res); | |
2887 | } | |
70439f0d | 2888 | else |
b1b6836e | 2889 | { |
a844293d | 2890 | gcall *call; |
b1b6836e | 2891 | if (ifn != IFN_LAST) |
a844293d | 2892 | call = gimple_build_call_internal_vec (ifn, vargs); |
b1b6836e | 2893 | else |
a844293d RS |
2894 | call = gimple_build_call_vec (fndecl, vargs); |
2895 | new_temp = make_ssa_name (vec_dest, call); | |
2896 | gimple_call_set_lhs (call, new_temp); | |
2897 | gimple_call_set_nothrow (call, true); | |
2898 | new_stmt = call; | |
b1b6836e | 2899 | } |
190c2236 | 2900 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 2901 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2902 | } |
2903 | ||
2904 | for (i = 0; i < nargs; i++) | |
2905 | { | |
37b5ec8f | 2906 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2907 | vec_oprndsi.release (); |
190c2236 | 2908 | } |
190c2236 JJ |
2909 | continue; |
2910 | } | |
2911 | ||
ebfd146a IR |
2912 | for (i = 0; i < nargs; i++) |
2913 | { | |
2914 | op = gimple_call_arg (stmt, i); | |
2915 | if (j == 0) | |
2916 | vec_oprnd0 | |
81c40241 | 2917 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 2918 | else |
63827fb8 IR |
2919 | { |
2920 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
2921 | vec_oprnd0 | |
2922 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
2923 | } | |
ebfd146a | 2924 | |
9771b263 | 2925 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
2926 | } |
2927 | ||
74bf76ed JJ |
2928 | if (gimple_call_internal_p (stmt) |
2929 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2930 | { | |
794e3180 RS |
2931 | auto_vec<tree, 32> v (nunits_out); |
2932 | for (int k = 0; k < nunits_out; ++k) | |
2933 | v.quick_push (build_int_cst (unsigned_type_node, | |
2934 | j * nunits_out + k)); | |
74bf76ed JJ |
2935 | tree cst = build_vector (vectype_out, v); |
2936 | tree new_var | |
0e22bb5a | 2937 | = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_"); |
355fe088 | 2938 | gimple *init_stmt = gimple_build_assign (new_var, cst); |
74bf76ed | 2939 | vect_init_vector_1 (stmt, init_stmt, NULL); |
b731b390 | 2940 | new_temp = make_ssa_name (vec_dest); |
0e22bb5a | 2941 | new_stmt = gimple_build_assign (new_temp, new_var); |
74bf76ed | 2942 | } |
b1b6836e RS |
2943 | else if (modifier == NARROW) |
2944 | { | |
2945 | tree half_res = make_ssa_name (vectype_in); | |
a844293d RS |
2946 | gcall *call = gimple_build_call_internal_vec (ifn, vargs); |
2947 | gimple_call_set_lhs (call, half_res); | |
2948 | gimple_call_set_nothrow (call, true); | |
2949 | new_stmt = call; | |
b1b6836e RS |
2950 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2951 | if ((j & 1) == 0) | |
2952 | { | |
2953 | prev_res = half_res; | |
2954 | continue; | |
2955 | } | |
2956 | new_temp = make_ssa_name (vec_dest); | |
2957 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
2958 | prev_res, half_res); | |
2959 | } | |
74bf76ed JJ |
2960 | else |
2961 | { | |
a844293d | 2962 | gcall *call; |
70439f0d | 2963 | if (ifn != IFN_LAST) |
a844293d | 2964 | call = gimple_build_call_internal_vec (ifn, vargs); |
70439f0d | 2965 | else |
a844293d | 2966 | call = gimple_build_call_vec (fndecl, vargs); |
74bf76ed | 2967 | new_temp = make_ssa_name (vec_dest, new_stmt); |
a844293d RS |
2968 | gimple_call_set_lhs (call, new_temp); |
2969 | gimple_call_set_nothrow (call, true); | |
2970 | new_stmt = call; | |
74bf76ed | 2971 | } |
ebfd146a IR |
2972 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2973 | ||
b1b6836e | 2974 | if (j == (modifier == NARROW ? 1 : 0)) |
ebfd146a IR |
2975 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; |
2976 | else | |
2977 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2978 | ||
2979 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2980 | } | |
b1b6836e RS |
2981 | } |
2982 | else if (modifier == NARROW) | |
2983 | { | |
ebfd146a IR |
2984 | for (j = 0; j < ncopies; ++j) |
2985 | { | |
2986 | /* Build argument list for the vectorized call. */ | |
2987 | if (j == 0) | |
9771b263 | 2988 | vargs.create (nargs * 2); |
ebfd146a | 2989 | else |
9771b263 | 2990 | vargs.truncate (0); |
ebfd146a | 2991 | |
190c2236 JJ |
2992 | if (slp_node) |
2993 | { | |
ef062b13 | 2994 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2995 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2996 | |
2997 | for (i = 0; i < nargs; i++) | |
9771b263 | 2998 | vargs.quick_push (gimple_call_arg (stmt, i)); |
306b0c92 | 2999 | vect_get_slp_defs (vargs, slp_node, &vec_defs); |
37b5ec8f | 3000 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
3001 | |
3002 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3003 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
3004 | { |
3005 | size_t k; | |
9771b263 | 3006 | vargs.truncate (0); |
190c2236 JJ |
3007 | for (k = 0; k < nargs; k++) |
3008 | { | |
37b5ec8f | 3009 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
3010 | vargs.quick_push (vec_oprndsk[i]); |
3011 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 | 3012 | } |
a844293d | 3013 | gcall *call; |
70439f0d | 3014 | if (ifn != IFN_LAST) |
a844293d | 3015 | call = gimple_build_call_internal_vec (ifn, vargs); |
70439f0d | 3016 | else |
a844293d RS |
3017 | call = gimple_build_call_vec (fndecl, vargs); |
3018 | new_temp = make_ssa_name (vec_dest, call); | |
3019 | gimple_call_set_lhs (call, new_temp); | |
3020 | gimple_call_set_nothrow (call, true); | |
3021 | new_stmt = call; | |
190c2236 | 3022 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 3023 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
3024 | } |
3025 | ||
3026 | for (i = 0; i < nargs; i++) | |
3027 | { | |
37b5ec8f | 3028 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 3029 | vec_oprndsi.release (); |
190c2236 | 3030 | } |
190c2236 JJ |
3031 | continue; |
3032 | } | |
3033 | ||
ebfd146a IR |
3034 | for (i = 0; i < nargs; i++) |
3035 | { | |
3036 | op = gimple_call_arg (stmt, i); | |
3037 | if (j == 0) | |
3038 | { | |
3039 | vec_oprnd0 | |
81c40241 | 3040 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 3041 | vec_oprnd1 |
63827fb8 | 3042 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
3043 | } |
3044 | else | |
3045 | { | |
336ecb65 | 3046 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 3047 | vec_oprnd0 |
63827fb8 | 3048 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 3049 | vec_oprnd1 |
63827fb8 | 3050 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
3051 | } |
3052 | ||
9771b263 DN |
3053 | vargs.quick_push (vec_oprnd0); |
3054 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
3055 | } |
3056 | ||
b1b6836e | 3057 | new_stmt = gimple_build_call_vec (fndecl, vargs); |
ebfd146a IR |
3058 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3059 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
3060 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3061 | ||
3062 | if (j == 0) | |
3063 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3064 | else | |
3065 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3066 | ||
3067 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3068 | } | |
3069 | ||
3070 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a | 3071 | } |
b1b6836e RS |
3072 | else |
3073 | /* No current target implements this case. */ | |
3074 | return false; | |
ebfd146a | 3075 | |
9771b263 | 3076 | vargs.release (); |
ebfd146a | 3077 | |
ebfd146a IR |
3078 | /* The call in STMT might prevent it from being removed in dce. |
3079 | We however cannot remove it here, due to the way the ssa name | |
3080 | it defines is mapped to the new definition. So just replace | |
3081 | rhs of the statement with something harmless. */ | |
3082 | ||
dd34c087 JJ |
3083 | if (slp_node) |
3084 | return true; | |
3085 | ||
ebfd146a | 3086 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
3087 | if (is_pattern_stmt_p (stmt_info)) |
3088 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
3089 | else | |
3090 | lhs = gimple_call_lhs (stmt); | |
3cc2fa2a | 3091 | |
9d5e7640 | 3092 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); |
ebfd146a | 3093 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 3094 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
3095 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
3096 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
3097 | |
3098 | return true; | |
3099 | } | |
3100 | ||
3101 | ||
0136f8f0 AH |
3102 | struct simd_call_arg_info |
3103 | { | |
3104 | tree vectype; | |
3105 | tree op; | |
0136f8f0 | 3106 | HOST_WIDE_INT linear_step; |
34e82342 | 3107 | enum vect_def_type dt; |
0136f8f0 | 3108 | unsigned int align; |
17b658af | 3109 | bool simd_lane_linear; |
0136f8f0 AH |
3110 | }; |
3111 | ||
17b658af JJ |
3112 | /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME, |
3113 | is linear within simd lane (but not within whole loop), note it in | |
3114 | *ARGINFO. */ | |
3115 | ||
3116 | static void | |
3117 | vect_simd_lane_linear (tree op, struct loop *loop, | |
3118 | struct simd_call_arg_info *arginfo) | |
3119 | { | |
355fe088 | 3120 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
17b658af JJ |
3121 | |
3122 | if (!is_gimple_assign (def_stmt) | |
3123 | || gimple_assign_rhs_code (def_stmt) != POINTER_PLUS_EXPR | |
3124 | || !is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))) | |
3125 | return; | |
3126 | ||
3127 | tree base = gimple_assign_rhs1 (def_stmt); | |
3128 | HOST_WIDE_INT linear_step = 0; | |
3129 | tree v = gimple_assign_rhs2 (def_stmt); | |
3130 | while (TREE_CODE (v) == SSA_NAME) | |
3131 | { | |
3132 | tree t; | |
3133 | def_stmt = SSA_NAME_DEF_STMT (v); | |
3134 | if (is_gimple_assign (def_stmt)) | |
3135 | switch (gimple_assign_rhs_code (def_stmt)) | |
3136 | { | |
3137 | case PLUS_EXPR: | |
3138 | t = gimple_assign_rhs2 (def_stmt); | |
3139 | if (linear_step || TREE_CODE (t) != INTEGER_CST) | |
3140 | return; | |
3141 | base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t); | |
3142 | v = gimple_assign_rhs1 (def_stmt); | |
3143 | continue; | |
3144 | case MULT_EXPR: | |
3145 | t = gimple_assign_rhs2 (def_stmt); | |
3146 | if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t)) | |
3147 | return; | |
3148 | linear_step = tree_to_shwi (t); | |
3149 | v = gimple_assign_rhs1 (def_stmt); | |
3150 | continue; | |
3151 | CASE_CONVERT: | |
3152 | t = gimple_assign_rhs1 (def_stmt); | |
3153 | if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE | |
3154 | || (TYPE_PRECISION (TREE_TYPE (v)) | |
3155 | < TYPE_PRECISION (TREE_TYPE (t)))) | |
3156 | return; | |
3157 | if (!linear_step) | |
3158 | linear_step = 1; | |
3159 | v = t; | |
3160 | continue; | |
3161 | default: | |
3162 | return; | |
3163 | } | |
8e4284d0 | 3164 | else if (gimple_call_internal_p (def_stmt, IFN_GOMP_SIMD_LANE) |
17b658af JJ |
3165 | && loop->simduid |
3166 | && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME | |
3167 | && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0)) | |
3168 | == loop->simduid)) | |
3169 | { | |
3170 | if (!linear_step) | |
3171 | linear_step = 1; | |
3172 | arginfo->linear_step = linear_step; | |
3173 | arginfo->op = base; | |
3174 | arginfo->simd_lane_linear = true; | |
3175 | return; | |
3176 | } | |
3177 | } | |
3178 | } | |
3179 | ||
0136f8f0 AH |
3180 | /* Function vectorizable_simd_clone_call. |
3181 | ||
3182 | Check if STMT performs a function call that can be vectorized | |
3183 | by calling a simd clone of the function. | |
3184 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3185 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3186 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3187 | ||
3188 | static bool | |
355fe088 TS |
3189 | vectorizable_simd_clone_call (gimple *stmt, gimple_stmt_iterator *gsi, |
3190 | gimple **vec_stmt, slp_tree slp_node) | |
0136f8f0 AH |
3191 | { |
3192 | tree vec_dest; | |
3193 | tree scalar_dest; | |
3194 | tree op, type; | |
3195 | tree vec_oprnd0 = NULL_TREE; | |
3196 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
3197 | tree vectype; | |
3198 | unsigned int nunits; | |
3199 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3200 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
310213d4 | 3201 | vec_info *vinfo = stmt_info->vinfo; |
0136f8f0 | 3202 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
81c40241 | 3203 | tree fndecl, new_temp; |
355fe088 TS |
3204 | gimple *def_stmt; |
3205 | gimple *new_stmt = NULL; | |
0136f8f0 | 3206 | int ncopies, j; |
00426f9a | 3207 | auto_vec<simd_call_arg_info> arginfo; |
0136f8f0 AH |
3208 | vec<tree> vargs = vNULL; |
3209 | size_t i, nargs; | |
3210 | tree lhs, rtype, ratype; | |
3211 | vec<constructor_elt, va_gc> *ret_ctor_elts; | |
3212 | ||
3213 | /* Is STMT a vectorizable call? */ | |
3214 | if (!is_gimple_call (stmt)) | |
3215 | return false; | |
3216 | ||
3217 | fndecl = gimple_call_fndecl (stmt); | |
3218 | if (fndecl == NULL_TREE) | |
3219 | return false; | |
3220 | ||
d52f5295 | 3221 | struct cgraph_node *node = cgraph_node::get (fndecl); |
0136f8f0 AH |
3222 | if (node == NULL || node->simd_clones == NULL) |
3223 | return false; | |
3224 | ||
3225 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
3226 | return false; | |
3227 | ||
66c16fd9 RB |
3228 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3229 | && ! vec_stmt) | |
0136f8f0 AH |
3230 | return false; |
3231 | ||
3232 | if (gimple_call_lhs (stmt) | |
3233 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
3234 | return false; | |
3235 | ||
3236 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); | |
3237 | ||
3238 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3239 | ||
3240 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt)) | |
3241 | return false; | |
3242 | ||
3243 | /* FORNOW */ | |
fce57248 | 3244 | if (slp_node) |
0136f8f0 AH |
3245 | return false; |
3246 | ||
3247 | /* Process function arguments. */ | |
3248 | nargs = gimple_call_num_args (stmt); | |
3249 | ||
3250 | /* Bail out if the function has zero arguments. */ | |
3251 | if (nargs == 0) | |
3252 | return false; | |
3253 | ||
00426f9a | 3254 | arginfo.reserve (nargs, true); |
0136f8f0 AH |
3255 | |
3256 | for (i = 0; i < nargs; i++) | |
3257 | { | |
3258 | simd_call_arg_info thisarginfo; | |
3259 | affine_iv iv; | |
3260 | ||
3261 | thisarginfo.linear_step = 0; | |
3262 | thisarginfo.align = 0; | |
3263 | thisarginfo.op = NULL_TREE; | |
17b658af | 3264 | thisarginfo.simd_lane_linear = false; |
0136f8f0 AH |
3265 | |
3266 | op = gimple_call_arg (stmt, i); | |
81c40241 RB |
3267 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &thisarginfo.dt, |
3268 | &thisarginfo.vectype) | |
0136f8f0 AH |
3269 | || thisarginfo.dt == vect_uninitialized_def) |
3270 | { | |
3271 | if (dump_enabled_p ()) | |
3272 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3273 | "use not simple.\n"); | |
0136f8f0 AH |
3274 | return false; |
3275 | } | |
3276 | ||
3277 | if (thisarginfo.dt == vect_constant_def | |
3278 | || thisarginfo.dt == vect_external_def) | |
3279 | gcc_assert (thisarginfo.vectype == NULL_TREE); | |
3280 | else | |
3281 | gcc_assert (thisarginfo.vectype != NULL_TREE); | |
3282 | ||
6c9e85fb JJ |
3283 | /* For linear arguments, the analyze phase should have saved |
3284 | the base and step in STMT_VINFO_SIMD_CLONE_INFO. */ | |
17b658af JJ |
3285 | if (i * 3 + 4 <= STMT_VINFO_SIMD_CLONE_INFO (stmt_info).length () |
3286 | && STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]) | |
6c9e85fb JJ |
3287 | { |
3288 | gcc_assert (vec_stmt); | |
3289 | thisarginfo.linear_step | |
17b658af | 3290 | = tree_to_shwi (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]); |
6c9e85fb | 3291 | thisarginfo.op |
17b658af JJ |
3292 | = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 1]; |
3293 | thisarginfo.simd_lane_linear | |
3294 | = (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 3] | |
3295 | == boolean_true_node); | |
6c9e85fb JJ |
3296 | /* If loop has been peeled for alignment, we need to adjust it. */ |
3297 | tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo); | |
3298 | tree n2 = LOOP_VINFO_NITERS (loop_vinfo); | |
17b658af | 3299 | if (n1 != n2 && !thisarginfo.simd_lane_linear) |
6c9e85fb JJ |
3300 | { |
3301 | tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2); | |
17b658af | 3302 | tree step = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]; |
6c9e85fb JJ |
3303 | tree opt = TREE_TYPE (thisarginfo.op); |
3304 | bias = fold_convert (TREE_TYPE (step), bias); | |
3305 | bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step); | |
3306 | thisarginfo.op | |
3307 | = fold_build2 (POINTER_TYPE_P (opt) | |
3308 | ? POINTER_PLUS_EXPR : PLUS_EXPR, opt, | |
3309 | thisarginfo.op, bias); | |
3310 | } | |
3311 | } | |
3312 | else if (!vec_stmt | |
3313 | && thisarginfo.dt != vect_constant_def | |
3314 | && thisarginfo.dt != vect_external_def | |
3315 | && loop_vinfo | |
3316 | && TREE_CODE (op) == SSA_NAME | |
3317 | && simple_iv (loop, loop_containing_stmt (stmt), op, | |
3318 | &iv, false) | |
3319 | && tree_fits_shwi_p (iv.step)) | |
0136f8f0 AH |
3320 | { |
3321 | thisarginfo.linear_step = tree_to_shwi (iv.step); | |
3322 | thisarginfo.op = iv.base; | |
3323 | } | |
3324 | else if ((thisarginfo.dt == vect_constant_def | |
3325 | || thisarginfo.dt == vect_external_def) | |
3326 | && POINTER_TYPE_P (TREE_TYPE (op))) | |
3327 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; | |
17b658af JJ |
3328 | /* Addresses of array elements indexed by GOMP_SIMD_LANE are |
3329 | linear too. */ | |
3330 | if (POINTER_TYPE_P (TREE_TYPE (op)) | |
3331 | && !thisarginfo.linear_step | |
3332 | && !vec_stmt | |
3333 | && thisarginfo.dt != vect_constant_def | |
3334 | && thisarginfo.dt != vect_external_def | |
3335 | && loop_vinfo | |
3336 | && !slp_node | |
3337 | && TREE_CODE (op) == SSA_NAME) | |
3338 | vect_simd_lane_linear (op, loop, &thisarginfo); | |
0136f8f0 AH |
3339 | |
3340 | arginfo.quick_push (thisarginfo); | |
3341 | } | |
3342 | ||
3343 | unsigned int badness = 0; | |
3344 | struct cgraph_node *bestn = NULL; | |
6c9e85fb JJ |
3345 | if (STMT_VINFO_SIMD_CLONE_INFO (stmt_info).exists ()) |
3346 | bestn = cgraph_node::get (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[0]); | |
0136f8f0 AH |
3347 | else |
3348 | for (struct cgraph_node *n = node->simd_clones; n != NULL; | |
3349 | n = n->simdclone->next_clone) | |
3350 | { | |
3351 | unsigned int this_badness = 0; | |
3352 | if (n->simdclone->simdlen | |
3353 | > (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
3354 | || n->simdclone->nargs != nargs) | |
3355 | continue; | |
3356 | if (n->simdclone->simdlen | |
3357 | < (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
3358 | this_badness += (exact_log2 (LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
3359 | - exact_log2 (n->simdclone->simdlen)) * 1024; | |
3360 | if (n->simdclone->inbranch) | |
3361 | this_badness += 2048; | |
3362 | int target_badness = targetm.simd_clone.usable (n); | |
3363 | if (target_badness < 0) | |
3364 | continue; | |
3365 | this_badness += target_badness * 512; | |
3366 | /* FORNOW: Have to add code to add the mask argument. */ | |
3367 | if (n->simdclone->inbranch) | |
3368 | continue; | |
3369 | for (i = 0; i < nargs; i++) | |
3370 | { | |
3371 | switch (n->simdclone->args[i].arg_type) | |
3372 | { | |
3373 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3374 | if (!useless_type_conversion_p | |
3375 | (n->simdclone->args[i].orig_type, | |
3376 | TREE_TYPE (gimple_call_arg (stmt, i)))) | |
3377 | i = -1; | |
3378 | else if (arginfo[i].dt == vect_constant_def | |
3379 | || arginfo[i].dt == vect_external_def | |
3380 | || arginfo[i].linear_step) | |
3381 | this_badness += 64; | |
3382 | break; | |
3383 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3384 | if (arginfo[i].dt != vect_constant_def | |
3385 | && arginfo[i].dt != vect_external_def) | |
3386 | i = -1; | |
3387 | break; | |
3388 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
d9a6bd32 | 3389 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
3390 | if (arginfo[i].dt == vect_constant_def |
3391 | || arginfo[i].dt == vect_external_def | |
3392 | || (arginfo[i].linear_step | |
3393 | != n->simdclone->args[i].linear_step)) | |
3394 | i = -1; | |
3395 | break; | |
3396 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
d9a6bd32 JJ |
3397 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
3398 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
e01d41e5 JJ |
3399 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3400 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3401 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3402 | /* FORNOW */ |
3403 | i = -1; | |
3404 | break; | |
3405 | case SIMD_CLONE_ARG_TYPE_MASK: | |
3406 | gcc_unreachable (); | |
3407 | } | |
3408 | if (i == (size_t) -1) | |
3409 | break; | |
3410 | if (n->simdclone->args[i].alignment > arginfo[i].align) | |
3411 | { | |
3412 | i = -1; | |
3413 | break; | |
3414 | } | |
3415 | if (arginfo[i].align) | |
3416 | this_badness += (exact_log2 (arginfo[i].align) | |
3417 | - exact_log2 (n->simdclone->args[i].alignment)); | |
3418 | } | |
3419 | if (i == (size_t) -1) | |
3420 | continue; | |
3421 | if (bestn == NULL || this_badness < badness) | |
3422 | { | |
3423 | bestn = n; | |
3424 | badness = this_badness; | |
3425 | } | |
3426 | } | |
3427 | ||
3428 | if (bestn == NULL) | |
00426f9a | 3429 | return false; |
0136f8f0 AH |
3430 | |
3431 | for (i = 0; i < nargs; i++) | |
3432 | if ((arginfo[i].dt == vect_constant_def | |
3433 | || arginfo[i].dt == vect_external_def) | |
3434 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
3435 | { | |
3436 | arginfo[i].vectype | |
3437 | = get_vectype_for_scalar_type (TREE_TYPE (gimple_call_arg (stmt, | |
3438 | i))); | |
3439 | if (arginfo[i].vectype == NULL | |
3440 | || (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
3441 | > bestn->simdclone->simdlen)) | |
00426f9a | 3442 | return false; |
0136f8f0 AH |
3443 | } |
3444 | ||
3445 | fndecl = bestn->decl; | |
3446 | nunits = bestn->simdclone->simdlen; | |
3447 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
3448 | ||
3449 | /* If the function isn't const, only allow it in simd loops where user | |
3450 | has asserted that at least nunits consecutive iterations can be | |
3451 | performed using SIMD instructions. */ | |
3452 | if ((loop == NULL || (unsigned) loop->safelen < nunits) | |
3453 | && gimple_vuse (stmt)) | |
00426f9a | 3454 | return false; |
0136f8f0 AH |
3455 | |
3456 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
3457 | needs to be generated. */ | |
3458 | gcc_assert (ncopies >= 1); | |
3459 | ||
3460 | if (!vec_stmt) /* transformation not required. */ | |
3461 | { | |
6c9e85fb JJ |
3462 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (bestn->decl); |
3463 | for (i = 0; i < nargs; i++) | |
7adb26f2 JJ |
3464 | if ((bestn->simdclone->args[i].arg_type |
3465 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
3466 | || (bestn->simdclone->args[i].arg_type | |
3467 | == SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP)) | |
6c9e85fb | 3468 | { |
17b658af | 3469 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_grow_cleared (i * 3 |
6c9e85fb JJ |
3470 | + 1); |
3471 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (arginfo[i].op); | |
3472 | tree lst = POINTER_TYPE_P (TREE_TYPE (arginfo[i].op)) | |
3473 | ? size_type_node : TREE_TYPE (arginfo[i].op); | |
3474 | tree ls = build_int_cst (lst, arginfo[i].linear_step); | |
3475 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (ls); | |
17b658af JJ |
3476 | tree sll = arginfo[i].simd_lane_linear |
3477 | ? boolean_true_node : boolean_false_node; | |
3478 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (sll); | |
6c9e85fb | 3479 | } |
0136f8f0 AH |
3480 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; |
3481 | if (dump_enabled_p ()) | |
3482 | dump_printf_loc (MSG_NOTE, vect_location, | |
3483 | "=== vectorizable_simd_clone_call ===\n"); | |
3484 | /* vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); */ | |
0136f8f0 AH |
3485 | return true; |
3486 | } | |
3487 | ||
67b8dbac | 3488 | /* Transform. */ |
0136f8f0 AH |
3489 | |
3490 | if (dump_enabled_p ()) | |
3491 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); | |
3492 | ||
3493 | /* Handle def. */ | |
3494 | scalar_dest = gimple_call_lhs (stmt); | |
3495 | vec_dest = NULL_TREE; | |
3496 | rtype = NULL_TREE; | |
3497 | ratype = NULL_TREE; | |
3498 | if (scalar_dest) | |
3499 | { | |
3500 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3501 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); | |
3502 | if (TREE_CODE (rtype) == ARRAY_TYPE) | |
3503 | { | |
3504 | ratype = rtype; | |
3505 | rtype = TREE_TYPE (ratype); | |
3506 | } | |
3507 | } | |
3508 | ||
3509 | prev_stmt_info = NULL; | |
3510 | for (j = 0; j < ncopies; ++j) | |
3511 | { | |
3512 | /* Build argument list for the vectorized call. */ | |
3513 | if (j == 0) | |
3514 | vargs.create (nargs); | |
3515 | else | |
3516 | vargs.truncate (0); | |
3517 | ||
3518 | for (i = 0; i < nargs; i++) | |
3519 | { | |
3520 | unsigned int k, l, m, o; | |
3521 | tree atype; | |
3522 | op = gimple_call_arg (stmt, i); | |
3523 | switch (bestn->simdclone->args[i].arg_type) | |
3524 | { | |
3525 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3526 | atype = bestn->simdclone->args[i].vector_type; | |
3527 | o = nunits / TYPE_VECTOR_SUBPARTS (atype); | |
3528 | for (m = j * o; m < (j + 1) * o; m++) | |
3529 | { | |
3530 | if (TYPE_VECTOR_SUBPARTS (atype) | |
3531 | < TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)) | |
3532 | { | |
3533 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); | |
3534 | k = (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
3535 | / TYPE_VECTOR_SUBPARTS (atype)); | |
3536 | gcc_assert ((k & (k - 1)) == 0); | |
3537 | if (m == 0) | |
3538 | vec_oprnd0 | |
81c40241 | 3539 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3540 | else |
3541 | { | |
3542 | vec_oprnd0 = arginfo[i].op; | |
3543 | if ((m & (k - 1)) == 0) | |
3544 | vec_oprnd0 | |
3545 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3546 | vec_oprnd0); | |
3547 | } | |
3548 | arginfo[i].op = vec_oprnd0; | |
3549 | vec_oprnd0 | |
3550 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, | |
92e29a5e | 3551 | bitsize_int (prec), |
0136f8f0 AH |
3552 | bitsize_int ((m & (k - 1)) * prec)); |
3553 | new_stmt | |
b731b390 | 3554 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3555 | vec_oprnd0); |
3556 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3557 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3558 | } | |
3559 | else | |
3560 | { | |
3561 | k = (TYPE_VECTOR_SUBPARTS (atype) | |
3562 | / TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)); | |
3563 | gcc_assert ((k & (k - 1)) == 0); | |
3564 | vec<constructor_elt, va_gc> *ctor_elts; | |
3565 | if (k != 1) | |
3566 | vec_alloc (ctor_elts, k); | |
3567 | else | |
3568 | ctor_elts = NULL; | |
3569 | for (l = 0; l < k; l++) | |
3570 | { | |
3571 | if (m == 0 && l == 0) | |
3572 | vec_oprnd0 | |
81c40241 | 3573 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3574 | else |
3575 | vec_oprnd0 | |
3576 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3577 | arginfo[i].op); | |
3578 | arginfo[i].op = vec_oprnd0; | |
3579 | if (k == 1) | |
3580 | break; | |
3581 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, | |
3582 | vec_oprnd0); | |
3583 | } | |
3584 | if (k == 1) | |
3585 | vargs.safe_push (vec_oprnd0); | |
3586 | else | |
3587 | { | |
3588 | vec_oprnd0 = build_constructor (atype, ctor_elts); | |
3589 | new_stmt | |
b731b390 | 3590 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3591 | vec_oprnd0); |
3592 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3593 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3594 | } | |
3595 | } | |
3596 | } | |
3597 | break; | |
3598 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3599 | vargs.safe_push (op); | |
3600 | break; | |
3601 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
7adb26f2 | 3602 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
3603 | if (j == 0) |
3604 | { | |
3605 | gimple_seq stmts; | |
3606 | arginfo[i].op | |
3607 | = force_gimple_operand (arginfo[i].op, &stmts, true, | |
3608 | NULL_TREE); | |
3609 | if (stmts != NULL) | |
3610 | { | |
3611 | basic_block new_bb; | |
3612 | edge pe = loop_preheader_edge (loop); | |
3613 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
3614 | gcc_assert (!new_bb); | |
3615 | } | |
17b658af JJ |
3616 | if (arginfo[i].simd_lane_linear) |
3617 | { | |
3618 | vargs.safe_push (arginfo[i].op); | |
3619 | break; | |
3620 | } | |
b731b390 | 3621 | tree phi_res = copy_ssa_name (op); |
538dd0b7 | 3622 | gphi *new_phi = create_phi_node (phi_res, loop->header); |
0136f8f0 | 3623 | set_vinfo_for_stmt (new_phi, |
310213d4 | 3624 | new_stmt_vec_info (new_phi, loop_vinfo)); |
0136f8f0 AH |
3625 | add_phi_arg (new_phi, arginfo[i].op, |
3626 | loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
3627 | enum tree_code code | |
3628 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3629 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3630 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3631 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3632 | widest_int cst |
3633 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3634 | ncopies * nunits); | |
3635 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3636 | tree phi_arg = copy_ssa_name (op); |
0d0e4a03 JJ |
3637 | new_stmt |
3638 | = gimple_build_assign (phi_arg, code, phi_res, tcst); | |
0136f8f0 AH |
3639 | gimple_stmt_iterator si = gsi_after_labels (loop->header); |
3640 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); | |
3641 | set_vinfo_for_stmt (new_stmt, | |
310213d4 | 3642 | new_stmt_vec_info (new_stmt, loop_vinfo)); |
0136f8f0 AH |
3643 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), |
3644 | UNKNOWN_LOCATION); | |
3645 | arginfo[i].op = phi_res; | |
3646 | vargs.safe_push (phi_res); | |
3647 | } | |
3648 | else | |
3649 | { | |
3650 | enum tree_code code | |
3651 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3652 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3653 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3654 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3655 | widest_int cst |
3656 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3657 | j * nunits); | |
3658 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3659 | new_temp = make_ssa_name (TREE_TYPE (op)); |
0d0e4a03 JJ |
3660 | new_stmt = gimple_build_assign (new_temp, code, |
3661 | arginfo[i].op, tcst); | |
0136f8f0 AH |
3662 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3663 | vargs.safe_push (new_temp); | |
3664 | } | |
3665 | break; | |
7adb26f2 JJ |
3666 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
3667 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
0136f8f0 | 3668 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: |
e01d41e5 JJ |
3669 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3670 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3671 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3672 | default: |
3673 | gcc_unreachable (); | |
3674 | } | |
3675 | } | |
3676 | ||
3677 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
3678 | if (vec_dest) | |
3679 | { | |
3680 | gcc_assert (ratype || TYPE_VECTOR_SUBPARTS (rtype) == nunits); | |
3681 | if (ratype) | |
b731b390 | 3682 | new_temp = create_tmp_var (ratype); |
0136f8f0 AH |
3683 | else if (TYPE_VECTOR_SUBPARTS (vectype) |
3684 | == TYPE_VECTOR_SUBPARTS (rtype)) | |
3685 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3686 | else | |
3687 | new_temp = make_ssa_name (rtype, new_stmt); | |
3688 | gimple_call_set_lhs (new_stmt, new_temp); | |
3689 | } | |
3690 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3691 | ||
3692 | if (vec_dest) | |
3693 | { | |
3694 | if (TYPE_VECTOR_SUBPARTS (vectype) < nunits) | |
3695 | { | |
3696 | unsigned int k, l; | |
3697 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); | |
3698 | k = nunits / TYPE_VECTOR_SUBPARTS (vectype); | |
3699 | gcc_assert ((k & (k - 1)) == 0); | |
3700 | for (l = 0; l < k; l++) | |
3701 | { | |
3702 | tree t; | |
3703 | if (ratype) | |
3704 | { | |
3705 | t = build_fold_addr_expr (new_temp); | |
3706 | t = build2 (MEM_REF, vectype, t, | |
3707 | build_int_cst (TREE_TYPE (t), | |
3708 | l * prec / BITS_PER_UNIT)); | |
3709 | } | |
3710 | else | |
3711 | t = build3 (BIT_FIELD_REF, vectype, new_temp, | |
92e29a5e | 3712 | bitsize_int (prec), bitsize_int (l * prec)); |
0136f8f0 | 3713 | new_stmt |
b731b390 | 3714 | = gimple_build_assign (make_ssa_name (vectype), t); |
0136f8f0 AH |
3715 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3716 | if (j == 0 && l == 0) | |
3717 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3718 | else | |
3719 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3720 | ||
3721 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3722 | } | |
3723 | ||
3724 | if (ratype) | |
3725 | { | |
3726 | tree clobber = build_constructor (ratype, NULL); | |
3727 | TREE_THIS_VOLATILE (clobber) = 1; | |
3728 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3729 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3730 | } | |
3731 | continue; | |
3732 | } | |
3733 | else if (TYPE_VECTOR_SUBPARTS (vectype) > nunits) | |
3734 | { | |
3735 | unsigned int k = (TYPE_VECTOR_SUBPARTS (vectype) | |
3736 | / TYPE_VECTOR_SUBPARTS (rtype)); | |
3737 | gcc_assert ((k & (k - 1)) == 0); | |
3738 | if ((j & (k - 1)) == 0) | |
3739 | vec_alloc (ret_ctor_elts, k); | |
3740 | if (ratype) | |
3741 | { | |
3742 | unsigned int m, o = nunits / TYPE_VECTOR_SUBPARTS (rtype); | |
3743 | for (m = 0; m < o; m++) | |
3744 | { | |
3745 | tree tem = build4 (ARRAY_REF, rtype, new_temp, | |
3746 | size_int (m), NULL_TREE, NULL_TREE); | |
3747 | new_stmt | |
b731b390 | 3748 | = gimple_build_assign (make_ssa_name (rtype), tem); |
0136f8f0 AH |
3749 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3750 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, | |
3751 | gimple_assign_lhs (new_stmt)); | |
3752 | } | |
3753 | tree clobber = build_constructor (ratype, NULL); | |
3754 | TREE_THIS_VOLATILE (clobber) = 1; | |
3755 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3756 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3757 | } | |
3758 | else | |
3759 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); | |
3760 | if ((j & (k - 1)) != k - 1) | |
3761 | continue; | |
3762 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); | |
3763 | new_stmt | |
b731b390 | 3764 | = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0); |
0136f8f0 AH |
3765 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3766 | ||
3767 | if ((unsigned) j == k - 1) | |
3768 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3769 | else | |
3770 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3771 | ||
3772 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3773 | continue; | |
3774 | } | |
3775 | else if (ratype) | |
3776 | { | |
3777 | tree t = build_fold_addr_expr (new_temp); | |
3778 | t = build2 (MEM_REF, vectype, t, | |
3779 | build_int_cst (TREE_TYPE (t), 0)); | |
3780 | new_stmt | |
b731b390 | 3781 | = gimple_build_assign (make_ssa_name (vec_dest), t); |
0136f8f0 AH |
3782 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3783 | tree clobber = build_constructor (ratype, NULL); | |
3784 | TREE_THIS_VOLATILE (clobber) = 1; | |
3785 | vect_finish_stmt_generation (stmt, | |
3786 | gimple_build_assign (new_temp, | |
3787 | clobber), gsi); | |
3788 | } | |
3789 | } | |
3790 | ||
3791 | if (j == 0) | |
3792 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3793 | else | |
3794 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3795 | ||
3796 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3797 | } | |
3798 | ||
3799 | vargs.release (); | |
3800 | ||
3801 | /* The call in STMT might prevent it from being removed in dce. | |
3802 | We however cannot remove it here, due to the way the ssa name | |
3803 | it defines is mapped to the new definition. So just replace | |
3804 | rhs of the statement with something harmless. */ | |
3805 | ||
3806 | if (slp_node) | |
3807 | return true; | |
3808 | ||
3809 | if (scalar_dest) | |
3810 | { | |
3811 | type = TREE_TYPE (scalar_dest); | |
3812 | if (is_pattern_stmt_p (stmt_info)) | |
3813 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
3814 | else | |
3815 | lhs = gimple_call_lhs (stmt); | |
3816 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
3817 | } | |
3818 | else | |
3819 | new_stmt = gimple_build_nop (); | |
3820 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
3821 | set_vinfo_for_stmt (stmt, NULL); | |
3822 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2865f32a | 3823 | gsi_replace (gsi, new_stmt, true); |
0136f8f0 AH |
3824 | unlink_stmt_vdef (stmt); |
3825 | ||
3826 | return true; | |
3827 | } | |
3828 | ||
3829 | ||
ebfd146a IR |
3830 | /* Function vect_gen_widened_results_half |
3831 | ||
3832 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 3833 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 3834 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
3835 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
3836 | needs to be created (DECL is a function-decl of a target-builtin). | |
3837 | STMT is the original scalar stmt that we are vectorizing. */ | |
3838 | ||
355fe088 | 3839 | static gimple * |
ebfd146a IR |
3840 | vect_gen_widened_results_half (enum tree_code code, |
3841 | tree decl, | |
3842 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
3843 | tree vec_dest, gimple_stmt_iterator *gsi, | |
355fe088 | 3844 | gimple *stmt) |
b8698a0f | 3845 | { |
355fe088 | 3846 | gimple *new_stmt; |
b8698a0f L |
3847 | tree new_temp; |
3848 | ||
3849 | /* Generate half of the widened result: */ | |
3850 | if (code == CALL_EXPR) | |
3851 | { | |
3852 | /* Target specific support */ | |
ebfd146a IR |
3853 | if (op_type == binary_op) |
3854 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
3855 | else | |
3856 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
3857 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3858 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
3859 | } |
3860 | else | |
ebfd146a | 3861 | { |
b8698a0f L |
3862 | /* Generic support */ |
3863 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
3864 | if (op_type != binary_op) |
3865 | vec_oprnd1 = NULL; | |
0d0e4a03 | 3866 | new_stmt = gimple_build_assign (vec_dest, code, vec_oprnd0, vec_oprnd1); |
ebfd146a IR |
3867 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3868 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 3869 | } |
ebfd146a IR |
3870 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3871 | ||
ebfd146a IR |
3872 | return new_stmt; |
3873 | } | |
3874 | ||
4a00c761 JJ |
3875 | |
3876 | /* Get vectorized definitions for loop-based vectorization. For the first | |
3877 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
3878 | scalar operand), and for the rest we get a copy with | |
3879 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
3880 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
3881 | The vectors are collected into VEC_OPRNDS. */ | |
3882 | ||
3883 | static void | |
355fe088 | 3884 | vect_get_loop_based_defs (tree *oprnd, gimple *stmt, enum vect_def_type dt, |
9771b263 | 3885 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
3886 | { |
3887 | tree vec_oprnd; | |
3888 | ||
3889 | /* Get first vector operand. */ | |
3890 | /* All the vector operands except the very first one (that is scalar oprnd) | |
3891 | are stmt copies. */ | |
3892 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
81c40241 | 3893 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt); |
4a00c761 JJ |
3894 | else |
3895 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
3896 | ||
9771b263 | 3897 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3898 | |
3899 | /* Get second vector operand. */ | |
3900 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 3901 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3902 | |
3903 | *oprnd = vec_oprnd; | |
3904 | ||
3905 | /* For conversion in multiple steps, continue to get operands | |
3906 | recursively. */ | |
3907 | if (multi_step_cvt) | |
3908 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
3909 | } | |
3910 | ||
3911 | ||
3912 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
3913 | For multi-step conversions store the resulting vectors and call the function | |
3914 | recursively. */ | |
3915 | ||
3916 | static void | |
9771b263 | 3917 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
355fe088 | 3918 | int multi_step_cvt, gimple *stmt, |
9771b263 | 3919 | vec<tree> vec_dsts, |
4a00c761 JJ |
3920 | gimple_stmt_iterator *gsi, |
3921 | slp_tree slp_node, enum tree_code code, | |
3922 | stmt_vec_info *prev_stmt_info) | |
3923 | { | |
3924 | unsigned int i; | |
3925 | tree vop0, vop1, new_tmp, vec_dest; | |
355fe088 | 3926 | gimple *new_stmt; |
4a00c761 JJ |
3927 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
3928 | ||
9771b263 | 3929 | vec_dest = vec_dsts.pop (); |
4a00c761 | 3930 | |
9771b263 | 3931 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
3932 | { |
3933 | /* Create demotion operation. */ | |
9771b263 DN |
3934 | vop0 = (*vec_oprnds)[i]; |
3935 | vop1 = (*vec_oprnds)[i + 1]; | |
0d0e4a03 | 3936 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
4a00c761 JJ |
3937 | new_tmp = make_ssa_name (vec_dest, new_stmt); |
3938 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
3939 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3940 | ||
3941 | if (multi_step_cvt) | |
3942 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 3943 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
3944 | else |
3945 | { | |
3946 | /* This is the last step of the conversion sequence. Store the | |
3947 | vectors in SLP_NODE or in vector info of the scalar statement | |
3948 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
3949 | if (slp_node) | |
9771b263 | 3950 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 3951 | else |
c689ce1e RB |
3952 | { |
3953 | if (!*prev_stmt_info) | |
3954 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3955 | else | |
3956 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
4a00c761 | 3957 | |
c689ce1e RB |
3958 | *prev_stmt_info = vinfo_for_stmt (new_stmt); |
3959 | } | |
4a00c761 JJ |
3960 | } |
3961 | } | |
3962 | ||
3963 | /* For multi-step demotion operations we first generate demotion operations | |
3964 | from the source type to the intermediate types, and then combine the | |
3965 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
3966 | type. */ | |
3967 | if (multi_step_cvt) | |
3968 | { | |
3969 | /* At each level of recursion we have half of the operands we had at the | |
3970 | previous level. */ | |
9771b263 | 3971 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
3972 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
3973 | stmt, vec_dsts, gsi, slp_node, | |
3974 | VEC_PACK_TRUNC_EXPR, | |
3975 | prev_stmt_info); | |
3976 | } | |
3977 | ||
9771b263 | 3978 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3979 | } |
3980 | ||
3981 | ||
3982 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
3983 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
3984 | the resulting vectors and call the function recursively. */ | |
3985 | ||
3986 | static void | |
9771b263 DN |
3987 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
3988 | vec<tree> *vec_oprnds1, | |
355fe088 | 3989 | gimple *stmt, tree vec_dest, |
4a00c761 JJ |
3990 | gimple_stmt_iterator *gsi, |
3991 | enum tree_code code1, | |
3992 | enum tree_code code2, tree decl1, | |
3993 | tree decl2, int op_type) | |
3994 | { | |
3995 | int i; | |
3996 | tree vop0, vop1, new_tmp1, new_tmp2; | |
355fe088 | 3997 | gimple *new_stmt1, *new_stmt2; |
6e1aa848 | 3998 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 3999 | |
9771b263 DN |
4000 | vec_tmp.create (vec_oprnds0->length () * 2); |
4001 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
4002 | { |
4003 | if (op_type == binary_op) | |
9771b263 | 4004 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
4005 | else |
4006 | vop1 = NULL_TREE; | |
4007 | ||
4008 | /* Generate the two halves of promotion operation. */ | |
4009 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
4010 | op_type, vec_dest, gsi, stmt); | |
4011 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
4012 | op_type, vec_dest, gsi, stmt); | |
4013 | if (is_gimple_call (new_stmt1)) | |
4014 | { | |
4015 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
4016 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
4017 | } | |
4018 | else | |
4019 | { | |
4020 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
4021 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
4022 | } | |
4023 | ||
4024 | /* Store the results for the next step. */ | |
9771b263 DN |
4025 | vec_tmp.quick_push (new_tmp1); |
4026 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
4027 | } |
4028 | ||
689eaba3 | 4029 | vec_oprnds0->release (); |
4a00c761 JJ |
4030 | *vec_oprnds0 = vec_tmp; |
4031 | } | |
4032 | ||
4033 | ||
b8698a0f L |
4034 | /* Check if STMT performs a conversion operation, that can be vectorized. |
4035 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 4036 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
4037 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
4038 | ||
4039 | static bool | |
355fe088 TS |
4040 | vectorizable_conversion (gimple *stmt, gimple_stmt_iterator *gsi, |
4041 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
4042 | { |
4043 | tree vec_dest; | |
4044 | tree scalar_dest; | |
4a00c761 | 4045 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
4046 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
4047 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4048 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4049 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 4050 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
4051 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
4052 | tree new_temp; | |
355fe088 | 4053 | gimple *def_stmt; |
ebfd146a | 4054 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 4055 | int ndts = 2; |
355fe088 | 4056 | gimple *new_stmt = NULL; |
ebfd146a IR |
4057 | stmt_vec_info prev_stmt_info; |
4058 | int nunits_in; | |
4059 | int nunits_out; | |
4060 | tree vectype_out, vectype_in; | |
4a00c761 JJ |
4061 | int ncopies, i, j; |
4062 | tree lhs_type, rhs_type; | |
ebfd146a | 4063 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
4064 | vec<tree> vec_oprnds0 = vNULL; |
4065 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 4066 | tree vop0; |
4a00c761 | 4067 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4068 | vec_info *vinfo = stmt_info->vinfo; |
4a00c761 | 4069 | int multi_step_cvt = 0; |
6e1aa848 | 4070 | vec<tree> interm_types = vNULL; |
4a00c761 JJ |
4071 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
4072 | int op_type; | |
4a00c761 | 4073 | unsigned short fltsz; |
ebfd146a IR |
4074 | |
4075 | /* Is STMT a vectorizable conversion? */ | |
4076 | ||
4a00c761 | 4077 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4078 | return false; |
4079 | ||
66c16fd9 RB |
4080 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4081 | && ! vec_stmt) | |
ebfd146a IR |
4082 | return false; |
4083 | ||
4084 | if (!is_gimple_assign (stmt)) | |
4085 | return false; | |
4086 | ||
4087 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4088 | return false; | |
4089 | ||
4090 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
4091 | if (!CONVERT_EXPR_CODE_P (code) |
4092 | && code != FIX_TRUNC_EXPR | |
4093 | && code != FLOAT_EXPR | |
4094 | && code != WIDEN_MULT_EXPR | |
4095 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
4096 | return false; |
4097 | ||
4a00c761 JJ |
4098 | op_type = TREE_CODE_LENGTH (code); |
4099 | ||
ebfd146a | 4100 | /* Check types of lhs and rhs. */ |
b690cc0f | 4101 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 4102 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
4103 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
4104 | ||
ebfd146a IR |
4105 | op0 = gimple_assign_rhs1 (stmt); |
4106 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
4107 | |
4108 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4109 | && !((INTEGRAL_TYPE_P (lhs_type) | |
4110 | && INTEGRAL_TYPE_P (rhs_type)) | |
4111 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
4112 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
4113 | return false; | |
4114 | ||
e6f5c25d IE |
4115 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4116 | && ((INTEGRAL_TYPE_P (lhs_type) | |
2be65d9e | 4117 | && !type_has_mode_precision_p (lhs_type)) |
e6f5c25d | 4118 | || (INTEGRAL_TYPE_P (rhs_type) |
2be65d9e | 4119 | && !type_has_mode_precision_p (rhs_type)))) |
4a00c761 | 4120 | { |
73fbfcad | 4121 | if (dump_enabled_p ()) |
78c60e3d | 4122 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
4123 | "type conversion to/from bit-precision unsupported." |
4124 | "\n"); | |
4a00c761 JJ |
4125 | return false; |
4126 | } | |
4127 | ||
b690cc0f | 4128 | /* Check the operands of the operation. */ |
81c40241 | 4129 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype_in)) |
b690cc0f | 4130 | { |
73fbfcad | 4131 | if (dump_enabled_p ()) |
78c60e3d | 4132 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4133 | "use not simple.\n"); |
b690cc0f RG |
4134 | return false; |
4135 | } | |
4a00c761 JJ |
4136 | if (op_type == binary_op) |
4137 | { | |
4138 | bool ok; | |
4139 | ||
4140 | op1 = gimple_assign_rhs2 (stmt); | |
4141 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
4142 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
4143 | OP1. */ | |
4144 | if (CONSTANT_CLASS_P (op0)) | |
81c40241 | 4145 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &vectype_in); |
4a00c761 | 4146 | else |
81c40241 | 4147 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1]); |
4a00c761 JJ |
4148 | |
4149 | if (!ok) | |
4150 | { | |
73fbfcad | 4151 | if (dump_enabled_p ()) |
78c60e3d | 4152 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4153 | "use not simple.\n"); |
4a00c761 JJ |
4154 | return false; |
4155 | } | |
4156 | } | |
4157 | ||
b690cc0f RG |
4158 | /* If op0 is an external or constant defs use a vector type of |
4159 | the same size as the output vector type. */ | |
ebfd146a | 4160 | if (!vectype_in) |
b690cc0f | 4161 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
4162 | if (vec_stmt) |
4163 | gcc_assert (vectype_in); | |
4164 | if (!vectype_in) | |
4165 | { | |
73fbfcad | 4166 | if (dump_enabled_p ()) |
4a00c761 | 4167 | { |
78c60e3d SS |
4168 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4169 | "no vectype for scalar type "); | |
4170 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 4171 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 4172 | } |
7d8930a0 IR |
4173 | |
4174 | return false; | |
4175 | } | |
ebfd146a | 4176 | |
e6f5c25d IE |
4177 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4178 | && !VECTOR_BOOLEAN_TYPE_P (vectype_in)) | |
4179 | { | |
4180 | if (dump_enabled_p ()) | |
4181 | { | |
4182 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
4183 | "can't convert between boolean and non " | |
4184 | "boolean vectors"); | |
4185 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
4186 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
4187 | } | |
4188 | ||
4189 | return false; | |
4190 | } | |
4191 | ||
b690cc0f RG |
4192 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
4193 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4a00c761 | 4194 | if (nunits_in < nunits_out) |
ebfd146a IR |
4195 | modifier = NARROW; |
4196 | else if (nunits_out == nunits_in) | |
4197 | modifier = NONE; | |
ebfd146a | 4198 | else |
4a00c761 | 4199 | modifier = WIDEN; |
ebfd146a | 4200 | |
ff802fa1 IR |
4201 | /* Multiple types in SLP are handled by creating the appropriate number of |
4202 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4203 | case of SLP. */ | |
fce57248 | 4204 | if (slp_node) |
ebfd146a | 4205 | ncopies = 1; |
4a00c761 JJ |
4206 | else if (modifier == NARROW) |
4207 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
4208 | else | |
4209 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
b8698a0f | 4210 | |
ebfd146a IR |
4211 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
4212 | needs to be generated. */ | |
4213 | gcc_assert (ncopies >= 1); | |
4214 | ||
16d22000 RS |
4215 | bool found_mode = false; |
4216 | scalar_mode lhs_mode = SCALAR_TYPE_MODE (lhs_type); | |
4217 | scalar_mode rhs_mode = SCALAR_TYPE_MODE (rhs_type); | |
4218 | opt_scalar_mode rhs_mode_iter; | |
b397965c | 4219 | |
ebfd146a | 4220 | /* Supportable by target? */ |
4a00c761 | 4221 | switch (modifier) |
ebfd146a | 4222 | { |
4a00c761 JJ |
4223 | case NONE: |
4224 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4225 | return false; | |
4226 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
4227 | &decl1, &code1)) | |
4228 | break; | |
4229 | /* FALLTHRU */ | |
4230 | unsupported: | |
73fbfcad | 4231 | if (dump_enabled_p ()) |
78c60e3d | 4232 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4233 | "conversion not supported by target.\n"); |
ebfd146a | 4234 | return false; |
ebfd146a | 4235 | |
4a00c761 JJ |
4236 | case WIDEN: |
4237 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
4238 | &code1, &code2, &multi_step_cvt, |
4239 | &interm_types)) | |
4a00c761 JJ |
4240 | { |
4241 | /* Binary widening operation can only be supported directly by the | |
4242 | architecture. */ | |
4243 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
4244 | break; | |
4245 | } | |
4246 | ||
4247 | if (code != FLOAT_EXPR | |
b397965c | 4248 | || GET_MODE_SIZE (lhs_mode) <= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4249 | goto unsupported; |
4250 | ||
b397965c | 4251 | fltsz = GET_MODE_SIZE (lhs_mode); |
16d22000 | 4252 | FOR_EACH_2XWIDER_MODE (rhs_mode_iter, rhs_mode) |
4a00c761 | 4253 | { |
16d22000 | 4254 | rhs_mode = rhs_mode_iter.require (); |
c94843d2 RS |
4255 | if (GET_MODE_SIZE (rhs_mode) > fltsz) |
4256 | break; | |
4257 | ||
4a00c761 JJ |
4258 | cvt_type |
4259 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4260 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4261 | if (cvt_type == NULL_TREE) | |
4262 | goto unsupported; | |
4263 | ||
4264 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4265 | { | |
4266 | if (!supportable_convert_operation (code, vectype_out, | |
4267 | cvt_type, &decl1, &codecvt1)) | |
4268 | goto unsupported; | |
4269 | } | |
4270 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
4271 | cvt_type, &codecvt1, |
4272 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
4273 | &interm_types)) |
4274 | continue; | |
4275 | else | |
4276 | gcc_assert (multi_step_cvt == 0); | |
4277 | ||
4278 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
4279 | vectype_in, &code1, &code2, |
4280 | &multi_step_cvt, &interm_types)) | |
16d22000 RS |
4281 | { |
4282 | found_mode = true; | |
4283 | break; | |
4284 | } | |
4a00c761 JJ |
4285 | } |
4286 | ||
16d22000 | 4287 | if (!found_mode) |
4a00c761 JJ |
4288 | goto unsupported; |
4289 | ||
4290 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4291 | codecvt2 = ERROR_MARK; | |
4292 | else | |
4293 | { | |
4294 | multi_step_cvt++; | |
9771b263 | 4295 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
4296 | cvt_type = NULL_TREE; |
4297 | } | |
4298 | break; | |
4299 | ||
4300 | case NARROW: | |
4301 | gcc_assert (op_type == unary_op); | |
4302 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
4303 | &code1, &multi_step_cvt, | |
4304 | &interm_types)) | |
4305 | break; | |
4306 | ||
4307 | if (code != FIX_TRUNC_EXPR | |
b397965c | 4308 | || GET_MODE_SIZE (lhs_mode) >= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4309 | goto unsupported; |
4310 | ||
4a00c761 JJ |
4311 | cvt_type |
4312 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4313 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4314 | if (cvt_type == NULL_TREE) | |
4315 | goto unsupported; | |
4316 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
4317 | &decl1, &codecvt1)) | |
4318 | goto unsupported; | |
4319 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
4320 | &code1, &multi_step_cvt, | |
4321 | &interm_types)) | |
4322 | break; | |
4323 | goto unsupported; | |
4324 | ||
4325 | default: | |
4326 | gcc_unreachable (); | |
ebfd146a IR |
4327 | } |
4328 | ||
4329 | if (!vec_stmt) /* transformation not required. */ | |
4330 | { | |
73fbfcad | 4331 | if (dump_enabled_p ()) |
78c60e3d | 4332 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4333 | "=== vectorizable_conversion ===\n"); |
4a00c761 | 4334 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
4335 | { |
4336 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
4fc5ebf1 | 4337 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); |
8bd37302 | 4338 | } |
4a00c761 JJ |
4339 | else if (modifier == NARROW) |
4340 | { | |
4341 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
8bd37302 | 4342 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 JJ |
4343 | } |
4344 | else | |
4345 | { | |
4346 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
8bd37302 | 4347 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 | 4348 | } |
9771b263 | 4349 | interm_types.release (); |
ebfd146a IR |
4350 | return true; |
4351 | } | |
4352 | ||
67b8dbac | 4353 | /* Transform. */ |
73fbfcad | 4354 | if (dump_enabled_p ()) |
78c60e3d | 4355 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4356 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 4357 | |
4a00c761 JJ |
4358 | if (op_type == binary_op) |
4359 | { | |
4360 | if (CONSTANT_CLASS_P (op0)) | |
4361 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
4362 | else if (CONSTANT_CLASS_P (op1)) | |
4363 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
4364 | } | |
4365 | ||
4366 | /* In case of multi-step conversion, we first generate conversion operations | |
4367 | to the intermediate types, and then from that types to the final one. | |
4368 | We create vector destinations for the intermediate type (TYPES) received | |
4369 | from supportable_*_operation, and store them in the correct order | |
4370 | for future use in vect_create_vectorized_*_stmts (). */ | |
8c681247 | 4371 | auto_vec<tree> vec_dsts (multi_step_cvt + 1); |
82294ec1 JJ |
4372 | vec_dest = vect_create_destination_var (scalar_dest, |
4373 | (cvt_type && modifier == WIDEN) | |
4374 | ? cvt_type : vectype_out); | |
9771b263 | 4375 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4376 | |
4377 | if (multi_step_cvt) | |
4378 | { | |
9771b263 DN |
4379 | for (i = interm_types.length () - 1; |
4380 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
4381 | { |
4382 | vec_dest = vect_create_destination_var (scalar_dest, | |
4383 | intermediate_type); | |
9771b263 | 4384 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4385 | } |
4386 | } | |
ebfd146a | 4387 | |
4a00c761 | 4388 | if (cvt_type) |
82294ec1 JJ |
4389 | vec_dest = vect_create_destination_var (scalar_dest, |
4390 | modifier == WIDEN | |
4391 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
4392 | |
4393 | if (!slp_node) | |
4394 | { | |
30862efc | 4395 | if (modifier == WIDEN) |
4a00c761 | 4396 | { |
c3284718 | 4397 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 4398 | if (op_type == binary_op) |
9771b263 | 4399 | vec_oprnds1.create (1); |
4a00c761 | 4400 | } |
30862efc | 4401 | else if (modifier == NARROW) |
9771b263 DN |
4402 | vec_oprnds0.create ( |
4403 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
4404 | } |
4405 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 4406 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 4407 | |
4a00c761 | 4408 | last_oprnd = op0; |
ebfd146a IR |
4409 | prev_stmt_info = NULL; |
4410 | switch (modifier) | |
4411 | { | |
4412 | case NONE: | |
4413 | for (j = 0; j < ncopies; j++) | |
4414 | { | |
ebfd146a | 4415 | if (j == 0) |
306b0c92 | 4416 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node); |
ebfd146a IR |
4417 | else |
4418 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
4419 | ||
9771b263 | 4420 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4421 | { |
4422 | /* Arguments are ready, create the new vector stmt. */ | |
4423 | if (code1 == CALL_EXPR) | |
4424 | { | |
4425 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4426 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4427 | gimple_call_set_lhs (new_stmt, new_temp); | |
4428 | } | |
4429 | else | |
4430 | { | |
4431 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
0d0e4a03 | 4432 | new_stmt = gimple_build_assign (vec_dest, code1, vop0); |
4a00c761 JJ |
4433 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4434 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4435 | } | |
4436 | ||
4437 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4438 | if (slp_node) | |
9771b263 | 4439 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
225ce44b RB |
4440 | else |
4441 | { | |
4442 | if (!prev_stmt_info) | |
4443 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4444 | else | |
4445 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4446 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4447 | } | |
4a00c761 | 4448 | } |
ebfd146a IR |
4449 | } |
4450 | break; | |
4451 | ||
4452 | case WIDEN: | |
4453 | /* In case the vectorization factor (VF) is bigger than the number | |
4454 | of elements that we can fit in a vectype (nunits), we have to | |
4455 | generate more than one vector stmt - i.e - we need to "unroll" | |
4456 | the vector stmt by a factor VF/nunits. */ | |
4457 | for (j = 0; j < ncopies; j++) | |
4458 | { | |
4a00c761 | 4459 | /* Handle uses. */ |
ebfd146a | 4460 | if (j == 0) |
4a00c761 JJ |
4461 | { |
4462 | if (slp_node) | |
4463 | { | |
4464 | if (code == WIDEN_LSHIFT_EXPR) | |
4465 | { | |
4466 | unsigned int k; | |
ebfd146a | 4467 | |
4a00c761 JJ |
4468 | vec_oprnd1 = op1; |
4469 | /* Store vec_oprnd1 for every vector stmt to be created | |
4470 | for SLP_NODE. We check during the analysis that all | |
4471 | the shift arguments are the same. */ | |
4472 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4473 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4474 | |
4475 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 4476 | slp_node); |
4a00c761 JJ |
4477 | } |
4478 | else | |
4479 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
306b0c92 | 4480 | &vec_oprnds1, slp_node); |
4a00c761 JJ |
4481 | } |
4482 | else | |
4483 | { | |
81c40241 | 4484 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 4485 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
4486 | if (op_type == binary_op) |
4487 | { | |
4488 | if (code == WIDEN_LSHIFT_EXPR) | |
4489 | vec_oprnd1 = op1; | |
4490 | else | |
81c40241 | 4491 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 4492 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4493 | } |
4494 | } | |
4495 | } | |
ebfd146a | 4496 | else |
4a00c761 JJ |
4497 | { |
4498 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
4499 | vec_oprnds0.truncate (0); |
4500 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
4501 | if (op_type == binary_op) |
4502 | { | |
4503 | if (code == WIDEN_LSHIFT_EXPR) | |
4504 | vec_oprnd1 = op1; | |
4505 | else | |
4506 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
4507 | vec_oprnd1); | |
9771b263 DN |
4508 | vec_oprnds1.truncate (0); |
4509 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
4510 | } |
4511 | } | |
ebfd146a | 4512 | |
4a00c761 JJ |
4513 | /* Arguments are ready. Create the new vector stmts. */ |
4514 | for (i = multi_step_cvt; i >= 0; i--) | |
4515 | { | |
9771b263 | 4516 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
4517 | enum tree_code c1 = code1, c2 = code2; |
4518 | if (i == 0 && codecvt2 != ERROR_MARK) | |
4519 | { | |
4520 | c1 = codecvt1; | |
4521 | c2 = codecvt2; | |
4522 | } | |
4523 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
4524 | &vec_oprnds1, | |
4525 | stmt, this_dest, gsi, | |
4526 | c1, c2, decl1, decl2, | |
4527 | op_type); | |
4528 | } | |
4529 | ||
9771b263 | 4530 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4531 | { |
4532 | if (cvt_type) | |
4533 | { | |
4534 | if (codecvt1 == CALL_EXPR) | |
4535 | { | |
4536 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4537 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4538 | gimple_call_set_lhs (new_stmt, new_temp); | |
4539 | } | |
4540 | else | |
4541 | { | |
4542 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4543 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4544 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4545 | vop0); | |
4a00c761 JJ |
4546 | } |
4547 | ||
4548 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4549 | } | |
4550 | else | |
4551 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
4552 | ||
4553 | if (slp_node) | |
9771b263 | 4554 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 4555 | else |
c689ce1e RB |
4556 | { |
4557 | if (!prev_stmt_info) | |
4558 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
4559 | else | |
4560 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4561 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4562 | } | |
4a00c761 | 4563 | } |
ebfd146a | 4564 | } |
4a00c761 JJ |
4565 | |
4566 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
4567 | break; |
4568 | ||
4569 | case NARROW: | |
4570 | /* In case the vectorization factor (VF) is bigger than the number | |
4571 | of elements that we can fit in a vectype (nunits), we have to | |
4572 | generate more than one vector stmt - i.e - we need to "unroll" | |
4573 | the vector stmt by a factor VF/nunits. */ | |
4574 | for (j = 0; j < ncopies; j++) | |
4575 | { | |
4576 | /* Handle uses. */ | |
4a00c761 JJ |
4577 | if (slp_node) |
4578 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 4579 | slp_node); |
ebfd146a IR |
4580 | else |
4581 | { | |
9771b263 | 4582 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
4583 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
4584 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
4585 | } |
4586 | ||
4a00c761 JJ |
4587 | /* Arguments are ready. Create the new vector stmts. */ |
4588 | if (cvt_type) | |
9771b263 | 4589 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4590 | { |
4591 | if (codecvt1 == CALL_EXPR) | |
4592 | { | |
4593 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4594 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4595 | gimple_call_set_lhs (new_stmt, new_temp); | |
4596 | } | |
4597 | else | |
4598 | { | |
4599 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4600 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4601 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4602 | vop0); | |
4a00c761 | 4603 | } |
ebfd146a | 4604 | |
4a00c761 | 4605 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 4606 | vec_oprnds0[i] = new_temp; |
4a00c761 | 4607 | } |
ebfd146a | 4608 | |
4a00c761 JJ |
4609 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
4610 | stmt, vec_dsts, gsi, | |
4611 | slp_node, code1, | |
4612 | &prev_stmt_info); | |
ebfd146a IR |
4613 | } |
4614 | ||
4615 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 4616 | break; |
ebfd146a IR |
4617 | } |
4618 | ||
9771b263 DN |
4619 | vec_oprnds0.release (); |
4620 | vec_oprnds1.release (); | |
9771b263 | 4621 | interm_types.release (); |
ebfd146a IR |
4622 | |
4623 | return true; | |
4624 | } | |
ff802fa1 IR |
4625 | |
4626 | ||
ebfd146a IR |
4627 | /* Function vectorizable_assignment. |
4628 | ||
b8698a0f L |
4629 | Check if STMT performs an assignment (copy) that can be vectorized. |
4630 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4631 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4632 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4633 | ||
4634 | static bool | |
355fe088 TS |
4635 | vectorizable_assignment (gimple *stmt, gimple_stmt_iterator *gsi, |
4636 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
4637 | { |
4638 | tree vec_dest; | |
4639 | tree scalar_dest; | |
4640 | tree op; | |
4641 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a IR |
4642 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4643 | tree new_temp; | |
355fe088 | 4644 | gimple *def_stmt; |
4fc5ebf1 JG |
4645 | enum vect_def_type dt[1] = {vect_unknown_def_type}; |
4646 | int ndts = 1; | |
ebfd146a | 4647 | int ncopies; |
f18b55bd | 4648 | int i, j; |
6e1aa848 | 4649 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 4650 | tree vop; |
a70d6342 | 4651 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4652 | vec_info *vinfo = stmt_info->vinfo; |
355fe088 | 4653 | gimple *new_stmt = NULL; |
f18b55bd | 4654 | stmt_vec_info prev_stmt_info = NULL; |
fde9c428 RG |
4655 | enum tree_code code; |
4656 | tree vectype_in; | |
ebfd146a | 4657 | |
a70d6342 | 4658 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4659 | return false; |
4660 | ||
66c16fd9 RB |
4661 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4662 | && ! vec_stmt) | |
ebfd146a IR |
4663 | return false; |
4664 | ||
4665 | /* Is vectorizable assignment? */ | |
4666 | if (!is_gimple_assign (stmt)) | |
4667 | return false; | |
4668 | ||
4669 | scalar_dest = gimple_assign_lhs (stmt); | |
4670 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4671 | return false; | |
4672 | ||
fde9c428 | 4673 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 4674 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
4675 | || code == PAREN_EXPR |
4676 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
4677 | op = gimple_assign_rhs1 (stmt); |
4678 | else | |
4679 | return false; | |
4680 | ||
7b7ec6c5 RG |
4681 | if (code == VIEW_CONVERT_EXPR) |
4682 | op = TREE_OPERAND (op, 0); | |
4683 | ||
465c8c19 JJ |
4684 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
4685 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4686 | ||
4687 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4688 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4689 | case of SLP. */ | |
fce57248 | 4690 | if (slp_node) |
465c8c19 JJ |
4691 | ncopies = 1; |
4692 | else | |
4693 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4694 | ||
4695 | gcc_assert (ncopies >= 1); | |
4696 | ||
81c40241 | 4697 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[0], &vectype_in)) |
ebfd146a | 4698 | { |
73fbfcad | 4699 | if (dump_enabled_p ()) |
78c60e3d | 4700 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4701 | "use not simple.\n"); |
ebfd146a IR |
4702 | return false; |
4703 | } | |
4704 | ||
fde9c428 RG |
4705 | /* We can handle NOP_EXPR conversions that do not change the number |
4706 | of elements or the vector size. */ | |
7b7ec6c5 RG |
4707 | if ((CONVERT_EXPR_CODE_P (code) |
4708 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 RG |
4709 | && (!vectype_in |
4710 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
4711 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
4712 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
4713 | return false; | |
4714 | ||
7b7b1813 RG |
4715 | /* We do not handle bit-precision changes. */ |
4716 | if ((CONVERT_EXPR_CODE_P (code) | |
4717 | || code == VIEW_CONVERT_EXPR) | |
4718 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
2be65d9e RS |
4719 | && (!type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
4720 | || !type_has_mode_precision_p (TREE_TYPE (op))) | |
7b7b1813 RG |
4721 | /* But a conversion that does not change the bit-pattern is ok. */ |
4722 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4723 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2dab46d5 IE |
4724 | && TYPE_UNSIGNED (TREE_TYPE (op))) |
4725 | /* Conversion between boolean types of different sizes is | |
4726 | a simple assignment in case their vectypes are same | |
4727 | boolean vectors. */ | |
4728 | && (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
4729 | || !VECTOR_BOOLEAN_TYPE_P (vectype_in))) | |
7b7b1813 | 4730 | { |
73fbfcad | 4731 | if (dump_enabled_p ()) |
78c60e3d SS |
4732 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4733 | "type conversion to/from bit-precision " | |
e645e942 | 4734 | "unsupported.\n"); |
7b7b1813 RG |
4735 | return false; |
4736 | } | |
4737 | ||
ebfd146a IR |
4738 | if (!vec_stmt) /* transformation not required. */ |
4739 | { | |
4740 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 4741 | if (dump_enabled_p ()) |
78c60e3d | 4742 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4743 | "=== vectorizable_assignment ===\n"); |
4fc5ebf1 | 4744 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); |
ebfd146a IR |
4745 | return true; |
4746 | } | |
4747 | ||
67b8dbac | 4748 | /* Transform. */ |
73fbfcad | 4749 | if (dump_enabled_p ()) |
e645e942 | 4750 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
4751 | |
4752 | /* Handle def. */ | |
4753 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4754 | ||
4755 | /* Handle use. */ | |
f18b55bd | 4756 | for (j = 0; j < ncopies; j++) |
ebfd146a | 4757 | { |
f18b55bd IR |
4758 | /* Handle uses. */ |
4759 | if (j == 0) | |
306b0c92 | 4760 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); |
f18b55bd IR |
4761 | else |
4762 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
4763 | ||
4764 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 4765 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 4766 | { |
7b7ec6c5 RG |
4767 | if (CONVERT_EXPR_CODE_P (code) |
4768 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 4769 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
4770 | new_stmt = gimple_build_assign (vec_dest, vop); |
4771 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4772 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4773 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4774 | if (slp_node) | |
9771b263 | 4775 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 4776 | } |
ebfd146a IR |
4777 | |
4778 | if (slp_node) | |
f18b55bd IR |
4779 | continue; |
4780 | ||
4781 | if (j == 0) | |
4782 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4783 | else | |
4784 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4785 | ||
4786 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4787 | } | |
b8698a0f | 4788 | |
9771b263 | 4789 | vec_oprnds.release (); |
ebfd146a IR |
4790 | return true; |
4791 | } | |
4792 | ||
9dc3f7de | 4793 | |
1107f3ae IR |
4794 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
4795 | either as shift by a scalar or by a vector. */ | |
4796 | ||
4797 | bool | |
4798 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
4799 | { | |
4800 | ||
ef4bddc2 | 4801 | machine_mode vec_mode; |
1107f3ae IR |
4802 | optab optab; |
4803 | int icode; | |
4804 | tree vectype; | |
4805 | ||
4806 | vectype = get_vectype_for_scalar_type (scalar_type); | |
4807 | if (!vectype) | |
4808 | return false; | |
4809 | ||
4810 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4811 | if (!optab | |
4812 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
4813 | { | |
4814 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4815 | if (!optab | |
4816 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
4817 | == CODE_FOR_nothing)) | |
4818 | return false; | |
4819 | } | |
4820 | ||
4821 | vec_mode = TYPE_MODE (vectype); | |
4822 | icode = (int) optab_handler (optab, vec_mode); | |
4823 | if (icode == CODE_FOR_nothing) | |
4824 | return false; | |
4825 | ||
4826 | return true; | |
4827 | } | |
4828 | ||
4829 | ||
9dc3f7de IR |
4830 | /* Function vectorizable_shift. |
4831 | ||
4832 | Check if STMT performs a shift operation that can be vectorized. | |
4833 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4834 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
4835 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4836 | ||
4837 | static bool | |
355fe088 TS |
4838 | vectorizable_shift (gimple *stmt, gimple_stmt_iterator *gsi, |
4839 | gimple **vec_stmt, slp_tree slp_node) | |
9dc3f7de IR |
4840 | { |
4841 | tree vec_dest; | |
4842 | tree scalar_dest; | |
4843 | tree op0, op1 = NULL; | |
4844 | tree vec_oprnd1 = NULL_TREE; | |
4845 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4846 | tree vectype; | |
4847 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4848 | enum tree_code code; | |
ef4bddc2 | 4849 | machine_mode vec_mode; |
9dc3f7de IR |
4850 | tree new_temp; |
4851 | optab optab; | |
4852 | int icode; | |
ef4bddc2 | 4853 | machine_mode optab_op2_mode; |
355fe088 | 4854 | gimple *def_stmt; |
9dc3f7de | 4855 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 4856 | int ndts = 2; |
355fe088 | 4857 | gimple *new_stmt = NULL; |
9dc3f7de IR |
4858 | stmt_vec_info prev_stmt_info; |
4859 | int nunits_in; | |
4860 | int nunits_out; | |
4861 | tree vectype_out; | |
cede2577 | 4862 | tree op1_vectype; |
9dc3f7de IR |
4863 | int ncopies; |
4864 | int j, i; | |
6e1aa848 DN |
4865 | vec<tree> vec_oprnds0 = vNULL; |
4866 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
4867 | tree vop0, vop1; |
4868 | unsigned int k; | |
49eab32e | 4869 | bool scalar_shift_arg = true; |
9dc3f7de | 4870 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4871 | vec_info *vinfo = stmt_info->vinfo; |
9dc3f7de IR |
4872 | int vf; |
4873 | ||
4874 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
4875 | return false; | |
4876 | ||
66c16fd9 RB |
4877 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4878 | && ! vec_stmt) | |
9dc3f7de IR |
4879 | return false; |
4880 | ||
4881 | /* Is STMT a vectorizable binary/unary operation? */ | |
4882 | if (!is_gimple_assign (stmt)) | |
4883 | return false; | |
4884 | ||
4885 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4886 | return false; | |
4887 | ||
4888 | code = gimple_assign_rhs_code (stmt); | |
4889 | ||
4890 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
4891 | || code == RROTATE_EXPR)) | |
4892 | return false; | |
4893 | ||
4894 | scalar_dest = gimple_assign_lhs (stmt); | |
4895 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2be65d9e | 4896 | if (!type_has_mode_precision_p (TREE_TYPE (scalar_dest))) |
7b7b1813 | 4897 | { |
73fbfcad | 4898 | if (dump_enabled_p ()) |
78c60e3d | 4899 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4900 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
4901 | return false; |
4902 | } | |
9dc3f7de IR |
4903 | |
4904 | op0 = gimple_assign_rhs1 (stmt); | |
81c40241 | 4905 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
9dc3f7de | 4906 | { |
73fbfcad | 4907 | if (dump_enabled_p ()) |
78c60e3d | 4908 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4909 | "use not simple.\n"); |
9dc3f7de IR |
4910 | return false; |
4911 | } | |
4912 | /* If op0 is an external or constant def use a vector type with | |
4913 | the same size as the output vector type. */ | |
4914 | if (!vectype) | |
4915 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
4916 | if (vec_stmt) | |
4917 | gcc_assert (vectype); | |
4918 | if (!vectype) | |
4919 | { | |
73fbfcad | 4920 | if (dump_enabled_p ()) |
78c60e3d | 4921 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4922 | "no vectype for scalar type\n"); |
9dc3f7de IR |
4923 | return false; |
4924 | } | |
4925 | ||
4926 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4927 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4928 | if (nunits_out != nunits_in) | |
4929 | return false; | |
4930 | ||
4931 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 4932 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &op1_vectype)) |
9dc3f7de | 4933 | { |
73fbfcad | 4934 | if (dump_enabled_p ()) |
78c60e3d | 4935 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4936 | "use not simple.\n"); |
9dc3f7de IR |
4937 | return false; |
4938 | } | |
4939 | ||
4940 | if (loop_vinfo) | |
4941 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4942 | else | |
4943 | vf = 1; | |
4944 | ||
4945 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4946 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4947 | case of SLP. */ | |
fce57248 | 4948 | if (slp_node) |
9dc3f7de IR |
4949 | ncopies = 1; |
4950 | else | |
4951 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
4952 | ||
4953 | gcc_assert (ncopies >= 1); | |
4954 | ||
4955 | /* Determine whether the shift amount is a vector, or scalar. If the | |
4956 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
4957 | ||
dbfa87aa YR |
4958 | if ((dt[1] == vect_internal_def |
4959 | || dt[1] == vect_induction_def) | |
4960 | && !slp_node) | |
49eab32e JJ |
4961 | scalar_shift_arg = false; |
4962 | else if (dt[1] == vect_constant_def | |
4963 | || dt[1] == vect_external_def | |
4964 | || dt[1] == vect_internal_def) | |
4965 | { | |
4966 | /* In SLP, need to check whether the shift count is the same, | |
4967 | in loops if it is a constant or invariant, it is always | |
4968 | a scalar shift. */ | |
4969 | if (slp_node) | |
4970 | { | |
355fe088 TS |
4971 | vec<gimple *> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
4972 | gimple *slpstmt; | |
49eab32e | 4973 | |
9771b263 | 4974 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
4975 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
4976 | scalar_shift_arg = false; | |
4977 | } | |
60d393e8 RB |
4978 | |
4979 | /* If the shift amount is computed by a pattern stmt we cannot | |
4980 | use the scalar amount directly thus give up and use a vector | |
4981 | shift. */ | |
4982 | if (dt[1] == vect_internal_def) | |
4983 | { | |
4984 | gimple *def = SSA_NAME_DEF_STMT (op1); | |
4985 | if (is_pattern_stmt_p (vinfo_for_stmt (def))) | |
4986 | scalar_shift_arg = false; | |
4987 | } | |
49eab32e JJ |
4988 | } |
4989 | else | |
4990 | { | |
73fbfcad | 4991 | if (dump_enabled_p ()) |
78c60e3d | 4992 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4993 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
4994 | return false; |
4995 | } | |
4996 | ||
9dc3f7de | 4997 | /* Vector shifted by vector. */ |
49eab32e | 4998 | if (!scalar_shift_arg) |
9dc3f7de IR |
4999 | { |
5000 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 5001 | if (dump_enabled_p ()) |
78c60e3d | 5002 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5003 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 5004 | |
aa948027 JJ |
5005 | if (!op1_vectype) |
5006 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
5007 | if (op1_vectype == NULL_TREE | |
5008 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 5009 | { |
73fbfcad | 5010 | if (dump_enabled_p ()) |
78c60e3d SS |
5011 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5012 | "unusable type for last operand in" | |
e645e942 | 5013 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
5014 | return false; |
5015 | } | |
9dc3f7de IR |
5016 | } |
5017 | /* See if the machine has a vector shifted by scalar insn and if not | |
5018 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 5019 | else |
9dc3f7de IR |
5020 | { |
5021 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
5022 | if (optab | |
5023 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
5024 | { | |
73fbfcad | 5025 | if (dump_enabled_p ()) |
78c60e3d | 5026 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5027 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
5028 | } |
5029 | else | |
5030 | { | |
5031 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
5032 | if (optab | |
5033 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
5034 | != CODE_FOR_nothing)) | |
5035 | { | |
49eab32e JJ |
5036 | scalar_shift_arg = false; |
5037 | ||
73fbfcad | 5038 | if (dump_enabled_p ()) |
78c60e3d | 5039 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5040 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
5041 | |
5042 | /* Unlike the other binary operators, shifts/rotates have | |
5043 | the rhs being int, instead of the same type as the lhs, | |
5044 | so make sure the scalar is the right type if we are | |
aa948027 | 5045 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
5046 | if (dt[1] == vect_constant_def) |
5047 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
5048 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
5049 | TREE_TYPE (op1))) | |
5050 | { | |
5051 | if (slp_node | |
5052 | && TYPE_MODE (TREE_TYPE (vectype)) | |
5053 | != TYPE_MODE (TREE_TYPE (op1))) | |
5054 | { | |
73fbfcad | 5055 | if (dump_enabled_p ()) |
78c60e3d SS |
5056 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5057 | "unusable type for last operand in" | |
e645e942 | 5058 | " vector/vector shift/rotate.\n"); |
21c0a521 | 5059 | return false; |
aa948027 JJ |
5060 | } |
5061 | if (vec_stmt && !slp_node) | |
5062 | { | |
5063 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
5064 | op1 = vect_init_vector (stmt, op1, | |
5065 | TREE_TYPE (vectype), NULL); | |
5066 | } | |
5067 | } | |
9dc3f7de IR |
5068 | } |
5069 | } | |
5070 | } | |
9dc3f7de IR |
5071 | |
5072 | /* Supportable by target? */ | |
5073 | if (!optab) | |
5074 | { | |
73fbfcad | 5075 | if (dump_enabled_p ()) |
78c60e3d | 5076 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5077 | "no optab.\n"); |
9dc3f7de IR |
5078 | return false; |
5079 | } | |
5080 | vec_mode = TYPE_MODE (vectype); | |
5081 | icode = (int) optab_handler (optab, vec_mode); | |
5082 | if (icode == CODE_FOR_nothing) | |
5083 | { | |
73fbfcad | 5084 | if (dump_enabled_p ()) |
78c60e3d | 5085 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5086 | "op not supported by target.\n"); |
9dc3f7de IR |
5087 | /* Check only during analysis. */ |
5088 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
5089 | || (vf < vect_min_worthwhile_factor (code) | |
5090 | && !vec_stmt)) | |
5091 | return false; | |
73fbfcad | 5092 | if (dump_enabled_p ()) |
e645e942 TJ |
5093 | dump_printf_loc (MSG_NOTE, vect_location, |
5094 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
5095 | } |
5096 | ||
5097 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
5098 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
5099 | && vf < vect_min_worthwhile_factor (code) | |
5100 | && !vec_stmt) | |
5101 | { | |
73fbfcad | 5102 | if (dump_enabled_p ()) |
78c60e3d | 5103 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5104 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
5105 | return false; |
5106 | } | |
5107 | ||
5108 | if (!vec_stmt) /* transformation not required. */ | |
5109 | { | |
5110 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 5111 | if (dump_enabled_p ()) |
e645e942 TJ |
5112 | dump_printf_loc (MSG_NOTE, vect_location, |
5113 | "=== vectorizable_shift ===\n"); | |
4fc5ebf1 | 5114 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); |
9dc3f7de IR |
5115 | return true; |
5116 | } | |
5117 | ||
67b8dbac | 5118 | /* Transform. */ |
9dc3f7de | 5119 | |
73fbfcad | 5120 | if (dump_enabled_p ()) |
78c60e3d | 5121 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5122 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
5123 | |
5124 | /* Handle def. */ | |
5125 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5126 | ||
9dc3f7de IR |
5127 | prev_stmt_info = NULL; |
5128 | for (j = 0; j < ncopies; j++) | |
5129 | { | |
5130 | /* Handle uses. */ | |
5131 | if (j == 0) | |
5132 | { | |
5133 | if (scalar_shift_arg) | |
5134 | { | |
5135 | /* Vector shl and shr insn patterns can be defined with scalar | |
5136 | operand 2 (shift operand). In this case, use constant or loop | |
5137 | invariant op1 directly, without extending it to vector mode | |
5138 | first. */ | |
5139 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
5140 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
5141 | { | |
73fbfcad | 5142 | if (dump_enabled_p ()) |
78c60e3d | 5143 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5144 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 5145 | vec_oprnd1 = op1; |
8930f723 | 5146 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 5147 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5148 | if (slp_node) |
5149 | { | |
5150 | /* Store vec_oprnd1 for every vector stmt to be created | |
5151 | for SLP_NODE. We check during the analysis that all | |
5152 | the shift arguments are the same. | |
5153 | TODO: Allow different constants for different vector | |
5154 | stmts generated for an SLP instance. */ | |
5155 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 5156 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5157 | } |
5158 | } | |
5159 | } | |
5160 | ||
5161 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
5162 | (a special case for certain kind of vector shifts); otherwise, | |
5163 | operand 1 should be of a vector type (the usual case). */ | |
5164 | if (vec_oprnd1) | |
5165 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5166 | slp_node); |
9dc3f7de IR |
5167 | else |
5168 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
306b0c92 | 5169 | slp_node); |
9dc3f7de IR |
5170 | } |
5171 | else | |
5172 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5173 | ||
5174 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 5175 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 5176 | { |
9771b263 | 5177 | vop1 = vec_oprnds1[i]; |
0d0e4a03 | 5178 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
9dc3f7de IR |
5179 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5180 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5181 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5182 | if (slp_node) | |
9771b263 | 5183 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
5184 | } |
5185 | ||
5186 | if (slp_node) | |
5187 | continue; | |
5188 | ||
5189 | if (j == 0) | |
5190 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5191 | else | |
5192 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5193 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5194 | } | |
5195 | ||
9771b263 DN |
5196 | vec_oprnds0.release (); |
5197 | vec_oprnds1.release (); | |
9dc3f7de IR |
5198 | |
5199 | return true; | |
5200 | } | |
5201 | ||
5202 | ||
ebfd146a IR |
5203 | /* Function vectorizable_operation. |
5204 | ||
16949072 RG |
5205 | Check if STMT performs a binary, unary or ternary operation that can |
5206 | be vectorized. | |
b8698a0f | 5207 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
5208 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5209 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5210 | ||
5211 | static bool | |
355fe088 TS |
5212 | vectorizable_operation (gimple *stmt, gimple_stmt_iterator *gsi, |
5213 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a | 5214 | { |
00f07b86 | 5215 | tree vec_dest; |
ebfd146a | 5216 | tree scalar_dest; |
16949072 | 5217 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 5218 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 5219 | tree vectype; |
ebfd146a IR |
5220 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
5221 | enum tree_code code; | |
ef4bddc2 | 5222 | machine_mode vec_mode; |
ebfd146a IR |
5223 | tree new_temp; |
5224 | int op_type; | |
00f07b86 | 5225 | optab optab; |
523ba738 | 5226 | bool target_support_p; |
355fe088 | 5227 | gimple *def_stmt; |
16949072 RG |
5228 | enum vect_def_type dt[3] |
5229 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 5230 | int ndts = 3; |
355fe088 | 5231 | gimple *new_stmt = NULL; |
ebfd146a | 5232 | stmt_vec_info prev_stmt_info; |
b690cc0f | 5233 | int nunits_in; |
ebfd146a IR |
5234 | int nunits_out; |
5235 | tree vectype_out; | |
5236 | int ncopies; | |
5237 | int j, i; | |
6e1aa848 DN |
5238 | vec<tree> vec_oprnds0 = vNULL; |
5239 | vec<tree> vec_oprnds1 = vNULL; | |
5240 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 5241 | tree vop0, vop1, vop2; |
a70d6342 | 5242 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5243 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 IR |
5244 | int vf; |
5245 | ||
a70d6342 | 5246 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5247 | return false; |
5248 | ||
66c16fd9 RB |
5249 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5250 | && ! vec_stmt) | |
ebfd146a IR |
5251 | return false; |
5252 | ||
5253 | /* Is STMT a vectorizable binary/unary operation? */ | |
5254 | if (!is_gimple_assign (stmt)) | |
5255 | return false; | |
5256 | ||
5257 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
5258 | return false; | |
5259 | ||
ebfd146a IR |
5260 | code = gimple_assign_rhs_code (stmt); |
5261 | ||
5262 | /* For pointer addition, we should use the normal plus for | |
5263 | the vector addition. */ | |
5264 | if (code == POINTER_PLUS_EXPR) | |
5265 | code = PLUS_EXPR; | |
5266 | ||
5267 | /* Support only unary or binary operations. */ | |
5268 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 5269 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 5270 | { |
73fbfcad | 5271 | if (dump_enabled_p ()) |
78c60e3d | 5272 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5273 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 5274 | op_type); |
ebfd146a IR |
5275 | return false; |
5276 | } | |
5277 | ||
b690cc0f RG |
5278 | scalar_dest = gimple_assign_lhs (stmt); |
5279 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
5280 | ||
7b7b1813 RG |
5281 | /* Most operations cannot handle bit-precision types without extra |
5282 | truncations. */ | |
045c1278 | 5283 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
2be65d9e | 5284 | && !type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
7b7b1813 RG |
5285 | /* Exception are bitwise binary operations. */ |
5286 | && code != BIT_IOR_EXPR | |
5287 | && code != BIT_XOR_EXPR | |
5288 | && code != BIT_AND_EXPR) | |
5289 | { | |
73fbfcad | 5290 | if (dump_enabled_p ()) |
78c60e3d | 5291 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5292 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
5293 | return false; |
5294 | } | |
5295 | ||
ebfd146a | 5296 | op0 = gimple_assign_rhs1 (stmt); |
81c40241 | 5297 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
ebfd146a | 5298 | { |
73fbfcad | 5299 | if (dump_enabled_p ()) |
78c60e3d | 5300 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5301 | "use not simple.\n"); |
ebfd146a IR |
5302 | return false; |
5303 | } | |
b690cc0f RG |
5304 | /* If op0 is an external or constant def use a vector type with |
5305 | the same size as the output vector type. */ | |
5306 | if (!vectype) | |
b036c6c5 IE |
5307 | { |
5308 | /* For boolean type we cannot determine vectype by | |
5309 | invariant value (don't know whether it is a vector | |
5310 | of booleans or vector of integers). We use output | |
5311 | vectype because operations on boolean don't change | |
5312 | type. */ | |
2568d8a1 | 5313 | if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op0))) |
b036c6c5 | 5314 | { |
2568d8a1 | 5315 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (scalar_dest))) |
b036c6c5 IE |
5316 | { |
5317 | if (dump_enabled_p ()) | |
5318 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5319 | "not supported operation on bool value.\n"); | |
5320 | return false; | |
5321 | } | |
5322 | vectype = vectype_out; | |
5323 | } | |
5324 | else | |
5325 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
5326 | } | |
7d8930a0 IR |
5327 | if (vec_stmt) |
5328 | gcc_assert (vectype); | |
5329 | if (!vectype) | |
5330 | { | |
73fbfcad | 5331 | if (dump_enabled_p ()) |
7d8930a0 | 5332 | { |
78c60e3d SS |
5333 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5334 | "no vectype for scalar type "); | |
5335 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
5336 | TREE_TYPE (op0)); | |
e645e942 | 5337 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
5338 | } |
5339 | ||
5340 | return false; | |
5341 | } | |
b690cc0f RG |
5342 | |
5343 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
5344 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
5345 | if (nunits_out != nunits_in) | |
5346 | return false; | |
ebfd146a | 5347 | |
16949072 | 5348 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
5349 | { |
5350 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 5351 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1])) |
ebfd146a | 5352 | { |
73fbfcad | 5353 | if (dump_enabled_p ()) |
78c60e3d | 5354 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5355 | "use not simple.\n"); |
ebfd146a IR |
5356 | return false; |
5357 | } | |
5358 | } | |
16949072 RG |
5359 | if (op_type == ternary_op) |
5360 | { | |
5361 | op2 = gimple_assign_rhs3 (stmt); | |
81c40241 | 5362 | if (!vect_is_simple_use (op2, vinfo, &def_stmt, &dt[2])) |
16949072 | 5363 | { |
73fbfcad | 5364 | if (dump_enabled_p ()) |
78c60e3d | 5365 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5366 | "use not simple.\n"); |
16949072 RG |
5367 | return false; |
5368 | } | |
5369 | } | |
ebfd146a | 5370 | |
b690cc0f RG |
5371 | if (loop_vinfo) |
5372 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
5373 | else | |
5374 | vf = 1; | |
5375 | ||
5376 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 5377 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 5378 | case of SLP. */ |
fce57248 | 5379 | if (slp_node) |
b690cc0f RG |
5380 | ncopies = 1; |
5381 | else | |
5382 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
5383 | ||
5384 | gcc_assert (ncopies >= 1); | |
5385 | ||
9dc3f7de | 5386 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
5387 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
5388 | || code == RROTATE_EXPR) | |
9dc3f7de | 5389 | return false; |
ebfd146a | 5390 | |
ebfd146a | 5391 | /* Supportable by target? */ |
00f07b86 RH |
5392 | |
5393 | vec_mode = TYPE_MODE (vectype); | |
5394 | if (code == MULT_HIGHPART_EXPR) | |
523ba738 | 5395 | target_support_p = can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype)); |
00f07b86 RH |
5396 | else |
5397 | { | |
5398 | optab = optab_for_tree_code (code, vectype, optab_default); | |
5399 | if (!optab) | |
5deb57cb | 5400 | { |
73fbfcad | 5401 | if (dump_enabled_p ()) |
78c60e3d | 5402 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5403 | "no optab.\n"); |
00f07b86 | 5404 | return false; |
5deb57cb | 5405 | } |
523ba738 RS |
5406 | target_support_p = (optab_handler (optab, vec_mode) |
5407 | != CODE_FOR_nothing); | |
5deb57cb JJ |
5408 | } |
5409 | ||
523ba738 | 5410 | if (!target_support_p) |
ebfd146a | 5411 | { |
73fbfcad | 5412 | if (dump_enabled_p ()) |
78c60e3d | 5413 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5414 | "op not supported by target.\n"); |
ebfd146a IR |
5415 | /* Check only during analysis. */ |
5416 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
5deb57cb | 5417 | || (!vec_stmt && vf < vect_min_worthwhile_factor (code))) |
ebfd146a | 5418 | return false; |
73fbfcad | 5419 | if (dump_enabled_p ()) |
e645e942 TJ |
5420 | dump_printf_loc (MSG_NOTE, vect_location, |
5421 | "proceeding using word mode.\n"); | |
383d9c83 IR |
5422 | } |
5423 | ||
4a00c761 | 5424 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
5425 | if (!VECTOR_MODE_P (vec_mode) |
5426 | && !vec_stmt | |
5427 | && vf < vect_min_worthwhile_factor (code)) | |
7d8930a0 | 5428 | { |
73fbfcad | 5429 | if (dump_enabled_p ()) |
78c60e3d | 5430 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5431 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 5432 | return false; |
7d8930a0 | 5433 | } |
ebfd146a | 5434 | |
ebfd146a IR |
5435 | if (!vec_stmt) /* transformation not required. */ |
5436 | { | |
4a00c761 | 5437 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 5438 | if (dump_enabled_p ()) |
78c60e3d | 5439 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5440 | "=== vectorizable_operation ===\n"); |
4fc5ebf1 | 5441 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); |
ebfd146a IR |
5442 | return true; |
5443 | } | |
5444 | ||
67b8dbac | 5445 | /* Transform. */ |
ebfd146a | 5446 | |
73fbfcad | 5447 | if (dump_enabled_p ()) |
78c60e3d | 5448 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5449 | "transform binary/unary operation.\n"); |
383d9c83 | 5450 | |
ebfd146a | 5451 | /* Handle def. */ |
00f07b86 | 5452 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 5453 | |
ebfd146a IR |
5454 | /* In case the vectorization factor (VF) is bigger than the number |
5455 | of elements that we can fit in a vectype (nunits), we have to generate | |
5456 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
5457 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
5458 | from one copy of the vector stmt to the next, in the field | |
5459 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
5460 | stages to find the correct vector defs to be used when vectorizing | |
5461 | stmts that use the defs of the current stmt. The example below | |
5462 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
5463 | we need to create 4 vectorized stmts): | |
5464 | ||
5465 | before vectorization: | |
5466 | RELATED_STMT VEC_STMT | |
5467 | S1: x = memref - - | |
5468 | S2: z = x + 1 - - | |
5469 | ||
5470 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
5471 | there): | |
5472 | RELATED_STMT VEC_STMT | |
5473 | VS1_0: vx0 = memref0 VS1_1 - | |
5474 | VS1_1: vx1 = memref1 VS1_2 - | |
5475 | VS1_2: vx2 = memref2 VS1_3 - | |
5476 | VS1_3: vx3 = memref3 - - | |
5477 | S1: x = load - VS1_0 | |
5478 | S2: z = x + 1 - - | |
5479 | ||
5480 | step2: vectorize stmt S2 (done here): | |
5481 | To vectorize stmt S2 we first need to find the relevant vector | |
5482 | def for the first operand 'x'. This is, as usual, obtained from | |
5483 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
5484 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
5485 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
5486 | the vector stmt VS2_0, and as usual, record it in the | |
5487 | STMT_VINFO_VEC_STMT of stmt S2. | |
5488 | When creating the second copy (VS2_1), we obtain the relevant vector | |
5489 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
5490 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
5491 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
5492 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
5493 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
5494 | chain of stmts and pointers: | |
5495 | RELATED_STMT VEC_STMT | |
5496 | VS1_0: vx0 = memref0 VS1_1 - | |
5497 | VS1_1: vx1 = memref1 VS1_2 - | |
5498 | VS1_2: vx2 = memref2 VS1_3 - | |
5499 | VS1_3: vx3 = memref3 - - | |
5500 | S1: x = load - VS1_0 | |
5501 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
5502 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
5503 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
5504 | VS2_3: vz3 = vx3 + v1 - - | |
5505 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
5506 | |
5507 | prev_stmt_info = NULL; | |
5508 | for (j = 0; j < ncopies; j++) | |
5509 | { | |
5510 | /* Handle uses. */ | |
5511 | if (j == 0) | |
4a00c761 JJ |
5512 | { |
5513 | if (op_type == binary_op || op_type == ternary_op) | |
5514 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
306b0c92 | 5515 | slp_node); |
4a00c761 JJ |
5516 | else |
5517 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5518 | slp_node); |
4a00c761 | 5519 | if (op_type == ternary_op) |
c392943c | 5520 | vect_get_vec_defs (op2, NULL_TREE, stmt, &vec_oprnds2, NULL, |
306b0c92 | 5521 | slp_node); |
4a00c761 | 5522 | } |
ebfd146a | 5523 | else |
4a00c761 JJ |
5524 | { |
5525 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5526 | if (op_type == ternary_op) | |
5527 | { | |
9771b263 DN |
5528 | tree vec_oprnd = vec_oprnds2.pop (); |
5529 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
5530 | vec_oprnd)); | |
4a00c761 JJ |
5531 | } |
5532 | } | |
5533 | ||
5534 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 5535 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 5536 | { |
4a00c761 | 5537 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 5538 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 5539 | vop2 = ((op_type == ternary_op) |
9771b263 | 5540 | ? vec_oprnds2[i] : NULL_TREE); |
0d0e4a03 | 5541 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1, vop2); |
4a00c761 JJ |
5542 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5543 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5544 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5545 | if (slp_node) | |
9771b263 | 5546 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
5547 | } |
5548 | ||
4a00c761 JJ |
5549 | if (slp_node) |
5550 | continue; | |
5551 | ||
5552 | if (j == 0) | |
5553 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5554 | else | |
5555 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5556 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
5557 | } |
5558 | ||
9771b263 DN |
5559 | vec_oprnds0.release (); |
5560 | vec_oprnds1.release (); | |
5561 | vec_oprnds2.release (); | |
ebfd146a | 5562 | |
ebfd146a IR |
5563 | return true; |
5564 | } | |
5565 | ||
c716e67f XDL |
5566 | /* A helper function to ensure data reference DR's base alignment |
5567 | for STMT_INFO. */ | |
5568 | ||
5569 | static void | |
5570 | ensure_base_align (stmt_vec_info stmt_info, struct data_reference *dr) | |
5571 | { | |
5572 | if (!dr->aux) | |
5573 | return; | |
5574 | ||
52639a61 | 5575 | if (DR_VECT_AUX (dr)->base_misaligned) |
c716e67f XDL |
5576 | { |
5577 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
52639a61 | 5578 | tree base_decl = DR_VECT_AUX (dr)->base_decl; |
c716e67f | 5579 | |
428f0c67 JH |
5580 | if (decl_in_symtab_p (base_decl)) |
5581 | symtab_node::get (base_decl)->increase_alignment (TYPE_ALIGN (vectype)); | |
5582 | else | |
5583 | { | |
fe37c7af | 5584 | SET_DECL_ALIGN (base_decl, TYPE_ALIGN (vectype)); |
428f0c67 JH |
5585 | DECL_USER_ALIGN (base_decl) = 1; |
5586 | } | |
52639a61 | 5587 | DR_VECT_AUX (dr)->base_misaligned = false; |
c716e67f XDL |
5588 | } |
5589 | } | |
5590 | ||
ebfd146a | 5591 | |
44fc7854 BE |
5592 | /* Function get_group_alias_ptr_type. |
5593 | ||
5594 | Return the alias type for the group starting at FIRST_STMT. */ | |
5595 | ||
5596 | static tree | |
5597 | get_group_alias_ptr_type (gimple *first_stmt) | |
5598 | { | |
5599 | struct data_reference *first_dr, *next_dr; | |
5600 | gimple *next_stmt; | |
5601 | ||
5602 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
5603 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (first_stmt)); | |
5604 | while (next_stmt) | |
5605 | { | |
5606 | next_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (next_stmt)); | |
5607 | if (get_alias_set (DR_REF (first_dr)) | |
5608 | != get_alias_set (DR_REF (next_dr))) | |
5609 | { | |
5610 | if (dump_enabled_p ()) | |
5611 | dump_printf_loc (MSG_NOTE, vect_location, | |
5612 | "conflicting alias set types.\n"); | |
5613 | return ptr_type_node; | |
5614 | } | |
5615 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); | |
5616 | } | |
5617 | return reference_alias_ptr_type (DR_REF (first_dr)); | |
5618 | } | |
5619 | ||
5620 | ||
ebfd146a IR |
5621 | /* Function vectorizable_store. |
5622 | ||
b8698a0f L |
5623 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
5624 | can be vectorized. | |
5625 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
5626 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5627 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5628 | ||
5629 | static bool | |
355fe088 | 5630 | vectorizable_store (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 5631 | slp_tree slp_node) |
ebfd146a IR |
5632 | { |
5633 | tree scalar_dest; | |
5634 | tree data_ref; | |
5635 | tree op; | |
5636 | tree vec_oprnd = NULL_TREE; | |
5637 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5638 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
272c6793 | 5639 | tree elem_type; |
ebfd146a | 5640 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 5641 | struct loop *loop = NULL; |
ef4bddc2 | 5642 | machine_mode vec_mode; |
ebfd146a IR |
5643 | tree dummy; |
5644 | enum dr_alignment_support alignment_support_scheme; | |
355fe088 | 5645 | gimple *def_stmt; |
ebfd146a IR |
5646 | enum vect_def_type dt; |
5647 | stmt_vec_info prev_stmt_info = NULL; | |
5648 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 5649 | tree dataref_offset = NULL_TREE; |
355fe088 | 5650 | gimple *ptr_incr = NULL; |
ebfd146a IR |
5651 | int ncopies; |
5652 | int j; | |
2de001ee RS |
5653 | gimple *next_stmt, *first_stmt; |
5654 | bool grouped_store; | |
ebfd146a | 5655 | unsigned int group_size, i; |
6e1aa848 DN |
5656 | vec<tree> oprnds = vNULL; |
5657 | vec<tree> result_chain = vNULL; | |
ebfd146a | 5658 | bool inv_p; |
09dfa495 | 5659 | tree offset = NULL_TREE; |
6e1aa848 | 5660 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 5661 | bool slp = (slp_node != NULL); |
ebfd146a | 5662 | unsigned int vec_num; |
a70d6342 | 5663 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5664 | vec_info *vinfo = stmt_info->vinfo; |
272c6793 | 5665 | tree aggr_type; |
134c85ca | 5666 | gather_scatter_info gs_info; |
3bab6342 | 5667 | enum vect_def_type scatter_src_dt = vect_unknown_def_type; |
355fe088 | 5668 | gimple *new_stmt; |
b17dc4d4 | 5669 | int vf; |
2de001ee | 5670 | vec_load_store_type vls_type; |
44fc7854 | 5671 | tree ref_type; |
a70d6342 | 5672 | |
a70d6342 | 5673 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5674 | return false; |
5675 | ||
66c16fd9 RB |
5676 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5677 | && ! vec_stmt) | |
ebfd146a IR |
5678 | return false; |
5679 | ||
5680 | /* Is vectorizable store? */ | |
5681 | ||
5682 | if (!is_gimple_assign (stmt)) | |
5683 | return false; | |
5684 | ||
5685 | scalar_dest = gimple_assign_lhs (stmt); | |
ab0ef706 JJ |
5686 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
5687 | && is_pattern_stmt_p (stmt_info)) | |
5688 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
ebfd146a | 5689 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
38000232 | 5690 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF |
ebfd146a | 5691 | && TREE_CODE (scalar_dest) != INDIRECT_REF |
e9dbe7bb IR |
5692 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
5693 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
5694 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
5695 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
5696 | return false; |
5697 | ||
fce57248 RS |
5698 | /* Cannot have hybrid store SLP -- that would mean storing to the |
5699 | same location twice. */ | |
5700 | gcc_assert (slp == PURE_SLP_STMT (stmt_info)); | |
5701 | ||
ebfd146a | 5702 | gcc_assert (gimple_assign_single_p (stmt)); |
465c8c19 | 5703 | |
f4d09712 | 5704 | tree vectype = STMT_VINFO_VECTYPE (stmt_info), rhs_vectype = NULL_TREE; |
465c8c19 JJ |
5705 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
5706 | ||
5707 | if (loop_vinfo) | |
b17dc4d4 RB |
5708 | { |
5709 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
5710 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
5711 | } | |
5712 | else | |
5713 | vf = 1; | |
465c8c19 JJ |
5714 | |
5715 | /* Multiple types in SLP are handled by creating the appropriate number of | |
5716 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5717 | case of SLP. */ | |
fce57248 | 5718 | if (slp) |
465c8c19 JJ |
5719 | ncopies = 1; |
5720 | else | |
5721 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
5722 | ||
5723 | gcc_assert (ncopies >= 1); | |
5724 | ||
5725 | /* FORNOW. This restriction should be relaxed. */ | |
5726 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) | |
5727 | { | |
5728 | if (dump_enabled_p ()) | |
5729 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5730 | "multiple types in nested loop.\n"); | |
5731 | return false; | |
5732 | } | |
5733 | ||
ebfd146a | 5734 | op = gimple_assign_rhs1 (stmt); |
f4d09712 | 5735 | |
11a82e25 RB |
5736 | /* In the case this is a store from a STRING_CST make sure |
5737 | native_encode_expr can handle it. */ | |
5738 | if (TREE_CODE (op) == STRING_CST | |
5739 | && ! can_native_encode_string_p (op)) | |
5740 | return false; | |
5741 | ||
f4d09712 | 5742 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt, &rhs_vectype)) |
ebfd146a | 5743 | { |
73fbfcad | 5744 | if (dump_enabled_p ()) |
78c60e3d | 5745 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5746 | "use not simple.\n"); |
ebfd146a IR |
5747 | return false; |
5748 | } | |
5749 | ||
2de001ee RS |
5750 | if (dt == vect_constant_def || dt == vect_external_def) |
5751 | vls_type = VLS_STORE_INVARIANT; | |
5752 | else | |
5753 | vls_type = VLS_STORE; | |
5754 | ||
f4d09712 KY |
5755 | if (rhs_vectype && !useless_type_conversion_p (vectype, rhs_vectype)) |
5756 | return false; | |
5757 | ||
272c6793 | 5758 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 5759 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 5760 | |
ebfd146a IR |
5761 | /* FORNOW. In some cases can vectorize even if data-type not supported |
5762 | (e.g. - array initialization with 0). */ | |
947131ba | 5763 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
5764 | return false; |
5765 | ||
5766 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
5767 | return false; | |
5768 | ||
2de001ee | 5769 | vect_memory_access_type memory_access_type; |
62da9e14 | 5770 | if (!get_load_store_type (stmt, vectype, slp, vls_type, ncopies, |
2de001ee RS |
5771 | &memory_access_type, &gs_info)) |
5772 | return false; | |
3bab6342 | 5773 | |
ebfd146a IR |
5774 | if (!vec_stmt) /* transformation not required. */ |
5775 | { | |
2de001ee | 5776 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; |
ebfd146a | 5777 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; |
2e8ab70c RB |
5778 | /* The SLP costs are calculated during SLP analysis. */ |
5779 | if (!PURE_SLP_STMT (stmt_info)) | |
2de001ee | 5780 | vect_model_store_cost (stmt_info, ncopies, memory_access_type, dt, |
2e8ab70c | 5781 | NULL, NULL, NULL); |
ebfd146a IR |
5782 | return true; |
5783 | } | |
2de001ee | 5784 | gcc_assert (memory_access_type == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); |
ebfd146a | 5785 | |
67b8dbac | 5786 | /* Transform. */ |
ebfd146a | 5787 | |
c716e67f XDL |
5788 | ensure_base_align (stmt_info, dr); |
5789 | ||
2de001ee | 5790 | if (memory_access_type == VMAT_GATHER_SCATTER) |
3bab6342 AT |
5791 | { |
5792 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE, op, src; | |
134c85ca | 5793 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
5794 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; |
5795 | tree ptr, mask, var, scale, perm_mask = NULL_TREE; | |
5796 | edge pe = loop_preheader_edge (loop); | |
5797 | gimple_seq seq; | |
5798 | basic_block new_bb; | |
5799 | enum { NARROW, NONE, WIDEN } modifier; | |
134c85ca | 5800 | int scatter_off_nunits = TYPE_VECTOR_SUBPARTS (gs_info.offset_vectype); |
3bab6342 AT |
5801 | |
5802 | if (nunits == (unsigned int) scatter_off_nunits) | |
5803 | modifier = NONE; | |
5804 | else if (nunits == (unsigned int) scatter_off_nunits / 2) | |
5805 | { | |
3bab6342 AT |
5806 | modifier = WIDEN; |
5807 | ||
908a1a16 | 5808 | auto_vec_perm_indices sel (scatter_off_nunits); |
3bab6342 | 5809 | for (i = 0; i < (unsigned int) scatter_off_nunits; ++i) |
908a1a16 | 5810 | sel.quick_push (i | nunits); |
3bab6342 | 5811 | |
134c85ca | 5812 | perm_mask = vect_gen_perm_mask_checked (gs_info.offset_vectype, sel); |
3bab6342 AT |
5813 | gcc_assert (perm_mask != NULL_TREE); |
5814 | } | |
5815 | else if (nunits == (unsigned int) scatter_off_nunits * 2) | |
5816 | { | |
3bab6342 AT |
5817 | modifier = NARROW; |
5818 | ||
908a1a16 | 5819 | auto_vec_perm_indices sel (nunits); |
3bab6342 | 5820 | for (i = 0; i < (unsigned int) nunits; ++i) |
908a1a16 | 5821 | sel.quick_push (i | scatter_off_nunits); |
3bab6342 AT |
5822 | |
5823 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); | |
5824 | gcc_assert (perm_mask != NULL_TREE); | |
5825 | ncopies *= 2; | |
5826 | } | |
5827 | else | |
5828 | gcc_unreachable (); | |
5829 | ||
134c85ca | 5830 | rettype = TREE_TYPE (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
5831 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
5832 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5833 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5834 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5835 | scaletype = TREE_VALUE (arglist); | |
5836 | ||
5837 | gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE | |
5838 | && TREE_CODE (rettype) == VOID_TYPE); | |
5839 | ||
134c85ca | 5840 | ptr = fold_convert (ptrtype, gs_info.base); |
3bab6342 AT |
5841 | if (!is_gimple_min_invariant (ptr)) |
5842 | { | |
5843 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
5844 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
5845 | gcc_assert (!new_bb); | |
5846 | } | |
5847 | ||
5848 | /* Currently we support only unconditional scatter stores, | |
5849 | so mask should be all ones. */ | |
5850 | mask = build_int_cst (masktype, -1); | |
5851 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
5852 | ||
134c85ca | 5853 | scale = build_int_cst (scaletype, gs_info.scale); |
3bab6342 AT |
5854 | |
5855 | prev_stmt_info = NULL; | |
5856 | for (j = 0; j < ncopies; ++j) | |
5857 | { | |
5858 | if (j == 0) | |
5859 | { | |
5860 | src = vec_oprnd1 | |
81c40241 | 5861 | = vect_get_vec_def_for_operand (gimple_assign_rhs1 (stmt), stmt); |
3bab6342 | 5862 | op = vec_oprnd0 |
134c85ca | 5863 | = vect_get_vec_def_for_operand (gs_info.offset, stmt); |
3bab6342 AT |
5864 | } |
5865 | else if (modifier != NONE && (j & 1)) | |
5866 | { | |
5867 | if (modifier == WIDEN) | |
5868 | { | |
5869 | src = vec_oprnd1 | |
5870 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5871 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, perm_mask, | |
5872 | stmt, gsi); | |
5873 | } | |
5874 | else if (modifier == NARROW) | |
5875 | { | |
5876 | src = permute_vec_elements (vec_oprnd1, vec_oprnd1, perm_mask, | |
5877 | stmt, gsi); | |
5878 | op = vec_oprnd0 | |
134c85ca RS |
5879 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
5880 | vec_oprnd0); | |
3bab6342 AT |
5881 | } |
5882 | else | |
5883 | gcc_unreachable (); | |
5884 | } | |
5885 | else | |
5886 | { | |
5887 | src = vec_oprnd1 | |
5888 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5889 | op = vec_oprnd0 | |
134c85ca RS |
5890 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
5891 | vec_oprnd0); | |
3bab6342 AT |
5892 | } |
5893 | ||
5894 | if (!useless_type_conversion_p (srctype, TREE_TYPE (src))) | |
5895 | { | |
5896 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)) | |
5897 | == TYPE_VECTOR_SUBPARTS (srctype)); | |
0e22bb5a | 5898 | var = vect_get_new_ssa_name (srctype, vect_simple_var); |
3bab6342 AT |
5899 | src = build1 (VIEW_CONVERT_EXPR, srctype, src); |
5900 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, src); | |
5901 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5902 | src = var; | |
5903 | } | |
5904 | ||
5905 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
5906 | { | |
5907 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
5908 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 5909 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
3bab6342 AT |
5910 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
5911 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
5912 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5913 | op = var; | |
5914 | } | |
5915 | ||
5916 | new_stmt | |
134c85ca | 5917 | = gimple_build_call (gs_info.decl, 5, ptr, mask, op, src, scale); |
3bab6342 AT |
5918 | |
5919 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5920 | ||
5921 | if (prev_stmt_info == NULL) | |
5922 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5923 | else | |
5924 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5925 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5926 | } | |
5927 | return true; | |
5928 | } | |
5929 | ||
2de001ee | 5930 | grouped_store = STMT_VINFO_GROUPED_ACCESS (stmt_info); |
0d0293ac | 5931 | if (grouped_store) |
ebfd146a | 5932 | { |
2de001ee | 5933 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a | 5934 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
e14c1050 | 5935 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 5936 | |
e14c1050 | 5937 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
5938 | |
5939 | /* FORNOW */ | |
a70d6342 | 5940 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
5941 | |
5942 | /* We vectorize all the stmts of the interleaving group when we | |
5943 | reach the last stmt in the group. */ | |
e14c1050 IR |
5944 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
5945 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
5946 | && !slp) |
5947 | { | |
5948 | *vec_stmt = NULL; | |
5949 | return true; | |
5950 | } | |
5951 | ||
5952 | if (slp) | |
4b5caab7 | 5953 | { |
0d0293ac | 5954 | grouped_store = false; |
4b5caab7 IR |
5955 | /* VEC_NUM is the number of vect stmts to be created for this |
5956 | group. */ | |
5957 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 5958 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
52eab378 | 5959 | gcc_assert (GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_stmt)) == first_stmt); |
4b5caab7 | 5960 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
d092494c | 5961 | op = gimple_assign_rhs1 (first_stmt); |
4b5caab7 | 5962 | } |
ebfd146a | 5963 | else |
4b5caab7 IR |
5964 | /* VEC_NUM is the number of vect stmts to be created for this |
5965 | group. */ | |
ebfd146a | 5966 | vec_num = group_size; |
44fc7854 BE |
5967 | |
5968 | ref_type = get_group_alias_ptr_type (first_stmt); | |
ebfd146a | 5969 | } |
b8698a0f | 5970 | else |
ebfd146a IR |
5971 | { |
5972 | first_stmt = stmt; | |
5973 | first_dr = dr; | |
5974 | group_size = vec_num = 1; | |
44fc7854 | 5975 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); |
ebfd146a | 5976 | } |
b8698a0f | 5977 | |
73fbfcad | 5978 | if (dump_enabled_p ()) |
78c60e3d | 5979 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5980 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 5981 | |
2de001ee RS |
5982 | if (memory_access_type == VMAT_ELEMENTWISE |
5983 | || memory_access_type == VMAT_STRIDED_SLP) | |
f2e2a985 MM |
5984 | { |
5985 | gimple_stmt_iterator incr_gsi; | |
5986 | bool insert_after; | |
355fe088 | 5987 | gimple *incr; |
f2e2a985 MM |
5988 | tree offvar; |
5989 | tree ivstep; | |
5990 | tree running_off; | |
5991 | gimple_seq stmts = NULL; | |
5992 | tree stride_base, stride_step, alias_off; | |
5993 | tree vec_oprnd; | |
f502d50e | 5994 | unsigned int g; |
f2e2a985 MM |
5995 | |
5996 | gcc_assert (!nested_in_vect_loop_p (loop, stmt)); | |
5997 | ||
5998 | stride_base | |
5999 | = fold_build_pointer_plus | |
f502d50e | 6000 | (unshare_expr (DR_BASE_ADDRESS (first_dr)), |
f2e2a985 | 6001 | size_binop (PLUS_EXPR, |
f502d50e | 6002 | convert_to_ptrofftype (unshare_expr (DR_OFFSET (first_dr))), |
44fc7854 | 6003 | convert_to_ptrofftype (DR_INIT (first_dr)))); |
f502d50e | 6004 | stride_step = fold_convert (sizetype, unshare_expr (DR_STEP (first_dr))); |
f2e2a985 MM |
6005 | |
6006 | /* For a store with loop-invariant (but other than power-of-2) | |
6007 | stride (i.e. not a grouped access) like so: | |
6008 | ||
6009 | for (i = 0; i < n; i += stride) | |
6010 | array[i] = ...; | |
6011 | ||
6012 | we generate a new induction variable and new stores from | |
6013 | the components of the (vectorized) rhs: | |
6014 | ||
6015 | for (j = 0; ; j += VF*stride) | |
6016 | vectemp = ...; | |
6017 | tmp1 = vectemp[0]; | |
6018 | array[j] = tmp1; | |
6019 | tmp2 = vectemp[1]; | |
6020 | array[j + stride] = tmp2; | |
6021 | ... | |
6022 | */ | |
6023 | ||
cee62fee | 6024 | unsigned nstores = nunits; |
b17dc4d4 | 6025 | unsigned lnel = 1; |
cee62fee | 6026 | tree ltype = elem_type; |
04199738 | 6027 | tree lvectype = vectype; |
cee62fee MM |
6028 | if (slp) |
6029 | { | |
b17dc4d4 RB |
6030 | if (group_size < nunits |
6031 | && nunits % group_size == 0) | |
6032 | { | |
6033 | nstores = nunits / group_size; | |
6034 | lnel = group_size; | |
6035 | ltype = build_vector_type (elem_type, group_size); | |
04199738 RB |
6036 | lvectype = vectype; |
6037 | ||
6038 | /* First check if vec_extract optab doesn't support extraction | |
6039 | of vector elts directly. */ | |
b397965c | 6040 | scalar_mode elmode = SCALAR_TYPE_MODE (elem_type); |
9da15d40 RS |
6041 | machine_mode vmode; |
6042 | if (!mode_for_vector (elmode, group_size).exists (&vmode) | |
6043 | || !VECTOR_MODE_P (vmode) | |
04199738 RB |
6044 | || (convert_optab_handler (vec_extract_optab, |
6045 | TYPE_MODE (vectype), vmode) | |
6046 | == CODE_FOR_nothing)) | |
6047 | { | |
6048 | /* Try to avoid emitting an extract of vector elements | |
6049 | by performing the extracts using an integer type of the | |
6050 | same size, extracting from a vector of those and then | |
6051 | re-interpreting it as the original vector type if | |
6052 | supported. */ | |
6053 | unsigned lsize | |
6054 | = group_size * GET_MODE_BITSIZE (elmode); | |
fffbab82 | 6055 | elmode = int_mode_for_size (lsize, 0).require (); |
04199738 RB |
6056 | /* If we can't construct such a vector fall back to |
6057 | element extracts from the original vector type and | |
6058 | element size stores. */ | |
9da15d40 RS |
6059 | if (mode_for_vector (elmode, |
6060 | nunits / group_size).exists (&vmode) | |
6061 | && VECTOR_MODE_P (vmode) | |
04199738 RB |
6062 | && (convert_optab_handler (vec_extract_optab, |
6063 | vmode, elmode) | |
6064 | != CODE_FOR_nothing)) | |
6065 | { | |
6066 | nstores = nunits / group_size; | |
6067 | lnel = group_size; | |
6068 | ltype = build_nonstandard_integer_type (lsize, 1); | |
6069 | lvectype = build_vector_type (ltype, nstores); | |
6070 | } | |
6071 | /* Else fall back to vector extraction anyway. | |
6072 | Fewer stores are more important than avoiding spilling | |
6073 | of the vector we extract from. Compared to the | |
6074 | construction case in vectorizable_load no store-forwarding | |
6075 | issue exists here for reasonable archs. */ | |
6076 | } | |
b17dc4d4 RB |
6077 | } |
6078 | else if (group_size >= nunits | |
6079 | && group_size % nunits == 0) | |
6080 | { | |
6081 | nstores = 1; | |
6082 | lnel = nunits; | |
6083 | ltype = vectype; | |
04199738 | 6084 | lvectype = vectype; |
b17dc4d4 | 6085 | } |
cee62fee MM |
6086 | ltype = build_aligned_type (ltype, TYPE_ALIGN (elem_type)); |
6087 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
6088 | } | |
6089 | ||
f2e2a985 MM |
6090 | ivstep = stride_step; |
6091 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
b17dc4d4 | 6092 | build_int_cst (TREE_TYPE (ivstep), vf)); |
f2e2a985 MM |
6093 | |
6094 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
6095 | ||
6096 | create_iv (stride_base, ivstep, NULL, | |
6097 | loop, &incr_gsi, insert_after, | |
6098 | &offvar, NULL); | |
6099 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 6100 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
f2e2a985 MM |
6101 | |
6102 | stride_step = force_gimple_operand (stride_step, &stmts, true, NULL_TREE); | |
6103 | if (stmts) | |
6104 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
6105 | ||
6106 | prev_stmt_info = NULL; | |
44fc7854 | 6107 | alias_off = build_int_cst (ref_type, 0); |
f502d50e MM |
6108 | next_stmt = first_stmt; |
6109 | for (g = 0; g < group_size; g++) | |
f2e2a985 | 6110 | { |
f502d50e MM |
6111 | running_off = offvar; |
6112 | if (g) | |
f2e2a985 | 6113 | { |
f502d50e MM |
6114 | tree size = TYPE_SIZE_UNIT (ltype); |
6115 | tree pos = fold_build2 (MULT_EXPR, sizetype, size_int (g), | |
f2e2a985 | 6116 | size); |
f502d50e | 6117 | tree newoff = copy_ssa_name (running_off, NULL); |
f2e2a985 | 6118 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
f502d50e | 6119 | running_off, pos); |
f2e2a985 | 6120 | vect_finish_stmt_generation (stmt, incr, gsi); |
f2e2a985 | 6121 | running_off = newoff; |
f502d50e | 6122 | } |
b17dc4d4 RB |
6123 | unsigned int group_el = 0; |
6124 | unsigned HOST_WIDE_INT | |
6125 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
f502d50e MM |
6126 | for (j = 0; j < ncopies; j++) |
6127 | { | |
6128 | /* We've set op and dt above, from gimple_assign_rhs1(stmt), | |
6129 | and first_stmt == stmt. */ | |
6130 | if (j == 0) | |
6131 | { | |
6132 | if (slp) | |
6133 | { | |
6134 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, NULL, | |
306b0c92 | 6135 | slp_node); |
f502d50e MM |
6136 | vec_oprnd = vec_oprnds[0]; |
6137 | } | |
6138 | else | |
6139 | { | |
6140 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
6141 | op = gimple_assign_rhs1 (next_stmt); | |
81c40241 | 6142 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
f502d50e MM |
6143 | } |
6144 | } | |
f2e2a985 | 6145 | else |
f502d50e MM |
6146 | { |
6147 | if (slp) | |
6148 | vec_oprnd = vec_oprnds[j]; | |
6149 | else | |
c079cbac | 6150 | { |
81c40241 | 6151 | vect_is_simple_use (vec_oprnd, vinfo, &def_stmt, &dt); |
c079cbac RB |
6152 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); |
6153 | } | |
f502d50e | 6154 | } |
04199738 RB |
6155 | /* Pun the vector to extract from if necessary. */ |
6156 | if (lvectype != vectype) | |
6157 | { | |
6158 | tree tem = make_ssa_name (lvectype); | |
6159 | gimple *pun | |
6160 | = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
6161 | lvectype, vec_oprnd)); | |
6162 | vect_finish_stmt_generation (stmt, pun, gsi); | |
6163 | vec_oprnd = tem; | |
6164 | } | |
f502d50e MM |
6165 | for (i = 0; i < nstores; i++) |
6166 | { | |
6167 | tree newref, newoff; | |
355fe088 | 6168 | gimple *incr, *assign; |
f502d50e MM |
6169 | tree size = TYPE_SIZE (ltype); |
6170 | /* Extract the i'th component. */ | |
6171 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, | |
6172 | bitsize_int (i), size); | |
6173 | tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd, | |
6174 | size, pos); | |
6175 | ||
6176 | elem = force_gimple_operand_gsi (gsi, elem, true, | |
6177 | NULL_TREE, true, | |
6178 | GSI_SAME_STMT); | |
6179 | ||
b17dc4d4 RB |
6180 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
6181 | group_el * elsz); | |
f502d50e | 6182 | newref = build2 (MEM_REF, ltype, |
b17dc4d4 | 6183 | running_off, this_off); |
f502d50e MM |
6184 | |
6185 | /* And store it to *running_off. */ | |
6186 | assign = gimple_build_assign (newref, elem); | |
6187 | vect_finish_stmt_generation (stmt, assign, gsi); | |
6188 | ||
b17dc4d4 RB |
6189 | group_el += lnel; |
6190 | if (! slp | |
6191 | || group_el == group_size) | |
6192 | { | |
6193 | newoff = copy_ssa_name (running_off, NULL); | |
6194 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
6195 | running_off, stride_step); | |
6196 | vect_finish_stmt_generation (stmt, incr, gsi); | |
f502d50e | 6197 | |
b17dc4d4 RB |
6198 | running_off = newoff; |
6199 | group_el = 0; | |
6200 | } | |
225ce44b RB |
6201 | if (g == group_size - 1 |
6202 | && !slp) | |
f502d50e MM |
6203 | { |
6204 | if (j == 0 && i == 0) | |
225ce44b RB |
6205 | STMT_VINFO_VEC_STMT (stmt_info) |
6206 | = *vec_stmt = assign; | |
f502d50e MM |
6207 | else |
6208 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = assign; | |
6209 | prev_stmt_info = vinfo_for_stmt (assign); | |
6210 | } | |
6211 | } | |
f2e2a985 | 6212 | } |
f502d50e | 6213 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
b17dc4d4 RB |
6214 | if (slp) |
6215 | break; | |
f2e2a985 | 6216 | } |
778dd3b6 RB |
6217 | |
6218 | vec_oprnds.release (); | |
f2e2a985 MM |
6219 | return true; |
6220 | } | |
6221 | ||
8c681247 | 6222 | auto_vec<tree> dr_chain (group_size); |
9771b263 | 6223 | oprnds.create (group_size); |
ebfd146a | 6224 | |
720f5239 | 6225 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 6226 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
6227 | /* Targets with store-lane instructions must not require explicit |
6228 | realignment. */ | |
2de001ee | 6229 | gcc_assert (memory_access_type != VMAT_LOAD_STORE_LANES |
272c6793 RS |
6230 | || alignment_support_scheme == dr_aligned |
6231 | || alignment_support_scheme == dr_unaligned_supported); | |
6232 | ||
62da9e14 RS |
6233 | if (memory_access_type == VMAT_CONTIGUOUS_DOWN |
6234 | || memory_access_type == VMAT_CONTIGUOUS_REVERSE) | |
09dfa495 BM |
6235 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
6236 | ||
2de001ee | 6237 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
272c6793 RS |
6238 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); |
6239 | else | |
6240 | aggr_type = vectype; | |
ebfd146a IR |
6241 | |
6242 | /* In case the vectorization factor (VF) is bigger than the number | |
6243 | of elements that we can fit in a vectype (nunits), we have to generate | |
6244 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 6245 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
6246 | vect_get_vec_def_for_copy_stmt. */ |
6247 | ||
0d0293ac | 6248 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
6249 | |
6250 | S1: &base + 2 = x2 | |
6251 | S2: &base = x0 | |
6252 | S3: &base + 1 = x1 | |
6253 | S4: &base + 3 = x3 | |
6254 | ||
6255 | We create vectorized stores starting from base address (the access of the | |
6256 | first stmt in the chain (S2 in the above example), when the last store stmt | |
6257 | of the chain (S4) is reached: | |
6258 | ||
6259 | VS1: &base = vx2 | |
6260 | VS2: &base + vec_size*1 = vx0 | |
6261 | VS3: &base + vec_size*2 = vx1 | |
6262 | VS4: &base + vec_size*3 = vx3 | |
6263 | ||
6264 | Then permutation statements are generated: | |
6265 | ||
3fcc1b55 JJ |
6266 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
6267 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 6268 | ... |
b8698a0f | 6269 | |
ebfd146a IR |
6270 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
6271 | (the order of the data-refs in the output of vect_permute_store_chain | |
6272 | corresponds to the order of scalar stmts in the interleaving chain - see | |
6273 | the documentation of vect_permute_store_chain()). | |
6274 | ||
6275 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 6276 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 6277 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 6278 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
6279 | */ |
6280 | ||
6281 | prev_stmt_info = NULL; | |
6282 | for (j = 0; j < ncopies; j++) | |
6283 | { | |
ebfd146a IR |
6284 | |
6285 | if (j == 0) | |
6286 | { | |
6287 | if (slp) | |
6288 | { | |
6289 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c | 6290 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
306b0c92 | 6291 | NULL, slp_node); |
ebfd146a | 6292 | |
9771b263 | 6293 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
6294 | } |
6295 | else | |
6296 | { | |
b8698a0f L |
6297 | /* For interleaved stores we collect vectorized defs for all the |
6298 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
6299 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
6300 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
6301 | ||
0d0293ac | 6302 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 6303 | OPRNDS are of size 1. */ |
b8698a0f | 6304 | next_stmt = first_stmt; |
ebfd146a IR |
6305 | for (i = 0; i < group_size; i++) |
6306 | { | |
b8698a0f L |
6307 | /* Since gaps are not supported for interleaved stores, |
6308 | GROUP_SIZE is the exact number of stmts in the chain. | |
6309 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
6310 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
6311 | iteration of the loop will be executed. */ |
6312 | gcc_assert (next_stmt | |
6313 | && gimple_assign_single_p (next_stmt)); | |
6314 | op = gimple_assign_rhs1 (next_stmt); | |
6315 | ||
81c40241 | 6316 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
9771b263 DN |
6317 | dr_chain.quick_push (vec_oprnd); |
6318 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 6319 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
6320 | } |
6321 | } | |
6322 | ||
6323 | /* We should have catched mismatched types earlier. */ | |
6324 | gcc_assert (useless_type_conversion_p (vectype, | |
6325 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
6326 | bool simd_lane_access_p |
6327 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
6328 | if (simd_lane_access_p | |
6329 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
6330 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
6331 | && integer_zerop (DR_OFFSET (first_dr)) | |
6332 | && integer_zerop (DR_INIT (first_dr)) | |
6333 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
44fc7854 | 6334 | get_alias_set (TREE_TYPE (ref_type)))) |
74bf76ed JJ |
6335 | { |
6336 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
44fc7854 | 6337 | dataref_offset = build_int_cst (ref_type, 0); |
8928eff3 | 6338 | inv_p = false; |
74bf76ed JJ |
6339 | } |
6340 | else | |
6341 | dataref_ptr | |
6342 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
6343 | simd_lane_access_p ? loop : NULL, | |
09dfa495 | 6344 | offset, &dummy, gsi, &ptr_incr, |
74bf76ed | 6345 | simd_lane_access_p, &inv_p); |
a70d6342 | 6346 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 6347 | } |
b8698a0f | 6348 | else |
ebfd146a | 6349 | { |
b8698a0f L |
6350 | /* For interleaved stores we created vectorized defs for all the |
6351 | defs stored in OPRNDS in the previous iteration (previous copy). | |
6352 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
6353 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
6354 | next copy. | |
0d0293ac | 6355 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
6356 | OPRNDS are of size 1. */ |
6357 | for (i = 0; i < group_size; i++) | |
6358 | { | |
9771b263 | 6359 | op = oprnds[i]; |
81c40241 | 6360 | vect_is_simple_use (op, vinfo, &def_stmt, &dt); |
b8698a0f | 6361 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
9771b263 DN |
6362 | dr_chain[i] = vec_oprnd; |
6363 | oprnds[i] = vec_oprnd; | |
ebfd146a | 6364 | } |
74bf76ed JJ |
6365 | if (dataref_offset) |
6366 | dataref_offset | |
6367 | = int_const_binop (PLUS_EXPR, dataref_offset, | |
6368 | TYPE_SIZE_UNIT (aggr_type)); | |
6369 | else | |
6370 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
6371 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
6372 | } |
6373 | ||
2de001ee | 6374 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
ebfd146a | 6375 | { |
272c6793 | 6376 | tree vec_array; |
267d3070 | 6377 | |
272c6793 RS |
6378 | /* Combine all the vectors into an array. */ |
6379 | vec_array = create_vector_array (vectype, vec_num); | |
6380 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 6381 | { |
9771b263 | 6382 | vec_oprnd = dr_chain[i]; |
272c6793 | 6383 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 6384 | } |
b8698a0f | 6385 | |
272c6793 RS |
6386 | /* Emit: |
6387 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
44fc7854 | 6388 | data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type); |
a844293d RS |
6389 | gcall *call = gimple_build_call_internal (IFN_STORE_LANES, 1, |
6390 | vec_array); | |
6391 | gimple_call_set_lhs (call, data_ref); | |
6392 | gimple_call_set_nothrow (call, true); | |
6393 | new_stmt = call; | |
267d3070 | 6394 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
6395 | } |
6396 | else | |
6397 | { | |
6398 | new_stmt = NULL; | |
0d0293ac | 6399 | if (grouped_store) |
272c6793 | 6400 | { |
b6b9227d JJ |
6401 | if (j == 0) |
6402 | result_chain.create (group_size); | |
272c6793 RS |
6403 | /* Permute. */ |
6404 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
6405 | &result_chain); | |
6406 | } | |
c2d7ab2a | 6407 | |
272c6793 RS |
6408 | next_stmt = first_stmt; |
6409 | for (i = 0; i < vec_num; i++) | |
6410 | { | |
644ffefd | 6411 | unsigned align, misalign; |
272c6793 RS |
6412 | |
6413 | if (i > 0) | |
6414 | /* Bump the vector pointer. */ | |
6415 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
6416 | stmt, NULL_TREE); | |
6417 | ||
6418 | if (slp) | |
9771b263 | 6419 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
6420 | else if (grouped_store) |
6421 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 6422 | vect_permute_store_chain(). */ |
9771b263 | 6423 | vec_oprnd = result_chain[i]; |
272c6793 | 6424 | |
69a2e8a1 | 6425 | data_ref = fold_build2 (MEM_REF, vectype, |
aed93b23 RB |
6426 | dataref_ptr, |
6427 | dataref_offset | |
6428 | ? dataref_offset | |
44fc7854 | 6429 | : build_int_cst (ref_type, 0)); |
644ffefd | 6430 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 | 6431 | if (aligned_access_p (first_dr)) |
644ffefd | 6432 | misalign = 0; |
272c6793 RS |
6433 | else if (DR_MISALIGNMENT (first_dr) == -1) |
6434 | { | |
25f68d90 | 6435 | align = dr_alignment (vect_dr_behavior (first_dr)); |
52639a61 | 6436 | misalign = 0; |
272c6793 RS |
6437 | TREE_TYPE (data_ref) |
6438 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 6439 | align * BITS_PER_UNIT); |
272c6793 RS |
6440 | } |
6441 | else | |
6442 | { | |
6443 | TREE_TYPE (data_ref) | |
6444 | = build_aligned_type (TREE_TYPE (data_ref), | |
6445 | TYPE_ALIGN (elem_type)); | |
644ffefd | 6446 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 6447 | } |
aed93b23 RB |
6448 | if (dataref_offset == NULL_TREE |
6449 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
6450 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
6451 | misalign); | |
c2d7ab2a | 6452 | |
62da9e14 | 6453 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
09dfa495 BM |
6454 | { |
6455 | tree perm_mask = perm_mask_for_reverse (vectype); | |
6456 | tree perm_dest | |
6457 | = vect_create_destination_var (gimple_assign_rhs1 (stmt), | |
6458 | vectype); | |
b731b390 | 6459 | tree new_temp = make_ssa_name (perm_dest); |
09dfa495 BM |
6460 | |
6461 | /* Generate the permute statement. */ | |
355fe088 | 6462 | gimple *perm_stmt |
0d0e4a03 JJ |
6463 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
6464 | vec_oprnd, perm_mask); | |
09dfa495 BM |
6465 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
6466 | ||
6467 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); | |
6468 | vec_oprnd = new_temp; | |
6469 | } | |
6470 | ||
272c6793 RS |
6471 | /* Arguments are ready. Create the new vector stmt. */ |
6472 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
6473 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
6474 | |
6475 | if (slp) | |
6476 | continue; | |
6477 | ||
e14c1050 | 6478 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
6479 | if (!next_stmt) |
6480 | break; | |
6481 | } | |
ebfd146a | 6482 | } |
1da0876c RS |
6483 | if (!slp) |
6484 | { | |
6485 | if (j == 0) | |
6486 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6487 | else | |
6488 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6489 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6490 | } | |
ebfd146a IR |
6491 | } |
6492 | ||
9771b263 DN |
6493 | oprnds.release (); |
6494 | result_chain.release (); | |
6495 | vec_oprnds.release (); | |
ebfd146a IR |
6496 | |
6497 | return true; | |
6498 | } | |
6499 | ||
557be5a8 AL |
6500 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
6501 | VECTOR_CST mask. No checks are made that the target platform supports the | |
6502 | mask, so callers may wish to test can_vec_perm_p separately, or use | |
6503 | vect_gen_perm_mask_checked. */ | |
a1e53f3f | 6504 | |
3fcc1b55 | 6505 | tree |
908a1a16 | 6506 | vect_gen_perm_mask_any (tree vectype, vec_perm_indices sel) |
a1e53f3f | 6507 | { |
794e3180 | 6508 | tree mask_elt_type, mask_type, mask_vec; |
a1e53f3f | 6509 | |
908a1a16 RS |
6510 | unsigned int nunits = sel.length (); |
6511 | gcc_checking_assert (nunits == TYPE_VECTOR_SUBPARTS (vectype)); | |
22e4dee7 | 6512 | |
96f9265a | 6513 | mask_elt_type = lang_hooks.types.type_for_mode |
304b9962 | 6514 | (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype))).require (), 1); |
22e4dee7 | 6515 | mask_type = get_vectype_for_scalar_type (mask_elt_type); |
a1e53f3f | 6516 | |
794e3180 | 6517 | auto_vec<tree, 32> mask_elts (nunits); |
908a1a16 | 6518 | for (unsigned int i = 0; i < nunits; ++i) |
794e3180 | 6519 | mask_elts.quick_push (build_int_cst (mask_elt_type, sel[i])); |
d2a12ae7 | 6520 | mask_vec = build_vector (mask_type, mask_elts); |
a1e53f3f | 6521 | |
2635892a | 6522 | return mask_vec; |
a1e53f3f L |
6523 | } |
6524 | ||
cf7aa6a3 AL |
6525 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_p, |
6526 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ | |
557be5a8 AL |
6527 | |
6528 | tree | |
908a1a16 | 6529 | vect_gen_perm_mask_checked (tree vectype, vec_perm_indices sel) |
557be5a8 | 6530 | { |
908a1a16 | 6531 | gcc_assert (can_vec_perm_p (TYPE_MODE (vectype), false, &sel)); |
557be5a8 AL |
6532 | return vect_gen_perm_mask_any (vectype, sel); |
6533 | } | |
6534 | ||
aec7ae7d JJ |
6535 | /* Given a vector variable X and Y, that was generated for the scalar |
6536 | STMT, generate instructions to permute the vector elements of X and Y | |
6537 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
6538 | permuted vector variable. */ | |
a1e53f3f L |
6539 | |
6540 | static tree | |
355fe088 | 6541 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple *stmt, |
aec7ae7d | 6542 | gimple_stmt_iterator *gsi) |
a1e53f3f L |
6543 | { |
6544 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 6545 | tree perm_dest, data_ref; |
355fe088 | 6546 | gimple *perm_stmt; |
a1e53f3f | 6547 | |
acdcd61b | 6548 | perm_dest = vect_create_destination_var (gimple_get_lhs (stmt), vectype); |
b731b390 | 6549 | data_ref = make_ssa_name (perm_dest); |
a1e53f3f L |
6550 | |
6551 | /* Generate the permute statement. */ | |
0d0e4a03 | 6552 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
a1e53f3f L |
6553 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
6554 | ||
6555 | return data_ref; | |
6556 | } | |
6557 | ||
6b916b36 RB |
6558 | /* Hoist the definitions of all SSA uses on STMT out of the loop LOOP, |
6559 | inserting them on the loops preheader edge. Returns true if we | |
6560 | were successful in doing so (and thus STMT can be moved then), | |
6561 | otherwise returns false. */ | |
6562 | ||
6563 | static bool | |
355fe088 | 6564 | hoist_defs_of_uses (gimple *stmt, struct loop *loop) |
6b916b36 RB |
6565 | { |
6566 | ssa_op_iter i; | |
6567 | tree op; | |
6568 | bool any = false; | |
6569 | ||
6570 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6571 | { | |
355fe088 | 6572 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6573 | if (!gimple_nop_p (def_stmt) |
6574 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6575 | { | |
6576 | /* Make sure we don't need to recurse. While we could do | |
6577 | so in simple cases when there are more complex use webs | |
6578 | we don't have an easy way to preserve stmt order to fulfil | |
6579 | dependencies within them. */ | |
6580 | tree op2; | |
6581 | ssa_op_iter i2; | |
d1417442 JJ |
6582 | if (gimple_code (def_stmt) == GIMPLE_PHI) |
6583 | return false; | |
6b916b36 RB |
6584 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
6585 | { | |
355fe088 | 6586 | gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2); |
6b916b36 RB |
6587 | if (!gimple_nop_p (def_stmt2) |
6588 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2))) | |
6589 | return false; | |
6590 | } | |
6591 | any = true; | |
6592 | } | |
6593 | } | |
6594 | ||
6595 | if (!any) | |
6596 | return true; | |
6597 | ||
6598 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6599 | { | |
355fe088 | 6600 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6601 | if (!gimple_nop_p (def_stmt) |
6602 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6603 | { | |
6604 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); | |
6605 | gsi_remove (&gsi, false); | |
6606 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); | |
6607 | } | |
6608 | } | |
6609 | ||
6610 | return true; | |
6611 | } | |
6612 | ||
ebfd146a IR |
6613 | /* vectorizable_load. |
6614 | ||
b8698a0f L |
6615 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
6616 | can be vectorized. | |
6617 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
6618 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
6619 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
6620 | ||
6621 | static bool | |
355fe088 | 6622 | vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 6623 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a IR |
6624 | { |
6625 | tree scalar_dest; | |
6626 | tree vec_dest = NULL; | |
6627 | tree data_ref = NULL; | |
6628 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 6629 | stmt_vec_info prev_stmt_info; |
ebfd146a | 6630 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 6631 | struct loop *loop = NULL; |
ebfd146a | 6632 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 6633 | bool nested_in_vect_loop = false; |
c716e67f | 6634 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
272c6793 | 6635 | tree elem_type; |
ebfd146a | 6636 | tree new_temp; |
ef4bddc2 | 6637 | machine_mode mode; |
355fe088 | 6638 | gimple *new_stmt = NULL; |
ebfd146a IR |
6639 | tree dummy; |
6640 | enum dr_alignment_support alignment_support_scheme; | |
6641 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 6642 | tree dataref_offset = NULL_TREE; |
355fe088 | 6643 | gimple *ptr_incr = NULL; |
ebfd146a | 6644 | int ncopies; |
44fc7854 | 6645 | int i, j, group_size, group_gap_adj; |
ebfd146a IR |
6646 | tree msq = NULL_TREE, lsq; |
6647 | tree offset = NULL_TREE; | |
356bbc4c | 6648 | tree byte_offset = NULL_TREE; |
ebfd146a | 6649 | tree realignment_token = NULL_TREE; |
538dd0b7 | 6650 | gphi *phi = NULL; |
6e1aa848 | 6651 | vec<tree> dr_chain = vNULL; |
0d0293ac | 6652 | bool grouped_load = false; |
355fe088 | 6653 | gimple *first_stmt; |
4f0a0218 | 6654 | gimple *first_stmt_for_drptr = NULL; |
ebfd146a IR |
6655 | bool inv_p; |
6656 | bool compute_in_loop = false; | |
6657 | struct loop *at_loop; | |
6658 | int vec_num; | |
6659 | bool slp = (slp_node != NULL); | |
6660 | bool slp_perm = false; | |
6661 | enum tree_code code; | |
a70d6342 IR |
6662 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
6663 | int vf; | |
272c6793 | 6664 | tree aggr_type; |
134c85ca | 6665 | gather_scatter_info gs_info; |
310213d4 | 6666 | vec_info *vinfo = stmt_info->vinfo; |
44fc7854 | 6667 | tree ref_type; |
a70d6342 | 6668 | |
465c8c19 JJ |
6669 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
6670 | return false; | |
6671 | ||
66c16fd9 RB |
6672 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
6673 | && ! vec_stmt) | |
465c8c19 JJ |
6674 | return false; |
6675 | ||
6676 | /* Is vectorizable load? */ | |
6677 | if (!is_gimple_assign (stmt)) | |
6678 | return false; | |
6679 | ||
6680 | scalar_dest = gimple_assign_lhs (stmt); | |
6681 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
6682 | return false; | |
6683 | ||
6684 | code = gimple_assign_rhs_code (stmt); | |
6685 | if (code != ARRAY_REF | |
6686 | && code != BIT_FIELD_REF | |
6687 | && code != INDIRECT_REF | |
6688 | && code != COMPONENT_REF | |
6689 | && code != IMAGPART_EXPR | |
6690 | && code != REALPART_EXPR | |
6691 | && code != MEM_REF | |
6692 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
6693 | return false; | |
6694 | ||
6695 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
6696 | return false; | |
6697 | ||
6698 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
6699 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
6700 | ||
a70d6342 IR |
6701 | if (loop_vinfo) |
6702 | { | |
6703 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
6704 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
6705 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
6706 | } | |
6707 | else | |
3533e503 | 6708 | vf = 1; |
ebfd146a IR |
6709 | |
6710 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 6711 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 6712 | case of SLP. */ |
fce57248 | 6713 | if (slp) |
ebfd146a IR |
6714 | ncopies = 1; |
6715 | else | |
6716 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
6717 | ||
6718 | gcc_assert (ncopies >= 1); | |
6719 | ||
6720 | /* FORNOW. This restriction should be relaxed. */ | |
6721 | if (nested_in_vect_loop && ncopies > 1) | |
6722 | { | |
73fbfcad | 6723 | if (dump_enabled_p ()) |
78c60e3d | 6724 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6725 | "multiple types in nested loop.\n"); |
ebfd146a IR |
6726 | return false; |
6727 | } | |
6728 | ||
f2556b68 RB |
6729 | /* Invalidate assumptions made by dependence analysis when vectorization |
6730 | on the unrolled body effectively re-orders stmts. */ | |
6731 | if (ncopies > 1 | |
6732 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
6733 | && ((unsigned)LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
6734 | > STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
6735 | { | |
6736 | if (dump_enabled_p ()) | |
6737 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6738 | "cannot perform implicit CSE when unrolling " | |
6739 | "with negative dependence distance\n"); | |
6740 | return false; | |
6741 | } | |
6742 | ||
7b7b1813 | 6743 | elem_type = TREE_TYPE (vectype); |
947131ba | 6744 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
6745 | |
6746 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
6747 | (e.g. - data copies). */ | |
947131ba | 6748 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 6749 | { |
73fbfcad | 6750 | if (dump_enabled_p ()) |
78c60e3d | 6751 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6752 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
6753 | return false; |
6754 | } | |
6755 | ||
ebfd146a | 6756 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 6757 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 6758 | { |
0d0293ac | 6759 | grouped_load = true; |
ebfd146a | 6760 | /* FORNOW */ |
2de001ee RS |
6761 | gcc_assert (!nested_in_vect_loop); |
6762 | gcc_assert (!STMT_VINFO_GATHER_SCATTER_P (stmt_info)); | |
ebfd146a | 6763 | |
e14c1050 | 6764 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
d3465d72 | 6765 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
d5f035ea | 6766 | |
b1af7da6 RB |
6767 | if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
6768 | slp_perm = true; | |
6769 | ||
f2556b68 RB |
6770 | /* Invalidate assumptions made by dependence analysis when vectorization |
6771 | on the unrolled body effectively re-orders stmts. */ | |
6772 | if (!PURE_SLP_STMT (stmt_info) | |
6773 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
6774 | && ((unsigned)LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
6775 | > STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
6776 | { | |
6777 | if (dump_enabled_p ()) | |
6778 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6779 | "cannot perform implicit CSE when performing " | |
6780 | "group loads with negative dependence distance\n"); | |
6781 | return false; | |
6782 | } | |
96bb56b2 RB |
6783 | |
6784 | /* Similarly when the stmt is a load that is both part of a SLP | |
6785 | instance and a loop vectorized stmt via the same-dr mechanism | |
6786 | we have to give up. */ | |
6787 | if (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info) | |
6788 | && (STMT_SLP_TYPE (stmt_info) | |
6789 | != STMT_SLP_TYPE (vinfo_for_stmt | |
6790 | (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info))))) | |
6791 | { | |
6792 | if (dump_enabled_p ()) | |
6793 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6794 | "conflicting SLP types for CSEd load\n"); | |
6795 | return false; | |
6796 | } | |
ebfd146a IR |
6797 | } |
6798 | ||
2de001ee | 6799 | vect_memory_access_type memory_access_type; |
62da9e14 | 6800 | if (!get_load_store_type (stmt, vectype, slp, VLS_LOAD, ncopies, |
2de001ee RS |
6801 | &memory_access_type, &gs_info)) |
6802 | return false; | |
a1e53f3f | 6803 | |
ebfd146a IR |
6804 | if (!vec_stmt) /* transformation not required. */ |
6805 | { | |
2de001ee RS |
6806 | if (!slp) |
6807 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; | |
ebfd146a | 6808 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; |
2e8ab70c RB |
6809 | /* The SLP costs are calculated during SLP analysis. */ |
6810 | if (!PURE_SLP_STMT (stmt_info)) | |
2de001ee | 6811 | vect_model_load_cost (stmt_info, ncopies, memory_access_type, |
2e8ab70c | 6812 | NULL, NULL, NULL); |
ebfd146a IR |
6813 | return true; |
6814 | } | |
6815 | ||
2de001ee RS |
6816 | if (!slp) |
6817 | gcc_assert (memory_access_type | |
6818 | == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); | |
6819 | ||
73fbfcad | 6820 | if (dump_enabled_p ()) |
78c60e3d | 6821 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6822 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a | 6823 | |
67b8dbac | 6824 | /* Transform. */ |
ebfd146a | 6825 | |
c716e67f XDL |
6826 | ensure_base_align (stmt_info, dr); |
6827 | ||
2de001ee | 6828 | if (memory_access_type == VMAT_GATHER_SCATTER) |
aec7ae7d JJ |
6829 | { |
6830 | tree vec_oprnd0 = NULL_TREE, op; | |
134c85ca | 6831 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); |
aec7ae7d | 6832 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; |
d3c2fee0 | 6833 | tree ptr, mask, var, scale, merge, perm_mask = NULL_TREE, prev_res = NULL_TREE; |
aec7ae7d JJ |
6834 | edge pe = loop_preheader_edge (loop); |
6835 | gimple_seq seq; | |
6836 | basic_block new_bb; | |
6837 | enum { NARROW, NONE, WIDEN } modifier; | |
134c85ca | 6838 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gs_info.offset_vectype); |
aec7ae7d JJ |
6839 | |
6840 | if (nunits == gather_off_nunits) | |
6841 | modifier = NONE; | |
6842 | else if (nunits == gather_off_nunits / 2) | |
6843 | { | |
aec7ae7d JJ |
6844 | modifier = WIDEN; |
6845 | ||
908a1a16 | 6846 | auto_vec_perm_indices sel (gather_off_nunits); |
aec7ae7d | 6847 | for (i = 0; i < gather_off_nunits; ++i) |
908a1a16 | 6848 | sel.quick_push (i | nunits); |
aec7ae7d | 6849 | |
134c85ca | 6850 | perm_mask = vect_gen_perm_mask_checked (gs_info.offset_vectype, sel); |
aec7ae7d JJ |
6851 | } |
6852 | else if (nunits == gather_off_nunits * 2) | |
6853 | { | |
aec7ae7d JJ |
6854 | modifier = NARROW; |
6855 | ||
908a1a16 | 6856 | auto_vec_perm_indices sel (nunits); |
aec7ae7d | 6857 | for (i = 0; i < nunits; ++i) |
908a1a16 RS |
6858 | sel.quick_push (i < gather_off_nunits |
6859 | ? i : i + nunits - gather_off_nunits); | |
aec7ae7d | 6860 | |
557be5a8 | 6861 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); |
aec7ae7d JJ |
6862 | ncopies *= 2; |
6863 | } | |
6864 | else | |
6865 | gcc_unreachable (); | |
6866 | ||
134c85ca | 6867 | rettype = TREE_TYPE (TREE_TYPE (gs_info.decl)); |
aec7ae7d JJ |
6868 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
6869 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6870 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6871 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6872 | scaletype = TREE_VALUE (arglist); | |
d3c2fee0 | 6873 | gcc_checking_assert (types_compatible_p (srctype, rettype)); |
aec7ae7d JJ |
6874 | |
6875 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
6876 | ||
134c85ca | 6877 | ptr = fold_convert (ptrtype, gs_info.base); |
aec7ae7d JJ |
6878 | if (!is_gimple_min_invariant (ptr)) |
6879 | { | |
6880 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
6881 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
6882 | gcc_assert (!new_bb); | |
6883 | } | |
6884 | ||
6885 | /* Currently we support only unconditional gather loads, | |
6886 | so mask should be all ones. */ | |
d3c2fee0 AI |
6887 | if (TREE_CODE (masktype) == INTEGER_TYPE) |
6888 | mask = build_int_cst (masktype, -1); | |
6889 | else if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
6890 | { | |
6891 | mask = build_int_cst (TREE_TYPE (masktype), -1); | |
6892 | mask = build_vector_from_val (masktype, mask); | |
03b9e8e4 | 6893 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
d3c2fee0 | 6894 | } |
aec7ae7d JJ |
6895 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) |
6896 | { | |
6897 | REAL_VALUE_TYPE r; | |
6898 | long tmp[6]; | |
6899 | for (j = 0; j < 6; ++j) | |
6900 | tmp[j] = -1; | |
6901 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
6902 | mask = build_real (TREE_TYPE (masktype), r); | |
d3c2fee0 | 6903 | mask = build_vector_from_val (masktype, mask); |
03b9e8e4 | 6904 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
aec7ae7d JJ |
6905 | } |
6906 | else | |
6907 | gcc_unreachable (); | |
aec7ae7d | 6908 | |
134c85ca | 6909 | scale = build_int_cst (scaletype, gs_info.scale); |
aec7ae7d | 6910 | |
d3c2fee0 AI |
6911 | if (TREE_CODE (TREE_TYPE (rettype)) == INTEGER_TYPE) |
6912 | merge = build_int_cst (TREE_TYPE (rettype), 0); | |
6913 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (rettype))) | |
6914 | { | |
6915 | REAL_VALUE_TYPE r; | |
6916 | long tmp[6]; | |
6917 | for (j = 0; j < 6; ++j) | |
6918 | tmp[j] = 0; | |
6919 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (rettype))); | |
6920 | merge = build_real (TREE_TYPE (rettype), r); | |
6921 | } | |
6922 | else | |
6923 | gcc_unreachable (); | |
6924 | merge = build_vector_from_val (rettype, merge); | |
6925 | merge = vect_init_vector (stmt, merge, rettype, NULL); | |
6926 | ||
aec7ae7d JJ |
6927 | prev_stmt_info = NULL; |
6928 | for (j = 0; j < ncopies; ++j) | |
6929 | { | |
6930 | if (modifier == WIDEN && (j & 1)) | |
6931 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
6932 | perm_mask, stmt, gsi); | |
6933 | else if (j == 0) | |
6934 | op = vec_oprnd0 | |
134c85ca | 6935 | = vect_get_vec_def_for_operand (gs_info.offset, stmt); |
aec7ae7d JJ |
6936 | else |
6937 | op = vec_oprnd0 | |
134c85ca | 6938 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, vec_oprnd0); |
aec7ae7d JJ |
6939 | |
6940 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
6941 | { | |
6942 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
6943 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 6944 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
aec7ae7d JJ |
6945 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
6946 | new_stmt | |
0d0e4a03 | 6947 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
6948 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6949 | op = var; | |
6950 | } | |
6951 | ||
6952 | new_stmt | |
134c85ca | 6953 | = gimple_build_call (gs_info.decl, 5, merge, ptr, op, mask, scale); |
aec7ae7d JJ |
6954 | |
6955 | if (!useless_type_conversion_p (vectype, rettype)) | |
6956 | { | |
6957 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
6958 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 6959 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
aec7ae7d JJ |
6960 | gimple_call_set_lhs (new_stmt, op); |
6961 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 6962 | var = make_ssa_name (vec_dest); |
aec7ae7d JJ |
6963 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
6964 | new_stmt | |
0d0e4a03 | 6965 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
6966 | } |
6967 | else | |
6968 | { | |
6969 | var = make_ssa_name (vec_dest, new_stmt); | |
6970 | gimple_call_set_lhs (new_stmt, var); | |
6971 | } | |
6972 | ||
6973 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6974 | ||
6975 | if (modifier == NARROW) | |
6976 | { | |
6977 | if ((j & 1) == 0) | |
6978 | { | |
6979 | prev_res = var; | |
6980 | continue; | |
6981 | } | |
6982 | var = permute_vec_elements (prev_res, var, | |
6983 | perm_mask, stmt, gsi); | |
6984 | new_stmt = SSA_NAME_DEF_STMT (var); | |
6985 | } | |
6986 | ||
6987 | if (prev_stmt_info == NULL) | |
6988 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6989 | else | |
6990 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6991 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6992 | } | |
6993 | return true; | |
6994 | } | |
2de001ee RS |
6995 | |
6996 | if (memory_access_type == VMAT_ELEMENTWISE | |
6997 | || memory_access_type == VMAT_STRIDED_SLP) | |
7d75abc8 MM |
6998 | { |
6999 | gimple_stmt_iterator incr_gsi; | |
7000 | bool insert_after; | |
355fe088 | 7001 | gimple *incr; |
7d75abc8 | 7002 | tree offvar; |
7d75abc8 MM |
7003 | tree ivstep; |
7004 | tree running_off; | |
9771b263 | 7005 | vec<constructor_elt, va_gc> *v = NULL; |
7d75abc8 | 7006 | gimple_seq stmts = NULL; |
14ac6aa2 RB |
7007 | tree stride_base, stride_step, alias_off; |
7008 | ||
7009 | gcc_assert (!nested_in_vect_loop); | |
7d75abc8 | 7010 | |
f502d50e | 7011 | if (slp && grouped_load) |
44fc7854 BE |
7012 | { |
7013 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); | |
7014 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
7015 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
7016 | ref_type = get_group_alias_ptr_type (first_stmt); | |
7017 | } | |
ab313a8c | 7018 | else |
44fc7854 BE |
7019 | { |
7020 | first_stmt = stmt; | |
7021 | first_dr = dr; | |
7022 | group_size = 1; | |
7023 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); | |
7024 | } | |
ab313a8c | 7025 | |
14ac6aa2 RB |
7026 | stride_base |
7027 | = fold_build_pointer_plus | |
ab313a8c | 7028 | (DR_BASE_ADDRESS (first_dr), |
14ac6aa2 | 7029 | size_binop (PLUS_EXPR, |
ab313a8c RB |
7030 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
7031 | convert_to_ptrofftype (DR_INIT (first_dr)))); | |
7032 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); | |
7d75abc8 MM |
7033 | |
7034 | /* For a load with loop-invariant (but other than power-of-2) | |
7035 | stride (i.e. not a grouped access) like so: | |
7036 | ||
7037 | for (i = 0; i < n; i += stride) | |
7038 | ... = array[i]; | |
7039 | ||
7040 | we generate a new induction variable and new accesses to | |
7041 | form a new vector (or vectors, depending on ncopies): | |
7042 | ||
7043 | for (j = 0; ; j += VF*stride) | |
7044 | tmp1 = array[j]; | |
7045 | tmp2 = array[j + stride]; | |
7046 | ... | |
7047 | vectemp = {tmp1, tmp2, ...} | |
7048 | */ | |
7049 | ||
ab313a8c RB |
7050 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step, |
7051 | build_int_cst (TREE_TYPE (stride_step), vf)); | |
7d75abc8 MM |
7052 | |
7053 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
7054 | ||
ab313a8c | 7055 | create_iv (unshare_expr (stride_base), unshare_expr (ivstep), NULL, |
7d75abc8 MM |
7056 | loop, &incr_gsi, insert_after, |
7057 | &offvar, NULL); | |
7058 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 7059 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
7d75abc8 | 7060 | |
ab313a8c RB |
7061 | stride_step = force_gimple_operand (unshare_expr (stride_step), |
7062 | &stmts, true, NULL_TREE); | |
7d75abc8 MM |
7063 | if (stmts) |
7064 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
7065 | ||
7066 | prev_stmt_info = NULL; | |
7067 | running_off = offvar; | |
44fc7854 | 7068 | alias_off = build_int_cst (ref_type, 0); |
7b5fc413 | 7069 | int nloads = nunits; |
e09b4c37 | 7070 | int lnel = 1; |
7b5fc413 | 7071 | tree ltype = TREE_TYPE (vectype); |
ea60dd34 | 7072 | tree lvectype = vectype; |
b266b968 | 7073 | auto_vec<tree> dr_chain; |
2de001ee | 7074 | if (memory_access_type == VMAT_STRIDED_SLP) |
7b5fc413 | 7075 | { |
2de001ee | 7076 | if (group_size < nunits) |
e09b4c37 | 7077 | { |
ff03930a JJ |
7078 | /* First check if vec_init optab supports construction from |
7079 | vector elts directly. */ | |
b397965c | 7080 | scalar_mode elmode = SCALAR_TYPE_MODE (TREE_TYPE (vectype)); |
9da15d40 RS |
7081 | machine_mode vmode; |
7082 | if (mode_for_vector (elmode, group_size).exists (&vmode) | |
7083 | && VECTOR_MODE_P (vmode) | |
ff03930a JJ |
7084 | && (convert_optab_handler (vec_init_optab, |
7085 | TYPE_MODE (vectype), vmode) | |
7086 | != CODE_FOR_nothing)) | |
ea60dd34 RB |
7087 | { |
7088 | nloads = nunits / group_size; | |
7089 | lnel = group_size; | |
ff03930a JJ |
7090 | ltype = build_vector_type (TREE_TYPE (vectype), group_size); |
7091 | } | |
7092 | else | |
7093 | { | |
7094 | /* Otherwise avoid emitting a constructor of vector elements | |
7095 | by performing the loads using an integer type of the same | |
7096 | size, constructing a vector of those and then | |
7097 | re-interpreting it as the original vector type. | |
7098 | This avoids a huge runtime penalty due to the general | |
7099 | inability to perform store forwarding from smaller stores | |
7100 | to a larger load. */ | |
7101 | unsigned lsize | |
7102 | = group_size * TYPE_PRECISION (TREE_TYPE (vectype)); | |
fffbab82 | 7103 | elmode = int_mode_for_size (lsize, 0).require (); |
ff03930a JJ |
7104 | /* If we can't construct such a vector fall back to |
7105 | element loads of the original vector type. */ | |
9da15d40 RS |
7106 | if (mode_for_vector (elmode, |
7107 | nunits / group_size).exists (&vmode) | |
7108 | && VECTOR_MODE_P (vmode) | |
ff03930a JJ |
7109 | && (convert_optab_handler (vec_init_optab, vmode, elmode) |
7110 | != CODE_FOR_nothing)) | |
7111 | { | |
7112 | nloads = nunits / group_size; | |
7113 | lnel = group_size; | |
7114 | ltype = build_nonstandard_integer_type (lsize, 1); | |
7115 | lvectype = build_vector_type (ltype, nloads); | |
7116 | } | |
ea60dd34 | 7117 | } |
e09b4c37 | 7118 | } |
2de001ee | 7119 | else |
e09b4c37 | 7120 | { |
ea60dd34 | 7121 | nloads = 1; |
e09b4c37 RB |
7122 | lnel = nunits; |
7123 | ltype = vectype; | |
e09b4c37 | 7124 | } |
2de001ee RS |
7125 | ltype = build_aligned_type (ltype, TYPE_ALIGN (TREE_TYPE (vectype))); |
7126 | } | |
7127 | if (slp) | |
7128 | { | |
66c16fd9 RB |
7129 | /* For SLP permutation support we need to load the whole group, |
7130 | not only the number of vector stmts the permutation result | |
7131 | fits in. */ | |
b266b968 | 7132 | if (slp_perm) |
66c16fd9 RB |
7133 | { |
7134 | ncopies = (group_size * vf + nunits - 1) / nunits; | |
7135 | dr_chain.create (ncopies); | |
7136 | } | |
7137 | else | |
7138 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
7b5fc413 | 7139 | } |
e09b4c37 RB |
7140 | int group_el = 0; |
7141 | unsigned HOST_WIDE_INT | |
7142 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
7d75abc8 MM |
7143 | for (j = 0; j < ncopies; j++) |
7144 | { | |
7b5fc413 | 7145 | if (nloads > 1) |
e09b4c37 RB |
7146 | vec_alloc (v, nloads); |
7147 | for (i = 0; i < nloads; i++) | |
7b5fc413 | 7148 | { |
e09b4c37 RB |
7149 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
7150 | group_el * elsz); | |
7151 | new_stmt = gimple_build_assign (make_ssa_name (ltype), | |
7152 | build2 (MEM_REF, ltype, | |
7153 | running_off, this_off)); | |
7154 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7155 | if (nloads > 1) | |
7156 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, | |
7157 | gimple_assign_lhs (new_stmt)); | |
7158 | ||
7159 | group_el += lnel; | |
7160 | if (! slp | |
7161 | || group_el == group_size) | |
7b5fc413 | 7162 | { |
e09b4c37 RB |
7163 | tree newoff = copy_ssa_name (running_off); |
7164 | gimple *incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
7165 | running_off, stride_step); | |
7b5fc413 RB |
7166 | vect_finish_stmt_generation (stmt, incr, gsi); |
7167 | ||
7168 | running_off = newoff; | |
e09b4c37 | 7169 | group_el = 0; |
7b5fc413 | 7170 | } |
7b5fc413 | 7171 | } |
e09b4c37 | 7172 | if (nloads > 1) |
7d75abc8 | 7173 | { |
ea60dd34 RB |
7174 | tree vec_inv = build_constructor (lvectype, v); |
7175 | new_temp = vect_init_vector (stmt, vec_inv, lvectype, gsi); | |
e09b4c37 | 7176 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
ea60dd34 RB |
7177 | if (lvectype != vectype) |
7178 | { | |
7179 | new_stmt = gimple_build_assign (make_ssa_name (vectype), | |
7180 | VIEW_CONVERT_EXPR, | |
7181 | build1 (VIEW_CONVERT_EXPR, | |
7182 | vectype, new_temp)); | |
7183 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7184 | } | |
7d75abc8 MM |
7185 | } |
7186 | ||
7b5fc413 | 7187 | if (slp) |
b266b968 | 7188 | { |
b266b968 RB |
7189 | if (slp_perm) |
7190 | dr_chain.quick_push (gimple_assign_lhs (new_stmt)); | |
66c16fd9 RB |
7191 | else |
7192 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
b266b968 | 7193 | } |
7d75abc8 | 7194 | else |
225ce44b RB |
7195 | { |
7196 | if (j == 0) | |
7197 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7198 | else | |
7199 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7200 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7201 | } | |
7d75abc8 | 7202 | } |
b266b968 | 7203 | if (slp_perm) |
29afecdf RB |
7204 | { |
7205 | unsigned n_perms; | |
7206 | vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, | |
7207 | slp_node_instance, false, &n_perms); | |
7208 | } | |
7d75abc8 MM |
7209 | return true; |
7210 | } | |
aec7ae7d | 7211 | |
0d0293ac | 7212 | if (grouped_load) |
ebfd146a | 7213 | { |
e14c1050 | 7214 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
44fc7854 | 7215 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
4f0a0218 | 7216 | /* For SLP vectorization we directly vectorize a subchain |
52eab378 RB |
7217 | without permutation. */ |
7218 | if (slp && ! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
4f0a0218 RB |
7219 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
7220 | /* For BB vectorization always use the first stmt to base | |
7221 | the data ref pointer on. */ | |
7222 | if (bb_vinfo) | |
7223 | first_stmt_for_drptr = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 7224 | |
ebfd146a | 7225 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
7226 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
7227 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
7228 | ??? But we can only do so if there is exactly one | |
7229 | as we have no way to get at the rest. Leave the CSE | |
7230 | opportunity alone. | |
7231 | ??? With the group load eventually participating | |
7232 | in multiple different permutations (having multiple | |
7233 | slp nodes which refer to the same group) the CSE | |
7234 | is even wrong code. See PR56270. */ | |
7235 | && !slp) | |
ebfd146a IR |
7236 | { |
7237 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
7238 | return true; | |
7239 | } | |
7240 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
9b999e8c | 7241 | group_gap_adj = 0; |
ebfd146a IR |
7242 | |
7243 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
7244 | if (slp) | |
7245 | { | |
0d0293ac | 7246 | grouped_load = false; |
91ff1504 RB |
7247 | /* For SLP permutation support we need to load the whole group, |
7248 | not only the number of vector stmts the permutation result | |
7249 | fits in. */ | |
7250 | if (slp_perm) | |
b267968e RB |
7251 | { |
7252 | vec_num = (group_size * vf + nunits - 1) / nunits; | |
7253 | group_gap_adj = vf * group_size - nunits * vec_num; | |
7254 | } | |
91ff1504 | 7255 | else |
b267968e RB |
7256 | { |
7257 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
796bd467 RB |
7258 | group_gap_adj |
7259 | = group_size - SLP_INSTANCE_GROUP_SIZE (slp_node_instance); | |
b267968e | 7260 | } |
a70d6342 | 7261 | } |
ebfd146a | 7262 | else |
9b999e8c | 7263 | vec_num = group_size; |
44fc7854 BE |
7264 | |
7265 | ref_type = get_group_alias_ptr_type (first_stmt); | |
ebfd146a IR |
7266 | } |
7267 | else | |
7268 | { | |
7269 | first_stmt = stmt; | |
7270 | first_dr = dr; | |
7271 | group_size = vec_num = 1; | |
9b999e8c | 7272 | group_gap_adj = 0; |
44fc7854 | 7273 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); |
ebfd146a IR |
7274 | } |
7275 | ||
720f5239 | 7276 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 7277 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
7278 | /* Targets with load-lane instructions must not require explicit |
7279 | realignment. */ | |
2de001ee | 7280 | gcc_assert (memory_access_type != VMAT_LOAD_STORE_LANES |
272c6793 RS |
7281 | || alignment_support_scheme == dr_aligned |
7282 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
7283 | |
7284 | /* In case the vectorization factor (VF) is bigger than the number | |
7285 | of elements that we can fit in a vectype (nunits), we have to generate | |
7286 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 7287 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 7288 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 7289 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 7290 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
7291 | stmts that use the defs of the current stmt. The example below |
7292 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
7293 | need to create 4 vectorized stmts): | |
ebfd146a IR |
7294 | |
7295 | before vectorization: | |
7296 | RELATED_STMT VEC_STMT | |
7297 | S1: x = memref - - | |
7298 | S2: z = x + 1 - - | |
7299 | ||
7300 | step 1: vectorize stmt S1: | |
7301 | We first create the vector stmt VS1_0, and, as usual, record a | |
7302 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
7303 | Next, we create the vector stmt VS1_1, and record a pointer to | |
7304 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 7305 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
7306 | stmts and pointers: |
7307 | RELATED_STMT VEC_STMT | |
7308 | VS1_0: vx0 = memref0 VS1_1 - | |
7309 | VS1_1: vx1 = memref1 VS1_2 - | |
7310 | VS1_2: vx2 = memref2 VS1_3 - | |
7311 | VS1_3: vx3 = memref3 - - | |
7312 | S1: x = load - VS1_0 | |
7313 | S2: z = x + 1 - - | |
7314 | ||
b8698a0f L |
7315 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
7316 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
7317 | stmt S2. */ |
7318 | ||
0d0293ac | 7319 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
7320 | |
7321 | S1: x2 = &base + 2 | |
7322 | S2: x0 = &base | |
7323 | S3: x1 = &base + 1 | |
7324 | S4: x3 = &base + 3 | |
7325 | ||
b8698a0f | 7326 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
7327 | starting from the access of the first stmt of the chain: |
7328 | ||
7329 | VS1: vx0 = &base | |
7330 | VS2: vx1 = &base + vec_size*1 | |
7331 | VS3: vx3 = &base + vec_size*2 | |
7332 | VS4: vx4 = &base + vec_size*3 | |
7333 | ||
7334 | Then permutation statements are generated: | |
7335 | ||
e2c83630 RH |
7336 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
7337 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
7338 | ... |
7339 | ||
7340 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
7341 | (the order of the data-refs in the output of vect_permute_load_chain | |
7342 | corresponds to the order of scalar stmts in the interleaving chain - see | |
7343 | the documentation of vect_permute_load_chain()). | |
7344 | The generation of permutation stmts and recording them in | |
0d0293ac | 7345 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 7346 | |
b8698a0f | 7347 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
7348 | permutation stmts above are created for every copy. The result vector |
7349 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
7350 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
7351 | |
7352 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
7353 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
7354 | we generate the following code: | |
7355 | p = initial_addr; | |
7356 | indx = 0; | |
7357 | loop { | |
7358 | p = p + indx * vectype_size; | |
7359 | vec_dest = *(p); | |
7360 | indx = indx + 1; | |
7361 | } | |
7362 | ||
7363 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 7364 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
7365 | then generate the following code, in which the data in each iteration is |
7366 | obtained by two vector loads, one from the previous iteration, and one | |
7367 | from the current iteration: | |
7368 | p1 = initial_addr; | |
7369 | msq_init = *(floor(p1)) | |
7370 | p2 = initial_addr + VS - 1; | |
7371 | realignment_token = call target_builtin; | |
7372 | indx = 0; | |
7373 | loop { | |
7374 | p2 = p2 + indx * vectype_size | |
7375 | lsq = *(floor(p2)) | |
7376 | vec_dest = realign_load (msq, lsq, realignment_token) | |
7377 | indx = indx + 1; | |
7378 | msq = lsq; | |
7379 | } */ | |
7380 | ||
7381 | /* If the misalignment remains the same throughout the execution of the | |
7382 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 7383 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
7384 | This can only occur when vectorizing memory accesses in the inner-loop |
7385 | nested within an outer-loop that is being vectorized. */ | |
7386 | ||
d1e4b493 | 7387 | if (nested_in_vect_loop |
832b4117 | 7388 | && (DR_STEP_ALIGNMENT (dr) % GET_MODE_SIZE (TYPE_MODE (vectype))) != 0) |
ebfd146a IR |
7389 | { |
7390 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
7391 | compute_in_loop = true; | |
7392 | } | |
7393 | ||
7394 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
7395 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 7396 | && !compute_in_loop) |
ebfd146a IR |
7397 | { |
7398 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
7399 | alignment_support_scheme, NULL_TREE, | |
7400 | &at_loop); | |
7401 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7402 | { | |
538dd0b7 | 7403 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (msq)); |
356bbc4c JJ |
7404 | byte_offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
7405 | size_one_node); | |
ebfd146a IR |
7406 | } |
7407 | } | |
7408 | else | |
7409 | at_loop = loop; | |
7410 | ||
62da9e14 | 7411 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
a1e53f3f L |
7412 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
7413 | ||
2de001ee | 7414 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
272c6793 RS |
7415 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); |
7416 | else | |
7417 | aggr_type = vectype; | |
7418 | ||
ebfd146a | 7419 | prev_stmt_info = NULL; |
b267968e | 7420 | int group_elt = 0; |
ebfd146a | 7421 | for (j = 0; j < ncopies; j++) |
b8698a0f | 7422 | { |
272c6793 | 7423 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 7424 | if (j == 0) |
74bf76ed JJ |
7425 | { |
7426 | bool simd_lane_access_p | |
7427 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
7428 | if (simd_lane_access_p | |
7429 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
7430 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
7431 | && integer_zerop (DR_OFFSET (first_dr)) | |
7432 | && integer_zerop (DR_INIT (first_dr)) | |
7433 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
44fc7854 | 7434 | get_alias_set (TREE_TYPE (ref_type))) |
74bf76ed JJ |
7435 | && (alignment_support_scheme == dr_aligned |
7436 | || alignment_support_scheme == dr_unaligned_supported)) | |
7437 | { | |
7438 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
44fc7854 | 7439 | dataref_offset = build_int_cst (ref_type, 0); |
8928eff3 | 7440 | inv_p = false; |
74bf76ed | 7441 | } |
4f0a0218 RB |
7442 | else if (first_stmt_for_drptr |
7443 | && first_stmt != first_stmt_for_drptr) | |
7444 | { | |
7445 | dataref_ptr | |
7446 | = vect_create_data_ref_ptr (first_stmt_for_drptr, aggr_type, | |
7447 | at_loop, offset, &dummy, gsi, | |
7448 | &ptr_incr, simd_lane_access_p, | |
7449 | &inv_p, byte_offset); | |
7450 | /* Adjust the pointer by the difference to first_stmt. */ | |
7451 | data_reference_p ptrdr | |
7452 | = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt_for_drptr)); | |
7453 | tree diff = fold_convert (sizetype, | |
7454 | size_binop (MINUS_EXPR, | |
7455 | DR_INIT (first_dr), | |
7456 | DR_INIT (ptrdr))); | |
7457 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
7458 | stmt, diff); | |
7459 | } | |
74bf76ed JJ |
7460 | else |
7461 | dataref_ptr | |
7462 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
7463 | offset, &dummy, gsi, &ptr_incr, | |
356bbc4c JJ |
7464 | simd_lane_access_p, &inv_p, |
7465 | byte_offset); | |
74bf76ed JJ |
7466 | } |
7467 | else if (dataref_offset) | |
7468 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
7469 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 7470 | else |
272c6793 RS |
7471 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
7472 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 7473 | |
0d0293ac | 7474 | if (grouped_load || slp_perm) |
9771b263 | 7475 | dr_chain.create (vec_num); |
5ce1ee7f | 7476 | |
2de001ee | 7477 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
ebfd146a | 7478 | { |
272c6793 RS |
7479 | tree vec_array; |
7480 | ||
7481 | vec_array = create_vector_array (vectype, vec_num); | |
7482 | ||
7483 | /* Emit: | |
7484 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
44fc7854 | 7485 | data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type); |
a844293d RS |
7486 | gcall *call = gimple_build_call_internal (IFN_LOAD_LANES, 1, |
7487 | data_ref); | |
7488 | gimple_call_set_lhs (call, vec_array); | |
7489 | gimple_call_set_nothrow (call, true); | |
7490 | new_stmt = call; | |
272c6793 | 7491 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a | 7492 | |
272c6793 RS |
7493 | /* Extract each vector into an SSA_NAME. */ |
7494 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 7495 | { |
272c6793 RS |
7496 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
7497 | vec_array, i); | |
9771b263 | 7498 | dr_chain.quick_push (new_temp); |
272c6793 RS |
7499 | } |
7500 | ||
7501 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 7502 | vect_record_grouped_load_vectors (stmt, dr_chain); |
272c6793 RS |
7503 | } |
7504 | else | |
7505 | { | |
7506 | for (i = 0; i < vec_num; i++) | |
7507 | { | |
7508 | if (i > 0) | |
7509 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
7510 | stmt, NULL_TREE); | |
7511 | ||
7512 | /* 2. Create the vector-load in the loop. */ | |
7513 | switch (alignment_support_scheme) | |
7514 | { | |
7515 | case dr_aligned: | |
7516 | case dr_unaligned_supported: | |
be1ac4ec | 7517 | { |
644ffefd MJ |
7518 | unsigned int align, misalign; |
7519 | ||
272c6793 | 7520 | data_ref |
aed93b23 RB |
7521 | = fold_build2 (MEM_REF, vectype, dataref_ptr, |
7522 | dataref_offset | |
7523 | ? dataref_offset | |
44fc7854 | 7524 | : build_int_cst (ref_type, 0)); |
644ffefd | 7525 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 RS |
7526 | if (alignment_support_scheme == dr_aligned) |
7527 | { | |
7528 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 7529 | misalign = 0; |
272c6793 RS |
7530 | } |
7531 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
7532 | { | |
25f68d90 | 7533 | align = dr_alignment (vect_dr_behavior (first_dr)); |
52639a61 | 7534 | misalign = 0; |
272c6793 RS |
7535 | TREE_TYPE (data_ref) |
7536 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 7537 | align * BITS_PER_UNIT); |
272c6793 RS |
7538 | } |
7539 | else | |
7540 | { | |
7541 | TREE_TYPE (data_ref) | |
7542 | = build_aligned_type (TREE_TYPE (data_ref), | |
7543 | TYPE_ALIGN (elem_type)); | |
644ffefd | 7544 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 7545 | } |
aed93b23 RB |
7546 | if (dataref_offset == NULL_TREE |
7547 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
7548 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), |
7549 | align, misalign); | |
272c6793 | 7550 | break; |
be1ac4ec | 7551 | } |
272c6793 | 7552 | case dr_explicit_realign: |
267d3070 | 7553 | { |
272c6793 | 7554 | tree ptr, bump; |
272c6793 | 7555 | |
d88981fc | 7556 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
272c6793 RS |
7557 | |
7558 | if (compute_in_loop) | |
7559 | msq = vect_setup_realignment (first_stmt, gsi, | |
7560 | &realignment_token, | |
7561 | dr_explicit_realign, | |
7562 | dataref_ptr, NULL); | |
7563 | ||
aed93b23 RB |
7564 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
7565 | ptr = copy_ssa_name (dataref_ptr); | |
7566 | else | |
7567 | ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
0d0e4a03 JJ |
7568 | new_stmt = gimple_build_assign |
7569 | (ptr, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
7570 | build_int_cst |
7571 | (TREE_TYPE (dataref_ptr), | |
7572 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
7573 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7574 | data_ref | |
7575 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 7576 | build_int_cst (ref_type, 0)); |
272c6793 RS |
7577 | vec_dest = vect_create_destination_var (scalar_dest, |
7578 | vectype); | |
7579 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
7580 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
7581 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7582 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
7583 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
7584 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7585 | msq = new_temp; | |
7586 | ||
d88981fc | 7587 | bump = size_binop (MULT_EXPR, vs, |
7b7b1813 | 7588 | TYPE_SIZE_UNIT (elem_type)); |
d88981fc | 7589 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
272c6793 | 7590 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
0d0e4a03 JJ |
7591 | new_stmt = gimple_build_assign |
7592 | (NULL_TREE, BIT_AND_EXPR, ptr, | |
272c6793 RS |
7593 | build_int_cst |
7594 | (TREE_TYPE (ptr), | |
7595 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
aed93b23 | 7596 | ptr = copy_ssa_name (ptr, new_stmt); |
272c6793 RS |
7597 | gimple_assign_set_lhs (new_stmt, ptr); |
7598 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7599 | data_ref | |
7600 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 7601 | build_int_cst (ref_type, 0)); |
272c6793 | 7602 | break; |
267d3070 | 7603 | } |
272c6793 | 7604 | case dr_explicit_realign_optimized: |
aed93b23 RB |
7605 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
7606 | new_temp = copy_ssa_name (dataref_ptr); | |
7607 | else | |
7608 | new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
0d0e4a03 JJ |
7609 | new_stmt = gimple_build_assign |
7610 | (new_temp, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
7611 | build_int_cst |
7612 | (TREE_TYPE (dataref_ptr), | |
7613 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
7614 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7615 | data_ref | |
7616 | = build2 (MEM_REF, vectype, new_temp, | |
44fc7854 | 7617 | build_int_cst (ref_type, 0)); |
272c6793 RS |
7618 | break; |
7619 | default: | |
7620 | gcc_unreachable (); | |
7621 | } | |
ebfd146a | 7622 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 7623 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
7624 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7625 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7626 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7627 | ||
272c6793 RS |
7628 | /* 3. Handle explicit realignment if necessary/supported. |
7629 | Create in loop: | |
7630 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
7631 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
7632 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 7633 | { |
272c6793 RS |
7634 | lsq = gimple_assign_lhs (new_stmt); |
7635 | if (!realignment_token) | |
7636 | realignment_token = dataref_ptr; | |
7637 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
0d0e4a03 JJ |
7638 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, |
7639 | msq, lsq, realignment_token); | |
272c6793 RS |
7640 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7641 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7642 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7643 | ||
7644 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7645 | { | |
7646 | gcc_assert (phi); | |
7647 | if (i == vec_num - 1 && j == ncopies - 1) | |
7648 | add_phi_arg (phi, lsq, | |
7649 | loop_latch_edge (containing_loop), | |
9e227d60 | 7650 | UNKNOWN_LOCATION); |
272c6793 RS |
7651 | msq = lsq; |
7652 | } | |
ebfd146a | 7653 | } |
ebfd146a | 7654 | |
59fd17e3 RB |
7655 | /* 4. Handle invariant-load. */ |
7656 | if (inv_p && !bb_vinfo) | |
7657 | { | |
59fd17e3 | 7658 | gcc_assert (!grouped_load); |
d1417442 JJ |
7659 | /* If we have versioned for aliasing or the loop doesn't |
7660 | have any data dependencies that would preclude this, | |
7661 | then we are sure this is a loop invariant load and | |
7662 | thus we can insert it on the preheader edge. */ | |
7663 | if (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) | |
7664 | && !nested_in_vect_loop | |
6b916b36 | 7665 | && hoist_defs_of_uses (stmt, loop)) |
a0e35eb0 RB |
7666 | { |
7667 | if (dump_enabled_p ()) | |
7668 | { | |
7669 | dump_printf_loc (MSG_NOTE, vect_location, | |
7670 | "hoisting out of the vectorized " | |
7671 | "loop: "); | |
7672 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
a0e35eb0 | 7673 | } |
b731b390 | 7674 | tree tem = copy_ssa_name (scalar_dest); |
a0e35eb0 RB |
7675 | gsi_insert_on_edge_immediate |
7676 | (loop_preheader_edge (loop), | |
7677 | gimple_build_assign (tem, | |
7678 | unshare_expr | |
7679 | (gimple_assign_rhs1 (stmt)))); | |
7680 | new_temp = vect_init_vector (stmt, tem, vectype, NULL); | |
34cd48e5 RB |
7681 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
7682 | set_vinfo_for_stmt (new_stmt, | |
7683 | new_stmt_vec_info (new_stmt, vinfo)); | |
a0e35eb0 RB |
7684 | } |
7685 | else | |
7686 | { | |
7687 | gimple_stmt_iterator gsi2 = *gsi; | |
7688 | gsi_next (&gsi2); | |
7689 | new_temp = vect_init_vector (stmt, scalar_dest, | |
7690 | vectype, &gsi2); | |
34cd48e5 | 7691 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
a0e35eb0 | 7692 | } |
59fd17e3 RB |
7693 | } |
7694 | ||
62da9e14 | 7695 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
272c6793 | 7696 | { |
aec7ae7d JJ |
7697 | tree perm_mask = perm_mask_for_reverse (vectype); |
7698 | new_temp = permute_vec_elements (new_temp, new_temp, | |
7699 | perm_mask, stmt, gsi); | |
ebfd146a IR |
7700 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
7701 | } | |
267d3070 | 7702 | |
272c6793 | 7703 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
7704 | vect_transform_grouped_load (). */ |
7705 | if (grouped_load || slp_perm) | |
9771b263 | 7706 | dr_chain.quick_push (new_temp); |
267d3070 | 7707 | |
272c6793 RS |
7708 | /* Store vector loads in the corresponding SLP_NODE. */ |
7709 | if (slp && !slp_perm) | |
9771b263 | 7710 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
b267968e RB |
7711 | |
7712 | /* With SLP permutation we load the gaps as well, without | |
7713 | we need to skip the gaps after we manage to fully load | |
7714 | all elements. group_gap_adj is GROUP_SIZE here. */ | |
7715 | group_elt += nunits; | |
7716 | if (group_gap_adj != 0 && ! slp_perm | |
7717 | && group_elt == group_size - group_gap_adj) | |
7718 | { | |
7719 | bool ovf; | |
7720 | tree bump | |
7721 | = wide_int_to_tree (sizetype, | |
7722 | wi::smul (TYPE_SIZE_UNIT (elem_type), | |
7723 | group_gap_adj, &ovf)); | |
7724 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
7725 | stmt, bump); | |
7726 | group_elt = 0; | |
7727 | } | |
272c6793 | 7728 | } |
9b999e8c RB |
7729 | /* Bump the vector pointer to account for a gap or for excess |
7730 | elements loaded for a permuted SLP load. */ | |
b267968e | 7731 | if (group_gap_adj != 0 && slp_perm) |
a64b9c26 | 7732 | { |
9b999e8c RB |
7733 | bool ovf; |
7734 | tree bump | |
7735 | = wide_int_to_tree (sizetype, | |
7736 | wi::smul (TYPE_SIZE_UNIT (elem_type), | |
7737 | group_gap_adj, &ovf)); | |
a64b9c26 RB |
7738 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
7739 | stmt, bump); | |
7740 | } | |
ebfd146a IR |
7741 | } |
7742 | ||
7743 | if (slp && !slp_perm) | |
7744 | continue; | |
7745 | ||
7746 | if (slp_perm) | |
7747 | { | |
29afecdf | 7748 | unsigned n_perms; |
01d8bf07 | 7749 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
29afecdf RB |
7750 | slp_node_instance, false, |
7751 | &n_perms)) | |
ebfd146a | 7752 | { |
9771b263 | 7753 | dr_chain.release (); |
ebfd146a IR |
7754 | return false; |
7755 | } | |
7756 | } | |
7757 | else | |
7758 | { | |
0d0293ac | 7759 | if (grouped_load) |
ebfd146a | 7760 | { |
2de001ee | 7761 | if (memory_access_type != VMAT_LOAD_STORE_LANES) |
0d0293ac | 7762 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 7763 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
7764 | } |
7765 | else | |
7766 | { | |
7767 | if (j == 0) | |
7768 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7769 | else | |
7770 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7771 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7772 | } | |
7773 | } | |
9771b263 | 7774 | dr_chain.release (); |
ebfd146a IR |
7775 | } |
7776 | ||
ebfd146a IR |
7777 | return true; |
7778 | } | |
7779 | ||
7780 | /* Function vect_is_simple_cond. | |
b8698a0f | 7781 | |
ebfd146a IR |
7782 | Input: |
7783 | LOOP - the loop that is being vectorized. | |
7784 | COND - Condition that is checked for simple use. | |
7785 | ||
e9e1d143 RG |
7786 | Output: |
7787 | *COMP_VECTYPE - the vector type for the comparison. | |
4fc5ebf1 | 7788 | *DTS - The def types for the arguments of the comparison |
e9e1d143 | 7789 | |
ebfd146a IR |
7790 | Returns whether a COND can be vectorized. Checks whether |
7791 | condition operands are supportable using vec_is_simple_use. */ | |
7792 | ||
87aab9b2 | 7793 | static bool |
4fc5ebf1 JG |
7794 | vect_is_simple_cond (tree cond, vec_info *vinfo, |
7795 | tree *comp_vectype, enum vect_def_type *dts) | |
ebfd146a IR |
7796 | { |
7797 | tree lhs, rhs; | |
e9e1d143 | 7798 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a | 7799 | |
a414c77f IE |
7800 | /* Mask case. */ |
7801 | if (TREE_CODE (cond) == SSA_NAME | |
2568d8a1 | 7802 | && VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (cond))) |
a414c77f IE |
7803 | { |
7804 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (cond); | |
7805 | if (!vect_is_simple_use (cond, vinfo, &lhs_def_stmt, | |
4fc5ebf1 | 7806 | &dts[0], comp_vectype) |
a414c77f IE |
7807 | || !*comp_vectype |
7808 | || !VECTOR_BOOLEAN_TYPE_P (*comp_vectype)) | |
7809 | return false; | |
7810 | return true; | |
7811 | } | |
7812 | ||
ebfd146a IR |
7813 | if (!COMPARISON_CLASS_P (cond)) |
7814 | return false; | |
7815 | ||
7816 | lhs = TREE_OPERAND (cond, 0); | |
7817 | rhs = TREE_OPERAND (cond, 1); | |
7818 | ||
7819 | if (TREE_CODE (lhs) == SSA_NAME) | |
7820 | { | |
355fe088 | 7821 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); |
4fc5ebf1 | 7822 | if (!vect_is_simple_use (lhs, vinfo, &lhs_def_stmt, &dts[0], &vectype1)) |
ebfd146a IR |
7823 | return false; |
7824 | } | |
4fc5ebf1 JG |
7825 | else if (TREE_CODE (lhs) == INTEGER_CST || TREE_CODE (lhs) == REAL_CST |
7826 | || TREE_CODE (lhs) == FIXED_CST) | |
7827 | dts[0] = vect_constant_def; | |
7828 | else | |
ebfd146a IR |
7829 | return false; |
7830 | ||
7831 | if (TREE_CODE (rhs) == SSA_NAME) | |
7832 | { | |
355fe088 | 7833 | gimple *rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); |
4fc5ebf1 | 7834 | if (!vect_is_simple_use (rhs, vinfo, &rhs_def_stmt, &dts[1], &vectype2)) |
ebfd146a IR |
7835 | return false; |
7836 | } | |
4fc5ebf1 JG |
7837 | else if (TREE_CODE (rhs) == INTEGER_CST || TREE_CODE (rhs) == REAL_CST |
7838 | || TREE_CODE (rhs) == FIXED_CST) | |
7839 | dts[1] = vect_constant_def; | |
7840 | else | |
ebfd146a IR |
7841 | return false; |
7842 | ||
28b33016 IE |
7843 | if (vectype1 && vectype2 |
7844 | && TYPE_VECTOR_SUBPARTS (vectype1) != TYPE_VECTOR_SUBPARTS (vectype2)) | |
7845 | return false; | |
7846 | ||
e9e1d143 | 7847 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
ebfd146a IR |
7848 | return true; |
7849 | } | |
7850 | ||
7851 | /* vectorizable_condition. | |
7852 | ||
b8698a0f L |
7853 | Check if STMT is conditional modify expression that can be vectorized. |
7854 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
7855 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
7856 | at GSI. |
7857 | ||
7858 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
7859 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
0ad23163 | 7860 | else clause if it is 2). |
ebfd146a IR |
7861 | |
7862 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7863 | ||
4bbe8262 | 7864 | bool |
355fe088 TS |
7865 | vectorizable_condition (gimple *stmt, gimple_stmt_iterator *gsi, |
7866 | gimple **vec_stmt, tree reduc_def, int reduc_index, | |
f7e531cf | 7867 | slp_tree slp_node) |
ebfd146a IR |
7868 | { |
7869 | tree scalar_dest = NULL_TREE; | |
7870 | tree vec_dest = NULL_TREE; | |
01216d27 JJ |
7871 | tree cond_expr, cond_expr0 = NULL_TREE, cond_expr1 = NULL_TREE; |
7872 | tree then_clause, else_clause; | |
ebfd146a | 7873 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
df11cc78 | 7874 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
7875 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
7876 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
5958f9e2 | 7877 | tree vec_compare; |
ebfd146a IR |
7878 | tree new_temp; |
7879 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4fc5ebf1 JG |
7880 | enum vect_def_type dts[4] |
7881 | = {vect_unknown_def_type, vect_unknown_def_type, | |
7882 | vect_unknown_def_type, vect_unknown_def_type}; | |
7883 | int ndts = 4; | |
f7e531cf | 7884 | int ncopies; |
01216d27 | 7885 | enum tree_code code, cond_code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
a855b1b1 | 7886 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
7887 | int i, j; |
7888 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
7889 | vec<tree> vec_oprnds0 = vNULL; |
7890 | vec<tree> vec_oprnds1 = vNULL; | |
7891 | vec<tree> vec_oprnds2 = vNULL; | |
7892 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 7893 | tree vec_cmp_type; |
a414c77f | 7894 | bool masked = false; |
b8698a0f | 7895 | |
f7e531cf IR |
7896 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
7897 | return false; | |
7898 | ||
af29617a AH |
7899 | if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) == TREE_CODE_REDUCTION) |
7900 | { | |
7901 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
7902 | return false; | |
ebfd146a | 7903 | |
af29617a AH |
7904 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7905 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
7906 | && reduc_def)) | |
7907 | return false; | |
ebfd146a | 7908 | |
af29617a AH |
7909 | /* FORNOW: not yet supported. */ |
7910 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
7911 | { | |
7912 | if (dump_enabled_p ()) | |
7913 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7914 | "value used after loop.\n"); | |
7915 | return false; | |
7916 | } | |
ebfd146a IR |
7917 | } |
7918 | ||
7919 | /* Is vectorizable conditional operation? */ | |
7920 | if (!is_gimple_assign (stmt)) | |
7921 | return false; | |
7922 | ||
7923 | code = gimple_assign_rhs_code (stmt); | |
7924 | ||
7925 | if (code != COND_EXPR) | |
7926 | return false; | |
7927 | ||
465c8c19 JJ |
7928 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
7929 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
2947d3b2 | 7930 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
465c8c19 | 7931 | |
fce57248 | 7932 | if (slp_node) |
465c8c19 JJ |
7933 | ncopies = 1; |
7934 | else | |
7935 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
7936 | ||
7937 | gcc_assert (ncopies >= 1); | |
7938 | if (reduc_index && ncopies > 1) | |
7939 | return false; /* FORNOW */ | |
7940 | ||
4e71066d RG |
7941 | cond_expr = gimple_assign_rhs1 (stmt); |
7942 | then_clause = gimple_assign_rhs2 (stmt); | |
7943 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 7944 | |
4fc5ebf1 JG |
7945 | if (!vect_is_simple_cond (cond_expr, stmt_info->vinfo, |
7946 | &comp_vectype, &dts[0]) | |
e9e1d143 | 7947 | || !comp_vectype) |
ebfd146a IR |
7948 | return false; |
7949 | ||
81c40241 | 7950 | gimple *def_stmt; |
4fc5ebf1 | 7951 | if (!vect_is_simple_use (then_clause, stmt_info->vinfo, &def_stmt, &dts[2], |
2947d3b2 IE |
7952 | &vectype1)) |
7953 | return false; | |
4fc5ebf1 | 7954 | if (!vect_is_simple_use (else_clause, stmt_info->vinfo, &def_stmt, &dts[3], |
2947d3b2 | 7955 | &vectype2)) |
ebfd146a | 7956 | return false; |
2947d3b2 IE |
7957 | |
7958 | if (vectype1 && !useless_type_conversion_p (vectype, vectype1)) | |
7959 | return false; | |
7960 | ||
7961 | if (vectype2 && !useless_type_conversion_p (vectype, vectype2)) | |
ebfd146a IR |
7962 | return false; |
7963 | ||
28b33016 IE |
7964 | masked = !COMPARISON_CLASS_P (cond_expr); |
7965 | vec_cmp_type = build_same_sized_truth_vector_type (comp_vectype); | |
7966 | ||
74946978 MP |
7967 | if (vec_cmp_type == NULL_TREE) |
7968 | return false; | |
784fb9b3 | 7969 | |
01216d27 JJ |
7970 | cond_code = TREE_CODE (cond_expr); |
7971 | if (!masked) | |
7972 | { | |
7973 | cond_expr0 = TREE_OPERAND (cond_expr, 0); | |
7974 | cond_expr1 = TREE_OPERAND (cond_expr, 1); | |
7975 | } | |
7976 | ||
7977 | if (!masked && VECTOR_BOOLEAN_TYPE_P (comp_vectype)) | |
7978 | { | |
7979 | /* Boolean values may have another representation in vectors | |
7980 | and therefore we prefer bit operations over comparison for | |
7981 | them (which also works for scalar masks). We store opcodes | |
7982 | to use in bitop1 and bitop2. Statement is vectorized as | |
7983 | BITOP2 (rhs1 BITOP1 rhs2) or rhs1 BITOP2 (BITOP1 rhs2) | |
7984 | depending on bitop1 and bitop2 arity. */ | |
7985 | switch (cond_code) | |
7986 | { | |
7987 | case GT_EXPR: | |
7988 | bitop1 = BIT_NOT_EXPR; | |
7989 | bitop2 = BIT_AND_EXPR; | |
7990 | break; | |
7991 | case GE_EXPR: | |
7992 | bitop1 = BIT_NOT_EXPR; | |
7993 | bitop2 = BIT_IOR_EXPR; | |
7994 | break; | |
7995 | case LT_EXPR: | |
7996 | bitop1 = BIT_NOT_EXPR; | |
7997 | bitop2 = BIT_AND_EXPR; | |
7998 | std::swap (cond_expr0, cond_expr1); | |
7999 | break; | |
8000 | case LE_EXPR: | |
8001 | bitop1 = BIT_NOT_EXPR; | |
8002 | bitop2 = BIT_IOR_EXPR; | |
8003 | std::swap (cond_expr0, cond_expr1); | |
8004 | break; | |
8005 | case NE_EXPR: | |
8006 | bitop1 = BIT_XOR_EXPR; | |
8007 | break; | |
8008 | case EQ_EXPR: | |
8009 | bitop1 = BIT_XOR_EXPR; | |
8010 | bitop2 = BIT_NOT_EXPR; | |
8011 | break; | |
8012 | default: | |
8013 | return false; | |
8014 | } | |
8015 | cond_code = SSA_NAME; | |
8016 | } | |
8017 | ||
b8698a0f | 8018 | if (!vec_stmt) |
ebfd146a IR |
8019 | { |
8020 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
01216d27 JJ |
8021 | if (bitop1 != NOP_EXPR) |
8022 | { | |
8023 | machine_mode mode = TYPE_MODE (comp_vectype); | |
8024 | optab optab; | |
8025 | ||
8026 | optab = optab_for_tree_code (bitop1, comp_vectype, optab_default); | |
8027 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8028 | return false; | |
8029 | ||
8030 | if (bitop2 != NOP_EXPR) | |
8031 | { | |
8032 | optab = optab_for_tree_code (bitop2, comp_vectype, | |
8033 | optab_default); | |
8034 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8035 | return false; | |
8036 | } | |
8037 | } | |
4fc5ebf1 JG |
8038 | if (expand_vec_cond_expr_p (vectype, comp_vectype, |
8039 | cond_code)) | |
8040 | { | |
8041 | vect_model_simple_cost (stmt_info, ncopies, dts, ndts, NULL, NULL); | |
8042 | return true; | |
8043 | } | |
8044 | return false; | |
ebfd146a IR |
8045 | } |
8046 | ||
f7e531cf IR |
8047 | /* Transform. */ |
8048 | ||
8049 | if (!slp_node) | |
8050 | { | |
9771b263 DN |
8051 | vec_oprnds0.create (1); |
8052 | vec_oprnds1.create (1); | |
8053 | vec_oprnds2.create (1); | |
8054 | vec_oprnds3.create (1); | |
f7e531cf | 8055 | } |
ebfd146a IR |
8056 | |
8057 | /* Handle def. */ | |
8058 | scalar_dest = gimple_assign_lhs (stmt); | |
8059 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
8060 | ||
8061 | /* Handle cond expr. */ | |
a855b1b1 MM |
8062 | for (j = 0; j < ncopies; j++) |
8063 | { | |
538dd0b7 | 8064 | gassign *new_stmt = NULL; |
a855b1b1 MM |
8065 | if (j == 0) |
8066 | { | |
f7e531cf IR |
8067 | if (slp_node) |
8068 | { | |
00f96dc9 TS |
8069 | auto_vec<tree, 4> ops; |
8070 | auto_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 8071 | |
a414c77f | 8072 | if (masked) |
01216d27 | 8073 | ops.safe_push (cond_expr); |
a414c77f IE |
8074 | else |
8075 | { | |
01216d27 JJ |
8076 | ops.safe_push (cond_expr0); |
8077 | ops.safe_push (cond_expr1); | |
a414c77f | 8078 | } |
9771b263 DN |
8079 | ops.safe_push (then_clause); |
8080 | ops.safe_push (else_clause); | |
306b0c92 | 8081 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
37b5ec8f JJ |
8082 | vec_oprnds3 = vec_defs.pop (); |
8083 | vec_oprnds2 = vec_defs.pop (); | |
a414c77f IE |
8084 | if (!masked) |
8085 | vec_oprnds1 = vec_defs.pop (); | |
37b5ec8f | 8086 | vec_oprnds0 = vec_defs.pop (); |
f7e531cf IR |
8087 | } |
8088 | else | |
8089 | { | |
355fe088 | 8090 | gimple *gtemp; |
a414c77f IE |
8091 | if (masked) |
8092 | { | |
8093 | vec_cond_lhs | |
8094 | = vect_get_vec_def_for_operand (cond_expr, stmt, | |
8095 | comp_vectype); | |
8096 | vect_is_simple_use (cond_expr, stmt_info->vinfo, | |
8097 | >emp, &dts[0]); | |
8098 | } | |
8099 | else | |
8100 | { | |
01216d27 JJ |
8101 | vec_cond_lhs |
8102 | = vect_get_vec_def_for_operand (cond_expr0, | |
8103 | stmt, comp_vectype); | |
8104 | vect_is_simple_use (cond_expr0, loop_vinfo, >emp, &dts[0]); | |
8105 | ||
8106 | vec_cond_rhs | |
8107 | = vect_get_vec_def_for_operand (cond_expr1, | |
8108 | stmt, comp_vectype); | |
8109 | vect_is_simple_use (cond_expr1, loop_vinfo, >emp, &dts[1]); | |
a414c77f | 8110 | } |
f7e531cf IR |
8111 | if (reduc_index == 1) |
8112 | vec_then_clause = reduc_def; | |
8113 | else | |
8114 | { | |
8115 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
81c40241 RB |
8116 | stmt); |
8117 | vect_is_simple_use (then_clause, loop_vinfo, | |
8118 | >emp, &dts[2]); | |
f7e531cf IR |
8119 | } |
8120 | if (reduc_index == 2) | |
8121 | vec_else_clause = reduc_def; | |
8122 | else | |
8123 | { | |
8124 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
81c40241 RB |
8125 | stmt); |
8126 | vect_is_simple_use (else_clause, loop_vinfo, >emp, &dts[3]); | |
f7e531cf | 8127 | } |
a855b1b1 MM |
8128 | } |
8129 | } | |
8130 | else | |
8131 | { | |
a414c77f IE |
8132 | vec_cond_lhs |
8133 | = vect_get_vec_def_for_stmt_copy (dts[0], | |
8134 | vec_oprnds0.pop ()); | |
8135 | if (!masked) | |
8136 | vec_cond_rhs | |
8137 | = vect_get_vec_def_for_stmt_copy (dts[1], | |
8138 | vec_oprnds1.pop ()); | |
8139 | ||
a855b1b1 | 8140 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 8141 | vec_oprnds2.pop ()); |
a855b1b1 | 8142 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 8143 | vec_oprnds3.pop ()); |
f7e531cf IR |
8144 | } |
8145 | ||
8146 | if (!slp_node) | |
8147 | { | |
9771b263 | 8148 | vec_oprnds0.quick_push (vec_cond_lhs); |
a414c77f IE |
8149 | if (!masked) |
8150 | vec_oprnds1.quick_push (vec_cond_rhs); | |
9771b263 DN |
8151 | vec_oprnds2.quick_push (vec_then_clause); |
8152 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
8153 | } |
8154 | ||
9dc3f7de | 8155 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 8156 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 8157 | { |
9771b263 DN |
8158 | vec_then_clause = vec_oprnds2[i]; |
8159 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 8160 | |
a414c77f IE |
8161 | if (masked) |
8162 | vec_compare = vec_cond_lhs; | |
8163 | else | |
8164 | { | |
8165 | vec_cond_rhs = vec_oprnds1[i]; | |
01216d27 JJ |
8166 | if (bitop1 == NOP_EXPR) |
8167 | vec_compare = build2 (cond_code, vec_cmp_type, | |
8168 | vec_cond_lhs, vec_cond_rhs); | |
8169 | else | |
8170 | { | |
8171 | new_temp = make_ssa_name (vec_cmp_type); | |
8172 | if (bitop1 == BIT_NOT_EXPR) | |
8173 | new_stmt = gimple_build_assign (new_temp, bitop1, | |
8174 | vec_cond_rhs); | |
8175 | else | |
8176 | new_stmt | |
8177 | = gimple_build_assign (new_temp, bitop1, vec_cond_lhs, | |
8178 | vec_cond_rhs); | |
8179 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8180 | if (bitop2 == NOP_EXPR) | |
8181 | vec_compare = new_temp; | |
8182 | else if (bitop2 == BIT_NOT_EXPR) | |
8183 | { | |
8184 | /* Instead of doing ~x ? y : z do x ? z : y. */ | |
8185 | vec_compare = new_temp; | |
8186 | std::swap (vec_then_clause, vec_else_clause); | |
8187 | } | |
8188 | else | |
8189 | { | |
8190 | vec_compare = make_ssa_name (vec_cmp_type); | |
8191 | new_stmt | |
8192 | = gimple_build_assign (vec_compare, bitop2, | |
8193 | vec_cond_lhs, new_temp); | |
8194 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8195 | } | |
8196 | } | |
a414c77f | 8197 | } |
5958f9e2 JJ |
8198 | new_temp = make_ssa_name (vec_dest); |
8199 | new_stmt = gimple_build_assign (new_temp, VEC_COND_EXPR, | |
8200 | vec_compare, vec_then_clause, | |
8201 | vec_else_clause); | |
f7e531cf IR |
8202 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
8203 | if (slp_node) | |
9771b263 | 8204 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
8205 | } |
8206 | ||
8207 | if (slp_node) | |
8208 | continue; | |
8209 | ||
8210 | if (j == 0) | |
8211 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
8212 | else | |
8213 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
8214 | ||
8215 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 8216 | } |
b8698a0f | 8217 | |
9771b263 DN |
8218 | vec_oprnds0.release (); |
8219 | vec_oprnds1.release (); | |
8220 | vec_oprnds2.release (); | |
8221 | vec_oprnds3.release (); | |
f7e531cf | 8222 | |
ebfd146a IR |
8223 | return true; |
8224 | } | |
8225 | ||
42fd8198 IE |
8226 | /* vectorizable_comparison. |
8227 | ||
8228 | Check if STMT is comparison expression that can be vectorized. | |
8229 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
8230 | comparison, put it in VEC_STMT, and insert it at GSI. | |
8231 | ||
8232 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
8233 | ||
fce57248 | 8234 | static bool |
42fd8198 IE |
8235 | vectorizable_comparison (gimple *stmt, gimple_stmt_iterator *gsi, |
8236 | gimple **vec_stmt, tree reduc_def, | |
8237 | slp_tree slp_node) | |
8238 | { | |
8239 | tree lhs, rhs1, rhs2; | |
8240 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8241 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; | |
8242 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
8243 | tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE; | |
8244 | tree new_temp; | |
8245 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
8246 | enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 8247 | int ndts = 2; |
42fd8198 IE |
8248 | unsigned nunits; |
8249 | int ncopies; | |
49e76ff1 | 8250 | enum tree_code code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
42fd8198 IE |
8251 | stmt_vec_info prev_stmt_info = NULL; |
8252 | int i, j; | |
8253 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
8254 | vec<tree> vec_oprnds0 = vNULL; | |
8255 | vec<tree> vec_oprnds1 = vNULL; | |
8256 | gimple *def_stmt; | |
8257 | tree mask_type; | |
8258 | tree mask; | |
8259 | ||
c245362b IE |
8260 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
8261 | return false; | |
8262 | ||
30480bcd | 8263 | if (!vectype || !VECTOR_BOOLEAN_TYPE_P (vectype)) |
42fd8198 IE |
8264 | return false; |
8265 | ||
8266 | mask_type = vectype; | |
8267 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
8268 | ||
fce57248 | 8269 | if (slp_node) |
42fd8198 IE |
8270 | ncopies = 1; |
8271 | else | |
8272 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
8273 | ||
8274 | gcc_assert (ncopies >= 1); | |
42fd8198 IE |
8275 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
8276 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
8277 | && reduc_def)) | |
8278 | return false; | |
8279 | ||
8280 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
8281 | { | |
8282 | if (dump_enabled_p ()) | |
8283 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
8284 | "value used after loop.\n"); | |
8285 | return false; | |
8286 | } | |
8287 | ||
8288 | if (!is_gimple_assign (stmt)) | |
8289 | return false; | |
8290 | ||
8291 | code = gimple_assign_rhs_code (stmt); | |
8292 | ||
8293 | if (TREE_CODE_CLASS (code) != tcc_comparison) | |
8294 | return false; | |
8295 | ||
8296 | rhs1 = gimple_assign_rhs1 (stmt); | |
8297 | rhs2 = gimple_assign_rhs2 (stmt); | |
8298 | ||
8299 | if (!vect_is_simple_use (rhs1, stmt_info->vinfo, &def_stmt, | |
8300 | &dts[0], &vectype1)) | |
8301 | return false; | |
8302 | ||
8303 | if (!vect_is_simple_use (rhs2, stmt_info->vinfo, &def_stmt, | |
8304 | &dts[1], &vectype2)) | |
8305 | return false; | |
8306 | ||
8307 | if (vectype1 && vectype2 | |
8308 | && TYPE_VECTOR_SUBPARTS (vectype1) != TYPE_VECTOR_SUBPARTS (vectype2)) | |
8309 | return false; | |
8310 | ||
8311 | vectype = vectype1 ? vectype1 : vectype2; | |
8312 | ||
8313 | /* Invariant comparison. */ | |
8314 | if (!vectype) | |
8315 | { | |
69a9a66f RB |
8316 | vectype = get_vectype_for_scalar_type (TREE_TYPE (rhs1)); |
8317 | if (TYPE_VECTOR_SUBPARTS (vectype) != nunits) | |
42fd8198 IE |
8318 | return false; |
8319 | } | |
8320 | else if (nunits != TYPE_VECTOR_SUBPARTS (vectype)) | |
8321 | return false; | |
8322 | ||
49e76ff1 IE |
8323 | /* Can't compare mask and non-mask types. */ |
8324 | if (vectype1 && vectype2 | |
8325 | && (VECTOR_BOOLEAN_TYPE_P (vectype1) ^ VECTOR_BOOLEAN_TYPE_P (vectype2))) | |
8326 | return false; | |
8327 | ||
8328 | /* Boolean values may have another representation in vectors | |
8329 | and therefore we prefer bit operations over comparison for | |
8330 | them (which also works for scalar masks). We store opcodes | |
8331 | to use in bitop1 and bitop2. Statement is vectorized as | |
8332 | BITOP2 (rhs1 BITOP1 rhs2) or | |
8333 | rhs1 BITOP2 (BITOP1 rhs2) | |
8334 | depending on bitop1 and bitop2 arity. */ | |
8335 | if (VECTOR_BOOLEAN_TYPE_P (vectype)) | |
8336 | { | |
8337 | if (code == GT_EXPR) | |
8338 | { | |
8339 | bitop1 = BIT_NOT_EXPR; | |
8340 | bitop2 = BIT_AND_EXPR; | |
8341 | } | |
8342 | else if (code == GE_EXPR) | |
8343 | { | |
8344 | bitop1 = BIT_NOT_EXPR; | |
8345 | bitop2 = BIT_IOR_EXPR; | |
8346 | } | |
8347 | else if (code == LT_EXPR) | |
8348 | { | |
8349 | bitop1 = BIT_NOT_EXPR; | |
8350 | bitop2 = BIT_AND_EXPR; | |
8351 | std::swap (rhs1, rhs2); | |
264d951a | 8352 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
8353 | } |
8354 | else if (code == LE_EXPR) | |
8355 | { | |
8356 | bitop1 = BIT_NOT_EXPR; | |
8357 | bitop2 = BIT_IOR_EXPR; | |
8358 | std::swap (rhs1, rhs2); | |
264d951a | 8359 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
8360 | } |
8361 | else | |
8362 | { | |
8363 | bitop1 = BIT_XOR_EXPR; | |
8364 | if (code == EQ_EXPR) | |
8365 | bitop2 = BIT_NOT_EXPR; | |
8366 | } | |
8367 | } | |
8368 | ||
42fd8198 IE |
8369 | if (!vec_stmt) |
8370 | { | |
8371 | STMT_VINFO_TYPE (stmt_info) = comparison_vec_info_type; | |
49e76ff1 | 8372 | vect_model_simple_cost (stmt_info, ncopies * (1 + (bitop2 != NOP_EXPR)), |
4fc5ebf1 | 8373 | dts, ndts, NULL, NULL); |
49e76ff1 | 8374 | if (bitop1 == NOP_EXPR) |
96592eed | 8375 | return expand_vec_cmp_expr_p (vectype, mask_type, code); |
49e76ff1 IE |
8376 | else |
8377 | { | |
8378 | machine_mode mode = TYPE_MODE (vectype); | |
8379 | optab optab; | |
8380 | ||
8381 | optab = optab_for_tree_code (bitop1, vectype, optab_default); | |
8382 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8383 | return false; | |
8384 | ||
8385 | if (bitop2 != NOP_EXPR) | |
8386 | { | |
8387 | optab = optab_for_tree_code (bitop2, vectype, optab_default); | |
8388 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8389 | return false; | |
8390 | } | |
8391 | return true; | |
8392 | } | |
42fd8198 IE |
8393 | } |
8394 | ||
8395 | /* Transform. */ | |
8396 | if (!slp_node) | |
8397 | { | |
8398 | vec_oprnds0.create (1); | |
8399 | vec_oprnds1.create (1); | |
8400 | } | |
8401 | ||
8402 | /* Handle def. */ | |
8403 | lhs = gimple_assign_lhs (stmt); | |
8404 | mask = vect_create_destination_var (lhs, mask_type); | |
8405 | ||
8406 | /* Handle cmp expr. */ | |
8407 | for (j = 0; j < ncopies; j++) | |
8408 | { | |
8409 | gassign *new_stmt = NULL; | |
8410 | if (j == 0) | |
8411 | { | |
8412 | if (slp_node) | |
8413 | { | |
8414 | auto_vec<tree, 2> ops; | |
8415 | auto_vec<vec<tree>, 2> vec_defs; | |
8416 | ||
8417 | ops.safe_push (rhs1); | |
8418 | ops.safe_push (rhs2); | |
306b0c92 | 8419 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
42fd8198 IE |
8420 | vec_oprnds1 = vec_defs.pop (); |
8421 | vec_oprnds0 = vec_defs.pop (); | |
8422 | } | |
8423 | else | |
8424 | { | |
e4af0bc4 IE |
8425 | vec_rhs1 = vect_get_vec_def_for_operand (rhs1, stmt, vectype); |
8426 | vec_rhs2 = vect_get_vec_def_for_operand (rhs2, stmt, vectype); | |
42fd8198 IE |
8427 | } |
8428 | } | |
8429 | else | |
8430 | { | |
8431 | vec_rhs1 = vect_get_vec_def_for_stmt_copy (dts[0], | |
8432 | vec_oprnds0.pop ()); | |
8433 | vec_rhs2 = vect_get_vec_def_for_stmt_copy (dts[1], | |
8434 | vec_oprnds1.pop ()); | |
8435 | } | |
8436 | ||
8437 | if (!slp_node) | |
8438 | { | |
8439 | vec_oprnds0.quick_push (vec_rhs1); | |
8440 | vec_oprnds1.quick_push (vec_rhs2); | |
8441 | } | |
8442 | ||
8443 | /* Arguments are ready. Create the new vector stmt. */ | |
8444 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1) | |
8445 | { | |
8446 | vec_rhs2 = vec_oprnds1[i]; | |
8447 | ||
8448 | new_temp = make_ssa_name (mask); | |
49e76ff1 IE |
8449 | if (bitop1 == NOP_EXPR) |
8450 | { | |
8451 | new_stmt = gimple_build_assign (new_temp, code, | |
8452 | vec_rhs1, vec_rhs2); | |
8453 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8454 | } | |
8455 | else | |
8456 | { | |
8457 | if (bitop1 == BIT_NOT_EXPR) | |
8458 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs2); | |
8459 | else | |
8460 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs1, | |
8461 | vec_rhs2); | |
8462 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8463 | if (bitop2 != NOP_EXPR) | |
8464 | { | |
8465 | tree res = make_ssa_name (mask); | |
8466 | if (bitop2 == BIT_NOT_EXPR) | |
8467 | new_stmt = gimple_build_assign (res, bitop2, new_temp); | |
8468 | else | |
8469 | new_stmt = gimple_build_assign (res, bitop2, vec_rhs1, | |
8470 | new_temp); | |
8471 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8472 | } | |
8473 | } | |
42fd8198 IE |
8474 | if (slp_node) |
8475 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
8476 | } | |
8477 | ||
8478 | if (slp_node) | |
8479 | continue; | |
8480 | ||
8481 | if (j == 0) | |
8482 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
8483 | else | |
8484 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
8485 | ||
8486 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
8487 | } | |
8488 | ||
8489 | vec_oprnds0.release (); | |
8490 | vec_oprnds1.release (); | |
8491 | ||
8492 | return true; | |
8493 | } | |
ebfd146a | 8494 | |
8644a673 | 8495 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
8496 | |
8497 | bool | |
891ad31c RB |
8498 | vect_analyze_stmt (gimple *stmt, bool *need_to_vectorize, slp_tree node, |
8499 | slp_instance node_instance) | |
ebfd146a | 8500 | { |
8644a673 | 8501 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 8502 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 8503 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 8504 | bool ok; |
355fe088 | 8505 | gimple *pattern_stmt; |
363477c0 | 8506 | gimple_seq pattern_def_seq; |
ebfd146a | 8507 | |
73fbfcad | 8508 | if (dump_enabled_p ()) |
ebfd146a | 8509 | { |
78c60e3d SS |
8510 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
8511 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
8644a673 | 8512 | } |
ebfd146a | 8513 | |
1825a1f3 | 8514 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 8515 | { |
73fbfcad | 8516 | if (dump_enabled_p ()) |
78c60e3d | 8517 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8518 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
8519 | |
8520 | return false; | |
8521 | } | |
b8698a0f L |
8522 | |
8523 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
8524 | to include: |
8525 | - the COND_EXPR which is the loop exit condition | |
8526 | - any LABEL_EXPRs in the loop | |
b8698a0f | 8527 | - computations that are used only for array indexing or loop control. |
8644a673 | 8528 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 8529 | instance, therefore, all the statements are relevant. |
ebfd146a | 8530 | |
d092494c | 8531 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 8532 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
8533 | statements. In basic blocks we are called from some SLP instance |
8534 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
8535 | already will be part of SLP instance. */ | |
83197f37 IR |
8536 | |
8537 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 8538 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 8539 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 8540 | { |
9d5e7640 | 8541 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 8542 | && pattern_stmt |
9d5e7640 IR |
8543 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
8544 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
8545 | { | |
83197f37 | 8546 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
8547 | stmt = pattern_stmt; |
8548 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 8549 | if (dump_enabled_p ()) |
9d5e7640 | 8550 | { |
78c60e3d SS |
8551 | dump_printf_loc (MSG_NOTE, vect_location, |
8552 | "==> examining pattern statement: "); | |
8553 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
9d5e7640 IR |
8554 | } |
8555 | } | |
8556 | else | |
8557 | { | |
73fbfcad | 8558 | if (dump_enabled_p ()) |
e645e942 | 8559 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 8560 | |
9d5e7640 IR |
8561 | return true; |
8562 | } | |
8644a673 | 8563 | } |
83197f37 | 8564 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 8565 | && node == NULL |
83197f37 IR |
8566 | && pattern_stmt |
8567 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
8568 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
8569 | { | |
8570 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 8571 | if (dump_enabled_p ()) |
83197f37 | 8572 | { |
78c60e3d SS |
8573 | dump_printf_loc (MSG_NOTE, vect_location, |
8574 | "==> examining pattern statement: "); | |
8575 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
83197f37 IR |
8576 | } |
8577 | ||
891ad31c RB |
8578 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node, |
8579 | node_instance)) | |
83197f37 IR |
8580 | return false; |
8581 | } | |
ebfd146a | 8582 | |
1107f3ae | 8583 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 8584 | && node == NULL |
363477c0 | 8585 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 8586 | { |
363477c0 | 8587 | gimple_stmt_iterator si; |
1107f3ae | 8588 | |
363477c0 JJ |
8589 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
8590 | { | |
355fe088 | 8591 | gimple *pattern_def_stmt = gsi_stmt (si); |
363477c0 JJ |
8592 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) |
8593 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
8594 | { | |
8595 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 8596 | if (dump_enabled_p ()) |
363477c0 | 8597 | { |
78c60e3d SS |
8598 | dump_printf_loc (MSG_NOTE, vect_location, |
8599 | "==> examining pattern def statement: "); | |
8600 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
363477c0 | 8601 | } |
1107f3ae | 8602 | |
363477c0 | 8603 | if (!vect_analyze_stmt (pattern_def_stmt, |
891ad31c | 8604 | need_to_vectorize, node, node_instance)) |
363477c0 JJ |
8605 | return false; |
8606 | } | |
8607 | } | |
8608 | } | |
1107f3ae | 8609 | |
8644a673 IR |
8610 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
8611 | { | |
8612 | case vect_internal_def: | |
8613 | break; | |
ebfd146a | 8614 | |
8644a673 | 8615 | case vect_reduction_def: |
7c5222ff | 8616 | case vect_nested_cycle: |
14a61437 RB |
8617 | gcc_assert (!bb_vinfo |
8618 | && (relevance == vect_used_in_outer | |
8619 | || relevance == vect_used_in_outer_by_reduction | |
8620 | || relevance == vect_used_by_reduction | |
b28ead45 AH |
8621 | || relevance == vect_unused_in_scope |
8622 | || relevance == vect_used_only_live)); | |
8644a673 IR |
8623 | break; |
8624 | ||
8625 | case vect_induction_def: | |
e7baeb39 RB |
8626 | gcc_assert (!bb_vinfo); |
8627 | break; | |
8628 | ||
8644a673 IR |
8629 | case vect_constant_def: |
8630 | case vect_external_def: | |
8631 | case vect_unknown_def_type: | |
8632 | default: | |
8633 | gcc_unreachable (); | |
8634 | } | |
ebfd146a | 8635 | |
8644a673 | 8636 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 8637 | { |
8644a673 | 8638 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
0136f8f0 AH |
8639 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
8640 | || (is_gimple_call (stmt) | |
8641 | && gimple_call_lhs (stmt) == NULL_TREE)); | |
8644a673 | 8642 | *need_to_vectorize = true; |
ebfd146a IR |
8643 | } |
8644 | ||
b1af7da6 RB |
8645 | if (PURE_SLP_STMT (stmt_info) && !node) |
8646 | { | |
8647 | dump_printf_loc (MSG_NOTE, vect_location, | |
8648 | "handled only by SLP analysis\n"); | |
8649 | return true; | |
8650 | } | |
8651 | ||
8652 | ok = true; | |
8653 | if (!bb_vinfo | |
8654 | && (STMT_VINFO_RELEVANT_P (stmt_info) | |
8655 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
8656 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
8657 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
8658 | || vectorizable_shift (stmt, NULL, NULL, node) | |
8659 | || vectorizable_operation (stmt, NULL, NULL, node) | |
8660 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
8661 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
8662 | || vectorizable_call (stmt, NULL, NULL, node) | |
8663 | || vectorizable_store (stmt, NULL, NULL, node) | |
891ad31c | 8664 | || vectorizable_reduction (stmt, NULL, NULL, node, node_instance) |
e7baeb39 | 8665 | || vectorizable_induction (stmt, NULL, NULL, node) |
42fd8198 IE |
8666 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
8667 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 RB |
8668 | else |
8669 | { | |
8670 | if (bb_vinfo) | |
8671 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
8672 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
8673 | || vectorizable_shift (stmt, NULL, NULL, node) | |
8674 | || vectorizable_operation (stmt, NULL, NULL, node) | |
8675 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
8676 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
8677 | || vectorizable_call (stmt, NULL, NULL, node) | |
8678 | || vectorizable_store (stmt, NULL, NULL, node) | |
42fd8198 IE |
8679 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
8680 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 | 8681 | } |
8644a673 IR |
8682 | |
8683 | if (!ok) | |
ebfd146a | 8684 | { |
73fbfcad | 8685 | if (dump_enabled_p ()) |
8644a673 | 8686 | { |
78c60e3d SS |
8687 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8688 | "not vectorized: relevant stmt not "); | |
8689 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
8690 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 8691 | } |
b8698a0f | 8692 | |
ebfd146a IR |
8693 | return false; |
8694 | } | |
8695 | ||
a70d6342 IR |
8696 | if (bb_vinfo) |
8697 | return true; | |
8698 | ||
8644a673 IR |
8699 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
8700 | need extra handling, except for vectorizable reductions. */ | |
8701 | if (STMT_VINFO_LIVE_P (stmt_info) | |
8702 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
b28ead45 | 8703 | ok = vectorizable_live_operation (stmt, NULL, NULL, -1, NULL); |
ebfd146a | 8704 | |
8644a673 | 8705 | if (!ok) |
ebfd146a | 8706 | { |
73fbfcad | 8707 | if (dump_enabled_p ()) |
8644a673 | 8708 | { |
78c60e3d SS |
8709 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8710 | "not vectorized: live stmt not "); | |
8711 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
8712 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 8713 | } |
b8698a0f | 8714 | |
8644a673 | 8715 | return false; |
ebfd146a IR |
8716 | } |
8717 | ||
ebfd146a IR |
8718 | return true; |
8719 | } | |
8720 | ||
8721 | ||
8722 | /* Function vect_transform_stmt. | |
8723 | ||
8724 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
8725 | ||
8726 | bool | |
355fe088 | 8727 | vect_transform_stmt (gimple *stmt, gimple_stmt_iterator *gsi, |
0d0293ac | 8728 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
8729 | slp_instance slp_node_instance) |
8730 | { | |
8731 | bool is_store = false; | |
355fe088 | 8732 | gimple *vec_stmt = NULL; |
ebfd146a | 8733 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
ebfd146a | 8734 | bool done; |
ebfd146a | 8735 | |
fce57248 | 8736 | gcc_assert (slp_node || !PURE_SLP_STMT (stmt_info)); |
355fe088 | 8737 | gimple *old_vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
225ce44b | 8738 | |
ebfd146a IR |
8739 | switch (STMT_VINFO_TYPE (stmt_info)) |
8740 | { | |
8741 | case type_demotion_vec_info_type: | |
ebfd146a | 8742 | case type_promotion_vec_info_type: |
ebfd146a IR |
8743 | case type_conversion_vec_info_type: |
8744 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
8745 | gcc_assert (done); | |
8746 | break; | |
8747 | ||
8748 | case induc_vec_info_type: | |
e7baeb39 | 8749 | done = vectorizable_induction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
8750 | gcc_assert (done); |
8751 | break; | |
8752 | ||
9dc3f7de IR |
8753 | case shift_vec_info_type: |
8754 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
8755 | gcc_assert (done); | |
8756 | break; | |
8757 | ||
ebfd146a IR |
8758 | case op_vec_info_type: |
8759 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
8760 | gcc_assert (done); | |
8761 | break; | |
8762 | ||
8763 | case assignment_vec_info_type: | |
8764 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
8765 | gcc_assert (done); | |
8766 | break; | |
8767 | ||
8768 | case load_vec_info_type: | |
b8698a0f | 8769 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
8770 | slp_node_instance); |
8771 | gcc_assert (done); | |
8772 | break; | |
8773 | ||
8774 | case store_vec_info_type: | |
8775 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
8776 | gcc_assert (done); | |
0d0293ac | 8777 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
8778 | { |
8779 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 8780 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
8781 | one are skipped, and there vec_stmt_info shouldn't be freed |
8782 | meanwhile. */ | |
0d0293ac | 8783 | *grouped_store = true; |
ebfd146a IR |
8784 | if (STMT_VINFO_VEC_STMT (stmt_info)) |
8785 | is_store = true; | |
8786 | } | |
8787 | else | |
8788 | is_store = true; | |
8789 | break; | |
8790 | ||
8791 | case condition_vec_info_type: | |
f7e531cf | 8792 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
8793 | gcc_assert (done); |
8794 | break; | |
8795 | ||
42fd8198 IE |
8796 | case comparison_vec_info_type: |
8797 | done = vectorizable_comparison (stmt, gsi, &vec_stmt, NULL, slp_node); | |
8798 | gcc_assert (done); | |
8799 | break; | |
8800 | ||
ebfd146a | 8801 | case call_vec_info_type: |
190c2236 | 8802 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 8803 | stmt = gsi_stmt (*gsi); |
8e4284d0 | 8804 | if (gimple_call_internal_p (stmt, IFN_MASK_STORE)) |
5ce9450f | 8805 | is_store = true; |
ebfd146a IR |
8806 | break; |
8807 | ||
0136f8f0 AH |
8808 | case call_simd_clone_vec_info_type: |
8809 | done = vectorizable_simd_clone_call (stmt, gsi, &vec_stmt, slp_node); | |
8810 | stmt = gsi_stmt (*gsi); | |
8811 | break; | |
8812 | ||
ebfd146a | 8813 | case reduc_vec_info_type: |
891ad31c RB |
8814 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node, |
8815 | slp_node_instance); | |
ebfd146a IR |
8816 | gcc_assert (done); |
8817 | break; | |
8818 | ||
8819 | default: | |
8820 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
8821 | { | |
73fbfcad | 8822 | if (dump_enabled_p ()) |
78c60e3d | 8823 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8824 | "stmt not supported.\n"); |
ebfd146a IR |
8825 | gcc_unreachable (); |
8826 | } | |
8827 | } | |
8828 | ||
225ce44b RB |
8829 | /* Verify SLP vectorization doesn't mess with STMT_VINFO_VEC_STMT. |
8830 | This would break hybrid SLP vectorization. */ | |
8831 | if (slp_node) | |
d90f8440 RB |
8832 | gcc_assert (!vec_stmt |
8833 | && STMT_VINFO_VEC_STMT (stmt_info) == old_vec_stmt); | |
225ce44b | 8834 | |
ebfd146a IR |
8835 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that |
8836 | is being vectorized, but outside the immediately enclosing loop. */ | |
8837 | if (vec_stmt | |
a70d6342 IR |
8838 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
8839 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
8840 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
8841 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
8842 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 8843 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 8844 | vect_used_in_outer_by_reduction)) |
ebfd146a | 8845 | { |
a70d6342 IR |
8846 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
8847 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
8848 | imm_use_iterator imm_iter; |
8849 | use_operand_p use_p; | |
8850 | tree scalar_dest; | |
355fe088 | 8851 | gimple *exit_phi; |
ebfd146a | 8852 | |
73fbfcad | 8853 | if (dump_enabled_p ()) |
78c60e3d | 8854 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 8855 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
8856 | |
8857 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
8858 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
8859 | STMT). */ | |
8860 | if (gimple_code (stmt) == GIMPLE_PHI) | |
8861 | scalar_dest = PHI_RESULT (stmt); | |
8862 | else | |
8863 | scalar_dest = gimple_assign_lhs (stmt); | |
8864 | ||
8865 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
8866 | { | |
8867 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
8868 | { | |
8869 | exit_phi = USE_STMT (use_p); | |
8870 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
8871 | } | |
8872 | } | |
8873 | } | |
8874 | ||
8875 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
8876 | being vectorized. */ | |
b28ead45 AH |
8877 | if (slp_node) |
8878 | { | |
8879 | gimple *slp_stmt; | |
8880 | int i; | |
bd2f172f RB |
8881 | if (STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) |
8882 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (slp_node), i, slp_stmt) | |
8883 | { | |
8884 | stmt_vec_info slp_stmt_info = vinfo_for_stmt (slp_stmt); | |
8885 | if (STMT_VINFO_LIVE_P (slp_stmt_info)) | |
8886 | { | |
8887 | done = vectorizable_live_operation (slp_stmt, gsi, slp_node, i, | |
8888 | &vec_stmt); | |
8889 | gcc_assert (done); | |
8890 | } | |
8891 | } | |
b28ead45 AH |
8892 | } |
8893 | else if (STMT_VINFO_LIVE_P (stmt_info) | |
bd2f172f | 8894 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) |
ebfd146a | 8895 | { |
b28ead45 | 8896 | done = vectorizable_live_operation (stmt, gsi, slp_node, -1, &vec_stmt); |
ebfd146a IR |
8897 | gcc_assert (done); |
8898 | } | |
8899 | ||
8900 | if (vec_stmt) | |
83197f37 | 8901 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 8902 | |
b8698a0f | 8903 | return is_store; |
ebfd146a IR |
8904 | } |
8905 | ||
8906 | ||
b8698a0f | 8907 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
8908 | stmt_vec_info. */ |
8909 | ||
8910 | void | |
355fe088 | 8911 | vect_remove_stores (gimple *first_stmt) |
ebfd146a | 8912 | { |
355fe088 TS |
8913 | gimple *next = first_stmt; |
8914 | gimple *tmp; | |
ebfd146a IR |
8915 | gimple_stmt_iterator next_si; |
8916 | ||
8917 | while (next) | |
8918 | { | |
78048b1c JJ |
8919 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
8920 | ||
8921 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
8922 | if (is_pattern_stmt_p (stmt_info)) | |
8923 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
8924 | /* Free the attached stmt_vec_info and remove the stmt. */ |
8925 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 8926 | unlink_stmt_vdef (next); |
ebfd146a | 8927 | gsi_remove (&next_si, true); |
3d3f2249 | 8928 | release_defs (next); |
ebfd146a IR |
8929 | free_stmt_vec_info (next); |
8930 | next = tmp; | |
8931 | } | |
8932 | } | |
8933 | ||
8934 | ||
8935 | /* Function new_stmt_vec_info. | |
8936 | ||
8937 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
8938 | ||
8939 | stmt_vec_info | |
310213d4 | 8940 | new_stmt_vec_info (gimple *stmt, vec_info *vinfo) |
ebfd146a IR |
8941 | { |
8942 | stmt_vec_info res; | |
8943 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
8944 | ||
8945 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
8946 | STMT_VINFO_STMT (res) = stmt; | |
310213d4 | 8947 | res->vinfo = vinfo; |
8644a673 | 8948 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
8949 | STMT_VINFO_LIVE_P (res) = false; |
8950 | STMT_VINFO_VECTYPE (res) = NULL; | |
8951 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 8952 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
8953 | STMT_VINFO_IN_PATTERN_P (res) = false; |
8954 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 8955 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a | 8956 | STMT_VINFO_DATA_REF (res) = NULL; |
af29617a | 8957 | STMT_VINFO_VEC_REDUCTION_TYPE (res) = TREE_CODE_REDUCTION; |
7e16ce79 | 8958 | STMT_VINFO_VEC_CONST_COND_REDUC_CODE (res) = ERROR_MARK; |
ebfd146a | 8959 | |
ebfd146a IR |
8960 | if (gimple_code (stmt) == GIMPLE_PHI |
8961 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
8962 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
8963 | else | |
8644a673 IR |
8964 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
8965 | ||
9771b263 | 8966 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 8967 | STMT_SLP_TYPE (res) = loop_vect; |
78810bd3 RB |
8968 | STMT_VINFO_NUM_SLP_USES (res) = 0; |
8969 | ||
e14c1050 IR |
8970 | GROUP_FIRST_ELEMENT (res) = NULL; |
8971 | GROUP_NEXT_ELEMENT (res) = NULL; | |
8972 | GROUP_SIZE (res) = 0; | |
8973 | GROUP_STORE_COUNT (res) = 0; | |
8974 | GROUP_GAP (res) = 0; | |
8975 | GROUP_SAME_DR_STMT (res) = NULL; | |
ebfd146a IR |
8976 | |
8977 | return res; | |
8978 | } | |
8979 | ||
8980 | ||
8981 | /* Create a hash table for stmt_vec_info. */ | |
8982 | ||
8983 | void | |
8984 | init_stmt_vec_info_vec (void) | |
8985 | { | |
9771b263 DN |
8986 | gcc_assert (!stmt_vec_info_vec.exists ()); |
8987 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
8988 | } |
8989 | ||
8990 | ||
8991 | /* Free hash table for stmt_vec_info. */ | |
8992 | ||
8993 | void | |
8994 | free_stmt_vec_info_vec (void) | |
8995 | { | |
93675444 | 8996 | unsigned int i; |
3161455c | 8997 | stmt_vec_info info; |
93675444 JJ |
8998 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) |
8999 | if (info != NULL) | |
3161455c | 9000 | free_stmt_vec_info (STMT_VINFO_STMT (info)); |
9771b263 DN |
9001 | gcc_assert (stmt_vec_info_vec.exists ()); |
9002 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
9003 | } |
9004 | ||
9005 | ||
9006 | /* Free stmt vectorization related info. */ | |
9007 | ||
9008 | void | |
355fe088 | 9009 | free_stmt_vec_info (gimple *stmt) |
ebfd146a IR |
9010 | { |
9011 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
9012 | ||
9013 | if (!stmt_info) | |
9014 | return; | |
9015 | ||
78048b1c JJ |
9016 | /* Check if this statement has a related "pattern stmt" |
9017 | (introduced by the vectorizer during the pattern recognition | |
9018 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
9019 | too. */ | |
9020 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
9021 | { | |
9022 | stmt_vec_info patt_info | |
9023 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
9024 | if (patt_info) | |
9025 | { | |
363477c0 | 9026 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
355fe088 | 9027 | gimple *patt_stmt = STMT_VINFO_STMT (patt_info); |
f0281fde RB |
9028 | gimple_set_bb (patt_stmt, NULL); |
9029 | tree lhs = gimple_get_lhs (patt_stmt); | |
e6f5c25d | 9030 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde | 9031 | release_ssa_name (lhs); |
363477c0 JJ |
9032 | if (seq) |
9033 | { | |
9034 | gimple_stmt_iterator si; | |
9035 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
f0281fde | 9036 | { |
355fe088 | 9037 | gimple *seq_stmt = gsi_stmt (si); |
f0281fde | 9038 | gimple_set_bb (seq_stmt, NULL); |
7532abf2 | 9039 | lhs = gimple_get_lhs (seq_stmt); |
e6f5c25d | 9040 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde RB |
9041 | release_ssa_name (lhs); |
9042 | free_stmt_vec_info (seq_stmt); | |
9043 | } | |
363477c0 | 9044 | } |
f0281fde | 9045 | free_stmt_vec_info (patt_stmt); |
78048b1c JJ |
9046 | } |
9047 | } | |
9048 | ||
9771b263 | 9049 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
6c9e85fb | 9050 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).release (); |
ebfd146a IR |
9051 | set_vinfo_for_stmt (stmt, NULL); |
9052 | free (stmt_info); | |
9053 | } | |
9054 | ||
9055 | ||
bb67d9c7 | 9056 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 9057 | |
bb67d9c7 | 9058 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
9059 | by the target. */ |
9060 | ||
bb67d9c7 RG |
9061 | static tree |
9062 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a | 9063 | { |
c7d97b28 | 9064 | tree orig_scalar_type = scalar_type; |
3bd8f481 | 9065 | scalar_mode inner_mode; |
ef4bddc2 | 9066 | machine_mode simd_mode; |
ebfd146a IR |
9067 | int nunits; |
9068 | tree vectype; | |
9069 | ||
3bd8f481 RS |
9070 | if (!is_int_mode (TYPE_MODE (scalar_type), &inner_mode) |
9071 | && !is_float_mode (TYPE_MODE (scalar_type), &inner_mode)) | |
ebfd146a IR |
9072 | return NULL_TREE; |
9073 | ||
3bd8f481 | 9074 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
48f2e373 | 9075 | |
7b7b1813 RG |
9076 | /* For vector types of elements whose mode precision doesn't |
9077 | match their types precision we use a element type of mode | |
9078 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
9079 | they support the proper result truncation/extension. |
9080 | We also make sure to build vector types with INTEGER_TYPE | |
9081 | component type only. */ | |
6d7971b8 | 9082 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
9083 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
9084 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
9085 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
9086 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 9087 | |
ccbf5bb4 RG |
9088 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
9089 | When the component mode passes the above test simply use a type | |
9090 | corresponding to that mode. The theory is that any use that | |
9091 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 9092 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 9093 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
9094 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
9095 | ||
9096 | /* We can't build a vector type of elements with alignment bigger than | |
9097 | their size. */ | |
dfc2e2ac | 9098 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
9099 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
9100 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 9101 | |
dfc2e2ac RB |
9102 | /* If we felt back to using the mode fail if there was |
9103 | no scalar type for it. */ | |
9104 | if (scalar_type == NULL_TREE) | |
9105 | return NULL_TREE; | |
9106 | ||
bb67d9c7 RG |
9107 | /* If no size was supplied use the mode the target prefers. Otherwise |
9108 | lookup a vector mode of the specified size. */ | |
9109 | if (size == 0) | |
9110 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
9da15d40 RS |
9111 | else if (!mode_for_vector (inner_mode, size / nbytes).exists (&simd_mode)) |
9112 | return NULL_TREE; | |
cc4b5170 | 9113 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
4c8fd8ac JB |
9114 | /* NOTE: nunits == 1 is allowed to support single element vector types. */ |
9115 | if (nunits < 1) | |
cc4b5170 | 9116 | return NULL_TREE; |
ebfd146a IR |
9117 | |
9118 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
9119 | |
9120 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
9121 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 9122 | return NULL_TREE; |
ebfd146a | 9123 | |
c7d97b28 RB |
9124 | /* Re-attach the address-space qualifier if we canonicalized the scalar |
9125 | type. */ | |
9126 | if (TYPE_ADDR_SPACE (orig_scalar_type) != TYPE_ADDR_SPACE (vectype)) | |
9127 | return build_qualified_type | |
9128 | (vectype, KEEP_QUAL_ADDR_SPACE (TYPE_QUALS (orig_scalar_type))); | |
9129 | ||
ebfd146a IR |
9130 | return vectype; |
9131 | } | |
9132 | ||
bb67d9c7 RG |
9133 | unsigned int current_vector_size; |
9134 | ||
9135 | /* Function get_vectype_for_scalar_type. | |
9136 | ||
9137 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
9138 | by the target. */ | |
9139 | ||
9140 | tree | |
9141 | get_vectype_for_scalar_type (tree scalar_type) | |
9142 | { | |
9143 | tree vectype; | |
9144 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
9145 | current_vector_size); | |
9146 | if (vectype | |
9147 | && current_vector_size == 0) | |
9148 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
9149 | return vectype; | |
9150 | } | |
9151 | ||
42fd8198 IE |
9152 | /* Function get_mask_type_for_scalar_type. |
9153 | ||
9154 | Returns the mask type corresponding to a result of comparison | |
9155 | of vectors of specified SCALAR_TYPE as supported by target. */ | |
9156 | ||
9157 | tree | |
9158 | get_mask_type_for_scalar_type (tree scalar_type) | |
9159 | { | |
9160 | tree vectype = get_vectype_for_scalar_type (scalar_type); | |
9161 | ||
9162 | if (!vectype) | |
9163 | return NULL; | |
9164 | ||
9165 | return build_truth_vector_type (TYPE_VECTOR_SUBPARTS (vectype), | |
9166 | current_vector_size); | |
9167 | } | |
9168 | ||
b690cc0f RG |
9169 | /* Function get_same_sized_vectype |
9170 | ||
9171 | Returns a vector type corresponding to SCALAR_TYPE of size | |
9172 | VECTOR_TYPE if supported by the target. */ | |
9173 | ||
9174 | tree | |
bb67d9c7 | 9175 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 9176 | { |
2568d8a1 | 9177 | if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type)) |
9f47c7e5 IE |
9178 | return build_same_sized_truth_vector_type (vector_type); |
9179 | ||
bb67d9c7 RG |
9180 | return get_vectype_for_scalar_type_and_size |
9181 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
9182 | } |
9183 | ||
ebfd146a IR |
9184 | /* Function vect_is_simple_use. |
9185 | ||
9186 | Input: | |
81c40241 RB |
9187 | VINFO - the vect info of the loop or basic block that is being vectorized. |
9188 | OPERAND - operand in the loop or bb. | |
9189 | Output: | |
9190 | DEF_STMT - the defining stmt in case OPERAND is an SSA_NAME. | |
9191 | DT - the type of definition | |
ebfd146a IR |
9192 | |
9193 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 9194 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 9195 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 9196 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
9197 | is the case in reduction/induction computations). |
9198 | For basic blocks, supportable operands are constants and bb invariants. | |
9199 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
9200 | |
9201 | bool | |
81c40241 RB |
9202 | vect_is_simple_use (tree operand, vec_info *vinfo, |
9203 | gimple **def_stmt, enum vect_def_type *dt) | |
b8698a0f | 9204 | { |
ebfd146a | 9205 | *def_stmt = NULL; |
3fc356dc | 9206 | *dt = vect_unknown_def_type; |
b8698a0f | 9207 | |
73fbfcad | 9208 | if (dump_enabled_p ()) |
ebfd146a | 9209 | { |
78c60e3d SS |
9210 | dump_printf_loc (MSG_NOTE, vect_location, |
9211 | "vect_is_simple_use: operand "); | |
9212 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
e645e942 | 9213 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 9214 | } |
b8698a0f | 9215 | |
b758f602 | 9216 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
9217 | { |
9218 | *dt = vect_constant_def; | |
9219 | return true; | |
9220 | } | |
b8698a0f | 9221 | |
ebfd146a IR |
9222 | if (is_gimple_min_invariant (operand)) |
9223 | { | |
8644a673 | 9224 | *dt = vect_external_def; |
ebfd146a IR |
9225 | return true; |
9226 | } | |
9227 | ||
ebfd146a IR |
9228 | if (TREE_CODE (operand) != SSA_NAME) |
9229 | { | |
73fbfcad | 9230 | if (dump_enabled_p ()) |
af29617a AH |
9231 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9232 | "not ssa-name.\n"); | |
ebfd146a IR |
9233 | return false; |
9234 | } | |
b8698a0f | 9235 | |
3fc356dc | 9236 | if (SSA_NAME_IS_DEFAULT_DEF (operand)) |
ebfd146a | 9237 | { |
3fc356dc RB |
9238 | *dt = vect_external_def; |
9239 | return true; | |
ebfd146a IR |
9240 | } |
9241 | ||
3fc356dc | 9242 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
73fbfcad | 9243 | if (dump_enabled_p ()) |
ebfd146a | 9244 | { |
78c60e3d SS |
9245 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
9246 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
ebfd146a IR |
9247 | } |
9248 | ||
61d371eb | 9249 | if (! vect_stmt_in_region_p (vinfo, *def_stmt)) |
8644a673 | 9250 | *dt = vect_external_def; |
ebfd146a IR |
9251 | else |
9252 | { | |
3fc356dc | 9253 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (*def_stmt); |
603cca93 | 9254 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
ebfd146a IR |
9255 | } |
9256 | ||
2e8ab70c RB |
9257 | if (dump_enabled_p ()) |
9258 | { | |
9259 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: "); | |
9260 | switch (*dt) | |
9261 | { | |
9262 | case vect_uninitialized_def: | |
9263 | dump_printf (MSG_NOTE, "uninitialized\n"); | |
9264 | break; | |
9265 | case vect_constant_def: | |
9266 | dump_printf (MSG_NOTE, "constant\n"); | |
9267 | break; | |
9268 | case vect_external_def: | |
9269 | dump_printf (MSG_NOTE, "external\n"); | |
9270 | break; | |
9271 | case vect_internal_def: | |
9272 | dump_printf (MSG_NOTE, "internal\n"); | |
9273 | break; | |
9274 | case vect_induction_def: | |
9275 | dump_printf (MSG_NOTE, "induction\n"); | |
9276 | break; | |
9277 | case vect_reduction_def: | |
9278 | dump_printf (MSG_NOTE, "reduction\n"); | |
9279 | break; | |
9280 | case vect_double_reduction_def: | |
9281 | dump_printf (MSG_NOTE, "double reduction\n"); | |
9282 | break; | |
9283 | case vect_nested_cycle: | |
9284 | dump_printf (MSG_NOTE, "nested cycle\n"); | |
9285 | break; | |
9286 | case vect_unknown_def_type: | |
9287 | dump_printf (MSG_NOTE, "unknown\n"); | |
9288 | break; | |
9289 | } | |
9290 | } | |
9291 | ||
81c40241 | 9292 | if (*dt == vect_unknown_def_type) |
ebfd146a | 9293 | { |
73fbfcad | 9294 | if (dump_enabled_p ()) |
78c60e3d | 9295 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9296 | "Unsupported pattern.\n"); |
ebfd146a IR |
9297 | return false; |
9298 | } | |
9299 | ||
ebfd146a IR |
9300 | switch (gimple_code (*def_stmt)) |
9301 | { | |
9302 | case GIMPLE_PHI: | |
ebfd146a | 9303 | case GIMPLE_ASSIGN: |
ebfd146a | 9304 | case GIMPLE_CALL: |
81c40241 | 9305 | break; |
ebfd146a | 9306 | default: |
73fbfcad | 9307 | if (dump_enabled_p ()) |
78c60e3d | 9308 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9309 | "unsupported defining stmt:\n"); |
ebfd146a IR |
9310 | return false; |
9311 | } | |
9312 | ||
9313 | return true; | |
9314 | } | |
9315 | ||
81c40241 | 9316 | /* Function vect_is_simple_use. |
b690cc0f | 9317 | |
81c40241 | 9318 | Same as vect_is_simple_use but also determines the vector operand |
b690cc0f RG |
9319 | type of OPERAND and stores it to *VECTYPE. If the definition of |
9320 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
9321 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
9322 | is responsible to compute the best suited vector type for the | |
9323 | scalar operand. */ | |
9324 | ||
9325 | bool | |
81c40241 RB |
9326 | vect_is_simple_use (tree operand, vec_info *vinfo, |
9327 | gimple **def_stmt, enum vect_def_type *dt, tree *vectype) | |
b690cc0f | 9328 | { |
81c40241 | 9329 | if (!vect_is_simple_use (operand, vinfo, def_stmt, dt)) |
b690cc0f RG |
9330 | return false; |
9331 | ||
9332 | /* Now get a vector type if the def is internal, otherwise supply | |
9333 | NULL_TREE and leave it up to the caller to figure out a proper | |
9334 | type for the use stmt. */ | |
9335 | if (*dt == vect_internal_def | |
9336 | || *dt == vect_induction_def | |
9337 | || *dt == vect_reduction_def | |
9338 | || *dt == vect_double_reduction_def | |
9339 | || *dt == vect_nested_cycle) | |
9340 | { | |
9341 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
9342 | |
9343 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
9344 | && !STMT_VINFO_RELEVANT (stmt_info) | |
9345 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 9346 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 9347 | |
b690cc0f RG |
9348 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
9349 | gcc_assert (*vectype != NULL_TREE); | |
9350 | } | |
9351 | else if (*dt == vect_uninitialized_def | |
9352 | || *dt == vect_constant_def | |
9353 | || *dt == vect_external_def) | |
9354 | *vectype = NULL_TREE; | |
9355 | else | |
9356 | gcc_unreachable (); | |
9357 | ||
9358 | return true; | |
9359 | } | |
9360 | ||
ebfd146a IR |
9361 | |
9362 | /* Function supportable_widening_operation | |
9363 | ||
b8698a0f L |
9364 | Check whether an operation represented by the code CODE is a |
9365 | widening operation that is supported by the target platform in | |
b690cc0f RG |
9366 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
9367 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 9368 | |
ebfd146a IR |
9369 | Widening operations we currently support are NOP (CONVERT), FLOAT |
9370 | and WIDEN_MULT. This function checks if these operations are supported | |
9371 | by the target platform either directly (via vector tree-codes), or via | |
9372 | target builtins. | |
9373 | ||
9374 | Output: | |
b8698a0f L |
9375 | - CODE1 and CODE2 are codes of vector operations to be used when |
9376 | vectorizing the operation, if available. | |
ebfd146a IR |
9377 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
9378 | case of multi-step conversion (like char->short->int - in that case | |
9379 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
9380 | - INTERM_TYPES contains the intermediate type required to perform the |
9381 | widening operation (short in the above example). */ | |
ebfd146a IR |
9382 | |
9383 | bool | |
355fe088 | 9384 | supportable_widening_operation (enum tree_code code, gimple *stmt, |
b690cc0f | 9385 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
9386 | enum tree_code *code1, enum tree_code *code2, |
9387 | int *multi_step_cvt, | |
9771b263 | 9388 | vec<tree> *interm_types) |
ebfd146a IR |
9389 | { |
9390 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
9391 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 9392 | struct loop *vect_loop = NULL; |
ef4bddc2 | 9393 | machine_mode vec_mode; |
81f40b79 | 9394 | enum insn_code icode1, icode2; |
ebfd146a | 9395 | optab optab1, optab2; |
b690cc0f RG |
9396 | tree vectype = vectype_in; |
9397 | tree wide_vectype = vectype_out; | |
ebfd146a | 9398 | enum tree_code c1, c2; |
4a00c761 JJ |
9399 | int i; |
9400 | tree prev_type, intermediate_type; | |
ef4bddc2 | 9401 | machine_mode intermediate_mode, prev_mode; |
4a00c761 | 9402 | optab optab3, optab4; |
ebfd146a | 9403 | |
4a00c761 | 9404 | *multi_step_cvt = 0; |
4ef69dfc IR |
9405 | if (loop_info) |
9406 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
9407 | ||
ebfd146a IR |
9408 | switch (code) |
9409 | { | |
9410 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
9411 | /* The result of a vectorized widening operation usually requires |
9412 | two vectors (because the widened results do not fit into one vector). | |
9413 | The generated vector results would normally be expected to be | |
9414 | generated in the same order as in the original scalar computation, | |
9415 | i.e. if 8 results are generated in each vector iteration, they are | |
9416 | to be organized as follows: | |
9417 | vect1: [res1,res2,res3,res4], | |
9418 | vect2: [res5,res6,res7,res8]. | |
9419 | ||
9420 | However, in the special case that the result of the widening | |
9421 | operation is used in a reduction computation only, the order doesn't | |
9422 | matter (because when vectorizing a reduction we change the order of | |
9423 | the computation). Some targets can take advantage of this and | |
9424 | generate more efficient code. For example, targets like Altivec, | |
9425 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
9426 | generate the following vectors: | |
9427 | vect1: [res1,res3,res5,res7], | |
9428 | vect2: [res2,res4,res6,res8]. | |
9429 | ||
9430 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
9431 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
9432 | iterations in parallel). We therefore don't allow to change the | |
9433 | order of the computation in the inner-loop during outer-loop | |
9434 | vectorization. */ | |
9435 | /* TODO: Another case in which order doesn't *really* matter is when we | |
9436 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
9437 | Normally, pack_trunc performs an even/odd permute, whereas the | |
9438 | repack from an even/odd expansion would be an interleave, which | |
9439 | would be significantly simpler for e.g. AVX2. */ | |
9440 | /* In any case, in order to avoid duplicating the code below, recurse | |
9441 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
9442 | are properly set up for the caller. If we fail, we'll continue with | |
9443 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
9444 | if (vect_loop | |
9445 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
9446 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
9447 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
9448 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
9449 | code1, code2, multi_step_cvt, |
9450 | interm_types)) | |
ebc047a2 CH |
9451 | { |
9452 | /* Elements in a vector with vect_used_by_reduction property cannot | |
9453 | be reordered if the use chain with this property does not have the | |
9454 | same operation. One such an example is s += a * b, where elements | |
9455 | in a and b cannot be reordered. Here we check if the vector defined | |
9456 | by STMT is only directly used in the reduction statement. */ | |
9457 | tree lhs = gimple_assign_lhs (stmt); | |
9458 | use_operand_p dummy; | |
355fe088 | 9459 | gimple *use_stmt; |
ebc047a2 CH |
9460 | stmt_vec_info use_stmt_info = NULL; |
9461 | if (single_imm_use (lhs, &dummy, &use_stmt) | |
9462 | && (use_stmt_info = vinfo_for_stmt (use_stmt)) | |
9463 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) | |
9464 | return true; | |
9465 | } | |
4a00c761 JJ |
9466 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
9467 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
9468 | break; |
9469 | ||
81c40241 RB |
9470 | case DOT_PROD_EXPR: |
9471 | c1 = DOT_PROD_EXPR; | |
9472 | c2 = DOT_PROD_EXPR; | |
9473 | break; | |
9474 | ||
9475 | case SAD_EXPR: | |
9476 | c1 = SAD_EXPR; | |
9477 | c2 = SAD_EXPR; | |
9478 | break; | |
9479 | ||
6ae6116f RH |
9480 | case VEC_WIDEN_MULT_EVEN_EXPR: |
9481 | /* Support the recursion induced just above. */ | |
9482 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
9483 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
9484 | break; | |
9485 | ||
36ba4aae | 9486 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
9487 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
9488 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
9489 | break; |
9490 | ||
ebfd146a | 9491 | CASE_CONVERT: |
4a00c761 JJ |
9492 | c1 = VEC_UNPACK_LO_EXPR; |
9493 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
9494 | break; |
9495 | ||
9496 | case FLOAT_EXPR: | |
4a00c761 JJ |
9497 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
9498 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
9499 | break; |
9500 | ||
9501 | case FIX_TRUNC_EXPR: | |
9502 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
9503 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
9504 | computing the operation. */ | |
9505 | return false; | |
9506 | ||
9507 | default: | |
9508 | gcc_unreachable (); | |
9509 | } | |
9510 | ||
6ae6116f | 9511 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
6b4db501 | 9512 | std::swap (c1, c2); |
4a00c761 | 9513 | |
ebfd146a IR |
9514 | if (code == FIX_TRUNC_EXPR) |
9515 | { | |
9516 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
9517 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
9518 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
9519 | } |
9520 | else | |
9521 | { | |
9522 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
9523 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
9524 | } | |
9525 | ||
9526 | if (!optab1 || !optab2) | |
9527 | return false; | |
9528 | ||
9529 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
9530 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
9531 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
9532 | return false; |
9533 | ||
4a00c761 JJ |
9534 | *code1 = c1; |
9535 | *code2 = c2; | |
9536 | ||
9537 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
9538 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff IE |
9539 | /* For scalar masks we may have different boolean |
9540 | vector types having the same QImode. Thus we | |
9541 | add additional check for elements number. */ | |
9542 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
9543 | || (TYPE_VECTOR_SUBPARTS (vectype) / 2 | |
9544 | == TYPE_VECTOR_SUBPARTS (wide_vectype))); | |
4a00c761 | 9545 | |
b8698a0f | 9546 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 9547 | types. */ |
ebfd146a | 9548 | |
4a00c761 JJ |
9549 | prev_type = vectype; |
9550 | prev_mode = vec_mode; | |
b8698a0f | 9551 | |
4a00c761 JJ |
9552 | if (!CONVERT_EXPR_CODE_P (code)) |
9553 | return false; | |
b8698a0f | 9554 | |
4a00c761 JJ |
9555 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
9556 | intermediate steps in promotion sequence. We try | |
9557 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
9558 | not. */ | |
9771b263 | 9559 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
9560 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
9561 | { | |
9562 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
9563 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
9564 | { | |
9565 | intermediate_type | |
9566 | = build_truth_vector_type (TYPE_VECTOR_SUBPARTS (prev_type) / 2, | |
9567 | current_vector_size); | |
9568 | if (intermediate_mode != TYPE_MODE (intermediate_type)) | |
9569 | return false; | |
9570 | } | |
9571 | else | |
9572 | intermediate_type | |
9573 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
9574 | TYPE_UNSIGNED (prev_type)); | |
9575 | ||
4a00c761 JJ |
9576 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); |
9577 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
9578 | ||
9579 | if (!optab3 || !optab4 | |
9580 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
9581 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
9582 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
9583 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
9584 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
9585 | == CODE_FOR_nothing) | |
9586 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
9587 | == CODE_FOR_nothing)) | |
9588 | break; | |
ebfd146a | 9589 | |
9771b263 | 9590 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
9591 | (*multi_step_cvt)++; |
9592 | ||
9593 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
9594 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff IE |
9595 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
9596 | || (TYPE_VECTOR_SUBPARTS (intermediate_type) / 2 | |
9597 | == TYPE_VECTOR_SUBPARTS (wide_vectype))); | |
4a00c761 JJ |
9598 | |
9599 | prev_type = intermediate_type; | |
9600 | prev_mode = intermediate_mode; | |
ebfd146a IR |
9601 | } |
9602 | ||
9771b263 | 9603 | interm_types->release (); |
4a00c761 | 9604 | return false; |
ebfd146a IR |
9605 | } |
9606 | ||
9607 | ||
9608 | /* Function supportable_narrowing_operation | |
9609 | ||
b8698a0f L |
9610 | Check whether an operation represented by the code CODE is a |
9611 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
9612 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
9613 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 9614 | |
ebfd146a | 9615 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 9616 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
9617 | the target platform directly via vector tree-codes. |
9618 | ||
9619 | Output: | |
b8698a0f L |
9620 | - CODE1 is the code of a vector operation to be used when |
9621 | vectorizing the operation, if available. | |
ebfd146a IR |
9622 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
9623 | case of multi-step conversion (like int->short->char - in that case | |
9624 | MULTI_STEP_CVT will be 1). | |
9625 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 9626 | narrowing operation (short in the above example). */ |
ebfd146a IR |
9627 | |
9628 | bool | |
9629 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 9630 | tree vectype_out, tree vectype_in, |
ebfd146a | 9631 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 9632 | vec<tree> *interm_types) |
ebfd146a | 9633 | { |
ef4bddc2 | 9634 | machine_mode vec_mode; |
ebfd146a IR |
9635 | enum insn_code icode1; |
9636 | optab optab1, interm_optab; | |
b690cc0f RG |
9637 | tree vectype = vectype_in; |
9638 | tree narrow_vectype = vectype_out; | |
ebfd146a | 9639 | enum tree_code c1; |
3ae0661a | 9640 | tree intermediate_type, prev_type; |
ef4bddc2 | 9641 | machine_mode intermediate_mode, prev_mode; |
ebfd146a | 9642 | int i; |
4a00c761 | 9643 | bool uns; |
ebfd146a | 9644 | |
4a00c761 | 9645 | *multi_step_cvt = 0; |
ebfd146a IR |
9646 | switch (code) |
9647 | { | |
9648 | CASE_CONVERT: | |
9649 | c1 = VEC_PACK_TRUNC_EXPR; | |
9650 | break; | |
9651 | ||
9652 | case FIX_TRUNC_EXPR: | |
9653 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
9654 | break; | |
9655 | ||
9656 | case FLOAT_EXPR: | |
9657 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
9658 | tree code and optabs used for computing the operation. */ | |
9659 | return false; | |
9660 | ||
9661 | default: | |
9662 | gcc_unreachable (); | |
9663 | } | |
9664 | ||
9665 | if (code == FIX_TRUNC_EXPR) | |
9666 | /* The signedness is determined from output operand. */ | |
b690cc0f | 9667 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
9668 | else |
9669 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
9670 | ||
9671 | if (!optab1) | |
9672 | return false; | |
9673 | ||
9674 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 9675 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
9676 | return false; |
9677 | ||
4a00c761 JJ |
9678 | *code1 = c1; |
9679 | ||
9680 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff IE |
9681 | /* For scalar masks we may have different boolean |
9682 | vector types having the same QImode. Thus we | |
9683 | add additional check for elements number. */ | |
9684 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
9685 | || (TYPE_VECTOR_SUBPARTS (vectype) * 2 | |
9686 | == TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 | 9687 | |
ebfd146a IR |
9688 | /* Check if it's a multi-step conversion that can be done using intermediate |
9689 | types. */ | |
4a00c761 | 9690 | prev_mode = vec_mode; |
3ae0661a | 9691 | prev_type = vectype; |
4a00c761 JJ |
9692 | if (code == FIX_TRUNC_EXPR) |
9693 | uns = TYPE_UNSIGNED (vectype_out); | |
9694 | else | |
9695 | uns = TYPE_UNSIGNED (vectype); | |
9696 | ||
9697 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
9698 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
9699 | costly than signed. */ | |
9700 | if (code == FIX_TRUNC_EXPR && uns) | |
9701 | { | |
9702 | enum insn_code icode2; | |
9703 | ||
9704 | intermediate_type | |
9705 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
9706 | interm_optab | |
9707 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 9708 | if (interm_optab != unknown_optab |
4a00c761 JJ |
9709 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
9710 | && insn_data[icode1].operand[0].mode | |
9711 | == insn_data[icode2].operand[0].mode) | |
9712 | { | |
9713 | uns = false; | |
9714 | optab1 = interm_optab; | |
9715 | icode1 = icode2; | |
9716 | } | |
9717 | } | |
ebfd146a | 9718 | |
4a00c761 JJ |
9719 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
9720 | intermediate steps in promotion sequence. We try | |
9721 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 9722 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
9723 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
9724 | { | |
9725 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
9726 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
9727 | { | |
9728 | intermediate_type | |
9729 | = build_truth_vector_type (TYPE_VECTOR_SUBPARTS (prev_type) * 2, | |
9730 | current_vector_size); | |
9731 | if (intermediate_mode != TYPE_MODE (intermediate_type)) | |
9732 | return false; | |
9733 | } | |
9734 | else | |
9735 | intermediate_type | |
9736 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
4a00c761 JJ |
9737 | interm_optab |
9738 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
9739 | optab_default); | |
9740 | if (!interm_optab | |
9741 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
9742 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
9743 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
9744 | == CODE_FOR_nothing)) | |
9745 | break; | |
9746 | ||
9771b263 | 9747 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
9748 | (*multi_step_cvt)++; |
9749 | ||
9750 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff IE |
9751 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
9752 | || (TYPE_VECTOR_SUBPARTS (intermediate_type) * 2 | |
9753 | == TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 JJ |
9754 | |
9755 | prev_mode = intermediate_mode; | |
3ae0661a | 9756 | prev_type = intermediate_type; |
4a00c761 | 9757 | optab1 = interm_optab; |
ebfd146a IR |
9758 | } |
9759 | ||
9771b263 | 9760 | interm_types->release (); |
4a00c761 | 9761 | return false; |
ebfd146a | 9762 | } |