]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
85ec4feb | 2 | Copyright (C) 2003-2018 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" |
5ebaa477 | 51 | #include "tree-vector-builder.h" |
f151c9e1 | 52 | #include "vec-perm-indices.h" |
7cfb4d93 RS |
53 | #include "tree-ssa-loop-niter.h" |
54 | #include "gimple-fold.h" | |
ebfd146a | 55 | |
7ee2468b SB |
56 | /* For lang_hooks.types.type_for_mode. */ |
57 | #include "langhooks.h" | |
ebfd146a | 58 | |
c3e7ee41 BS |
59 | /* Return the vectorized type for the given statement. */ |
60 | ||
61 | tree | |
62 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
63 | { | |
64 | return STMT_VINFO_VECTYPE (stmt_info); | |
65 | } | |
66 | ||
67 | /* Return TRUE iff the given statement is in an inner loop relative to | |
68 | the loop being vectorized. */ | |
69 | bool | |
70 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
71 | { | |
355fe088 | 72 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
c3e7ee41 BS |
73 | basic_block bb = gimple_bb (stmt); |
74 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
75 | struct loop* loop; | |
76 | ||
77 | if (!loop_vinfo) | |
78 | return false; | |
79 | ||
80 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
81 | ||
82 | return (bb->loop_father == loop->inner); | |
83 | } | |
84 | ||
85 | /* Record the cost of a statement, either by directly informing the | |
86 | target model or by saving it in a vector for later processing. | |
87 | Return a preliminary estimate of the statement's cost. */ | |
88 | ||
89 | unsigned | |
92345349 | 90 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 91 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 92 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 93 | { |
cc9fe6bb JH |
94 | if ((kind == vector_load || kind == unaligned_load) |
95 | && STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
96 | kind = vector_gather_load; | |
97 | if ((kind == vector_store || kind == unaligned_store) | |
98 | && STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
99 | kind = vector_scatter_store; | |
92345349 | 100 | if (body_cost_vec) |
c3e7ee41 | 101 | { |
92345349 | 102 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
ddf56386 RB |
103 | stmt_info_for_cost si = { count, kind, |
104 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
105 | misalign }; | |
106 | body_cost_vec->safe_push (si); | |
c3e7ee41 | 107 | return (unsigned) |
92345349 | 108 | (builtin_vectorization_cost (kind, vectype, misalign) * count); |
c3e7ee41 BS |
109 | } |
110 | else | |
310213d4 RB |
111 | return add_stmt_cost (stmt_info->vinfo->target_cost_data, |
112 | count, kind, stmt_info, misalign, where); | |
c3e7ee41 BS |
113 | } |
114 | ||
272c6793 RS |
115 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
116 | ||
117 | static tree | |
118 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
119 | { | |
120 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
121 | "vect_array"); | |
122 | } | |
123 | ||
124 | /* ARRAY is an array of vectors created by create_vector_array. | |
125 | Return an SSA_NAME for the vector in index N. The reference | |
126 | is part of the vectorization of STMT and the vector is associated | |
127 | with scalar destination SCALAR_DEST. */ | |
128 | ||
129 | static tree | |
355fe088 | 130 | read_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree scalar_dest, |
272c6793 RS |
131 | tree array, unsigned HOST_WIDE_INT n) |
132 | { | |
133 | tree vect_type, vect, vect_name, array_ref; | |
355fe088 | 134 | gimple *new_stmt; |
272c6793 RS |
135 | |
136 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
137 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
138 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
139 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
140 | build_int_cst (size_type_node, n), | |
141 | NULL_TREE, NULL_TREE); | |
142 | ||
143 | new_stmt = gimple_build_assign (vect, array_ref); | |
144 | vect_name = make_ssa_name (vect, new_stmt); | |
145 | gimple_assign_set_lhs (new_stmt, vect_name); | |
146 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
147 | |
148 | return vect_name; | |
149 | } | |
150 | ||
151 | /* ARRAY is an array of vectors created by create_vector_array. | |
152 | Emit code to store SSA_NAME VECT in index N of the array. | |
153 | The store is part of the vectorization of STMT. */ | |
154 | ||
155 | static void | |
355fe088 | 156 | write_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree vect, |
272c6793 RS |
157 | tree array, unsigned HOST_WIDE_INT n) |
158 | { | |
159 | tree array_ref; | |
355fe088 | 160 | gimple *new_stmt; |
272c6793 RS |
161 | |
162 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
163 | build_int_cst (size_type_node, n), | |
164 | NULL_TREE, NULL_TREE); | |
165 | ||
166 | new_stmt = gimple_build_assign (array_ref, vect); | |
167 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
168 | } |
169 | ||
170 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
171 | of *PTR. The memory reference replaces those in FIRST_DR | |
172 | (and its group). */ | |
173 | ||
174 | static tree | |
44fc7854 | 175 | create_array_ref (tree type, tree ptr, tree alias_ptr_type) |
272c6793 | 176 | { |
44fc7854 | 177 | tree mem_ref; |
272c6793 | 178 | |
272c6793 RS |
179 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); |
180 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 181 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
182 | return mem_ref; |
183 | } | |
184 | ||
ebfd146a IR |
185 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
186 | ||
187 | /* Function vect_mark_relevant. | |
188 | ||
189 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
190 | ||
191 | static void | |
355fe088 | 192 | vect_mark_relevant (vec<gimple *> *worklist, gimple *stmt, |
97ecdb46 | 193 | enum vect_relevant relevant, bool live_p) |
ebfd146a IR |
194 | { |
195 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
196 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
197 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
355fe088 | 198 | gimple *pattern_stmt; |
ebfd146a | 199 | |
73fbfcad | 200 | if (dump_enabled_p ()) |
66c16fd9 RB |
201 | { |
202 | dump_printf_loc (MSG_NOTE, vect_location, | |
203 | "mark relevant %d, live %d: ", relevant, live_p); | |
204 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
205 | } | |
ebfd146a | 206 | |
83197f37 IR |
207 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
208 | related pattern stmt instead of the original stmt. However, such stmts | |
209 | may have their own uses that are not in any pattern, in such cases the | |
210 | stmt itself should be marked. */ | |
ebfd146a IR |
211 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
212 | { | |
97ecdb46 JJ |
213 | /* This is the last stmt in a sequence that was detected as a |
214 | pattern that can potentially be vectorized. Don't mark the stmt | |
215 | as relevant/live because it's not going to be vectorized. | |
216 | Instead mark the pattern-stmt that replaces it. */ | |
83197f37 | 217 | |
97ecdb46 JJ |
218 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); |
219 | ||
220 | if (dump_enabled_p ()) | |
221 | dump_printf_loc (MSG_NOTE, vect_location, | |
222 | "last stmt in pattern. don't mark" | |
223 | " relevant/live.\n"); | |
224 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
225 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
226 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
227 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
228 | stmt = pattern_stmt; | |
ebfd146a IR |
229 | } |
230 | ||
231 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
232 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
233 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
234 | ||
235 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
236 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
237 | { | |
73fbfcad | 238 | if (dump_enabled_p ()) |
78c60e3d | 239 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 240 | "already marked relevant/live.\n"); |
ebfd146a IR |
241 | return; |
242 | } | |
243 | ||
9771b263 | 244 | worklist->safe_push (stmt); |
ebfd146a IR |
245 | } |
246 | ||
247 | ||
b28ead45 AH |
248 | /* Function is_simple_and_all_uses_invariant |
249 | ||
250 | Return true if STMT is simple and all uses of it are invariant. */ | |
251 | ||
252 | bool | |
253 | is_simple_and_all_uses_invariant (gimple *stmt, loop_vec_info loop_vinfo) | |
254 | { | |
255 | tree op; | |
256 | gimple *def_stmt; | |
257 | ssa_op_iter iter; | |
258 | ||
259 | if (!is_gimple_assign (stmt)) | |
260 | return false; | |
261 | ||
262 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) | |
263 | { | |
264 | enum vect_def_type dt = vect_uninitialized_def; | |
265 | ||
266 | if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt)) | |
267 | { | |
268 | if (dump_enabled_p ()) | |
269 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
270 | "use not simple.\n"); | |
271 | return false; | |
272 | } | |
273 | ||
274 | if (dt != vect_external_def && dt != vect_constant_def) | |
275 | return false; | |
276 | } | |
277 | return true; | |
278 | } | |
279 | ||
ebfd146a IR |
280 | /* Function vect_stmt_relevant_p. |
281 | ||
282 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
283 | "relevant for vectorization". | |
284 | ||
285 | A stmt is considered "relevant for vectorization" if: | |
286 | - it has uses outside the loop. | |
287 | - it has vdefs (it alters memory). | |
288 | - control stmts in the loop (except for the exit condition). | |
289 | ||
290 | CHECKME: what other side effects would the vectorizer allow? */ | |
291 | ||
292 | static bool | |
355fe088 | 293 | vect_stmt_relevant_p (gimple *stmt, loop_vec_info loop_vinfo, |
ebfd146a IR |
294 | enum vect_relevant *relevant, bool *live_p) |
295 | { | |
296 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
297 | ssa_op_iter op_iter; | |
298 | imm_use_iterator imm_iter; | |
299 | use_operand_p use_p; | |
300 | def_operand_p def_p; | |
301 | ||
8644a673 | 302 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
303 | *live_p = false; |
304 | ||
305 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
306 | if (is_ctrl_stmt (stmt) |
307 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
308 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 309 | *relevant = vect_used_in_scope; |
ebfd146a IR |
310 | |
311 | /* changing memory. */ | |
312 | if (gimple_code (stmt) != GIMPLE_PHI) | |
ac6aeab4 RB |
313 | if (gimple_vdef (stmt) |
314 | && !gimple_clobber_p (stmt)) | |
ebfd146a | 315 | { |
73fbfcad | 316 | if (dump_enabled_p ()) |
78c60e3d | 317 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 318 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 319 | *relevant = vect_used_in_scope; |
ebfd146a IR |
320 | } |
321 | ||
322 | /* uses outside the loop. */ | |
323 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
324 | { | |
325 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
326 | { | |
327 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
328 | if (!flow_bb_inside_loop_p (loop, bb)) | |
329 | { | |
73fbfcad | 330 | if (dump_enabled_p ()) |
78c60e3d | 331 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 332 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 333 | |
3157b0c2 AO |
334 | if (is_gimple_debug (USE_STMT (use_p))) |
335 | continue; | |
336 | ||
ebfd146a IR |
337 | /* We expect all such uses to be in the loop exit phis |
338 | (because of loop closed form) */ | |
339 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
340 | gcc_assert (bb == single_exit (loop)->dest); | |
341 | ||
342 | *live_p = true; | |
343 | } | |
344 | } | |
345 | } | |
346 | ||
3a2edf4c AH |
347 | if (*live_p && *relevant == vect_unused_in_scope |
348 | && !is_simple_and_all_uses_invariant (stmt, loop_vinfo)) | |
b28ead45 AH |
349 | { |
350 | if (dump_enabled_p ()) | |
351 | dump_printf_loc (MSG_NOTE, vect_location, | |
352 | "vec_stmt_relevant_p: stmt live but not relevant.\n"); | |
353 | *relevant = vect_used_only_live; | |
354 | } | |
355 | ||
ebfd146a IR |
356 | return (*live_p || *relevant); |
357 | } | |
358 | ||
359 | ||
b8698a0f | 360 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 361 | |
ff802fa1 | 362 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
363 | used in STMT for anything other than indexing an array. */ |
364 | ||
365 | static bool | |
355fe088 | 366 | exist_non_indexing_operands_for_use_p (tree use, gimple *stmt) |
ebfd146a IR |
367 | { |
368 | tree operand; | |
369 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 370 | |
ff802fa1 | 371 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
372 | reference in STMT, then any operand that corresponds to USE |
373 | is not indexing an array. */ | |
374 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
375 | return true; | |
59a05b0c | 376 | |
ebfd146a IR |
377 | /* STMT has a data_ref. FORNOW this means that its of one of |
378 | the following forms: | |
379 | -1- ARRAY_REF = var | |
380 | -2- var = ARRAY_REF | |
381 | (This should have been verified in analyze_data_refs). | |
382 | ||
383 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 384 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
385 | for array indexing. |
386 | ||
387 | Therefore, all we need to check is if STMT falls into the | |
388 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
389 | |
390 | if (!gimple_assign_copy_p (stmt)) | |
5ce9450f JJ |
391 | { |
392 | if (is_gimple_call (stmt) | |
393 | && gimple_call_internal_p (stmt)) | |
394 | switch (gimple_call_internal_fn (stmt)) | |
395 | { | |
396 | case IFN_MASK_STORE: | |
397 | operand = gimple_call_arg (stmt, 3); | |
398 | if (operand == use) | |
399 | return true; | |
400 | /* FALLTHRU */ | |
401 | case IFN_MASK_LOAD: | |
402 | operand = gimple_call_arg (stmt, 2); | |
403 | if (operand == use) | |
404 | return true; | |
405 | break; | |
406 | default: | |
407 | break; | |
408 | } | |
409 | return false; | |
410 | } | |
411 | ||
59a05b0c EB |
412 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
413 | return false; | |
ebfd146a | 414 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
415 | if (TREE_CODE (operand) != SSA_NAME) |
416 | return false; | |
417 | ||
418 | if (operand == use) | |
419 | return true; | |
420 | ||
421 | return false; | |
422 | } | |
423 | ||
424 | ||
b8698a0f | 425 | /* |
ebfd146a IR |
426 | Function process_use. |
427 | ||
428 | Inputs: | |
429 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b28ead45 | 430 | - RELEVANT - enum value to be set in the STMT_VINFO of the stmt |
ff802fa1 | 431 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 432 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
433 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
434 | be performed. | |
ebfd146a IR |
435 | |
436 | Outputs: | |
437 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
438 | relevance info of the DEF_STMT of this USE: | |
439 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
440 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
441 | Exceptions: | |
442 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 443 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 444 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
445 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
446 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
447 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
448 | be modified accordingly. | |
449 | ||
450 | Return true if everything is as expected. Return false otherwise. */ | |
451 | ||
452 | static bool | |
b28ead45 | 453 | process_use (gimple *stmt, tree use, loop_vec_info loop_vinfo, |
355fe088 | 454 | enum vect_relevant relevant, vec<gimple *> *worklist, |
aec7ae7d | 455 | bool force) |
ebfd146a IR |
456 | { |
457 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
458 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
459 | stmt_vec_info dstmt_vinfo; | |
460 | basic_block bb, def_bb; | |
355fe088 | 461 | gimple *def_stmt; |
ebfd146a IR |
462 | enum vect_def_type dt; |
463 | ||
b8698a0f | 464 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 465 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 466 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
467 | return true; |
468 | ||
81c40241 | 469 | if (!vect_is_simple_use (use, loop_vinfo, &def_stmt, &dt)) |
b8698a0f | 470 | { |
73fbfcad | 471 | if (dump_enabled_p ()) |
78c60e3d | 472 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 473 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
474 | return false; |
475 | } | |
476 | ||
477 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
478 | return true; | |
479 | ||
480 | def_bb = gimple_bb (def_stmt); | |
481 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
482 | { | |
73fbfcad | 483 | if (dump_enabled_p ()) |
e645e942 | 484 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n"); |
ebfd146a IR |
485 | return true; |
486 | } | |
487 | ||
b8698a0f L |
488 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
489 | DEF_STMT must have already been processed, because this should be the | |
490 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
491 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
492 | check that everything is as expected, and we are done. */ |
493 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
494 | bb = gimple_bb (stmt); | |
495 | if (gimple_code (stmt) == GIMPLE_PHI | |
496 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
497 | && gimple_code (def_stmt) != GIMPLE_PHI | |
498 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
499 | && bb->loop_father == def_bb->loop_father) | |
500 | { | |
73fbfcad | 501 | if (dump_enabled_p ()) |
78c60e3d | 502 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 503 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a IR |
504 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
505 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
506 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 507 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 508 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
509 | return true; |
510 | } | |
511 | ||
512 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
513 | outer-loop-header-bb: | |
514 | d = def_stmt | |
515 | inner-loop: | |
516 | stmt # use (d) | |
517 | outer-loop-tail-bb: | |
518 | ... */ | |
519 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
520 | { | |
73fbfcad | 521 | if (dump_enabled_p ()) |
78c60e3d | 522 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 523 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 524 | |
ebfd146a IR |
525 | switch (relevant) |
526 | { | |
8644a673 | 527 | case vect_unused_in_scope: |
7c5222ff IR |
528 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
529 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 530 | break; |
7c5222ff | 531 | |
ebfd146a | 532 | case vect_used_in_outer_by_reduction: |
7c5222ff | 533 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
534 | relevant = vect_used_by_reduction; |
535 | break; | |
7c5222ff | 536 | |
ebfd146a | 537 | case vect_used_in_outer: |
7c5222ff | 538 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 539 | relevant = vect_used_in_scope; |
ebfd146a | 540 | break; |
7c5222ff | 541 | |
8644a673 | 542 | case vect_used_in_scope: |
ebfd146a IR |
543 | break; |
544 | ||
545 | default: | |
546 | gcc_unreachable (); | |
b8698a0f | 547 | } |
ebfd146a IR |
548 | } |
549 | ||
550 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
551 | outer-loop-header-bb: | |
552 | ... | |
553 | inner-loop: | |
554 | d = def_stmt | |
06066f92 | 555 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
556 | stmt # use (d) */ |
557 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
558 | { | |
73fbfcad | 559 | if (dump_enabled_p ()) |
78c60e3d | 560 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 561 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 562 | |
ebfd146a IR |
563 | switch (relevant) |
564 | { | |
8644a673 | 565 | case vect_unused_in_scope: |
b8698a0f | 566 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 567 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 568 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
569 | break; |
570 | ||
ebfd146a | 571 | case vect_used_by_reduction: |
b28ead45 | 572 | case vect_used_only_live: |
ebfd146a IR |
573 | relevant = vect_used_in_outer_by_reduction; |
574 | break; | |
575 | ||
8644a673 | 576 | case vect_used_in_scope: |
ebfd146a IR |
577 | relevant = vect_used_in_outer; |
578 | break; | |
579 | ||
580 | default: | |
581 | gcc_unreachable (); | |
582 | } | |
583 | } | |
643a9684 RB |
584 | /* We are also not interested in uses on loop PHI backedges that are |
585 | inductions. Otherwise we'll needlessly vectorize the IV increment | |
e294f495 RB |
586 | and cause hybrid SLP for SLP inductions. Unless the PHI is live |
587 | of course. */ | |
643a9684 RB |
588 | else if (gimple_code (stmt) == GIMPLE_PHI |
589 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_induction_def | |
e294f495 | 590 | && ! STMT_VINFO_LIVE_P (stmt_vinfo) |
643a9684 RB |
591 | && (PHI_ARG_DEF_FROM_EDGE (stmt, loop_latch_edge (bb->loop_father)) |
592 | == use)) | |
593 | { | |
594 | if (dump_enabled_p ()) | |
595 | dump_printf_loc (MSG_NOTE, vect_location, | |
596 | "induction value on backedge.\n"); | |
597 | return true; | |
598 | } | |
599 | ||
ebfd146a | 600 | |
b28ead45 | 601 | vect_mark_relevant (worklist, def_stmt, relevant, false); |
ebfd146a IR |
602 | return true; |
603 | } | |
604 | ||
605 | ||
606 | /* Function vect_mark_stmts_to_be_vectorized. | |
607 | ||
608 | Not all stmts in the loop need to be vectorized. For example: | |
609 | ||
610 | for i... | |
611 | for j... | |
612 | 1. T0 = i + j | |
613 | 2. T1 = a[T0] | |
614 | ||
615 | 3. j = j + 1 | |
616 | ||
617 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
618 | addressing of vectorized data-refs are handled differently. | |
619 | ||
620 | This pass detects such stmts. */ | |
621 | ||
622 | bool | |
623 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
624 | { | |
ebfd146a IR |
625 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
626 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
627 | unsigned int nbbs = loop->num_nodes; | |
628 | gimple_stmt_iterator si; | |
355fe088 | 629 | gimple *stmt; |
ebfd146a IR |
630 | unsigned int i; |
631 | stmt_vec_info stmt_vinfo; | |
632 | basic_block bb; | |
355fe088 | 633 | gimple *phi; |
ebfd146a | 634 | bool live_p; |
b28ead45 | 635 | enum vect_relevant relevant; |
ebfd146a | 636 | |
73fbfcad | 637 | if (dump_enabled_p ()) |
78c60e3d | 638 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 639 | "=== vect_mark_stmts_to_be_vectorized ===\n"); |
ebfd146a | 640 | |
355fe088 | 641 | auto_vec<gimple *, 64> worklist; |
ebfd146a IR |
642 | |
643 | /* 1. Init worklist. */ | |
644 | for (i = 0; i < nbbs; i++) | |
645 | { | |
646 | bb = bbs[i]; | |
647 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 648 | { |
ebfd146a | 649 | phi = gsi_stmt (si); |
73fbfcad | 650 | if (dump_enabled_p ()) |
ebfd146a | 651 | { |
78c60e3d SS |
652 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
653 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
654 | } |
655 | ||
656 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 657 | vect_mark_relevant (&worklist, phi, relevant, live_p); |
ebfd146a IR |
658 | } |
659 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
660 | { | |
661 | stmt = gsi_stmt (si); | |
73fbfcad | 662 | if (dump_enabled_p ()) |
ebfd146a | 663 | { |
78c60e3d SS |
664 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
665 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
b8698a0f | 666 | } |
ebfd146a IR |
667 | |
668 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 669 | vect_mark_relevant (&worklist, stmt, relevant, live_p); |
ebfd146a IR |
670 | } |
671 | } | |
672 | ||
673 | /* 2. Process_worklist */ | |
9771b263 | 674 | while (worklist.length () > 0) |
ebfd146a IR |
675 | { |
676 | use_operand_p use_p; | |
677 | ssa_op_iter iter; | |
678 | ||
9771b263 | 679 | stmt = worklist.pop (); |
73fbfcad | 680 | if (dump_enabled_p ()) |
ebfd146a | 681 | { |
78c60e3d SS |
682 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
683 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
ebfd146a IR |
684 | } |
685 | ||
b8698a0f | 686 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
b28ead45 AH |
687 | (DEF_STMT) as relevant/irrelevant according to the relevance property |
688 | of STMT. */ | |
ebfd146a IR |
689 | stmt_vinfo = vinfo_for_stmt (stmt); |
690 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
ebfd146a | 691 | |
b28ead45 AH |
692 | /* Generally, the relevance property of STMT (in STMT_VINFO_RELEVANT) is |
693 | propagated as is to the DEF_STMTs of its USEs. | |
ebfd146a IR |
694 | |
695 | One exception is when STMT has been identified as defining a reduction | |
b28ead45 | 696 | variable; in this case we set the relevance to vect_used_by_reduction. |
ebfd146a | 697 | This is because we distinguish between two kinds of relevant stmts - |
b8698a0f | 698 | those that are used by a reduction computation, and those that are |
ff802fa1 | 699 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 700 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 701 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 702 | |
b28ead45 | 703 | switch (STMT_VINFO_DEF_TYPE (stmt_vinfo)) |
ebfd146a | 704 | { |
06066f92 | 705 | case vect_reduction_def: |
b28ead45 AH |
706 | gcc_assert (relevant != vect_unused_in_scope); |
707 | if (relevant != vect_unused_in_scope | |
708 | && relevant != vect_used_in_scope | |
709 | && relevant != vect_used_by_reduction | |
710 | && relevant != vect_used_only_live) | |
06066f92 | 711 | { |
b28ead45 AH |
712 | if (dump_enabled_p ()) |
713 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
714 | "unsupported use of reduction.\n"); | |
715 | return false; | |
06066f92 | 716 | } |
06066f92 | 717 | break; |
b8698a0f | 718 | |
06066f92 | 719 | case vect_nested_cycle: |
b28ead45 AH |
720 | if (relevant != vect_unused_in_scope |
721 | && relevant != vect_used_in_outer_by_reduction | |
722 | && relevant != vect_used_in_outer) | |
06066f92 | 723 | { |
73fbfcad | 724 | if (dump_enabled_p ()) |
78c60e3d | 725 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 726 | "unsupported use of nested cycle.\n"); |
7c5222ff | 727 | |
06066f92 IR |
728 | return false; |
729 | } | |
b8698a0f L |
730 | break; |
731 | ||
06066f92 | 732 | case vect_double_reduction_def: |
b28ead45 AH |
733 | if (relevant != vect_unused_in_scope |
734 | && relevant != vect_used_by_reduction | |
735 | && relevant != vect_used_only_live) | |
06066f92 | 736 | { |
73fbfcad | 737 | if (dump_enabled_p ()) |
78c60e3d | 738 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 739 | "unsupported use of double reduction.\n"); |
7c5222ff | 740 | |
7c5222ff | 741 | return false; |
06066f92 | 742 | } |
b8698a0f | 743 | break; |
7c5222ff | 744 | |
06066f92 IR |
745 | default: |
746 | break; | |
7c5222ff | 747 | } |
b8698a0f | 748 | |
aec7ae7d | 749 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
750 | { |
751 | /* Pattern statements are not inserted into the code, so | |
752 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
753 | have to scan the RHS or function arguments instead. */ | |
754 | if (is_gimple_assign (stmt)) | |
755 | { | |
69d2aade JJ |
756 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
757 | tree op = gimple_assign_rhs1 (stmt); | |
758 | ||
759 | i = 1; | |
760 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
761 | { | |
762 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
b28ead45 | 763 | relevant, &worklist, false) |
69d2aade | 764 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
b28ead45 | 765 | relevant, &worklist, false)) |
566d377a | 766 | return false; |
69d2aade JJ |
767 | i = 2; |
768 | } | |
769 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 770 | { |
69d2aade | 771 | op = gimple_op (stmt, i); |
afbe6325 | 772 | if (TREE_CODE (op) == SSA_NAME |
b28ead45 | 773 | && !process_use (stmt, op, loop_vinfo, relevant, |
afbe6325 | 774 | &worklist, false)) |
07687835 | 775 | return false; |
9d5e7640 IR |
776 | } |
777 | } | |
778 | else if (is_gimple_call (stmt)) | |
779 | { | |
780 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
781 | { | |
782 | tree arg = gimple_call_arg (stmt, i); | |
b28ead45 | 783 | if (!process_use (stmt, arg, loop_vinfo, relevant, |
aec7ae7d | 784 | &worklist, false)) |
07687835 | 785 | return false; |
9d5e7640 IR |
786 | } |
787 | } | |
788 | } | |
789 | else | |
790 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
791 | { | |
792 | tree op = USE_FROM_PTR (use_p); | |
b28ead45 | 793 | if (!process_use (stmt, op, loop_vinfo, relevant, |
aec7ae7d | 794 | &worklist, false)) |
07687835 | 795 | return false; |
9d5e7640 | 796 | } |
aec7ae7d | 797 | |
3bab6342 | 798 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo)) |
aec7ae7d | 799 | { |
134c85ca RS |
800 | gather_scatter_info gs_info; |
801 | if (!vect_check_gather_scatter (stmt, loop_vinfo, &gs_info)) | |
802 | gcc_unreachable (); | |
803 | if (!process_use (stmt, gs_info.offset, loop_vinfo, relevant, | |
804 | &worklist, true)) | |
566d377a | 805 | return false; |
aec7ae7d | 806 | } |
ebfd146a IR |
807 | } /* while worklist */ |
808 | ||
ebfd146a IR |
809 | return true; |
810 | } | |
811 | ||
812 | ||
b8698a0f | 813 | /* Function vect_model_simple_cost. |
ebfd146a | 814 | |
b8698a0f | 815 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
816 | single op. Right now, this does not account for multiple insns that could |
817 | be generated for the single vector op. We will handle that shortly. */ | |
818 | ||
819 | void | |
b8698a0f | 820 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 | 821 | enum vect_def_type *dt, |
4fc5ebf1 | 822 | int ndts, |
92345349 BS |
823 | stmt_vector_for_cost *prologue_cost_vec, |
824 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
825 | { |
826 | int i; | |
92345349 | 827 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
828 | |
829 | /* The SLP costs were already calculated during SLP tree build. */ | |
830 | if (PURE_SLP_STMT (stmt_info)) | |
831 | return; | |
832 | ||
4fc5ebf1 JG |
833 | /* Cost the "broadcast" of a scalar operand in to a vector operand. |
834 | Use scalar_to_vec to cost the broadcast, as elsewhere in the vector | |
835 | cost model. */ | |
836 | for (i = 0; i < ndts; i++) | |
92345349 | 837 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
4fc5ebf1 | 838 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
92345349 | 839 | stmt_info, 0, vect_prologue); |
c3e7ee41 BS |
840 | |
841 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
92345349 BS |
842 | inside_cost = record_stmt_cost (body_cost_vec, ncopies, vector_stmt, |
843 | stmt_info, 0, vect_body); | |
c3e7ee41 | 844 | |
73fbfcad | 845 | if (dump_enabled_p ()) |
78c60e3d SS |
846 | dump_printf_loc (MSG_NOTE, vect_location, |
847 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 848 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
849 | } |
850 | ||
851 | ||
8bd37302 BS |
852 | /* Model cost for type demotion and promotion operations. PWR is normally |
853 | zero for single-step promotions and demotions. It will be one if | |
854 | two-step promotion/demotion is required, and so on. Each additional | |
855 | step doubles the number of instructions required. */ | |
856 | ||
857 | static void | |
858 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
859 | enum vect_def_type *dt, int pwr) | |
860 | { | |
861 | int i, tmp; | |
92345349 | 862 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 BS |
863 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
864 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
865 | void *target_cost_data; | |
8bd37302 BS |
866 | |
867 | /* The SLP costs were already calculated during SLP tree build. */ | |
868 | if (PURE_SLP_STMT (stmt_info)) | |
869 | return; | |
870 | ||
c3e7ee41 BS |
871 | if (loop_vinfo) |
872 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
873 | else | |
874 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
875 | ||
8bd37302 BS |
876 | for (i = 0; i < pwr + 1; i++) |
877 | { | |
878 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
879 | (i + 1) : i; | |
c3e7ee41 | 880 | inside_cost += add_stmt_cost (target_cost_data, vect_pow2 (tmp), |
92345349 BS |
881 | vec_promote_demote, stmt_info, 0, |
882 | vect_body); | |
8bd37302 BS |
883 | } |
884 | ||
885 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
886 | for (i = 0; i < 2; i++) | |
92345349 BS |
887 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
888 | prologue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt, | |
889 | stmt_info, 0, vect_prologue); | |
8bd37302 | 890 | |
73fbfcad | 891 | if (dump_enabled_p ()) |
78c60e3d SS |
892 | dump_printf_loc (MSG_NOTE, vect_location, |
893 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 894 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
895 | } |
896 | ||
ebfd146a IR |
897 | /* Function vect_model_store_cost |
898 | ||
0d0293ac MM |
899 | Models cost for stores. In the case of grouped accesses, one access |
900 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a IR |
901 | |
902 | void | |
b8698a0f | 903 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
2de001ee | 904 | vect_memory_access_type memory_access_type, |
9ce4345a | 905 | vec_load_store_type vls_type, slp_tree slp_node, |
92345349 BS |
906 | stmt_vector_for_cost *prologue_cost_vec, |
907 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a | 908 | { |
92345349 | 909 | unsigned int inside_cost = 0, prologue_cost = 0; |
892a981f RS |
910 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
911 | gimple *first_stmt = STMT_VINFO_STMT (stmt_info); | |
912 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); | |
ebfd146a | 913 | |
9ce4345a | 914 | if (vls_type == VLS_STORE_INVARIANT) |
92345349 BS |
915 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
916 | stmt_info, 0, vect_prologue); | |
ebfd146a | 917 | |
892a981f RS |
918 | /* Grouped stores update all elements in the group at once, |
919 | so we want the DR for the first statement. */ | |
920 | if (!slp_node && grouped_access_p) | |
720f5239 | 921 | { |
892a981f RS |
922 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
923 | dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
720f5239 | 924 | } |
ebfd146a | 925 | |
892a981f RS |
926 | /* True if we should include any once-per-group costs as well as |
927 | the cost of the statement itself. For SLP we only get called | |
928 | once per group anyhow. */ | |
929 | bool first_stmt_p = (first_stmt == STMT_VINFO_STMT (stmt_info)); | |
930 | ||
272c6793 | 931 | /* We assume that the cost of a single store-lanes instruction is |
0d0293ac | 932 | equivalent to the cost of GROUP_SIZE separate stores. If a grouped |
272c6793 | 933 | access is instead being provided by a permute-and-store operation, |
2de001ee RS |
934 | include the cost of the permutes. */ |
935 | if (first_stmt_p | |
936 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 937 | { |
e1377713 ES |
938 | /* Uses a high and low interleave or shuffle operations for each |
939 | needed permute. */ | |
892a981f | 940 | int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
e1377713 | 941 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
92345349 BS |
942 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
943 | stmt_info, 0, vect_body); | |
ebfd146a | 944 | |
73fbfcad | 945 | if (dump_enabled_p ()) |
78c60e3d | 946 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 947 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 948 | group_size); |
ebfd146a IR |
949 | } |
950 | ||
cee62fee | 951 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a | 952 | /* Costs of the stores. */ |
067bc855 RB |
953 | if (memory_access_type == VMAT_ELEMENTWISE |
954 | || memory_access_type == VMAT_GATHER_SCATTER) | |
c5126ce8 RS |
955 | { |
956 | /* N scalar stores plus extracting the elements. */ | |
957 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); | |
958 | inside_cost += record_stmt_cost (body_cost_vec, | |
959 | ncopies * assumed_nunits, | |
960 | scalar_store, stmt_info, 0, vect_body); | |
961 | } | |
f2e2a985 | 962 | else |
892a981f | 963 | vect_get_store_cost (dr, ncopies, &inside_cost, body_cost_vec); |
ebfd146a | 964 | |
2de001ee RS |
965 | if (memory_access_type == VMAT_ELEMENTWISE |
966 | || memory_access_type == VMAT_STRIDED_SLP) | |
c5126ce8 RS |
967 | { |
968 | /* N scalar stores plus extracting the elements. */ | |
969 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); | |
970 | inside_cost += record_stmt_cost (body_cost_vec, | |
971 | ncopies * assumed_nunits, | |
972 | vec_to_scalar, stmt_info, 0, vect_body); | |
973 | } | |
cee62fee | 974 | |
73fbfcad | 975 | if (dump_enabled_p ()) |
78c60e3d SS |
976 | dump_printf_loc (MSG_NOTE, vect_location, |
977 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 978 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
979 | } |
980 | ||
981 | ||
720f5239 IR |
982 | /* Calculate cost of DR's memory access. */ |
983 | void | |
984 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 985 | unsigned int *inside_cost, |
92345349 | 986 | stmt_vector_for_cost *body_cost_vec) |
720f5239 IR |
987 | { |
988 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 989 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 990 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
991 | |
992 | switch (alignment_support_scheme) | |
993 | { | |
994 | case dr_aligned: | |
995 | { | |
92345349 BS |
996 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
997 | vector_store, stmt_info, 0, | |
998 | vect_body); | |
720f5239 | 999 | |
73fbfcad | 1000 | if (dump_enabled_p ()) |
78c60e3d | 1001 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1002 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
1003 | break; |
1004 | } | |
1005 | ||
1006 | case dr_unaligned_supported: | |
1007 | { | |
720f5239 | 1008 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1009 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1010 | unaligned_store, stmt_info, |
92345349 | 1011 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1012 | if (dump_enabled_p ()) |
78c60e3d SS |
1013 | dump_printf_loc (MSG_NOTE, vect_location, |
1014 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1015 | "hardware.\n"); |
720f5239 IR |
1016 | break; |
1017 | } | |
1018 | ||
38eec4c6 UW |
1019 | case dr_unaligned_unsupported: |
1020 | { | |
1021 | *inside_cost = VECT_MAX_COST; | |
1022 | ||
73fbfcad | 1023 | if (dump_enabled_p ()) |
78c60e3d | 1024 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1025 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1026 | break; |
1027 | } | |
1028 | ||
720f5239 IR |
1029 | default: |
1030 | gcc_unreachable (); | |
1031 | } | |
1032 | } | |
1033 | ||
1034 | ||
ebfd146a IR |
1035 | /* Function vect_model_load_cost |
1036 | ||
892a981f RS |
1037 | Models cost for loads. In the case of grouped accesses, one access has |
1038 | the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1039 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1040 | access scheme chosen. */ |
1041 | ||
1042 | void | |
92345349 | 1043 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, |
2de001ee RS |
1044 | vect_memory_access_type memory_access_type, |
1045 | slp_tree slp_node, | |
92345349 BS |
1046 | stmt_vector_for_cost *prologue_cost_vec, |
1047 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a | 1048 | { |
892a981f RS |
1049 | gimple *first_stmt = STMT_VINFO_STMT (stmt_info); |
1050 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
92345349 | 1051 | unsigned int inside_cost = 0, prologue_cost = 0; |
892a981f | 1052 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); |
ebfd146a | 1053 | |
892a981f RS |
1054 | /* Grouped loads read all elements in the group at once, |
1055 | so we want the DR for the first statement. */ | |
1056 | if (!slp_node && grouped_access_p) | |
ebfd146a | 1057 | { |
892a981f RS |
1058 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
1059 | dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
ebfd146a IR |
1060 | } |
1061 | ||
892a981f RS |
1062 | /* True if we should include any once-per-group costs as well as |
1063 | the cost of the statement itself. For SLP we only get called | |
1064 | once per group anyhow. */ | |
1065 | bool first_stmt_p = (first_stmt == STMT_VINFO_STMT (stmt_info)); | |
1066 | ||
272c6793 | 1067 | /* We assume that the cost of a single load-lanes instruction is |
0d0293ac | 1068 | equivalent to the cost of GROUP_SIZE separate loads. If a grouped |
272c6793 | 1069 | access is instead being provided by a load-and-permute operation, |
2de001ee RS |
1070 | include the cost of the permutes. */ |
1071 | if (first_stmt_p | |
1072 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 1073 | { |
2c23db6d ES |
1074 | /* Uses an even and odd extract operations or shuffle operations |
1075 | for each needed permute. */ | |
892a981f | 1076 | int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
2c23db6d ES |
1077 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
1078 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, | |
1079 | stmt_info, 0, vect_body); | |
ebfd146a | 1080 | |
73fbfcad | 1081 | if (dump_enabled_p ()) |
e645e942 TJ |
1082 | dump_printf_loc (MSG_NOTE, vect_location, |
1083 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1084 | group_size); |
ebfd146a IR |
1085 | } |
1086 | ||
1087 | /* The loads themselves. */ | |
067bc855 RB |
1088 | if (memory_access_type == VMAT_ELEMENTWISE |
1089 | || memory_access_type == VMAT_GATHER_SCATTER) | |
a82960aa | 1090 | { |
a21892ad BS |
1091 | /* N scalar loads plus gathering them into a vector. */ |
1092 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
c5126ce8 | 1093 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); |
92345349 | 1094 | inside_cost += record_stmt_cost (body_cost_vec, |
c5126ce8 | 1095 | ncopies * assumed_nunits, |
92345349 | 1096 | scalar_load, stmt_info, 0, vect_body); |
a82960aa RG |
1097 | } |
1098 | else | |
892a981f | 1099 | vect_get_load_cost (dr, ncopies, first_stmt_p, |
92345349 BS |
1100 | &inside_cost, &prologue_cost, |
1101 | prologue_cost_vec, body_cost_vec, true); | |
2de001ee RS |
1102 | if (memory_access_type == VMAT_ELEMENTWISE |
1103 | || memory_access_type == VMAT_STRIDED_SLP) | |
892a981f RS |
1104 | inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct, |
1105 | stmt_info, 0, vect_body); | |
720f5239 | 1106 | |
73fbfcad | 1107 | if (dump_enabled_p ()) |
78c60e3d SS |
1108 | dump_printf_loc (MSG_NOTE, vect_location, |
1109 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1110 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1111 | } |
1112 | ||
1113 | ||
1114 | /* Calculate cost of DR's memory access. */ | |
1115 | void | |
1116 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1117 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1118 | unsigned int *prologue_cost, |
1119 | stmt_vector_for_cost *prologue_cost_vec, | |
1120 | stmt_vector_for_cost *body_cost_vec, | |
1121 | bool record_prologue_costs) | |
720f5239 IR |
1122 | { |
1123 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 1124 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 1125 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
1126 | |
1127 | switch (alignment_support_scheme) | |
ebfd146a IR |
1128 | { |
1129 | case dr_aligned: | |
1130 | { | |
92345349 BS |
1131 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1132 | stmt_info, 0, vect_body); | |
ebfd146a | 1133 | |
73fbfcad | 1134 | if (dump_enabled_p ()) |
78c60e3d | 1135 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1136 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1137 | |
1138 | break; | |
1139 | } | |
1140 | case dr_unaligned_supported: | |
1141 | { | |
720f5239 | 1142 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1143 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1144 | unaligned_load, stmt_info, |
92345349 | 1145 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1146 | |
73fbfcad | 1147 | if (dump_enabled_p ()) |
78c60e3d SS |
1148 | dump_printf_loc (MSG_NOTE, vect_location, |
1149 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1150 | "hardware.\n"); |
ebfd146a IR |
1151 | |
1152 | break; | |
1153 | } | |
1154 | case dr_explicit_realign: | |
1155 | { | |
92345349 BS |
1156 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1157 | vector_load, stmt_info, 0, vect_body); | |
1158 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1159 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1160 | |
1161 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1162 | the containing loop, the following cost should be added to the | |
92345349 | 1163 | prologue costs. */ |
ebfd146a | 1164 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1165 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1166 | stmt_info, 0, vect_body); | |
ebfd146a | 1167 | |
73fbfcad | 1168 | if (dump_enabled_p ()) |
e645e942 TJ |
1169 | dump_printf_loc (MSG_NOTE, vect_location, |
1170 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1171 | |
ebfd146a IR |
1172 | break; |
1173 | } | |
1174 | case dr_explicit_realign_optimized: | |
1175 | { | |
73fbfcad | 1176 | if (dump_enabled_p ()) |
e645e942 | 1177 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1178 | "vect_model_load_cost: unaligned software " |
e645e942 | 1179 | "pipelined.\n"); |
ebfd146a IR |
1180 | |
1181 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1182 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1183 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1184 | access, then the above cost should only be considered for one |
ff802fa1 | 1185 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1186 | and a realignment op. */ |
1187 | ||
92345349 | 1188 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1189 | { |
92345349 BS |
1190 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1191 | vector_stmt, stmt_info, | |
1192 | 0, vect_prologue); | |
ebfd146a | 1193 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1194 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1195 | vector_stmt, stmt_info, | |
1196 | 0, vect_prologue); | |
ebfd146a IR |
1197 | } |
1198 | ||
92345349 BS |
1199 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1200 | stmt_info, 0, vect_body); | |
1201 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1202 | stmt_info, 0, vect_body); | |
8bd37302 | 1203 | |
73fbfcad | 1204 | if (dump_enabled_p ()) |
78c60e3d | 1205 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1206 | "vect_model_load_cost: explicit realign optimized" |
1207 | "\n"); | |
8bd37302 | 1208 | |
ebfd146a IR |
1209 | break; |
1210 | } | |
1211 | ||
38eec4c6 UW |
1212 | case dr_unaligned_unsupported: |
1213 | { | |
1214 | *inside_cost = VECT_MAX_COST; | |
1215 | ||
73fbfcad | 1216 | if (dump_enabled_p ()) |
78c60e3d | 1217 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1218 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1219 | break; |
1220 | } | |
1221 | ||
ebfd146a IR |
1222 | default: |
1223 | gcc_unreachable (); | |
1224 | } | |
ebfd146a IR |
1225 | } |
1226 | ||
418b7df3 RG |
1227 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1228 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1229 | |
418b7df3 | 1230 | static void |
355fe088 | 1231 | vect_init_vector_1 (gimple *stmt, gimple *new_stmt, gimple_stmt_iterator *gsi) |
ebfd146a | 1232 | { |
ebfd146a | 1233 | if (gsi) |
418b7df3 | 1234 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1235 | else |
1236 | { | |
418b7df3 | 1237 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1238 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1239 | |
a70d6342 IR |
1240 | if (loop_vinfo) |
1241 | { | |
1242 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1243 | basic_block new_bb; |
1244 | edge pe; | |
a70d6342 IR |
1245 | |
1246 | if (nested_in_vect_loop_p (loop, stmt)) | |
1247 | loop = loop->inner; | |
b8698a0f | 1248 | |
a70d6342 | 1249 | pe = loop_preheader_edge (loop); |
418b7df3 | 1250 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1251 | gcc_assert (!new_bb); |
1252 | } | |
1253 | else | |
1254 | { | |
1255 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1256 | basic_block bb; | |
1257 | gimple_stmt_iterator gsi_bb_start; | |
1258 | ||
1259 | gcc_assert (bb_vinfo); | |
1260 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1261 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1262 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1263 | } |
ebfd146a IR |
1264 | } |
1265 | ||
73fbfcad | 1266 | if (dump_enabled_p ()) |
ebfd146a | 1267 | { |
78c60e3d SS |
1268 | dump_printf_loc (MSG_NOTE, vect_location, |
1269 | "created new init_stmt: "); | |
1270 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
ebfd146a | 1271 | } |
418b7df3 RG |
1272 | } |
1273 | ||
1274 | /* Function vect_init_vector. | |
ebfd146a | 1275 | |
5467ee52 RG |
1276 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1277 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1278 | vector type a vector with all elements equal to VAL is created first. | |
1279 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1280 | initialization at the loop preheader. | |
418b7df3 RG |
1281 | Return the DEF of INIT_STMT. |
1282 | It will be used in the vectorization of STMT. */ | |
1283 | ||
1284 | tree | |
355fe088 | 1285 | vect_init_vector (gimple *stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 | 1286 | { |
355fe088 | 1287 | gimple *init_stmt; |
418b7df3 RG |
1288 | tree new_temp; |
1289 | ||
e412ece4 RB |
1290 | /* We abuse this function to push sth to a SSA name with initial 'val'. */ |
1291 | if (! useless_type_conversion_p (type, TREE_TYPE (val))) | |
418b7df3 | 1292 | { |
e412ece4 RB |
1293 | gcc_assert (TREE_CODE (type) == VECTOR_TYPE); |
1294 | if (! types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) | |
418b7df3 | 1295 | { |
5a308cf1 IE |
1296 | /* Scalar boolean value should be transformed into |
1297 | all zeros or all ones value before building a vector. */ | |
1298 | if (VECTOR_BOOLEAN_TYPE_P (type)) | |
1299 | { | |
b3d51f23 IE |
1300 | tree true_val = build_all_ones_cst (TREE_TYPE (type)); |
1301 | tree false_val = build_zero_cst (TREE_TYPE (type)); | |
5a308cf1 IE |
1302 | |
1303 | if (CONSTANT_CLASS_P (val)) | |
1304 | val = integer_zerop (val) ? false_val : true_val; | |
1305 | else | |
1306 | { | |
1307 | new_temp = make_ssa_name (TREE_TYPE (type)); | |
1308 | init_stmt = gimple_build_assign (new_temp, COND_EXPR, | |
1309 | val, true_val, false_val); | |
1310 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
1311 | val = new_temp; | |
1312 | } | |
1313 | } | |
1314 | else if (CONSTANT_CLASS_P (val)) | |
42fd8198 | 1315 | val = fold_convert (TREE_TYPE (type), val); |
418b7df3 RG |
1316 | else |
1317 | { | |
b731b390 | 1318 | new_temp = make_ssa_name (TREE_TYPE (type)); |
e412ece4 RB |
1319 | if (! INTEGRAL_TYPE_P (TREE_TYPE (val))) |
1320 | init_stmt = gimple_build_assign (new_temp, | |
1321 | fold_build1 (VIEW_CONVERT_EXPR, | |
1322 | TREE_TYPE (type), | |
1323 | val)); | |
1324 | else | |
1325 | init_stmt = gimple_build_assign (new_temp, NOP_EXPR, val); | |
418b7df3 | 1326 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1327 | val = new_temp; |
418b7df3 RG |
1328 | } |
1329 | } | |
5467ee52 | 1330 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1331 | } |
1332 | ||
0e22bb5a RB |
1333 | new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_"); |
1334 | init_stmt = gimple_build_assign (new_temp, val); | |
418b7df3 | 1335 | vect_init_vector_1 (stmt, init_stmt, gsi); |
0e22bb5a | 1336 | return new_temp; |
ebfd146a IR |
1337 | } |
1338 | ||
c83a894c | 1339 | /* Function vect_get_vec_def_for_operand_1. |
a70d6342 | 1340 | |
c83a894c AH |
1341 | For a defining stmt DEF_STMT of a scalar stmt, return a vector def with type |
1342 | DT that will be used in the vectorized stmt. */ | |
ebfd146a IR |
1343 | |
1344 | tree | |
c83a894c | 1345 | vect_get_vec_def_for_operand_1 (gimple *def_stmt, enum vect_def_type dt) |
ebfd146a IR |
1346 | { |
1347 | tree vec_oprnd; | |
355fe088 | 1348 | gimple *vec_stmt; |
ebfd146a | 1349 | stmt_vec_info def_stmt_info = NULL; |
ebfd146a IR |
1350 | |
1351 | switch (dt) | |
1352 | { | |
81c40241 | 1353 | /* operand is a constant or a loop invariant. */ |
ebfd146a | 1354 | case vect_constant_def: |
81c40241 | 1355 | case vect_external_def: |
c83a894c AH |
1356 | /* Code should use vect_get_vec_def_for_operand. */ |
1357 | gcc_unreachable (); | |
ebfd146a | 1358 | |
81c40241 | 1359 | /* operand is defined inside the loop. */ |
8644a673 | 1360 | case vect_internal_def: |
ebfd146a | 1361 | { |
ebfd146a IR |
1362 | /* Get the def from the vectorized stmt. */ |
1363 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1364 | |
ebfd146a | 1365 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1366 | /* Get vectorized pattern statement. */ |
1367 | if (!vec_stmt | |
1368 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1369 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1370 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1371 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1372 | gcc_assert (vec_stmt); |
1373 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1374 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1375 | else if (is_gimple_call (vec_stmt)) | |
1376 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1377 | else | |
1378 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1379 | return vec_oprnd; | |
1380 | } | |
1381 | ||
c78e3652 | 1382 | /* operand is defined by a loop header phi. */ |
ebfd146a | 1383 | case vect_reduction_def: |
06066f92 | 1384 | case vect_double_reduction_def: |
7c5222ff | 1385 | case vect_nested_cycle: |
ebfd146a IR |
1386 | case vect_induction_def: |
1387 | { | |
1388 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1389 | ||
1390 | /* Get the def from the vectorized stmt. */ | |
1391 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1392 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1393 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1394 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1395 | else | |
1396 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1397 | return vec_oprnd; |
1398 | } | |
1399 | ||
1400 | default: | |
1401 | gcc_unreachable (); | |
1402 | } | |
1403 | } | |
1404 | ||
1405 | ||
c83a894c AH |
1406 | /* Function vect_get_vec_def_for_operand. |
1407 | ||
1408 | OP is an operand in STMT. This function returns a (vector) def that will be | |
1409 | used in the vectorized stmt for STMT. | |
1410 | ||
1411 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1412 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1413 | ||
1414 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1415 | needs to be introduced. VECTYPE may be used to specify a required type for | |
1416 | vector invariant. */ | |
1417 | ||
1418 | tree | |
1419 | vect_get_vec_def_for_operand (tree op, gimple *stmt, tree vectype) | |
1420 | { | |
1421 | gimple *def_stmt; | |
1422 | enum vect_def_type dt; | |
1423 | bool is_simple_use; | |
1424 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
1425 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
1426 | ||
1427 | if (dump_enabled_p ()) | |
1428 | { | |
1429 | dump_printf_loc (MSG_NOTE, vect_location, | |
1430 | "vect_get_vec_def_for_operand: "); | |
1431 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
1432 | dump_printf (MSG_NOTE, "\n"); | |
1433 | } | |
1434 | ||
1435 | is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt); | |
1436 | gcc_assert (is_simple_use); | |
1437 | if (def_stmt && dump_enabled_p ()) | |
1438 | { | |
1439 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1440 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
1441 | } | |
1442 | ||
1443 | if (dt == vect_constant_def || dt == vect_external_def) | |
1444 | { | |
1445 | tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo); | |
1446 | tree vector_type; | |
1447 | ||
1448 | if (vectype) | |
1449 | vector_type = vectype; | |
2568d8a1 | 1450 | else if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op)) |
c83a894c AH |
1451 | && VECTOR_BOOLEAN_TYPE_P (stmt_vectype)) |
1452 | vector_type = build_same_sized_truth_vector_type (stmt_vectype); | |
1453 | else | |
1454 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
1455 | ||
1456 | gcc_assert (vector_type); | |
1457 | return vect_init_vector (stmt, op, vector_type, NULL); | |
1458 | } | |
1459 | else | |
1460 | return vect_get_vec_def_for_operand_1 (def_stmt, dt); | |
1461 | } | |
1462 | ||
1463 | ||
ebfd146a IR |
1464 | /* Function vect_get_vec_def_for_stmt_copy |
1465 | ||
ff802fa1 | 1466 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1467 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1468 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1469 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1470 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1471 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1472 | DT is the type of the vector def VEC_OPRND. |
1473 | ||
1474 | Context: | |
1475 | In case the vectorization factor (VF) is bigger than the number | |
1476 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1477 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1478 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1479 | smallest data-type determines the VF, and as a result, when vectorizing |
1480 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1481 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1482 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1483 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1484 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1485 | ||
1486 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1487 | |
ebfd146a IR |
1488 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1489 | VS1.1: vx.1 = memref1 VS1.2 | |
1490 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1491 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1492 | |
1493 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1494 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1495 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1496 | VSnew.3: vz3 = vx.3 + ... | |
1497 | ||
1498 | The vectorization of S1 is explained in vectorizable_load. | |
1499 | The vectorization of S2: | |
b8698a0f L |
1500 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1501 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1502 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1503 | returns the vector-def 'vx.0'. |
1504 | ||
b8698a0f L |
1505 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1506 | function is called to get the relevant vector-def for each operand. It is | |
1507 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1508 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1509 | ||
b8698a0f L |
1510 | For example, to obtain the vector-def 'vx.1' in order to create the |
1511 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1512 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1513 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1514 | and return its def ('vx.1'). | |
1515 | Overall, to create the above sequence this function will be called 3 times: | |
1516 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1517 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1518 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1519 | ||
1520 | tree | |
1521 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1522 | { | |
355fe088 | 1523 | gimple *vec_stmt_for_operand; |
ebfd146a IR |
1524 | stmt_vec_info def_stmt_info; |
1525 | ||
1526 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1527 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1528 | return vec_oprnd; |
1529 | ||
1530 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1531 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1532 | gcc_assert (def_stmt_info); | |
1533 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1534 | gcc_assert (vec_stmt_for_operand); | |
ebfd146a IR |
1535 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) |
1536 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1537 | else | |
1538 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1539 | return vec_oprnd; | |
1540 | } | |
1541 | ||
1542 | ||
1543 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1544 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a | 1545 | |
c78e3652 | 1546 | void |
b8698a0f | 1547 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1548 | vec<tree> *vec_oprnds0, |
1549 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1550 | { |
9771b263 | 1551 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1552 | |
1553 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1554 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1555 | |
9771b263 | 1556 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1557 | { |
9771b263 | 1558 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1559 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1560 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1561 | } |
1562 | } | |
1563 | ||
1564 | ||
c78e3652 | 1565 | /* Get vectorized definitions for OP0 and OP1. */ |
ebfd146a | 1566 | |
c78e3652 | 1567 | void |
355fe088 | 1568 | vect_get_vec_defs (tree op0, tree op1, gimple *stmt, |
9771b263 DN |
1569 | vec<tree> *vec_oprnds0, |
1570 | vec<tree> *vec_oprnds1, | |
306b0c92 | 1571 | slp_tree slp_node) |
ebfd146a IR |
1572 | { |
1573 | if (slp_node) | |
d092494c IR |
1574 | { |
1575 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1576 | auto_vec<tree> ops (nops); |
1577 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1578 | |
9771b263 | 1579 | ops.quick_push (op0); |
d092494c | 1580 | if (op1) |
9771b263 | 1581 | ops.quick_push (op1); |
d092494c | 1582 | |
306b0c92 | 1583 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
d092494c | 1584 | |
37b5ec8f | 1585 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1586 | if (op1) |
37b5ec8f | 1587 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1588 | } |
ebfd146a IR |
1589 | else |
1590 | { | |
1591 | tree vec_oprnd; | |
1592 | ||
9771b263 | 1593 | vec_oprnds0->create (1); |
81c40241 | 1594 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 1595 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1596 | |
1597 | if (op1) | |
1598 | { | |
9771b263 | 1599 | vec_oprnds1->create (1); |
81c40241 | 1600 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 1601 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1602 | } |
1603 | } | |
1604 | } | |
1605 | ||
1606 | ||
1607 | /* Function vect_finish_stmt_generation. | |
1608 | ||
1609 | Insert a new stmt. */ | |
1610 | ||
1611 | void | |
355fe088 | 1612 | vect_finish_stmt_generation (gimple *stmt, gimple *vec_stmt, |
ebfd146a IR |
1613 | gimple_stmt_iterator *gsi) |
1614 | { | |
1615 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
310213d4 | 1616 | vec_info *vinfo = stmt_info->vinfo; |
ebfd146a IR |
1617 | |
1618 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1619 | ||
54e8e2c3 RG |
1620 | if (!gsi_end_p (*gsi) |
1621 | && gimple_has_mem_ops (vec_stmt)) | |
1622 | { | |
355fe088 | 1623 | gimple *at_stmt = gsi_stmt (*gsi); |
54e8e2c3 RG |
1624 | tree vuse = gimple_vuse (at_stmt); |
1625 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1626 | { | |
1627 | tree vdef = gimple_vdef (at_stmt); | |
1628 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1629 | /* If we have an SSA vuse and insert a store, update virtual | |
1630 | SSA form to avoid triggering the renamer. Do so only | |
1631 | if we can easily see all uses - which is what almost always | |
1632 | happens with the way vectorized stmts are inserted. */ | |
1633 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1634 | && ((is_gimple_assign (vec_stmt) | |
1635 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1636 | || (is_gimple_call (vec_stmt) | |
1637 | && !(gimple_call_flags (vec_stmt) | |
1638 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1639 | { | |
1640 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1641 | gimple_set_vdef (vec_stmt, new_vdef); | |
1642 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1643 | } | |
1644 | } | |
1645 | } | |
ebfd146a IR |
1646 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1647 | ||
310213d4 | 1648 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, vinfo)); |
ebfd146a | 1649 | |
73fbfcad | 1650 | if (dump_enabled_p ()) |
ebfd146a | 1651 | { |
78c60e3d SS |
1652 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); |
1653 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
ebfd146a IR |
1654 | } |
1655 | ||
ad885386 | 1656 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
8e91d222 JJ |
1657 | |
1658 | /* While EH edges will generally prevent vectorization, stmt might | |
1659 | e.g. be in a must-not-throw region. Ensure newly created stmts | |
1660 | that could throw are part of the same region. */ | |
1661 | int lp_nr = lookup_stmt_eh_lp (stmt); | |
1662 | if (lp_nr != 0 && stmt_could_throw_p (vec_stmt)) | |
1663 | add_stmt_to_eh_lp (vec_stmt, lp_nr); | |
ebfd146a IR |
1664 | } |
1665 | ||
70439f0d RS |
1666 | /* We want to vectorize a call to combined function CFN with function |
1667 | decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN | |
1668 | as the types of all inputs. Check whether this is possible using | |
1669 | an internal function, returning its code if so or IFN_LAST if not. */ | |
ebfd146a | 1670 | |
70439f0d RS |
1671 | static internal_fn |
1672 | vectorizable_internal_function (combined_fn cfn, tree fndecl, | |
1673 | tree vectype_out, tree vectype_in) | |
ebfd146a | 1674 | { |
70439f0d RS |
1675 | internal_fn ifn; |
1676 | if (internal_fn_p (cfn)) | |
1677 | ifn = as_internal_fn (cfn); | |
1678 | else | |
1679 | ifn = associated_internal_fn (fndecl); | |
1680 | if (ifn != IFN_LAST && direct_internal_fn_p (ifn)) | |
1681 | { | |
1682 | const direct_internal_fn_info &info = direct_internal_fn (ifn); | |
1683 | if (info.vectorizable) | |
1684 | { | |
1685 | tree type0 = (info.type0 < 0 ? vectype_out : vectype_in); | |
1686 | tree type1 = (info.type1 < 0 ? vectype_out : vectype_in); | |
d95ab70a RS |
1687 | if (direct_internal_fn_supported_p (ifn, tree_pair (type0, type1), |
1688 | OPTIMIZE_FOR_SPEED)) | |
70439f0d RS |
1689 | return ifn; |
1690 | } | |
1691 | } | |
1692 | return IFN_LAST; | |
ebfd146a IR |
1693 | } |
1694 | ||
5ce9450f | 1695 | |
355fe088 | 1696 | static tree permute_vec_elements (tree, tree, tree, gimple *, |
5ce9450f JJ |
1697 | gimple_stmt_iterator *); |
1698 | ||
7cfb4d93 RS |
1699 | /* Check whether a load or store statement in the loop described by |
1700 | LOOP_VINFO is possible in a fully-masked loop. This is testing | |
1701 | whether the vectorizer pass has the appropriate support, as well as | |
1702 | whether the target does. | |
1703 | ||
1704 | VLS_TYPE says whether the statement is a load or store and VECTYPE | |
1705 | is the type of the vector being loaded or stored. MEMORY_ACCESS_TYPE | |
1706 | says how the load or store is going to be implemented and GROUP_SIZE | |
1707 | is the number of load or store statements in the containing group. | |
1708 | ||
1709 | Clear LOOP_VINFO_CAN_FULLY_MASK_P if a fully-masked loop is not | |
1710 | supported, otherwise record the required mask types. */ | |
1711 | ||
1712 | static void | |
1713 | check_load_store_masking (loop_vec_info loop_vinfo, tree vectype, | |
1714 | vec_load_store_type vls_type, int group_size, | |
1715 | vect_memory_access_type memory_access_type) | |
1716 | { | |
1717 | /* Invariant loads need no special support. */ | |
1718 | if (memory_access_type == VMAT_INVARIANT) | |
1719 | return; | |
1720 | ||
1721 | vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); | |
1722 | machine_mode vecmode = TYPE_MODE (vectype); | |
1723 | bool is_load = (vls_type == VLS_LOAD); | |
1724 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
1725 | { | |
1726 | if (is_load | |
1727 | ? !vect_load_lanes_supported (vectype, group_size, true) | |
1728 | : !vect_store_lanes_supported (vectype, group_size, true)) | |
1729 | { | |
1730 | if (dump_enabled_p ()) | |
1731 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1732 | "can't use a fully-masked loop because the" | |
1733 | " target doesn't have an appropriate masked" | |
1734 | " load/store-lanes instruction.\n"); | |
1735 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1736 | return; | |
1737 | } | |
1738 | unsigned int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
1739 | vect_record_loop_mask (loop_vinfo, masks, ncopies, vectype); | |
1740 | return; | |
1741 | } | |
1742 | ||
1743 | if (memory_access_type != VMAT_CONTIGUOUS | |
1744 | && memory_access_type != VMAT_CONTIGUOUS_PERMUTE) | |
1745 | { | |
1746 | /* Element X of the data must come from iteration i * VF + X of the | |
1747 | scalar loop. We need more work to support other mappings. */ | |
1748 | if (dump_enabled_p ()) | |
1749 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1750 | "can't use a fully-masked loop because an access" | |
1751 | " isn't contiguous.\n"); | |
1752 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1753 | return; | |
1754 | } | |
1755 | ||
1756 | machine_mode mask_mode; | |
1757 | if (!(targetm.vectorize.get_mask_mode | |
1758 | (GET_MODE_NUNITS (vecmode), | |
1759 | GET_MODE_SIZE (vecmode)).exists (&mask_mode)) | |
1760 | || !can_vec_mask_load_store_p (vecmode, mask_mode, is_load)) | |
1761 | { | |
1762 | if (dump_enabled_p ()) | |
1763 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1764 | "can't use a fully-masked loop because the target" | |
1765 | " doesn't have the appropriate masked load or" | |
1766 | " store.\n"); | |
1767 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1768 | return; | |
1769 | } | |
1770 | /* We might load more scalars than we need for permuting SLP loads. | |
1771 | We checked in get_group_load_store_type that the extra elements | |
1772 | don't leak into a new vector. */ | |
1773 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
1774 | poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
1775 | unsigned int nvectors; | |
1776 | if (can_div_away_from_zero_p (group_size * vf, nunits, &nvectors)) | |
1777 | vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype); | |
1778 | else | |
1779 | gcc_unreachable (); | |
1780 | } | |
1781 | ||
1782 | /* Return the mask input to a masked load or store. VEC_MASK is the vectorized | |
1783 | form of the scalar mask condition and LOOP_MASK, if nonnull, is the mask | |
1784 | that needs to be applied to all loads and stores in a vectorized loop. | |
1785 | Return VEC_MASK if LOOP_MASK is null, otherwise return VEC_MASK & LOOP_MASK. | |
1786 | ||
1787 | MASK_TYPE is the type of both masks. If new statements are needed, | |
1788 | insert them before GSI. */ | |
1789 | ||
1790 | static tree | |
1791 | prepare_load_store_mask (tree mask_type, tree loop_mask, tree vec_mask, | |
1792 | gimple_stmt_iterator *gsi) | |
1793 | { | |
1794 | gcc_assert (useless_type_conversion_p (mask_type, TREE_TYPE (vec_mask))); | |
1795 | if (!loop_mask) | |
1796 | return vec_mask; | |
1797 | ||
1798 | gcc_assert (TREE_TYPE (loop_mask) == mask_type); | |
1799 | tree and_res = make_temp_ssa_name (mask_type, NULL, "vec_mask_and"); | |
1800 | gimple *and_stmt = gimple_build_assign (and_res, BIT_AND_EXPR, | |
1801 | vec_mask, loop_mask); | |
1802 | gsi_insert_before (gsi, and_stmt, GSI_SAME_STMT); | |
1803 | return and_res; | |
1804 | } | |
1805 | ||
62da9e14 RS |
1806 | /* STMT is a non-strided load or store, meaning that it accesses |
1807 | elements with a known constant step. Return -1 if that step | |
1808 | is negative, 0 if it is zero, and 1 if it is greater than zero. */ | |
1809 | ||
1810 | static int | |
1811 | compare_step_with_zero (gimple *stmt) | |
1812 | { | |
1813 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3f5e8a76 RS |
1814 | data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
1815 | return tree_int_cst_compare (vect_dr_behavior (dr)->step, | |
1816 | size_zero_node); | |
62da9e14 RS |
1817 | } |
1818 | ||
1819 | /* If the target supports a permute mask that reverses the elements in | |
1820 | a vector of type VECTYPE, return that mask, otherwise return null. */ | |
1821 | ||
1822 | static tree | |
1823 | perm_mask_for_reverse (tree vectype) | |
1824 | { | |
928686b1 | 1825 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
62da9e14 | 1826 | |
d980067b RS |
1827 | /* The encoding has a single stepped pattern. */ |
1828 | vec_perm_builder sel (nunits, 1, 3); | |
928686b1 | 1829 | for (int i = 0; i < 3; ++i) |
908a1a16 | 1830 | sel.quick_push (nunits - 1 - i); |
62da9e14 | 1831 | |
e3342de4 RS |
1832 | vec_perm_indices indices (sel, 1, nunits); |
1833 | if (!can_vec_perm_const_p (TYPE_MODE (vectype), indices)) | |
62da9e14 | 1834 | return NULL_TREE; |
e3342de4 | 1835 | return vect_gen_perm_mask_checked (vectype, indices); |
62da9e14 | 1836 | } |
5ce9450f | 1837 | |
c3a8f964 RS |
1838 | /* STMT is either a masked or unconditional store. Return the value |
1839 | being stored. */ | |
1840 | ||
1841 | static tree | |
1842 | vect_get_store_rhs (gimple *stmt) | |
1843 | { | |
1844 | if (gassign *assign = dyn_cast <gassign *> (stmt)) | |
1845 | { | |
1846 | gcc_assert (gimple_assign_single_p (assign)); | |
1847 | return gimple_assign_rhs1 (assign); | |
1848 | } | |
1849 | if (gcall *call = dyn_cast <gcall *> (stmt)) | |
1850 | { | |
1851 | internal_fn ifn = gimple_call_internal_fn (call); | |
1852 | gcc_assert (ifn == IFN_MASK_STORE); | |
1853 | return gimple_call_arg (stmt, 3); | |
1854 | } | |
1855 | gcc_unreachable (); | |
1856 | } | |
1857 | ||
2de001ee RS |
1858 | /* A subroutine of get_load_store_type, with a subset of the same |
1859 | arguments. Handle the case where STMT is part of a grouped load | |
1860 | or store. | |
1861 | ||
1862 | For stores, the statements in the group are all consecutive | |
1863 | and there is no gap at the end. For loads, the statements in the | |
1864 | group might not be consecutive; there can be gaps between statements | |
1865 | as well as at the end. */ | |
1866 | ||
1867 | static bool | |
1868 | get_group_load_store_type (gimple *stmt, tree vectype, bool slp, | |
7e11fc7f | 1869 | bool masked_p, vec_load_store_type vls_type, |
2de001ee RS |
1870 | vect_memory_access_type *memory_access_type) |
1871 | { | |
1872 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1873 | vec_info *vinfo = stmt_info->vinfo; | |
1874 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1875 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; | |
1876 | gimple *first_stmt = GROUP_FIRST_ELEMENT (stmt_info); | |
f702e7d4 | 1877 | data_reference *first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
2de001ee RS |
1878 | unsigned int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
1879 | bool single_element_p = (stmt == first_stmt | |
1880 | && !GROUP_NEXT_ELEMENT (stmt_info)); | |
1881 | unsigned HOST_WIDE_INT gap = GROUP_GAP (vinfo_for_stmt (first_stmt)); | |
928686b1 | 1882 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2de001ee RS |
1883 | |
1884 | /* True if the vectorized statements would access beyond the last | |
1885 | statement in the group. */ | |
1886 | bool overrun_p = false; | |
1887 | ||
1888 | /* True if we can cope with such overrun by peeling for gaps, so that | |
1889 | there is at least one final scalar iteration after the vector loop. */ | |
7e11fc7f RS |
1890 | bool can_overrun_p = (!masked_p |
1891 | && vls_type == VLS_LOAD | |
1892 | && loop_vinfo | |
1893 | && !loop->inner); | |
2de001ee RS |
1894 | |
1895 | /* There can only be a gap at the end of the group if the stride is | |
1896 | known at compile time. */ | |
1897 | gcc_assert (!STMT_VINFO_STRIDED_P (stmt_info) || gap == 0); | |
1898 | ||
1899 | /* Stores can't yet have gaps. */ | |
1900 | gcc_assert (slp || vls_type == VLS_LOAD || gap == 0); | |
1901 | ||
1902 | if (slp) | |
1903 | { | |
1904 | if (STMT_VINFO_STRIDED_P (stmt_info)) | |
1905 | { | |
1906 | /* Try to use consecutive accesses of GROUP_SIZE elements, | |
1907 | separated by the stride, until we have a complete vector. | |
1908 | Fall back to scalar accesses if that isn't possible. */ | |
928686b1 | 1909 | if (multiple_p (nunits, group_size)) |
2de001ee RS |
1910 | *memory_access_type = VMAT_STRIDED_SLP; |
1911 | else | |
1912 | *memory_access_type = VMAT_ELEMENTWISE; | |
1913 | } | |
1914 | else | |
1915 | { | |
1916 | overrun_p = loop_vinfo && gap != 0; | |
1917 | if (overrun_p && vls_type != VLS_LOAD) | |
1918 | { | |
1919 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1920 | "Grouped store with gaps requires" | |
1921 | " non-consecutive accesses\n"); | |
1922 | return false; | |
1923 | } | |
f702e7d4 RS |
1924 | /* An overrun is fine if the trailing elements are smaller |
1925 | than the alignment boundary B. Every vector access will | |
1926 | be a multiple of B and so we are guaranteed to access a | |
1927 | non-gap element in the same B-sized block. */ | |
f9ef2c76 | 1928 | if (overrun_p |
f702e7d4 RS |
1929 | && gap < (vect_known_alignment_in_bytes (first_dr) |
1930 | / vect_get_scalar_dr_size (first_dr))) | |
f9ef2c76 | 1931 | overrun_p = false; |
2de001ee RS |
1932 | if (overrun_p && !can_overrun_p) |
1933 | { | |
1934 | if (dump_enabled_p ()) | |
1935 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1936 | "Peeling for outer loop is not supported\n"); | |
1937 | return false; | |
1938 | } | |
1939 | *memory_access_type = VMAT_CONTIGUOUS; | |
1940 | } | |
1941 | } | |
1942 | else | |
1943 | { | |
1944 | /* We can always handle this case using elementwise accesses, | |
1945 | but see if something more efficient is available. */ | |
1946 | *memory_access_type = VMAT_ELEMENTWISE; | |
1947 | ||
1948 | /* If there is a gap at the end of the group then these optimizations | |
1949 | would access excess elements in the last iteration. */ | |
1950 | bool would_overrun_p = (gap != 0); | |
f702e7d4 RS |
1951 | /* An overrun is fine if the trailing elements are smaller than the |
1952 | alignment boundary B. Every vector access will be a multiple of B | |
1953 | and so we are guaranteed to access a non-gap element in the | |
1954 | same B-sized block. */ | |
f9ef2c76 | 1955 | if (would_overrun_p |
7e11fc7f | 1956 | && !masked_p |
f702e7d4 RS |
1957 | && gap < (vect_known_alignment_in_bytes (first_dr) |
1958 | / vect_get_scalar_dr_size (first_dr))) | |
f9ef2c76 | 1959 | would_overrun_p = false; |
f702e7d4 | 1960 | |
2de001ee | 1961 | if (!STMT_VINFO_STRIDED_P (stmt_info) |
62da9e14 RS |
1962 | && (can_overrun_p || !would_overrun_p) |
1963 | && compare_step_with_zero (stmt) > 0) | |
2de001ee | 1964 | { |
6737facb RS |
1965 | /* First cope with the degenerate case of a single-element |
1966 | vector. */ | |
1967 | if (known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U)) | |
1968 | *memory_access_type = VMAT_CONTIGUOUS; | |
1969 | ||
1970 | /* Otherwise try using LOAD/STORE_LANES. */ | |
1971 | if (*memory_access_type == VMAT_ELEMENTWISE | |
1972 | && (vls_type == VLS_LOAD | |
7e11fc7f RS |
1973 | ? vect_load_lanes_supported (vectype, group_size, masked_p) |
1974 | : vect_store_lanes_supported (vectype, group_size, | |
1975 | masked_p))) | |
2de001ee RS |
1976 | { |
1977 | *memory_access_type = VMAT_LOAD_STORE_LANES; | |
1978 | overrun_p = would_overrun_p; | |
1979 | } | |
1980 | ||
1981 | /* If that fails, try using permuting loads. */ | |
1982 | if (*memory_access_type == VMAT_ELEMENTWISE | |
1983 | && (vls_type == VLS_LOAD | |
1984 | ? vect_grouped_load_supported (vectype, single_element_p, | |
1985 | group_size) | |
1986 | : vect_grouped_store_supported (vectype, group_size))) | |
1987 | { | |
1988 | *memory_access_type = VMAT_CONTIGUOUS_PERMUTE; | |
1989 | overrun_p = would_overrun_p; | |
1990 | } | |
1991 | } | |
1992 | } | |
1993 | ||
1994 | if (vls_type != VLS_LOAD && first_stmt == stmt) | |
1995 | { | |
1996 | /* STMT is the leader of the group. Check the operands of all the | |
1997 | stmts of the group. */ | |
1998 | gimple *next_stmt = GROUP_NEXT_ELEMENT (stmt_info); | |
1999 | while (next_stmt) | |
2000 | { | |
7e11fc7f | 2001 | tree op = vect_get_store_rhs (next_stmt); |
2de001ee RS |
2002 | gimple *def_stmt; |
2003 | enum vect_def_type dt; | |
2004 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt)) | |
2005 | { | |
2006 | if (dump_enabled_p ()) | |
2007 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2008 | "use not simple.\n"); | |
2009 | return false; | |
2010 | } | |
2011 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); | |
2012 | } | |
2013 | } | |
2014 | ||
2015 | if (overrun_p) | |
2016 | { | |
2017 | gcc_assert (can_overrun_p); | |
2018 | if (dump_enabled_p ()) | |
2019 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2020 | "Data access with gaps requires scalar " | |
2021 | "epilogue loop\n"); | |
2022 | LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true; | |
2023 | } | |
2024 | ||
2025 | return true; | |
2026 | } | |
2027 | ||
62da9e14 RS |
2028 | /* A subroutine of get_load_store_type, with a subset of the same |
2029 | arguments. Handle the case where STMT is a load or store that | |
2030 | accesses consecutive elements with a negative step. */ | |
2031 | ||
2032 | static vect_memory_access_type | |
2033 | get_negative_load_store_type (gimple *stmt, tree vectype, | |
2034 | vec_load_store_type vls_type, | |
2035 | unsigned int ncopies) | |
2036 | { | |
2037 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2038 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
2039 | dr_alignment_support alignment_support_scheme; | |
2040 | ||
2041 | if (ncopies > 1) | |
2042 | { | |
2043 | if (dump_enabled_p ()) | |
2044 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2045 | "multiple types with negative step.\n"); | |
2046 | return VMAT_ELEMENTWISE; | |
2047 | } | |
2048 | ||
2049 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
2050 | if (alignment_support_scheme != dr_aligned | |
2051 | && alignment_support_scheme != dr_unaligned_supported) | |
2052 | { | |
2053 | if (dump_enabled_p ()) | |
2054 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2055 | "negative step but alignment required.\n"); | |
2056 | return VMAT_ELEMENTWISE; | |
2057 | } | |
2058 | ||
2059 | if (vls_type == VLS_STORE_INVARIANT) | |
2060 | { | |
2061 | if (dump_enabled_p ()) | |
2062 | dump_printf_loc (MSG_NOTE, vect_location, | |
2063 | "negative step with invariant source;" | |
2064 | " no permute needed.\n"); | |
2065 | return VMAT_CONTIGUOUS_DOWN; | |
2066 | } | |
2067 | ||
2068 | if (!perm_mask_for_reverse (vectype)) | |
2069 | { | |
2070 | if (dump_enabled_p ()) | |
2071 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2072 | "negative step and reversing not supported.\n"); | |
2073 | return VMAT_ELEMENTWISE; | |
2074 | } | |
2075 | ||
2076 | return VMAT_CONTIGUOUS_REVERSE; | |
2077 | } | |
2078 | ||
2de001ee RS |
2079 | /* Analyze load or store statement STMT of type VLS_TYPE. Return true |
2080 | if there is a memory access type that the vectorized form can use, | |
2081 | storing it in *MEMORY_ACCESS_TYPE if so. If we decide to use gathers | |
2082 | or scatters, fill in GS_INFO accordingly. | |
2083 | ||
2084 | SLP says whether we're performing SLP rather than loop vectorization. | |
7e11fc7f | 2085 | MASKED_P is true if the statement is conditional on a vectorized mask. |
62da9e14 RS |
2086 | VECTYPE is the vector type that the vectorized statements will use. |
2087 | NCOPIES is the number of vector statements that will be needed. */ | |
2de001ee RS |
2088 | |
2089 | static bool | |
7e11fc7f | 2090 | get_load_store_type (gimple *stmt, tree vectype, bool slp, bool masked_p, |
62da9e14 | 2091 | vec_load_store_type vls_type, unsigned int ncopies, |
2de001ee RS |
2092 | vect_memory_access_type *memory_access_type, |
2093 | gather_scatter_info *gs_info) | |
2094 | { | |
2095 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2096 | vec_info *vinfo = stmt_info->vinfo; | |
2097 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4d694b27 | 2098 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2de001ee RS |
2099 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
2100 | { | |
2101 | *memory_access_type = VMAT_GATHER_SCATTER; | |
2102 | gimple *def_stmt; | |
2103 | if (!vect_check_gather_scatter (stmt, loop_vinfo, gs_info)) | |
2104 | gcc_unreachable (); | |
2105 | else if (!vect_is_simple_use (gs_info->offset, vinfo, &def_stmt, | |
2106 | &gs_info->offset_dt, | |
2107 | &gs_info->offset_vectype)) | |
2108 | { | |
2109 | if (dump_enabled_p ()) | |
2110 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2111 | "%s index use not simple.\n", | |
2112 | vls_type == VLS_LOAD ? "gather" : "scatter"); | |
2113 | return false; | |
2114 | } | |
2115 | } | |
2116 | else if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
2117 | { | |
7e11fc7f | 2118 | if (!get_group_load_store_type (stmt, vectype, slp, masked_p, vls_type, |
2de001ee RS |
2119 | memory_access_type)) |
2120 | return false; | |
2121 | } | |
2122 | else if (STMT_VINFO_STRIDED_P (stmt_info)) | |
2123 | { | |
2124 | gcc_assert (!slp); | |
2125 | *memory_access_type = VMAT_ELEMENTWISE; | |
2126 | } | |
2127 | else | |
62da9e14 RS |
2128 | { |
2129 | int cmp = compare_step_with_zero (stmt); | |
2130 | if (cmp < 0) | |
2131 | *memory_access_type = get_negative_load_store_type | |
2132 | (stmt, vectype, vls_type, ncopies); | |
2133 | else if (cmp == 0) | |
2134 | { | |
2135 | gcc_assert (vls_type == VLS_LOAD); | |
2136 | *memory_access_type = VMAT_INVARIANT; | |
2137 | } | |
2138 | else | |
2139 | *memory_access_type = VMAT_CONTIGUOUS; | |
2140 | } | |
2de001ee | 2141 | |
4d694b27 RS |
2142 | if ((*memory_access_type == VMAT_ELEMENTWISE |
2143 | || *memory_access_type == VMAT_STRIDED_SLP) | |
2144 | && !nunits.is_constant ()) | |
2145 | { | |
2146 | if (dump_enabled_p ()) | |
2147 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2148 | "Not using elementwise accesses due to variable " | |
2149 | "vectorization factor.\n"); | |
2150 | return false; | |
2151 | } | |
2152 | ||
2de001ee RS |
2153 | /* FIXME: At the moment the cost model seems to underestimate the |
2154 | cost of using elementwise accesses. This check preserves the | |
2155 | traditional behavior until that can be fixed. */ | |
2156 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2157 | && !STMT_VINFO_STRIDED_P (stmt_info)) | |
2158 | { | |
2159 | if (dump_enabled_p ()) | |
2160 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2161 | "not falling back to elementwise accesses\n"); | |
2162 | return false; | |
2163 | } | |
2164 | return true; | |
2165 | } | |
2166 | ||
aaeefd88 RS |
2167 | /* Return true if boolean argument MASK is suitable for vectorizing |
2168 | conditional load or store STMT. When returning true, store the | |
2169 | type of the vectorized mask in *MASK_VECTYPE_OUT. */ | |
2170 | ||
2171 | static bool | |
2172 | vect_check_load_store_mask (gimple *stmt, tree mask, tree *mask_vectype_out) | |
2173 | { | |
2174 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (mask))) | |
2175 | { | |
2176 | if (dump_enabled_p ()) | |
2177 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2178 | "mask argument is not a boolean.\n"); | |
2179 | return false; | |
2180 | } | |
2181 | ||
2182 | if (TREE_CODE (mask) != SSA_NAME) | |
2183 | { | |
2184 | if (dump_enabled_p ()) | |
2185 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2186 | "mask argument is not an SSA name.\n"); | |
2187 | return false; | |
2188 | } | |
2189 | ||
2190 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2191 | gimple *def_stmt; | |
2192 | enum vect_def_type dt; | |
2193 | tree mask_vectype; | |
2194 | if (!vect_is_simple_use (mask, stmt_info->vinfo, &def_stmt, &dt, | |
2195 | &mask_vectype)) | |
2196 | { | |
2197 | if (dump_enabled_p ()) | |
2198 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2199 | "mask use not simple.\n"); | |
2200 | return false; | |
2201 | } | |
2202 | ||
2203 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2204 | if (!mask_vectype) | |
2205 | mask_vectype = get_mask_type_for_scalar_type (TREE_TYPE (vectype)); | |
2206 | ||
2207 | if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype)) | |
2208 | { | |
2209 | if (dump_enabled_p ()) | |
2210 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2211 | "could not find an appropriate vector mask type.\n"); | |
2212 | return false; | |
2213 | } | |
2214 | ||
2215 | if (maybe_ne (TYPE_VECTOR_SUBPARTS (mask_vectype), | |
2216 | TYPE_VECTOR_SUBPARTS (vectype))) | |
2217 | { | |
2218 | if (dump_enabled_p ()) | |
2219 | { | |
2220 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2221 | "vector mask type "); | |
2222 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, mask_vectype); | |
2223 | dump_printf (MSG_MISSED_OPTIMIZATION, | |
2224 | " does not match vector data type "); | |
2225 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, vectype); | |
2226 | dump_printf (MSG_MISSED_OPTIMIZATION, ".\n"); | |
2227 | } | |
2228 | return false; | |
2229 | } | |
2230 | ||
2231 | *mask_vectype_out = mask_vectype; | |
2232 | return true; | |
2233 | } | |
2234 | ||
3133c3b6 RS |
2235 | /* Return true if stored value RHS is suitable for vectorizing store |
2236 | statement STMT. When returning true, store the type of the | |
2237 | vectorized store value in *RHS_VECTYPE_OUT and the type of the | |
2238 | store in *VLS_TYPE_OUT. */ | |
2239 | ||
2240 | static bool | |
2241 | vect_check_store_rhs (gimple *stmt, tree rhs, tree *rhs_vectype_out, | |
2242 | vec_load_store_type *vls_type_out) | |
2243 | { | |
2244 | /* In the case this is a store from a constant make sure | |
2245 | native_encode_expr can handle it. */ | |
2246 | if (CONSTANT_CLASS_P (rhs) && native_encode_expr (rhs, NULL, 64) == 0) | |
2247 | { | |
2248 | if (dump_enabled_p ()) | |
2249 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2250 | "cannot encode constant as a byte sequence.\n"); | |
2251 | return false; | |
2252 | } | |
2253 | ||
2254 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2255 | gimple *def_stmt; | |
2256 | enum vect_def_type dt; | |
2257 | tree rhs_vectype; | |
2258 | if (!vect_is_simple_use (rhs, stmt_info->vinfo, &def_stmt, &dt, | |
2259 | &rhs_vectype)) | |
2260 | { | |
2261 | if (dump_enabled_p ()) | |
2262 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2263 | "use not simple.\n"); | |
2264 | return false; | |
2265 | } | |
2266 | ||
2267 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2268 | if (rhs_vectype && !useless_type_conversion_p (vectype, rhs_vectype)) | |
2269 | { | |
2270 | if (dump_enabled_p ()) | |
2271 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2272 | "incompatible vector types.\n"); | |
2273 | return false; | |
2274 | } | |
2275 | ||
2276 | *rhs_vectype_out = rhs_vectype; | |
2277 | if (dt == vect_constant_def || dt == vect_external_def) | |
2278 | *vls_type_out = VLS_STORE_INVARIANT; | |
2279 | else | |
2280 | *vls_type_out = VLS_STORE; | |
2281 | return true; | |
2282 | } | |
2283 | ||
bc9587eb RS |
2284 | /* Build an all-ones vector mask of type MASKTYPE while vectorizing STMT. |
2285 | Note that we support masks with floating-point type, in which case the | |
2286 | floats are interpreted as a bitmask. */ | |
2287 | ||
2288 | static tree | |
2289 | vect_build_all_ones_mask (gimple *stmt, tree masktype) | |
2290 | { | |
2291 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
2292 | return build_int_cst (masktype, -1); | |
2293 | else if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
2294 | { | |
2295 | tree mask = build_int_cst (TREE_TYPE (masktype), -1); | |
2296 | mask = build_vector_from_val (masktype, mask); | |
2297 | return vect_init_vector (stmt, mask, masktype, NULL); | |
2298 | } | |
2299 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) | |
2300 | { | |
2301 | REAL_VALUE_TYPE r; | |
2302 | long tmp[6]; | |
2303 | for (int j = 0; j < 6; ++j) | |
2304 | tmp[j] = -1; | |
2305 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
2306 | tree mask = build_real (TREE_TYPE (masktype), r); | |
2307 | mask = build_vector_from_val (masktype, mask); | |
2308 | return vect_init_vector (stmt, mask, masktype, NULL); | |
2309 | } | |
2310 | gcc_unreachable (); | |
2311 | } | |
2312 | ||
2313 | /* Build an all-zero merge value of type VECTYPE while vectorizing | |
2314 | STMT as a gather load. */ | |
2315 | ||
2316 | static tree | |
2317 | vect_build_zero_merge_argument (gimple *stmt, tree vectype) | |
2318 | { | |
2319 | tree merge; | |
2320 | if (TREE_CODE (TREE_TYPE (vectype)) == INTEGER_TYPE) | |
2321 | merge = build_int_cst (TREE_TYPE (vectype), 0); | |
2322 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (vectype))) | |
2323 | { | |
2324 | REAL_VALUE_TYPE r; | |
2325 | long tmp[6]; | |
2326 | for (int j = 0; j < 6; ++j) | |
2327 | tmp[j] = 0; | |
2328 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (vectype))); | |
2329 | merge = build_real (TREE_TYPE (vectype), r); | |
2330 | } | |
2331 | else | |
2332 | gcc_unreachable (); | |
2333 | merge = build_vector_from_val (vectype, merge); | |
2334 | return vect_init_vector (stmt, merge, vectype, NULL); | |
2335 | } | |
2336 | ||
c48d2d35 RS |
2337 | /* Build a gather load call while vectorizing STMT. Insert new instructions |
2338 | before GSI and add them to VEC_STMT. GS_INFO describes the gather load | |
2339 | operation. If the load is conditional, MASK is the unvectorized | |
2340 | condition, otherwise MASK is null. */ | |
2341 | ||
2342 | static void | |
2343 | vect_build_gather_load_calls (gimple *stmt, gimple_stmt_iterator *gsi, | |
2344 | gimple **vec_stmt, gather_scatter_info *gs_info, | |
2345 | tree mask) | |
2346 | { | |
2347 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2348 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2349 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2350 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2351 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
2352 | int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
2353 | edge pe = loop_preheader_edge (loop); | |
2354 | enum { NARROW, NONE, WIDEN } modifier; | |
2355 | poly_uint64 gather_off_nunits | |
2356 | = TYPE_VECTOR_SUBPARTS (gs_info->offset_vectype); | |
2357 | ||
2358 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info->decl)); | |
2359 | tree rettype = TREE_TYPE (TREE_TYPE (gs_info->decl)); | |
2360 | tree srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2361 | tree ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2362 | tree idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2363 | tree masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2364 | tree scaletype = TREE_VALUE (arglist); | |
2365 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
2366 | && (!mask || types_compatible_p (srctype, masktype))); | |
2367 | ||
2368 | tree perm_mask = NULL_TREE; | |
2369 | tree mask_perm_mask = NULL_TREE; | |
2370 | if (known_eq (nunits, gather_off_nunits)) | |
2371 | modifier = NONE; | |
2372 | else if (known_eq (nunits * 2, gather_off_nunits)) | |
2373 | { | |
2374 | modifier = WIDEN; | |
2375 | ||
2376 | /* Currently widening gathers and scatters are only supported for | |
2377 | fixed-length vectors. */ | |
2378 | int count = gather_off_nunits.to_constant (); | |
2379 | vec_perm_builder sel (count, count, 1); | |
2380 | for (int i = 0; i < count; ++i) | |
2381 | sel.quick_push (i | (count / 2)); | |
2382 | ||
2383 | vec_perm_indices indices (sel, 1, count); | |
2384 | perm_mask = vect_gen_perm_mask_checked (gs_info->offset_vectype, | |
2385 | indices); | |
2386 | } | |
2387 | else if (known_eq (nunits, gather_off_nunits * 2)) | |
2388 | { | |
2389 | modifier = NARROW; | |
2390 | ||
2391 | /* Currently narrowing gathers and scatters are only supported for | |
2392 | fixed-length vectors. */ | |
2393 | int count = nunits.to_constant (); | |
2394 | vec_perm_builder sel (count, count, 1); | |
2395 | sel.quick_grow (count); | |
2396 | for (int i = 0; i < count; ++i) | |
2397 | sel[i] = i < count / 2 ? i : i + count / 2; | |
2398 | vec_perm_indices indices (sel, 2, count); | |
2399 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); | |
2400 | ||
2401 | ncopies *= 2; | |
2402 | ||
2403 | if (mask) | |
2404 | { | |
2405 | for (int i = 0; i < count; ++i) | |
2406 | sel[i] = i | (count / 2); | |
2407 | indices.new_vector (sel, 2, count); | |
2408 | mask_perm_mask = vect_gen_perm_mask_checked (masktype, indices); | |
2409 | } | |
2410 | } | |
2411 | else | |
2412 | gcc_unreachable (); | |
2413 | ||
2414 | tree vec_dest = vect_create_destination_var (gimple_get_lhs (stmt), | |
2415 | vectype); | |
2416 | ||
2417 | tree ptr = fold_convert (ptrtype, gs_info->base); | |
2418 | if (!is_gimple_min_invariant (ptr)) | |
2419 | { | |
2420 | gimple_seq seq; | |
2421 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
2422 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
2423 | gcc_assert (!new_bb); | |
2424 | } | |
2425 | ||
2426 | tree scale = build_int_cst (scaletype, gs_info->scale); | |
2427 | ||
2428 | tree vec_oprnd0 = NULL_TREE; | |
2429 | tree vec_mask = NULL_TREE; | |
2430 | tree src_op = NULL_TREE; | |
2431 | tree mask_op = NULL_TREE; | |
2432 | tree prev_res = NULL_TREE; | |
2433 | stmt_vec_info prev_stmt_info = NULL; | |
2434 | ||
2435 | if (!mask) | |
2436 | { | |
2437 | src_op = vect_build_zero_merge_argument (stmt, rettype); | |
2438 | mask_op = vect_build_all_ones_mask (stmt, masktype); | |
2439 | } | |
2440 | ||
2441 | for (int j = 0; j < ncopies; ++j) | |
2442 | { | |
2443 | tree op, var; | |
2444 | gimple *new_stmt; | |
2445 | if (modifier == WIDEN && (j & 1)) | |
2446 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
2447 | perm_mask, stmt, gsi); | |
2448 | else if (j == 0) | |
2449 | op = vec_oprnd0 | |
2450 | = vect_get_vec_def_for_operand (gs_info->offset, stmt); | |
2451 | else | |
2452 | op = vec_oprnd0 | |
2453 | = vect_get_vec_def_for_stmt_copy (gs_info->offset_dt, vec_oprnd0); | |
2454 | ||
2455 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
2456 | { | |
2457 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)), | |
2458 | TYPE_VECTOR_SUBPARTS (idxtype))); | |
2459 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); | |
2460 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); | |
2461 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
2462 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2463 | op = var; | |
2464 | } | |
2465 | ||
2466 | if (mask) | |
2467 | { | |
2468 | if (mask_perm_mask && (j & 1)) | |
2469 | mask_op = permute_vec_elements (mask_op, mask_op, | |
2470 | mask_perm_mask, stmt, gsi); | |
2471 | else | |
2472 | { | |
2473 | if (j == 0) | |
2474 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); | |
2475 | else | |
2476 | { | |
2477 | gimple *def_stmt; | |
2478 | enum vect_def_type dt; | |
2479 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); | |
2480 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); | |
2481 | } | |
2482 | ||
2483 | mask_op = vec_mask; | |
2484 | if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask))) | |
2485 | { | |
2486 | gcc_assert | |
2487 | (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)), | |
2488 | TYPE_VECTOR_SUBPARTS (masktype))); | |
2489 | var = vect_get_new_ssa_name (masktype, vect_simple_var); | |
2490 | mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op); | |
2491 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, | |
2492 | mask_op); | |
2493 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2494 | mask_op = var; | |
2495 | } | |
2496 | } | |
2497 | src_op = mask_op; | |
2498 | } | |
2499 | ||
2500 | new_stmt = gimple_build_call (gs_info->decl, 5, src_op, ptr, op, | |
2501 | mask_op, scale); | |
2502 | ||
2503 | if (!useless_type_conversion_p (vectype, rettype)) | |
2504 | { | |
2505 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (vectype), | |
2506 | TYPE_VECTOR_SUBPARTS (rettype))); | |
2507 | op = vect_get_new_ssa_name (rettype, vect_simple_var); | |
2508 | gimple_call_set_lhs (new_stmt, op); | |
2509 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2510 | var = make_ssa_name (vec_dest); | |
2511 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); | |
2512 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
2513 | } | |
2514 | else | |
2515 | { | |
2516 | var = make_ssa_name (vec_dest, new_stmt); | |
2517 | gimple_call_set_lhs (new_stmt, var); | |
2518 | } | |
2519 | ||
2520 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2521 | ||
2522 | if (modifier == NARROW) | |
2523 | { | |
2524 | if ((j & 1) == 0) | |
2525 | { | |
2526 | prev_res = var; | |
2527 | continue; | |
2528 | } | |
2529 | var = permute_vec_elements (prev_res, var, perm_mask, stmt, gsi); | |
2530 | new_stmt = SSA_NAME_DEF_STMT (var); | |
2531 | } | |
2532 | ||
2533 | if (prev_stmt_info == NULL) | |
2534 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2535 | else | |
2536 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2537 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2538 | } | |
2539 | } | |
2540 | ||
37b14185 RB |
2541 | /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64}. */ |
2542 | ||
2543 | static bool | |
2544 | vectorizable_bswap (gimple *stmt, gimple_stmt_iterator *gsi, | |
2545 | gimple **vec_stmt, slp_tree slp_node, | |
2546 | tree vectype_in, enum vect_def_type *dt) | |
2547 | { | |
2548 | tree op, vectype; | |
2549 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2550 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
928686b1 RS |
2551 | unsigned ncopies; |
2552 | unsigned HOST_WIDE_INT nunits, num_bytes; | |
37b14185 RB |
2553 | |
2554 | op = gimple_call_arg (stmt, 0); | |
2555 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
928686b1 RS |
2556 | |
2557 | if (!TYPE_VECTOR_SUBPARTS (vectype).is_constant (&nunits)) | |
2558 | return false; | |
37b14185 RB |
2559 | |
2560 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2561 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2562 | case of SLP. */ | |
2563 | if (slp_node) | |
2564 | ncopies = 1; | |
2565 | else | |
e8f142e2 | 2566 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
37b14185 RB |
2567 | |
2568 | gcc_assert (ncopies >= 1); | |
2569 | ||
2570 | tree char_vectype = get_same_sized_vectype (char_type_node, vectype_in); | |
2571 | if (! char_vectype) | |
2572 | return false; | |
2573 | ||
928686b1 RS |
2574 | if (!TYPE_VECTOR_SUBPARTS (char_vectype).is_constant (&num_bytes)) |
2575 | return false; | |
2576 | ||
794e3180 | 2577 | unsigned word_bytes = num_bytes / nunits; |
908a1a16 | 2578 | |
d980067b RS |
2579 | /* The encoding uses one stepped pattern for each byte in the word. */ |
2580 | vec_perm_builder elts (num_bytes, word_bytes, 3); | |
2581 | for (unsigned i = 0; i < 3; ++i) | |
37b14185 | 2582 | for (unsigned j = 0; j < word_bytes; ++j) |
908a1a16 | 2583 | elts.quick_push ((i + 1) * word_bytes - j - 1); |
37b14185 | 2584 | |
e3342de4 RS |
2585 | vec_perm_indices indices (elts, 1, num_bytes); |
2586 | if (!can_vec_perm_const_p (TYPE_MODE (char_vectype), indices)) | |
37b14185 RB |
2587 | return false; |
2588 | ||
2589 | if (! vec_stmt) | |
2590 | { | |
2591 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
2592 | if (dump_enabled_p ()) | |
2593 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_bswap ===" | |
2594 | "\n"); | |
2595 | if (! PURE_SLP_STMT (stmt_info)) | |
2596 | { | |
2597 | add_stmt_cost (stmt_info->vinfo->target_cost_data, | |
2598 | 1, vector_stmt, stmt_info, 0, vect_prologue); | |
2599 | add_stmt_cost (stmt_info->vinfo->target_cost_data, | |
2600 | ncopies, vec_perm, stmt_info, 0, vect_body); | |
2601 | } | |
2602 | return true; | |
2603 | } | |
2604 | ||
736d0f28 | 2605 | tree bswap_vconst = vec_perm_indices_to_tree (char_vectype, indices); |
37b14185 RB |
2606 | |
2607 | /* Transform. */ | |
2608 | vec<tree> vec_oprnds = vNULL; | |
2609 | gimple *new_stmt = NULL; | |
2610 | stmt_vec_info prev_stmt_info = NULL; | |
2611 | for (unsigned j = 0; j < ncopies; j++) | |
2612 | { | |
2613 | /* Handle uses. */ | |
2614 | if (j == 0) | |
306b0c92 | 2615 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); |
37b14185 RB |
2616 | else |
2617 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
2618 | ||
2619 | /* Arguments are ready. create the new vector stmt. */ | |
2620 | unsigned i; | |
2621 | tree vop; | |
2622 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) | |
2623 | { | |
2624 | tree tem = make_ssa_name (char_vectype); | |
2625 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
2626 | char_vectype, vop)); | |
2627 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2628 | tree tem2 = make_ssa_name (char_vectype); | |
2629 | new_stmt = gimple_build_assign (tem2, VEC_PERM_EXPR, | |
2630 | tem, tem, bswap_vconst); | |
2631 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2632 | tem = make_ssa_name (vectype); | |
2633 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
2634 | vectype, tem2)); | |
2635 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2636 | if (slp_node) | |
2637 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
2638 | } | |
2639 | ||
2640 | if (slp_node) | |
2641 | continue; | |
2642 | ||
2643 | if (j == 0) | |
2644 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2645 | else | |
2646 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2647 | ||
2648 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2649 | } | |
2650 | ||
2651 | vec_oprnds.release (); | |
2652 | return true; | |
2653 | } | |
2654 | ||
b1b6836e RS |
2655 | /* Return true if vector types VECTYPE_IN and VECTYPE_OUT have |
2656 | integer elements and if we can narrow VECTYPE_IN to VECTYPE_OUT | |
2657 | in a single step. On success, store the binary pack code in | |
2658 | *CONVERT_CODE. */ | |
2659 | ||
2660 | static bool | |
2661 | simple_integer_narrowing (tree vectype_out, tree vectype_in, | |
2662 | tree_code *convert_code) | |
2663 | { | |
2664 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_out)) | |
2665 | || !INTEGRAL_TYPE_P (TREE_TYPE (vectype_in))) | |
2666 | return false; | |
2667 | ||
2668 | tree_code code; | |
2669 | int multi_step_cvt = 0; | |
2670 | auto_vec <tree, 8> interm_types; | |
2671 | if (!supportable_narrowing_operation (NOP_EXPR, vectype_out, vectype_in, | |
2672 | &code, &multi_step_cvt, | |
2673 | &interm_types) | |
2674 | || multi_step_cvt) | |
2675 | return false; | |
2676 | ||
2677 | *convert_code = code; | |
2678 | return true; | |
2679 | } | |
5ce9450f | 2680 | |
ebfd146a IR |
2681 | /* Function vectorizable_call. |
2682 | ||
538dd0b7 | 2683 | Check if GS performs a function call that can be vectorized. |
b8698a0f | 2684 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
2685 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2686 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2687 | ||
2688 | static bool | |
355fe088 | 2689 | vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
190c2236 | 2690 | slp_tree slp_node) |
ebfd146a | 2691 | { |
538dd0b7 | 2692 | gcall *stmt; |
ebfd146a IR |
2693 | tree vec_dest; |
2694 | tree scalar_dest; | |
2695 | tree op, type; | |
2696 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
538dd0b7 | 2697 | stmt_vec_info stmt_info = vinfo_for_stmt (gs), prev_stmt_info; |
ebfd146a | 2698 | tree vectype_out, vectype_in; |
c7bda0f4 RS |
2699 | poly_uint64 nunits_in; |
2700 | poly_uint64 nunits_out; | |
ebfd146a | 2701 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
190c2236 | 2702 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 2703 | vec_info *vinfo = stmt_info->vinfo; |
81c40241 | 2704 | tree fndecl, new_temp, rhs_type; |
355fe088 | 2705 | gimple *def_stmt; |
0502fb85 UB |
2706 | enum vect_def_type dt[3] |
2707 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 2708 | int ndts = 3; |
355fe088 | 2709 | gimple *new_stmt = NULL; |
ebfd146a | 2710 | int ncopies, j; |
6e1aa848 | 2711 | vec<tree> vargs = vNULL; |
ebfd146a IR |
2712 | enum { NARROW, NONE, WIDEN } modifier; |
2713 | size_t i, nargs; | |
9d5e7640 | 2714 | tree lhs; |
ebfd146a | 2715 | |
190c2236 | 2716 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2717 | return false; |
2718 | ||
66c16fd9 RB |
2719 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
2720 | && ! vec_stmt) | |
ebfd146a IR |
2721 | return false; |
2722 | ||
538dd0b7 DM |
2723 | /* Is GS a vectorizable call? */ |
2724 | stmt = dyn_cast <gcall *> (gs); | |
2725 | if (!stmt) | |
ebfd146a IR |
2726 | return false; |
2727 | ||
5ce9450f JJ |
2728 | if (gimple_call_internal_p (stmt) |
2729 | && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD | |
2730 | || gimple_call_internal_fn (stmt) == IFN_MASK_STORE)) | |
c3a8f964 RS |
2731 | /* Handled by vectorizable_load and vectorizable_store. */ |
2732 | return false; | |
5ce9450f | 2733 | |
0136f8f0 AH |
2734 | if (gimple_call_lhs (stmt) == NULL_TREE |
2735 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
ebfd146a IR |
2736 | return false; |
2737 | ||
0136f8f0 | 2738 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); |
5a2c1986 | 2739 | |
b690cc0f RG |
2740 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
2741 | ||
ebfd146a IR |
2742 | /* Process function arguments. */ |
2743 | rhs_type = NULL_TREE; | |
b690cc0f | 2744 | vectype_in = NULL_TREE; |
ebfd146a IR |
2745 | nargs = gimple_call_num_args (stmt); |
2746 | ||
1b1562a5 MM |
2747 | /* Bail out if the function has more than three arguments, we do not have |
2748 | interesting builtin functions to vectorize with more than two arguments | |
2749 | except for fma. No arguments is also not good. */ | |
2750 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
2751 | return false; |
2752 | ||
74bf76ed JJ |
2753 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
2754 | if (gimple_call_internal_p (stmt) | |
2755 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2756 | { | |
2757 | nargs = 0; | |
2758 | rhs_type = unsigned_type_node; | |
2759 | } | |
2760 | ||
ebfd146a IR |
2761 | for (i = 0; i < nargs; i++) |
2762 | { | |
b690cc0f RG |
2763 | tree opvectype; |
2764 | ||
ebfd146a IR |
2765 | op = gimple_call_arg (stmt, i); |
2766 | ||
2767 | /* We can only handle calls with arguments of the same type. */ | |
2768 | if (rhs_type | |
8533c9d8 | 2769 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 2770 | { |
73fbfcad | 2771 | if (dump_enabled_p ()) |
78c60e3d | 2772 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2773 | "argument types differ.\n"); |
ebfd146a IR |
2774 | return false; |
2775 | } | |
b690cc0f RG |
2776 | if (!rhs_type) |
2777 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 2778 | |
81c40241 | 2779 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[i], &opvectype)) |
ebfd146a | 2780 | { |
73fbfcad | 2781 | if (dump_enabled_p ()) |
78c60e3d | 2782 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2783 | "use not simple.\n"); |
ebfd146a IR |
2784 | return false; |
2785 | } | |
ebfd146a | 2786 | |
b690cc0f RG |
2787 | if (!vectype_in) |
2788 | vectype_in = opvectype; | |
2789 | else if (opvectype | |
2790 | && opvectype != vectype_in) | |
2791 | { | |
73fbfcad | 2792 | if (dump_enabled_p ()) |
78c60e3d | 2793 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2794 | "argument vector types differ.\n"); |
b690cc0f RG |
2795 | return false; |
2796 | } | |
2797 | } | |
2798 | /* If all arguments are external or constant defs use a vector type with | |
2799 | the same size as the output vector type. */ | |
ebfd146a | 2800 | if (!vectype_in) |
b690cc0f | 2801 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
2802 | if (vec_stmt) |
2803 | gcc_assert (vectype_in); | |
2804 | if (!vectype_in) | |
2805 | { | |
73fbfcad | 2806 | if (dump_enabled_p ()) |
7d8930a0 | 2807 | { |
78c60e3d SS |
2808 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2809 | "no vectype for scalar type "); | |
2810 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 2811 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
2812 | } |
2813 | ||
2814 | return false; | |
2815 | } | |
ebfd146a IR |
2816 | |
2817 | /* FORNOW */ | |
b690cc0f RG |
2818 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
2819 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
c7bda0f4 | 2820 | if (known_eq (nunits_in * 2, nunits_out)) |
ebfd146a | 2821 | modifier = NARROW; |
c7bda0f4 | 2822 | else if (known_eq (nunits_out, nunits_in)) |
ebfd146a | 2823 | modifier = NONE; |
c7bda0f4 | 2824 | else if (known_eq (nunits_out * 2, nunits_in)) |
ebfd146a IR |
2825 | modifier = WIDEN; |
2826 | else | |
2827 | return false; | |
2828 | ||
70439f0d RS |
2829 | /* We only handle functions that do not read or clobber memory. */ |
2830 | if (gimple_vuse (stmt)) | |
2831 | { | |
2832 | if (dump_enabled_p ()) | |
2833 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2834 | "function reads from or writes to memory.\n"); | |
2835 | return false; | |
2836 | } | |
2837 | ||
ebfd146a IR |
2838 | /* For now, we only vectorize functions if a target specific builtin |
2839 | is available. TODO -- in some cases, it might be profitable to | |
2840 | insert the calls for pieces of the vector, in order to be able | |
2841 | to vectorize other operations in the loop. */ | |
70439f0d RS |
2842 | fndecl = NULL_TREE; |
2843 | internal_fn ifn = IFN_LAST; | |
2844 | combined_fn cfn = gimple_call_combined_fn (stmt); | |
2845 | tree callee = gimple_call_fndecl (stmt); | |
2846 | ||
2847 | /* First try using an internal function. */ | |
b1b6836e RS |
2848 | tree_code convert_code = ERROR_MARK; |
2849 | if (cfn != CFN_LAST | |
2850 | && (modifier == NONE | |
2851 | || (modifier == NARROW | |
2852 | && simple_integer_narrowing (vectype_out, vectype_in, | |
2853 | &convert_code)))) | |
70439f0d RS |
2854 | ifn = vectorizable_internal_function (cfn, callee, vectype_out, |
2855 | vectype_in); | |
2856 | ||
2857 | /* If that fails, try asking for a target-specific built-in function. */ | |
2858 | if (ifn == IFN_LAST) | |
2859 | { | |
2860 | if (cfn != CFN_LAST) | |
2861 | fndecl = targetm.vectorize.builtin_vectorized_function | |
2862 | (cfn, vectype_out, vectype_in); | |
2863 | else | |
2864 | fndecl = targetm.vectorize.builtin_md_vectorized_function | |
2865 | (callee, vectype_out, vectype_in); | |
2866 | } | |
2867 | ||
2868 | if (ifn == IFN_LAST && !fndecl) | |
ebfd146a | 2869 | { |
70439f0d | 2870 | if (cfn == CFN_GOMP_SIMD_LANE |
74bf76ed JJ |
2871 | && !slp_node |
2872 | && loop_vinfo | |
2873 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2874 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
2875 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2876 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
2877 | { | |
2878 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
2879 | { 0, 1, 2, ... vf - 1 } vector. */ | |
2880 | gcc_assert (nargs == 0); | |
2881 | } | |
37b14185 RB |
2882 | else if (modifier == NONE |
2883 | && (gimple_call_builtin_p (stmt, BUILT_IN_BSWAP16) | |
2884 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP32) | |
2885 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP64))) | |
2886 | return vectorizable_bswap (stmt, gsi, vec_stmt, slp_node, | |
2887 | vectype_in, dt); | |
74bf76ed JJ |
2888 | else |
2889 | { | |
2890 | if (dump_enabled_p ()) | |
2891 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 2892 | "function is not vectorizable.\n"); |
74bf76ed JJ |
2893 | return false; |
2894 | } | |
ebfd146a IR |
2895 | } |
2896 | ||
fce57248 | 2897 | if (slp_node) |
190c2236 | 2898 | ncopies = 1; |
b1b6836e | 2899 | else if (modifier == NARROW && ifn == IFN_LAST) |
e8f142e2 | 2900 | ncopies = vect_get_num_copies (loop_vinfo, vectype_out); |
ebfd146a | 2901 | else |
e8f142e2 | 2902 | ncopies = vect_get_num_copies (loop_vinfo, vectype_in); |
ebfd146a IR |
2903 | |
2904 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
2905 | needs to be generated. */ | |
2906 | gcc_assert (ncopies >= 1); | |
2907 | ||
2908 | if (!vec_stmt) /* transformation not required. */ | |
2909 | { | |
2910 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 2911 | if (dump_enabled_p ()) |
e645e942 TJ |
2912 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ===" |
2913 | "\n"); | |
4fc5ebf1 | 2914 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); |
b1b6836e RS |
2915 | if (ifn != IFN_LAST && modifier == NARROW && !slp_node) |
2916 | add_stmt_cost (stmt_info->vinfo->target_cost_data, ncopies / 2, | |
2917 | vec_promote_demote, stmt_info, 0, vect_body); | |
2918 | ||
ebfd146a IR |
2919 | return true; |
2920 | } | |
2921 | ||
67b8dbac | 2922 | /* Transform. */ |
ebfd146a | 2923 | |
73fbfcad | 2924 | if (dump_enabled_p ()) |
e645e942 | 2925 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
2926 | |
2927 | /* Handle def. */ | |
2928 | scalar_dest = gimple_call_lhs (stmt); | |
2929 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
2930 | ||
2931 | prev_stmt_info = NULL; | |
b1b6836e | 2932 | if (modifier == NONE || ifn != IFN_LAST) |
ebfd146a | 2933 | { |
b1b6836e | 2934 | tree prev_res = NULL_TREE; |
ebfd146a IR |
2935 | for (j = 0; j < ncopies; ++j) |
2936 | { | |
2937 | /* Build argument list for the vectorized call. */ | |
2938 | if (j == 0) | |
9771b263 | 2939 | vargs.create (nargs); |
ebfd146a | 2940 | else |
9771b263 | 2941 | vargs.truncate (0); |
ebfd146a | 2942 | |
190c2236 JJ |
2943 | if (slp_node) |
2944 | { | |
ef062b13 | 2945 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2946 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2947 | |
2948 | for (i = 0; i < nargs; i++) | |
9771b263 | 2949 | vargs.quick_push (gimple_call_arg (stmt, i)); |
306b0c92 | 2950 | vect_get_slp_defs (vargs, slp_node, &vec_defs); |
37b5ec8f | 2951 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2952 | |
2953 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2954 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
2955 | { |
2956 | size_t k; | |
2957 | for (k = 0; k < nargs; k++) | |
2958 | { | |
37b5ec8f | 2959 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 2960 | vargs[k] = vec_oprndsk[i]; |
190c2236 | 2961 | } |
b1b6836e RS |
2962 | if (modifier == NARROW) |
2963 | { | |
2964 | tree half_res = make_ssa_name (vectype_in); | |
a844293d RS |
2965 | gcall *call |
2966 | = gimple_build_call_internal_vec (ifn, vargs); | |
2967 | gimple_call_set_lhs (call, half_res); | |
2968 | gimple_call_set_nothrow (call, true); | |
2969 | new_stmt = call; | |
b1b6836e RS |
2970 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2971 | if ((i & 1) == 0) | |
2972 | { | |
2973 | prev_res = half_res; | |
2974 | continue; | |
2975 | } | |
2976 | new_temp = make_ssa_name (vec_dest); | |
2977 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
2978 | prev_res, half_res); | |
2979 | } | |
70439f0d | 2980 | else |
b1b6836e | 2981 | { |
a844293d | 2982 | gcall *call; |
b1b6836e | 2983 | if (ifn != IFN_LAST) |
a844293d | 2984 | call = gimple_build_call_internal_vec (ifn, vargs); |
b1b6836e | 2985 | else |
a844293d RS |
2986 | call = gimple_build_call_vec (fndecl, vargs); |
2987 | new_temp = make_ssa_name (vec_dest, call); | |
2988 | gimple_call_set_lhs (call, new_temp); | |
2989 | gimple_call_set_nothrow (call, true); | |
2990 | new_stmt = call; | |
b1b6836e | 2991 | } |
190c2236 | 2992 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 2993 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2994 | } |
2995 | ||
2996 | for (i = 0; i < nargs; i++) | |
2997 | { | |
37b5ec8f | 2998 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2999 | vec_oprndsi.release (); |
190c2236 | 3000 | } |
190c2236 JJ |
3001 | continue; |
3002 | } | |
3003 | ||
ebfd146a IR |
3004 | for (i = 0; i < nargs; i++) |
3005 | { | |
3006 | op = gimple_call_arg (stmt, i); | |
3007 | if (j == 0) | |
3008 | vec_oprnd0 | |
81c40241 | 3009 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 3010 | else |
63827fb8 IR |
3011 | { |
3012 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
3013 | vec_oprnd0 | |
3014 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
3015 | } | |
ebfd146a | 3016 | |
9771b263 | 3017 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
3018 | } |
3019 | ||
74bf76ed JJ |
3020 | if (gimple_call_internal_p (stmt) |
3021 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
3022 | { | |
c7bda0f4 | 3023 | tree cst = build_index_vector (vectype_out, j * nunits_out, 1); |
74bf76ed | 3024 | tree new_var |
0e22bb5a | 3025 | = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_"); |
355fe088 | 3026 | gimple *init_stmt = gimple_build_assign (new_var, cst); |
74bf76ed | 3027 | vect_init_vector_1 (stmt, init_stmt, NULL); |
b731b390 | 3028 | new_temp = make_ssa_name (vec_dest); |
0e22bb5a | 3029 | new_stmt = gimple_build_assign (new_temp, new_var); |
74bf76ed | 3030 | } |
b1b6836e RS |
3031 | else if (modifier == NARROW) |
3032 | { | |
3033 | tree half_res = make_ssa_name (vectype_in); | |
a844293d RS |
3034 | gcall *call = gimple_build_call_internal_vec (ifn, vargs); |
3035 | gimple_call_set_lhs (call, half_res); | |
3036 | gimple_call_set_nothrow (call, true); | |
3037 | new_stmt = call; | |
b1b6836e RS |
3038 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3039 | if ((j & 1) == 0) | |
3040 | { | |
3041 | prev_res = half_res; | |
3042 | continue; | |
3043 | } | |
3044 | new_temp = make_ssa_name (vec_dest); | |
3045 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
3046 | prev_res, half_res); | |
3047 | } | |
74bf76ed JJ |
3048 | else |
3049 | { | |
a844293d | 3050 | gcall *call; |
70439f0d | 3051 | if (ifn != IFN_LAST) |
a844293d | 3052 | call = gimple_build_call_internal_vec (ifn, vargs); |
70439f0d | 3053 | else |
a844293d | 3054 | call = gimple_build_call_vec (fndecl, vargs); |
74bf76ed | 3055 | new_temp = make_ssa_name (vec_dest, new_stmt); |
a844293d RS |
3056 | gimple_call_set_lhs (call, new_temp); |
3057 | gimple_call_set_nothrow (call, true); | |
3058 | new_stmt = call; | |
74bf76ed | 3059 | } |
ebfd146a IR |
3060 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3061 | ||
b1b6836e | 3062 | if (j == (modifier == NARROW ? 1 : 0)) |
ebfd146a IR |
3063 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = 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 | } | |
b1b6836e RS |
3069 | } |
3070 | else if (modifier == NARROW) | |
3071 | { | |
ebfd146a IR |
3072 | for (j = 0; j < ncopies; ++j) |
3073 | { | |
3074 | /* Build argument list for the vectorized call. */ | |
3075 | if (j == 0) | |
9771b263 | 3076 | vargs.create (nargs * 2); |
ebfd146a | 3077 | else |
9771b263 | 3078 | vargs.truncate (0); |
ebfd146a | 3079 | |
190c2236 JJ |
3080 | if (slp_node) |
3081 | { | |
ef062b13 | 3082 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 3083 | vec<tree> vec_oprnds0; |
190c2236 JJ |
3084 | |
3085 | for (i = 0; i < nargs; i++) | |
9771b263 | 3086 | vargs.quick_push (gimple_call_arg (stmt, i)); |
306b0c92 | 3087 | vect_get_slp_defs (vargs, slp_node, &vec_defs); |
37b5ec8f | 3088 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
3089 | |
3090 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3091 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
3092 | { |
3093 | size_t k; | |
9771b263 | 3094 | vargs.truncate (0); |
190c2236 JJ |
3095 | for (k = 0; k < nargs; k++) |
3096 | { | |
37b5ec8f | 3097 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
3098 | vargs.quick_push (vec_oprndsk[i]); |
3099 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 | 3100 | } |
a844293d | 3101 | gcall *call; |
70439f0d | 3102 | if (ifn != IFN_LAST) |
a844293d | 3103 | call = gimple_build_call_internal_vec (ifn, vargs); |
70439f0d | 3104 | else |
a844293d RS |
3105 | call = gimple_build_call_vec (fndecl, vargs); |
3106 | new_temp = make_ssa_name (vec_dest, call); | |
3107 | gimple_call_set_lhs (call, new_temp); | |
3108 | gimple_call_set_nothrow (call, true); | |
3109 | new_stmt = call; | |
190c2236 | 3110 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 3111 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
3112 | } |
3113 | ||
3114 | for (i = 0; i < nargs; i++) | |
3115 | { | |
37b5ec8f | 3116 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 3117 | vec_oprndsi.release (); |
190c2236 | 3118 | } |
190c2236 JJ |
3119 | continue; |
3120 | } | |
3121 | ||
ebfd146a IR |
3122 | for (i = 0; i < nargs; i++) |
3123 | { | |
3124 | op = gimple_call_arg (stmt, i); | |
3125 | if (j == 0) | |
3126 | { | |
3127 | vec_oprnd0 | |
81c40241 | 3128 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 3129 | vec_oprnd1 |
63827fb8 | 3130 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
3131 | } |
3132 | else | |
3133 | { | |
336ecb65 | 3134 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 3135 | vec_oprnd0 |
63827fb8 | 3136 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 3137 | vec_oprnd1 |
63827fb8 | 3138 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
3139 | } |
3140 | ||
9771b263 DN |
3141 | vargs.quick_push (vec_oprnd0); |
3142 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
3143 | } |
3144 | ||
b1b6836e | 3145 | new_stmt = gimple_build_call_vec (fndecl, vargs); |
ebfd146a IR |
3146 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3147 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
3148 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3149 | ||
3150 | if (j == 0) | |
3151 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3152 | else | |
3153 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3154 | ||
3155 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3156 | } | |
3157 | ||
3158 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a | 3159 | } |
b1b6836e RS |
3160 | else |
3161 | /* No current target implements this case. */ | |
3162 | return false; | |
ebfd146a | 3163 | |
9771b263 | 3164 | vargs.release (); |
ebfd146a | 3165 | |
ebfd146a IR |
3166 | /* The call in STMT might prevent it from being removed in dce. |
3167 | We however cannot remove it here, due to the way the ssa name | |
3168 | it defines is mapped to the new definition. So just replace | |
3169 | rhs of the statement with something harmless. */ | |
3170 | ||
dd34c087 JJ |
3171 | if (slp_node) |
3172 | return true; | |
3173 | ||
ebfd146a | 3174 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
3175 | if (is_pattern_stmt_p (stmt_info)) |
3176 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
3177 | else | |
3178 | lhs = gimple_call_lhs (stmt); | |
3cc2fa2a | 3179 | |
9d5e7640 | 3180 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); |
ebfd146a | 3181 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 3182 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
3183 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
3184 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
3185 | |
3186 | return true; | |
3187 | } | |
3188 | ||
3189 | ||
0136f8f0 AH |
3190 | struct simd_call_arg_info |
3191 | { | |
3192 | tree vectype; | |
3193 | tree op; | |
0136f8f0 | 3194 | HOST_WIDE_INT linear_step; |
34e82342 | 3195 | enum vect_def_type dt; |
0136f8f0 | 3196 | unsigned int align; |
17b658af | 3197 | bool simd_lane_linear; |
0136f8f0 AH |
3198 | }; |
3199 | ||
17b658af JJ |
3200 | /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME, |
3201 | is linear within simd lane (but not within whole loop), note it in | |
3202 | *ARGINFO. */ | |
3203 | ||
3204 | static void | |
3205 | vect_simd_lane_linear (tree op, struct loop *loop, | |
3206 | struct simd_call_arg_info *arginfo) | |
3207 | { | |
355fe088 | 3208 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
17b658af JJ |
3209 | |
3210 | if (!is_gimple_assign (def_stmt) | |
3211 | || gimple_assign_rhs_code (def_stmt) != POINTER_PLUS_EXPR | |
3212 | || !is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))) | |
3213 | return; | |
3214 | ||
3215 | tree base = gimple_assign_rhs1 (def_stmt); | |
3216 | HOST_WIDE_INT linear_step = 0; | |
3217 | tree v = gimple_assign_rhs2 (def_stmt); | |
3218 | while (TREE_CODE (v) == SSA_NAME) | |
3219 | { | |
3220 | tree t; | |
3221 | def_stmt = SSA_NAME_DEF_STMT (v); | |
3222 | if (is_gimple_assign (def_stmt)) | |
3223 | switch (gimple_assign_rhs_code (def_stmt)) | |
3224 | { | |
3225 | case PLUS_EXPR: | |
3226 | t = gimple_assign_rhs2 (def_stmt); | |
3227 | if (linear_step || TREE_CODE (t) != INTEGER_CST) | |
3228 | return; | |
3229 | base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t); | |
3230 | v = gimple_assign_rhs1 (def_stmt); | |
3231 | continue; | |
3232 | case MULT_EXPR: | |
3233 | t = gimple_assign_rhs2 (def_stmt); | |
3234 | if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t)) | |
3235 | return; | |
3236 | linear_step = tree_to_shwi (t); | |
3237 | v = gimple_assign_rhs1 (def_stmt); | |
3238 | continue; | |
3239 | CASE_CONVERT: | |
3240 | t = gimple_assign_rhs1 (def_stmt); | |
3241 | if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE | |
3242 | || (TYPE_PRECISION (TREE_TYPE (v)) | |
3243 | < TYPE_PRECISION (TREE_TYPE (t)))) | |
3244 | return; | |
3245 | if (!linear_step) | |
3246 | linear_step = 1; | |
3247 | v = t; | |
3248 | continue; | |
3249 | default: | |
3250 | return; | |
3251 | } | |
8e4284d0 | 3252 | else if (gimple_call_internal_p (def_stmt, IFN_GOMP_SIMD_LANE) |
17b658af JJ |
3253 | && loop->simduid |
3254 | && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME | |
3255 | && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0)) | |
3256 | == loop->simduid)) | |
3257 | { | |
3258 | if (!linear_step) | |
3259 | linear_step = 1; | |
3260 | arginfo->linear_step = linear_step; | |
3261 | arginfo->op = base; | |
3262 | arginfo->simd_lane_linear = true; | |
3263 | return; | |
3264 | } | |
3265 | } | |
3266 | } | |
3267 | ||
cf1b2ba4 RS |
3268 | /* Return the number of elements in vector type VECTYPE, which is associated |
3269 | with a SIMD clone. At present these vectors always have a constant | |
3270 | length. */ | |
3271 | ||
3272 | static unsigned HOST_WIDE_INT | |
3273 | simd_clone_subparts (tree vectype) | |
3274 | { | |
928686b1 | 3275 | return TYPE_VECTOR_SUBPARTS (vectype).to_constant (); |
cf1b2ba4 RS |
3276 | } |
3277 | ||
0136f8f0 AH |
3278 | /* Function vectorizable_simd_clone_call. |
3279 | ||
3280 | Check if STMT performs a function call that can be vectorized | |
3281 | by calling a simd clone of the function. | |
3282 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3283 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3284 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3285 | ||
3286 | static bool | |
355fe088 TS |
3287 | vectorizable_simd_clone_call (gimple *stmt, gimple_stmt_iterator *gsi, |
3288 | gimple **vec_stmt, slp_tree slp_node) | |
0136f8f0 AH |
3289 | { |
3290 | tree vec_dest; | |
3291 | tree scalar_dest; | |
3292 | tree op, type; | |
3293 | tree vec_oprnd0 = NULL_TREE; | |
3294 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
3295 | tree vectype; | |
3296 | unsigned int nunits; | |
3297 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3298 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
310213d4 | 3299 | vec_info *vinfo = stmt_info->vinfo; |
0136f8f0 | 3300 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
81c40241 | 3301 | tree fndecl, new_temp; |
355fe088 TS |
3302 | gimple *def_stmt; |
3303 | gimple *new_stmt = NULL; | |
0136f8f0 | 3304 | int ncopies, j; |
00426f9a | 3305 | auto_vec<simd_call_arg_info> arginfo; |
0136f8f0 AH |
3306 | vec<tree> vargs = vNULL; |
3307 | size_t i, nargs; | |
3308 | tree lhs, rtype, ratype; | |
e7a74006 | 3309 | vec<constructor_elt, va_gc> *ret_ctor_elts = NULL; |
0136f8f0 AH |
3310 | |
3311 | /* Is STMT a vectorizable call? */ | |
3312 | if (!is_gimple_call (stmt)) | |
3313 | return false; | |
3314 | ||
3315 | fndecl = gimple_call_fndecl (stmt); | |
3316 | if (fndecl == NULL_TREE) | |
3317 | return false; | |
3318 | ||
d52f5295 | 3319 | struct cgraph_node *node = cgraph_node::get (fndecl); |
0136f8f0 AH |
3320 | if (node == NULL || node->simd_clones == NULL) |
3321 | return false; | |
3322 | ||
3323 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
3324 | return false; | |
3325 | ||
66c16fd9 RB |
3326 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3327 | && ! vec_stmt) | |
0136f8f0 AH |
3328 | return false; |
3329 | ||
3330 | if (gimple_call_lhs (stmt) | |
3331 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
3332 | return false; | |
3333 | ||
3334 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); | |
3335 | ||
3336 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3337 | ||
3338 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt)) | |
3339 | return false; | |
3340 | ||
3341 | /* FORNOW */ | |
fce57248 | 3342 | if (slp_node) |
0136f8f0 AH |
3343 | return false; |
3344 | ||
3345 | /* Process function arguments. */ | |
3346 | nargs = gimple_call_num_args (stmt); | |
3347 | ||
3348 | /* Bail out if the function has zero arguments. */ | |
3349 | if (nargs == 0) | |
3350 | return false; | |
3351 | ||
00426f9a | 3352 | arginfo.reserve (nargs, true); |
0136f8f0 AH |
3353 | |
3354 | for (i = 0; i < nargs; i++) | |
3355 | { | |
3356 | simd_call_arg_info thisarginfo; | |
3357 | affine_iv iv; | |
3358 | ||
3359 | thisarginfo.linear_step = 0; | |
3360 | thisarginfo.align = 0; | |
3361 | thisarginfo.op = NULL_TREE; | |
17b658af | 3362 | thisarginfo.simd_lane_linear = false; |
0136f8f0 AH |
3363 | |
3364 | op = gimple_call_arg (stmt, i); | |
81c40241 RB |
3365 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &thisarginfo.dt, |
3366 | &thisarginfo.vectype) | |
0136f8f0 AH |
3367 | || thisarginfo.dt == vect_uninitialized_def) |
3368 | { | |
3369 | if (dump_enabled_p ()) | |
3370 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3371 | "use not simple.\n"); | |
0136f8f0 AH |
3372 | return false; |
3373 | } | |
3374 | ||
3375 | if (thisarginfo.dt == vect_constant_def | |
3376 | || thisarginfo.dt == vect_external_def) | |
3377 | gcc_assert (thisarginfo.vectype == NULL_TREE); | |
3378 | else | |
3379 | gcc_assert (thisarginfo.vectype != NULL_TREE); | |
3380 | ||
6c9e85fb JJ |
3381 | /* For linear arguments, the analyze phase should have saved |
3382 | the base and step in STMT_VINFO_SIMD_CLONE_INFO. */ | |
17b658af JJ |
3383 | if (i * 3 + 4 <= STMT_VINFO_SIMD_CLONE_INFO (stmt_info).length () |
3384 | && STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]) | |
6c9e85fb JJ |
3385 | { |
3386 | gcc_assert (vec_stmt); | |
3387 | thisarginfo.linear_step | |
17b658af | 3388 | = tree_to_shwi (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]); |
6c9e85fb | 3389 | thisarginfo.op |
17b658af JJ |
3390 | = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 1]; |
3391 | thisarginfo.simd_lane_linear | |
3392 | = (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 3] | |
3393 | == boolean_true_node); | |
6c9e85fb JJ |
3394 | /* If loop has been peeled for alignment, we need to adjust it. */ |
3395 | tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo); | |
3396 | tree n2 = LOOP_VINFO_NITERS (loop_vinfo); | |
17b658af | 3397 | if (n1 != n2 && !thisarginfo.simd_lane_linear) |
6c9e85fb JJ |
3398 | { |
3399 | tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2); | |
17b658af | 3400 | tree step = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]; |
6c9e85fb JJ |
3401 | tree opt = TREE_TYPE (thisarginfo.op); |
3402 | bias = fold_convert (TREE_TYPE (step), bias); | |
3403 | bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step); | |
3404 | thisarginfo.op | |
3405 | = fold_build2 (POINTER_TYPE_P (opt) | |
3406 | ? POINTER_PLUS_EXPR : PLUS_EXPR, opt, | |
3407 | thisarginfo.op, bias); | |
3408 | } | |
3409 | } | |
3410 | else if (!vec_stmt | |
3411 | && thisarginfo.dt != vect_constant_def | |
3412 | && thisarginfo.dt != vect_external_def | |
3413 | && loop_vinfo | |
3414 | && TREE_CODE (op) == SSA_NAME | |
3415 | && simple_iv (loop, loop_containing_stmt (stmt), op, | |
3416 | &iv, false) | |
3417 | && tree_fits_shwi_p (iv.step)) | |
0136f8f0 AH |
3418 | { |
3419 | thisarginfo.linear_step = tree_to_shwi (iv.step); | |
3420 | thisarginfo.op = iv.base; | |
3421 | } | |
3422 | else if ((thisarginfo.dt == vect_constant_def | |
3423 | || thisarginfo.dt == vect_external_def) | |
3424 | && POINTER_TYPE_P (TREE_TYPE (op))) | |
3425 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; | |
17b658af JJ |
3426 | /* Addresses of array elements indexed by GOMP_SIMD_LANE are |
3427 | linear too. */ | |
3428 | if (POINTER_TYPE_P (TREE_TYPE (op)) | |
3429 | && !thisarginfo.linear_step | |
3430 | && !vec_stmt | |
3431 | && thisarginfo.dt != vect_constant_def | |
3432 | && thisarginfo.dt != vect_external_def | |
3433 | && loop_vinfo | |
3434 | && !slp_node | |
3435 | && TREE_CODE (op) == SSA_NAME) | |
3436 | vect_simd_lane_linear (op, loop, &thisarginfo); | |
0136f8f0 AH |
3437 | |
3438 | arginfo.quick_push (thisarginfo); | |
3439 | } | |
3440 | ||
d9f21f6a RS |
3441 | unsigned HOST_WIDE_INT vf; |
3442 | if (!LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (&vf)) | |
3443 | { | |
3444 | if (dump_enabled_p ()) | |
3445 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3446 | "not considering SIMD clones; not yet supported" | |
3447 | " for variable-width vectors.\n"); | |
3448 | return NULL; | |
3449 | } | |
3450 | ||
0136f8f0 AH |
3451 | unsigned int badness = 0; |
3452 | struct cgraph_node *bestn = NULL; | |
6c9e85fb JJ |
3453 | if (STMT_VINFO_SIMD_CLONE_INFO (stmt_info).exists ()) |
3454 | bestn = cgraph_node::get (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[0]); | |
0136f8f0 AH |
3455 | else |
3456 | for (struct cgraph_node *n = node->simd_clones; n != NULL; | |
3457 | n = n->simdclone->next_clone) | |
3458 | { | |
3459 | unsigned int this_badness = 0; | |
d9f21f6a | 3460 | if (n->simdclone->simdlen > vf |
0136f8f0 AH |
3461 | || n->simdclone->nargs != nargs) |
3462 | continue; | |
d9f21f6a RS |
3463 | if (n->simdclone->simdlen < vf) |
3464 | this_badness += (exact_log2 (vf) | |
0136f8f0 AH |
3465 | - exact_log2 (n->simdclone->simdlen)) * 1024; |
3466 | if (n->simdclone->inbranch) | |
3467 | this_badness += 2048; | |
3468 | int target_badness = targetm.simd_clone.usable (n); | |
3469 | if (target_badness < 0) | |
3470 | continue; | |
3471 | this_badness += target_badness * 512; | |
3472 | /* FORNOW: Have to add code to add the mask argument. */ | |
3473 | if (n->simdclone->inbranch) | |
3474 | continue; | |
3475 | for (i = 0; i < nargs; i++) | |
3476 | { | |
3477 | switch (n->simdclone->args[i].arg_type) | |
3478 | { | |
3479 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3480 | if (!useless_type_conversion_p | |
3481 | (n->simdclone->args[i].orig_type, | |
3482 | TREE_TYPE (gimple_call_arg (stmt, i)))) | |
3483 | i = -1; | |
3484 | else if (arginfo[i].dt == vect_constant_def | |
3485 | || arginfo[i].dt == vect_external_def | |
3486 | || arginfo[i].linear_step) | |
3487 | this_badness += 64; | |
3488 | break; | |
3489 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3490 | if (arginfo[i].dt != vect_constant_def | |
3491 | && arginfo[i].dt != vect_external_def) | |
3492 | i = -1; | |
3493 | break; | |
3494 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
d9a6bd32 | 3495 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
3496 | if (arginfo[i].dt == vect_constant_def |
3497 | || arginfo[i].dt == vect_external_def | |
3498 | || (arginfo[i].linear_step | |
3499 | != n->simdclone->args[i].linear_step)) | |
3500 | i = -1; | |
3501 | break; | |
3502 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
d9a6bd32 JJ |
3503 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
3504 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
e01d41e5 JJ |
3505 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3506 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3507 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3508 | /* FORNOW */ |
3509 | i = -1; | |
3510 | break; | |
3511 | case SIMD_CLONE_ARG_TYPE_MASK: | |
3512 | gcc_unreachable (); | |
3513 | } | |
3514 | if (i == (size_t) -1) | |
3515 | break; | |
3516 | if (n->simdclone->args[i].alignment > arginfo[i].align) | |
3517 | { | |
3518 | i = -1; | |
3519 | break; | |
3520 | } | |
3521 | if (arginfo[i].align) | |
3522 | this_badness += (exact_log2 (arginfo[i].align) | |
3523 | - exact_log2 (n->simdclone->args[i].alignment)); | |
3524 | } | |
3525 | if (i == (size_t) -1) | |
3526 | continue; | |
3527 | if (bestn == NULL || this_badness < badness) | |
3528 | { | |
3529 | bestn = n; | |
3530 | badness = this_badness; | |
3531 | } | |
3532 | } | |
3533 | ||
3534 | if (bestn == NULL) | |
00426f9a | 3535 | return false; |
0136f8f0 AH |
3536 | |
3537 | for (i = 0; i < nargs; i++) | |
3538 | if ((arginfo[i].dt == vect_constant_def | |
3539 | || arginfo[i].dt == vect_external_def) | |
3540 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
3541 | { | |
3542 | arginfo[i].vectype | |
3543 | = get_vectype_for_scalar_type (TREE_TYPE (gimple_call_arg (stmt, | |
3544 | i))); | |
3545 | if (arginfo[i].vectype == NULL | |
cf1b2ba4 | 3546 | || (simd_clone_subparts (arginfo[i].vectype) |
0136f8f0 | 3547 | > bestn->simdclone->simdlen)) |
00426f9a | 3548 | return false; |
0136f8f0 AH |
3549 | } |
3550 | ||
3551 | fndecl = bestn->decl; | |
3552 | nunits = bestn->simdclone->simdlen; | |
d9f21f6a | 3553 | ncopies = vf / nunits; |
0136f8f0 AH |
3554 | |
3555 | /* If the function isn't const, only allow it in simd loops where user | |
3556 | has asserted that at least nunits consecutive iterations can be | |
3557 | performed using SIMD instructions. */ | |
3558 | if ((loop == NULL || (unsigned) loop->safelen < nunits) | |
3559 | && gimple_vuse (stmt)) | |
00426f9a | 3560 | return false; |
0136f8f0 AH |
3561 | |
3562 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
3563 | needs to be generated. */ | |
3564 | gcc_assert (ncopies >= 1); | |
3565 | ||
3566 | if (!vec_stmt) /* transformation not required. */ | |
3567 | { | |
6c9e85fb JJ |
3568 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (bestn->decl); |
3569 | for (i = 0; i < nargs; i++) | |
7adb26f2 JJ |
3570 | if ((bestn->simdclone->args[i].arg_type |
3571 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
3572 | || (bestn->simdclone->args[i].arg_type | |
3573 | == SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP)) | |
6c9e85fb | 3574 | { |
17b658af | 3575 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_grow_cleared (i * 3 |
6c9e85fb JJ |
3576 | + 1); |
3577 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (arginfo[i].op); | |
3578 | tree lst = POINTER_TYPE_P (TREE_TYPE (arginfo[i].op)) | |
3579 | ? size_type_node : TREE_TYPE (arginfo[i].op); | |
3580 | tree ls = build_int_cst (lst, arginfo[i].linear_step); | |
3581 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (ls); | |
17b658af JJ |
3582 | tree sll = arginfo[i].simd_lane_linear |
3583 | ? boolean_true_node : boolean_false_node; | |
3584 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (sll); | |
6c9e85fb | 3585 | } |
0136f8f0 AH |
3586 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; |
3587 | if (dump_enabled_p ()) | |
3588 | dump_printf_loc (MSG_NOTE, vect_location, | |
3589 | "=== vectorizable_simd_clone_call ===\n"); | |
3590 | /* vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); */ | |
0136f8f0 AH |
3591 | return true; |
3592 | } | |
3593 | ||
67b8dbac | 3594 | /* Transform. */ |
0136f8f0 AH |
3595 | |
3596 | if (dump_enabled_p ()) | |
3597 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); | |
3598 | ||
3599 | /* Handle def. */ | |
3600 | scalar_dest = gimple_call_lhs (stmt); | |
3601 | vec_dest = NULL_TREE; | |
3602 | rtype = NULL_TREE; | |
3603 | ratype = NULL_TREE; | |
3604 | if (scalar_dest) | |
3605 | { | |
3606 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3607 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); | |
3608 | if (TREE_CODE (rtype) == ARRAY_TYPE) | |
3609 | { | |
3610 | ratype = rtype; | |
3611 | rtype = TREE_TYPE (ratype); | |
3612 | } | |
3613 | } | |
3614 | ||
3615 | prev_stmt_info = NULL; | |
3616 | for (j = 0; j < ncopies; ++j) | |
3617 | { | |
3618 | /* Build argument list for the vectorized call. */ | |
3619 | if (j == 0) | |
3620 | vargs.create (nargs); | |
3621 | else | |
3622 | vargs.truncate (0); | |
3623 | ||
3624 | for (i = 0; i < nargs; i++) | |
3625 | { | |
3626 | unsigned int k, l, m, o; | |
3627 | tree atype; | |
3628 | op = gimple_call_arg (stmt, i); | |
3629 | switch (bestn->simdclone->args[i].arg_type) | |
3630 | { | |
3631 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3632 | atype = bestn->simdclone->args[i].vector_type; | |
cf1b2ba4 | 3633 | o = nunits / simd_clone_subparts (atype); |
0136f8f0 AH |
3634 | for (m = j * o; m < (j + 1) * o; m++) |
3635 | { | |
cf1b2ba4 RS |
3636 | if (simd_clone_subparts (atype) |
3637 | < simd_clone_subparts (arginfo[i].vectype)) | |
0136f8f0 | 3638 | { |
73a699ae | 3639 | poly_uint64 prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); |
cf1b2ba4 RS |
3640 | k = (simd_clone_subparts (arginfo[i].vectype) |
3641 | / simd_clone_subparts (atype)); | |
0136f8f0 AH |
3642 | gcc_assert ((k & (k - 1)) == 0); |
3643 | if (m == 0) | |
3644 | vec_oprnd0 | |
81c40241 | 3645 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3646 | else |
3647 | { | |
3648 | vec_oprnd0 = arginfo[i].op; | |
3649 | if ((m & (k - 1)) == 0) | |
3650 | vec_oprnd0 | |
3651 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3652 | vec_oprnd0); | |
3653 | } | |
3654 | arginfo[i].op = vec_oprnd0; | |
3655 | vec_oprnd0 | |
3656 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, | |
92e29a5e | 3657 | bitsize_int (prec), |
0136f8f0 AH |
3658 | bitsize_int ((m & (k - 1)) * prec)); |
3659 | new_stmt | |
b731b390 | 3660 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3661 | vec_oprnd0); |
3662 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3663 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3664 | } | |
3665 | else | |
3666 | { | |
cf1b2ba4 RS |
3667 | k = (simd_clone_subparts (atype) |
3668 | / simd_clone_subparts (arginfo[i].vectype)); | |
0136f8f0 AH |
3669 | gcc_assert ((k & (k - 1)) == 0); |
3670 | vec<constructor_elt, va_gc> *ctor_elts; | |
3671 | if (k != 1) | |
3672 | vec_alloc (ctor_elts, k); | |
3673 | else | |
3674 | ctor_elts = NULL; | |
3675 | for (l = 0; l < k; l++) | |
3676 | { | |
3677 | if (m == 0 && l == 0) | |
3678 | vec_oprnd0 | |
81c40241 | 3679 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3680 | else |
3681 | vec_oprnd0 | |
3682 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3683 | arginfo[i].op); | |
3684 | arginfo[i].op = vec_oprnd0; | |
3685 | if (k == 1) | |
3686 | break; | |
3687 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, | |
3688 | vec_oprnd0); | |
3689 | } | |
3690 | if (k == 1) | |
3691 | vargs.safe_push (vec_oprnd0); | |
3692 | else | |
3693 | { | |
3694 | vec_oprnd0 = build_constructor (atype, ctor_elts); | |
3695 | new_stmt | |
b731b390 | 3696 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3697 | vec_oprnd0); |
3698 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3699 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3700 | } | |
3701 | } | |
3702 | } | |
3703 | break; | |
3704 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3705 | vargs.safe_push (op); | |
3706 | break; | |
3707 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
7adb26f2 | 3708 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
3709 | if (j == 0) |
3710 | { | |
3711 | gimple_seq stmts; | |
3712 | arginfo[i].op | |
3713 | = force_gimple_operand (arginfo[i].op, &stmts, true, | |
3714 | NULL_TREE); | |
3715 | if (stmts != NULL) | |
3716 | { | |
3717 | basic_block new_bb; | |
3718 | edge pe = loop_preheader_edge (loop); | |
3719 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
3720 | gcc_assert (!new_bb); | |
3721 | } | |
17b658af JJ |
3722 | if (arginfo[i].simd_lane_linear) |
3723 | { | |
3724 | vargs.safe_push (arginfo[i].op); | |
3725 | break; | |
3726 | } | |
b731b390 | 3727 | tree phi_res = copy_ssa_name (op); |
538dd0b7 | 3728 | gphi *new_phi = create_phi_node (phi_res, loop->header); |
0136f8f0 | 3729 | set_vinfo_for_stmt (new_phi, |
310213d4 | 3730 | new_stmt_vec_info (new_phi, loop_vinfo)); |
0136f8f0 AH |
3731 | add_phi_arg (new_phi, arginfo[i].op, |
3732 | loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
3733 | enum tree_code code | |
3734 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3735 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3736 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3737 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3738 | widest_int cst |
3739 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3740 | ncopies * nunits); | |
3741 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3742 | tree phi_arg = copy_ssa_name (op); |
0d0e4a03 JJ |
3743 | new_stmt |
3744 | = gimple_build_assign (phi_arg, code, phi_res, tcst); | |
0136f8f0 AH |
3745 | gimple_stmt_iterator si = gsi_after_labels (loop->header); |
3746 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); | |
3747 | set_vinfo_for_stmt (new_stmt, | |
310213d4 | 3748 | new_stmt_vec_info (new_stmt, loop_vinfo)); |
0136f8f0 AH |
3749 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), |
3750 | UNKNOWN_LOCATION); | |
3751 | arginfo[i].op = phi_res; | |
3752 | vargs.safe_push (phi_res); | |
3753 | } | |
3754 | else | |
3755 | { | |
3756 | enum tree_code code | |
3757 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3758 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3759 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3760 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3761 | widest_int cst |
3762 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3763 | j * nunits); | |
3764 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3765 | new_temp = make_ssa_name (TREE_TYPE (op)); |
0d0e4a03 JJ |
3766 | new_stmt = gimple_build_assign (new_temp, code, |
3767 | arginfo[i].op, tcst); | |
0136f8f0 AH |
3768 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3769 | vargs.safe_push (new_temp); | |
3770 | } | |
3771 | break; | |
7adb26f2 JJ |
3772 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
3773 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
0136f8f0 | 3774 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: |
e01d41e5 JJ |
3775 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3776 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3777 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3778 | default: |
3779 | gcc_unreachable (); | |
3780 | } | |
3781 | } | |
3782 | ||
3783 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
3784 | if (vec_dest) | |
3785 | { | |
cf1b2ba4 | 3786 | gcc_assert (ratype || simd_clone_subparts (rtype) == nunits); |
0136f8f0 | 3787 | if (ratype) |
b731b390 | 3788 | new_temp = create_tmp_var (ratype); |
cf1b2ba4 RS |
3789 | else if (simd_clone_subparts (vectype) |
3790 | == simd_clone_subparts (rtype)) | |
0136f8f0 AH |
3791 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3792 | else | |
3793 | new_temp = make_ssa_name (rtype, new_stmt); | |
3794 | gimple_call_set_lhs (new_stmt, new_temp); | |
3795 | } | |
3796 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3797 | ||
3798 | if (vec_dest) | |
3799 | { | |
cf1b2ba4 | 3800 | if (simd_clone_subparts (vectype) < nunits) |
0136f8f0 AH |
3801 | { |
3802 | unsigned int k, l; | |
73a699ae RS |
3803 | poly_uint64 prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); |
3804 | poly_uint64 bytes = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
cf1b2ba4 | 3805 | k = nunits / simd_clone_subparts (vectype); |
0136f8f0 AH |
3806 | gcc_assert ((k & (k - 1)) == 0); |
3807 | for (l = 0; l < k; l++) | |
3808 | { | |
3809 | tree t; | |
3810 | if (ratype) | |
3811 | { | |
3812 | t = build_fold_addr_expr (new_temp); | |
3813 | t = build2 (MEM_REF, vectype, t, | |
73a699ae | 3814 | build_int_cst (TREE_TYPE (t), l * bytes)); |
0136f8f0 AH |
3815 | } |
3816 | else | |
3817 | t = build3 (BIT_FIELD_REF, vectype, new_temp, | |
92e29a5e | 3818 | bitsize_int (prec), bitsize_int (l * prec)); |
0136f8f0 | 3819 | new_stmt |
b731b390 | 3820 | = gimple_build_assign (make_ssa_name (vectype), t); |
0136f8f0 AH |
3821 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3822 | if (j == 0 && l == 0) | |
3823 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3824 | else | |
3825 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3826 | ||
3827 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3828 | } | |
3829 | ||
3830 | if (ratype) | |
3831 | { | |
3832 | tree clobber = build_constructor (ratype, NULL); | |
3833 | TREE_THIS_VOLATILE (clobber) = 1; | |
3834 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3835 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3836 | } | |
3837 | continue; | |
3838 | } | |
cf1b2ba4 | 3839 | else if (simd_clone_subparts (vectype) > nunits) |
0136f8f0 | 3840 | { |
cf1b2ba4 RS |
3841 | unsigned int k = (simd_clone_subparts (vectype) |
3842 | / simd_clone_subparts (rtype)); | |
0136f8f0 AH |
3843 | gcc_assert ((k & (k - 1)) == 0); |
3844 | if ((j & (k - 1)) == 0) | |
3845 | vec_alloc (ret_ctor_elts, k); | |
3846 | if (ratype) | |
3847 | { | |
cf1b2ba4 | 3848 | unsigned int m, o = nunits / simd_clone_subparts (rtype); |
0136f8f0 AH |
3849 | for (m = 0; m < o; m++) |
3850 | { | |
3851 | tree tem = build4 (ARRAY_REF, rtype, new_temp, | |
3852 | size_int (m), NULL_TREE, NULL_TREE); | |
3853 | new_stmt | |
b731b390 | 3854 | = gimple_build_assign (make_ssa_name (rtype), tem); |
0136f8f0 AH |
3855 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3856 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, | |
3857 | gimple_assign_lhs (new_stmt)); | |
3858 | } | |
3859 | tree clobber = build_constructor (ratype, NULL); | |
3860 | TREE_THIS_VOLATILE (clobber) = 1; | |
3861 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3862 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3863 | } | |
3864 | else | |
3865 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); | |
3866 | if ((j & (k - 1)) != k - 1) | |
3867 | continue; | |
3868 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); | |
3869 | new_stmt | |
b731b390 | 3870 | = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0); |
0136f8f0 AH |
3871 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3872 | ||
3873 | if ((unsigned) j == k - 1) | |
3874 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3875 | else | |
3876 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3877 | ||
3878 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3879 | continue; | |
3880 | } | |
3881 | else if (ratype) | |
3882 | { | |
3883 | tree t = build_fold_addr_expr (new_temp); | |
3884 | t = build2 (MEM_REF, vectype, t, | |
3885 | build_int_cst (TREE_TYPE (t), 0)); | |
3886 | new_stmt | |
b731b390 | 3887 | = gimple_build_assign (make_ssa_name (vec_dest), t); |
0136f8f0 AH |
3888 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3889 | tree clobber = build_constructor (ratype, NULL); | |
3890 | TREE_THIS_VOLATILE (clobber) = 1; | |
3891 | vect_finish_stmt_generation (stmt, | |
3892 | gimple_build_assign (new_temp, | |
3893 | clobber), gsi); | |
3894 | } | |
3895 | } | |
3896 | ||
3897 | if (j == 0) | |
3898 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3899 | else | |
3900 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3901 | ||
3902 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3903 | } | |
3904 | ||
3905 | vargs.release (); | |
3906 | ||
3907 | /* The call in STMT might prevent it from being removed in dce. | |
3908 | We however cannot remove it here, due to the way the ssa name | |
3909 | it defines is mapped to the new definition. So just replace | |
3910 | rhs of the statement with something harmless. */ | |
3911 | ||
3912 | if (slp_node) | |
3913 | return true; | |
3914 | ||
3915 | if (scalar_dest) | |
3916 | { | |
3917 | type = TREE_TYPE (scalar_dest); | |
3918 | if (is_pattern_stmt_p (stmt_info)) | |
3919 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
3920 | else | |
3921 | lhs = gimple_call_lhs (stmt); | |
3922 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
3923 | } | |
3924 | else | |
3925 | new_stmt = gimple_build_nop (); | |
3926 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
3927 | set_vinfo_for_stmt (stmt, NULL); | |
3928 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2865f32a | 3929 | gsi_replace (gsi, new_stmt, true); |
0136f8f0 AH |
3930 | unlink_stmt_vdef (stmt); |
3931 | ||
3932 | return true; | |
3933 | } | |
3934 | ||
3935 | ||
ebfd146a IR |
3936 | /* Function vect_gen_widened_results_half |
3937 | ||
3938 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 3939 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 3940 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
3941 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
3942 | needs to be created (DECL is a function-decl of a target-builtin). | |
3943 | STMT is the original scalar stmt that we are vectorizing. */ | |
3944 | ||
355fe088 | 3945 | static gimple * |
ebfd146a IR |
3946 | vect_gen_widened_results_half (enum tree_code code, |
3947 | tree decl, | |
3948 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
3949 | tree vec_dest, gimple_stmt_iterator *gsi, | |
355fe088 | 3950 | gimple *stmt) |
b8698a0f | 3951 | { |
355fe088 | 3952 | gimple *new_stmt; |
b8698a0f L |
3953 | tree new_temp; |
3954 | ||
3955 | /* Generate half of the widened result: */ | |
3956 | if (code == CALL_EXPR) | |
3957 | { | |
3958 | /* Target specific support */ | |
ebfd146a IR |
3959 | if (op_type == binary_op) |
3960 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
3961 | else | |
3962 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
3963 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3964 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
3965 | } |
3966 | else | |
ebfd146a | 3967 | { |
b8698a0f L |
3968 | /* Generic support */ |
3969 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
3970 | if (op_type != binary_op) |
3971 | vec_oprnd1 = NULL; | |
0d0e4a03 | 3972 | new_stmt = gimple_build_assign (vec_dest, code, vec_oprnd0, vec_oprnd1); |
ebfd146a IR |
3973 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3974 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 3975 | } |
ebfd146a IR |
3976 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3977 | ||
ebfd146a IR |
3978 | return new_stmt; |
3979 | } | |
3980 | ||
4a00c761 JJ |
3981 | |
3982 | /* Get vectorized definitions for loop-based vectorization. For the first | |
3983 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
3984 | scalar operand), and for the rest we get a copy with | |
3985 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
3986 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
3987 | The vectors are collected into VEC_OPRNDS. */ | |
3988 | ||
3989 | static void | |
355fe088 | 3990 | vect_get_loop_based_defs (tree *oprnd, gimple *stmt, enum vect_def_type dt, |
9771b263 | 3991 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
3992 | { |
3993 | tree vec_oprnd; | |
3994 | ||
3995 | /* Get first vector operand. */ | |
3996 | /* All the vector operands except the very first one (that is scalar oprnd) | |
3997 | are stmt copies. */ | |
3998 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
81c40241 | 3999 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt); |
4a00c761 JJ |
4000 | else |
4001 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
4002 | ||
9771b263 | 4003 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
4004 | |
4005 | /* Get second vector operand. */ | |
4006 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 4007 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
4008 | |
4009 | *oprnd = vec_oprnd; | |
4010 | ||
4011 | /* For conversion in multiple steps, continue to get operands | |
4012 | recursively. */ | |
4013 | if (multi_step_cvt) | |
4014 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
4015 | } | |
4016 | ||
4017 | ||
4018 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
4019 | For multi-step conversions store the resulting vectors and call the function | |
4020 | recursively. */ | |
4021 | ||
4022 | static void | |
9771b263 | 4023 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
355fe088 | 4024 | int multi_step_cvt, gimple *stmt, |
9771b263 | 4025 | vec<tree> vec_dsts, |
4a00c761 JJ |
4026 | gimple_stmt_iterator *gsi, |
4027 | slp_tree slp_node, enum tree_code code, | |
4028 | stmt_vec_info *prev_stmt_info) | |
4029 | { | |
4030 | unsigned int i; | |
4031 | tree vop0, vop1, new_tmp, vec_dest; | |
355fe088 | 4032 | gimple *new_stmt; |
4a00c761 JJ |
4033 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
4034 | ||
9771b263 | 4035 | vec_dest = vec_dsts.pop (); |
4a00c761 | 4036 | |
9771b263 | 4037 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
4038 | { |
4039 | /* Create demotion operation. */ | |
9771b263 DN |
4040 | vop0 = (*vec_oprnds)[i]; |
4041 | vop1 = (*vec_oprnds)[i + 1]; | |
0d0e4a03 | 4042 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
4a00c761 JJ |
4043 | new_tmp = make_ssa_name (vec_dest, new_stmt); |
4044 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
4045 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4046 | ||
4047 | if (multi_step_cvt) | |
4048 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 4049 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
4050 | else |
4051 | { | |
4052 | /* This is the last step of the conversion sequence. Store the | |
4053 | vectors in SLP_NODE or in vector info of the scalar statement | |
4054 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
4055 | if (slp_node) | |
9771b263 | 4056 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 4057 | else |
c689ce1e RB |
4058 | { |
4059 | if (!*prev_stmt_info) | |
4060 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
4061 | else | |
4062 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
4a00c761 | 4063 | |
c689ce1e RB |
4064 | *prev_stmt_info = vinfo_for_stmt (new_stmt); |
4065 | } | |
4a00c761 JJ |
4066 | } |
4067 | } | |
4068 | ||
4069 | /* For multi-step demotion operations we first generate demotion operations | |
4070 | from the source type to the intermediate types, and then combine the | |
4071 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
4072 | type. */ | |
4073 | if (multi_step_cvt) | |
4074 | { | |
4075 | /* At each level of recursion we have half of the operands we had at the | |
4076 | previous level. */ | |
9771b263 | 4077 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
4078 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
4079 | stmt, vec_dsts, gsi, slp_node, | |
4080 | VEC_PACK_TRUNC_EXPR, | |
4081 | prev_stmt_info); | |
4082 | } | |
4083 | ||
9771b263 | 4084 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4085 | } |
4086 | ||
4087 | ||
4088 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
4089 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
4090 | the resulting vectors and call the function recursively. */ | |
4091 | ||
4092 | static void | |
9771b263 DN |
4093 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
4094 | vec<tree> *vec_oprnds1, | |
355fe088 | 4095 | gimple *stmt, tree vec_dest, |
4a00c761 JJ |
4096 | gimple_stmt_iterator *gsi, |
4097 | enum tree_code code1, | |
4098 | enum tree_code code2, tree decl1, | |
4099 | tree decl2, int op_type) | |
4100 | { | |
4101 | int i; | |
4102 | tree vop0, vop1, new_tmp1, new_tmp2; | |
355fe088 | 4103 | gimple *new_stmt1, *new_stmt2; |
6e1aa848 | 4104 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 4105 | |
9771b263 DN |
4106 | vec_tmp.create (vec_oprnds0->length () * 2); |
4107 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
4108 | { |
4109 | if (op_type == binary_op) | |
9771b263 | 4110 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
4111 | else |
4112 | vop1 = NULL_TREE; | |
4113 | ||
4114 | /* Generate the two halves of promotion operation. */ | |
4115 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
4116 | op_type, vec_dest, gsi, stmt); | |
4117 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
4118 | op_type, vec_dest, gsi, stmt); | |
4119 | if (is_gimple_call (new_stmt1)) | |
4120 | { | |
4121 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
4122 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
4123 | } | |
4124 | else | |
4125 | { | |
4126 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
4127 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
4128 | } | |
4129 | ||
4130 | /* Store the results for the next step. */ | |
9771b263 DN |
4131 | vec_tmp.quick_push (new_tmp1); |
4132 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
4133 | } |
4134 | ||
689eaba3 | 4135 | vec_oprnds0->release (); |
4a00c761 JJ |
4136 | *vec_oprnds0 = vec_tmp; |
4137 | } | |
4138 | ||
4139 | ||
b8698a0f L |
4140 | /* Check if STMT performs a conversion operation, that can be vectorized. |
4141 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 4142 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
4143 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
4144 | ||
4145 | static bool | |
355fe088 TS |
4146 | vectorizable_conversion (gimple *stmt, gimple_stmt_iterator *gsi, |
4147 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
4148 | { |
4149 | tree vec_dest; | |
4150 | tree scalar_dest; | |
4a00c761 | 4151 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
4152 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
4153 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4154 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4155 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 4156 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
4157 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
4158 | tree new_temp; | |
355fe088 | 4159 | gimple *def_stmt; |
ebfd146a | 4160 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 4161 | int ndts = 2; |
355fe088 | 4162 | gimple *new_stmt = NULL; |
ebfd146a | 4163 | stmt_vec_info prev_stmt_info; |
062d5ccc RS |
4164 | poly_uint64 nunits_in; |
4165 | poly_uint64 nunits_out; | |
ebfd146a | 4166 | tree vectype_out, vectype_in; |
4a00c761 JJ |
4167 | int ncopies, i, j; |
4168 | tree lhs_type, rhs_type; | |
ebfd146a | 4169 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
4170 | vec<tree> vec_oprnds0 = vNULL; |
4171 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 4172 | tree vop0; |
4a00c761 | 4173 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4174 | vec_info *vinfo = stmt_info->vinfo; |
4a00c761 | 4175 | int multi_step_cvt = 0; |
6e1aa848 | 4176 | vec<tree> interm_types = vNULL; |
4a00c761 JJ |
4177 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
4178 | int op_type; | |
4a00c761 | 4179 | unsigned short fltsz; |
ebfd146a IR |
4180 | |
4181 | /* Is STMT a vectorizable conversion? */ | |
4182 | ||
4a00c761 | 4183 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4184 | return false; |
4185 | ||
66c16fd9 RB |
4186 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4187 | && ! vec_stmt) | |
ebfd146a IR |
4188 | return false; |
4189 | ||
4190 | if (!is_gimple_assign (stmt)) | |
4191 | return false; | |
4192 | ||
4193 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4194 | return false; | |
4195 | ||
4196 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
4197 | if (!CONVERT_EXPR_CODE_P (code) |
4198 | && code != FIX_TRUNC_EXPR | |
4199 | && code != FLOAT_EXPR | |
4200 | && code != WIDEN_MULT_EXPR | |
4201 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
4202 | return false; |
4203 | ||
4a00c761 JJ |
4204 | op_type = TREE_CODE_LENGTH (code); |
4205 | ||
ebfd146a | 4206 | /* Check types of lhs and rhs. */ |
b690cc0f | 4207 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 4208 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
4209 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
4210 | ||
ebfd146a IR |
4211 | op0 = gimple_assign_rhs1 (stmt); |
4212 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
4213 | |
4214 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4215 | && !((INTEGRAL_TYPE_P (lhs_type) | |
4216 | && INTEGRAL_TYPE_P (rhs_type)) | |
4217 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
4218 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
4219 | return false; | |
4220 | ||
e6f5c25d IE |
4221 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4222 | && ((INTEGRAL_TYPE_P (lhs_type) | |
2be65d9e | 4223 | && !type_has_mode_precision_p (lhs_type)) |
e6f5c25d | 4224 | || (INTEGRAL_TYPE_P (rhs_type) |
2be65d9e | 4225 | && !type_has_mode_precision_p (rhs_type)))) |
4a00c761 | 4226 | { |
73fbfcad | 4227 | if (dump_enabled_p ()) |
78c60e3d | 4228 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
4229 | "type conversion to/from bit-precision unsupported." |
4230 | "\n"); | |
4a00c761 JJ |
4231 | return false; |
4232 | } | |
4233 | ||
b690cc0f | 4234 | /* Check the operands of the operation. */ |
81c40241 | 4235 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype_in)) |
b690cc0f | 4236 | { |
73fbfcad | 4237 | if (dump_enabled_p ()) |
78c60e3d | 4238 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4239 | "use not simple.\n"); |
b690cc0f RG |
4240 | return false; |
4241 | } | |
4a00c761 JJ |
4242 | if (op_type == binary_op) |
4243 | { | |
4244 | bool ok; | |
4245 | ||
4246 | op1 = gimple_assign_rhs2 (stmt); | |
4247 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
4248 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
4249 | OP1. */ | |
4250 | if (CONSTANT_CLASS_P (op0)) | |
81c40241 | 4251 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &vectype_in); |
4a00c761 | 4252 | else |
81c40241 | 4253 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1]); |
4a00c761 JJ |
4254 | |
4255 | if (!ok) | |
4256 | { | |
73fbfcad | 4257 | if (dump_enabled_p ()) |
78c60e3d | 4258 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4259 | "use not simple.\n"); |
4a00c761 JJ |
4260 | return false; |
4261 | } | |
4262 | } | |
4263 | ||
b690cc0f RG |
4264 | /* If op0 is an external or constant defs use a vector type of |
4265 | the same size as the output vector type. */ | |
ebfd146a | 4266 | if (!vectype_in) |
b690cc0f | 4267 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
4268 | if (vec_stmt) |
4269 | gcc_assert (vectype_in); | |
4270 | if (!vectype_in) | |
4271 | { | |
73fbfcad | 4272 | if (dump_enabled_p ()) |
4a00c761 | 4273 | { |
78c60e3d SS |
4274 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4275 | "no vectype for scalar type "); | |
4276 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 4277 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 4278 | } |
7d8930a0 IR |
4279 | |
4280 | return false; | |
4281 | } | |
ebfd146a | 4282 | |
e6f5c25d IE |
4283 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4284 | && !VECTOR_BOOLEAN_TYPE_P (vectype_in)) | |
4285 | { | |
4286 | if (dump_enabled_p ()) | |
4287 | { | |
4288 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
4289 | "can't convert between boolean and non " | |
4290 | "boolean vectors"); | |
4291 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
4292 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
4293 | } | |
4294 | ||
4295 | return false; | |
4296 | } | |
4297 | ||
b690cc0f RG |
4298 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
4299 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
062d5ccc | 4300 | if (known_eq (nunits_out, nunits_in)) |
ebfd146a | 4301 | modifier = NONE; |
062d5ccc RS |
4302 | else if (multiple_p (nunits_out, nunits_in)) |
4303 | modifier = NARROW; | |
ebfd146a | 4304 | else |
062d5ccc RS |
4305 | { |
4306 | gcc_checking_assert (multiple_p (nunits_in, nunits_out)); | |
4307 | modifier = WIDEN; | |
4308 | } | |
ebfd146a | 4309 | |
ff802fa1 IR |
4310 | /* Multiple types in SLP are handled by creating the appropriate number of |
4311 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4312 | case of SLP. */ | |
fce57248 | 4313 | if (slp_node) |
ebfd146a | 4314 | ncopies = 1; |
4a00c761 | 4315 | else if (modifier == NARROW) |
e8f142e2 | 4316 | ncopies = vect_get_num_copies (loop_vinfo, vectype_out); |
4a00c761 | 4317 | else |
e8f142e2 | 4318 | ncopies = vect_get_num_copies (loop_vinfo, vectype_in); |
b8698a0f | 4319 | |
ebfd146a IR |
4320 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
4321 | needs to be generated. */ | |
4322 | gcc_assert (ncopies >= 1); | |
4323 | ||
16d22000 RS |
4324 | bool found_mode = false; |
4325 | scalar_mode lhs_mode = SCALAR_TYPE_MODE (lhs_type); | |
4326 | scalar_mode rhs_mode = SCALAR_TYPE_MODE (rhs_type); | |
4327 | opt_scalar_mode rhs_mode_iter; | |
b397965c | 4328 | |
ebfd146a | 4329 | /* Supportable by target? */ |
4a00c761 | 4330 | switch (modifier) |
ebfd146a | 4331 | { |
4a00c761 JJ |
4332 | case NONE: |
4333 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4334 | return false; | |
4335 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
4336 | &decl1, &code1)) | |
4337 | break; | |
4338 | /* FALLTHRU */ | |
4339 | unsupported: | |
73fbfcad | 4340 | if (dump_enabled_p ()) |
78c60e3d | 4341 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4342 | "conversion not supported by target.\n"); |
ebfd146a | 4343 | return false; |
ebfd146a | 4344 | |
4a00c761 JJ |
4345 | case WIDEN: |
4346 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
4347 | &code1, &code2, &multi_step_cvt, |
4348 | &interm_types)) | |
4a00c761 JJ |
4349 | { |
4350 | /* Binary widening operation can only be supported directly by the | |
4351 | architecture. */ | |
4352 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
4353 | break; | |
4354 | } | |
4355 | ||
4356 | if (code != FLOAT_EXPR | |
b397965c | 4357 | || GET_MODE_SIZE (lhs_mode) <= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4358 | goto unsupported; |
4359 | ||
b397965c | 4360 | fltsz = GET_MODE_SIZE (lhs_mode); |
16d22000 | 4361 | FOR_EACH_2XWIDER_MODE (rhs_mode_iter, rhs_mode) |
4a00c761 | 4362 | { |
16d22000 | 4363 | rhs_mode = rhs_mode_iter.require (); |
c94843d2 RS |
4364 | if (GET_MODE_SIZE (rhs_mode) > fltsz) |
4365 | break; | |
4366 | ||
4a00c761 JJ |
4367 | cvt_type |
4368 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4369 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4370 | if (cvt_type == NULL_TREE) | |
4371 | goto unsupported; | |
4372 | ||
4373 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4374 | { | |
4375 | if (!supportable_convert_operation (code, vectype_out, | |
4376 | cvt_type, &decl1, &codecvt1)) | |
4377 | goto unsupported; | |
4378 | } | |
4379 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
4380 | cvt_type, &codecvt1, |
4381 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
4382 | &interm_types)) |
4383 | continue; | |
4384 | else | |
4385 | gcc_assert (multi_step_cvt == 0); | |
4386 | ||
4387 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
4388 | vectype_in, &code1, &code2, |
4389 | &multi_step_cvt, &interm_types)) | |
16d22000 RS |
4390 | { |
4391 | found_mode = true; | |
4392 | break; | |
4393 | } | |
4a00c761 JJ |
4394 | } |
4395 | ||
16d22000 | 4396 | if (!found_mode) |
4a00c761 JJ |
4397 | goto unsupported; |
4398 | ||
4399 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4400 | codecvt2 = ERROR_MARK; | |
4401 | else | |
4402 | { | |
4403 | multi_step_cvt++; | |
9771b263 | 4404 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
4405 | cvt_type = NULL_TREE; |
4406 | } | |
4407 | break; | |
4408 | ||
4409 | case NARROW: | |
4410 | gcc_assert (op_type == unary_op); | |
4411 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
4412 | &code1, &multi_step_cvt, | |
4413 | &interm_types)) | |
4414 | break; | |
4415 | ||
4416 | if (code != FIX_TRUNC_EXPR | |
b397965c | 4417 | || GET_MODE_SIZE (lhs_mode) >= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4418 | goto unsupported; |
4419 | ||
4a00c761 JJ |
4420 | cvt_type |
4421 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4422 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4423 | if (cvt_type == NULL_TREE) | |
4424 | goto unsupported; | |
4425 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
4426 | &decl1, &codecvt1)) | |
4427 | goto unsupported; | |
4428 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
4429 | &code1, &multi_step_cvt, | |
4430 | &interm_types)) | |
4431 | break; | |
4432 | goto unsupported; | |
4433 | ||
4434 | default: | |
4435 | gcc_unreachable (); | |
ebfd146a IR |
4436 | } |
4437 | ||
4438 | if (!vec_stmt) /* transformation not required. */ | |
4439 | { | |
73fbfcad | 4440 | if (dump_enabled_p ()) |
78c60e3d | 4441 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4442 | "=== vectorizable_conversion ===\n"); |
4a00c761 | 4443 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
4444 | { |
4445 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
4fc5ebf1 | 4446 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); |
8bd37302 | 4447 | } |
4a00c761 JJ |
4448 | else if (modifier == NARROW) |
4449 | { | |
4450 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
8bd37302 | 4451 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 JJ |
4452 | } |
4453 | else | |
4454 | { | |
4455 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
8bd37302 | 4456 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 | 4457 | } |
9771b263 | 4458 | interm_types.release (); |
ebfd146a IR |
4459 | return true; |
4460 | } | |
4461 | ||
67b8dbac | 4462 | /* Transform. */ |
73fbfcad | 4463 | if (dump_enabled_p ()) |
78c60e3d | 4464 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4465 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 4466 | |
4a00c761 JJ |
4467 | if (op_type == binary_op) |
4468 | { | |
4469 | if (CONSTANT_CLASS_P (op0)) | |
4470 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
4471 | else if (CONSTANT_CLASS_P (op1)) | |
4472 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
4473 | } | |
4474 | ||
4475 | /* In case of multi-step conversion, we first generate conversion operations | |
4476 | to the intermediate types, and then from that types to the final one. | |
4477 | We create vector destinations for the intermediate type (TYPES) received | |
4478 | from supportable_*_operation, and store them in the correct order | |
4479 | for future use in vect_create_vectorized_*_stmts (). */ | |
8c681247 | 4480 | auto_vec<tree> vec_dsts (multi_step_cvt + 1); |
82294ec1 JJ |
4481 | vec_dest = vect_create_destination_var (scalar_dest, |
4482 | (cvt_type && modifier == WIDEN) | |
4483 | ? cvt_type : vectype_out); | |
9771b263 | 4484 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4485 | |
4486 | if (multi_step_cvt) | |
4487 | { | |
9771b263 DN |
4488 | for (i = interm_types.length () - 1; |
4489 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
4490 | { |
4491 | vec_dest = vect_create_destination_var (scalar_dest, | |
4492 | intermediate_type); | |
9771b263 | 4493 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4494 | } |
4495 | } | |
ebfd146a | 4496 | |
4a00c761 | 4497 | if (cvt_type) |
82294ec1 JJ |
4498 | vec_dest = vect_create_destination_var (scalar_dest, |
4499 | modifier == WIDEN | |
4500 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
4501 | |
4502 | if (!slp_node) | |
4503 | { | |
30862efc | 4504 | if (modifier == WIDEN) |
4a00c761 | 4505 | { |
c3284718 | 4506 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 4507 | if (op_type == binary_op) |
9771b263 | 4508 | vec_oprnds1.create (1); |
4a00c761 | 4509 | } |
30862efc | 4510 | else if (modifier == NARROW) |
9771b263 DN |
4511 | vec_oprnds0.create ( |
4512 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
4513 | } |
4514 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 4515 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 4516 | |
4a00c761 | 4517 | last_oprnd = op0; |
ebfd146a IR |
4518 | prev_stmt_info = NULL; |
4519 | switch (modifier) | |
4520 | { | |
4521 | case NONE: | |
4522 | for (j = 0; j < ncopies; j++) | |
4523 | { | |
ebfd146a | 4524 | if (j == 0) |
306b0c92 | 4525 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node); |
ebfd146a IR |
4526 | else |
4527 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
4528 | ||
9771b263 | 4529 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4530 | { |
4531 | /* Arguments are ready, create the new vector stmt. */ | |
4532 | if (code1 == CALL_EXPR) | |
4533 | { | |
4534 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4535 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4536 | gimple_call_set_lhs (new_stmt, new_temp); | |
4537 | } | |
4538 | else | |
4539 | { | |
4540 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
0d0e4a03 | 4541 | new_stmt = gimple_build_assign (vec_dest, code1, vop0); |
4a00c761 JJ |
4542 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4543 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4544 | } | |
4545 | ||
4546 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4547 | if (slp_node) | |
9771b263 | 4548 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
225ce44b RB |
4549 | else |
4550 | { | |
4551 | if (!prev_stmt_info) | |
4552 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4553 | else | |
4554 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4555 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4556 | } | |
4a00c761 | 4557 | } |
ebfd146a IR |
4558 | } |
4559 | break; | |
4560 | ||
4561 | case WIDEN: | |
4562 | /* In case the vectorization factor (VF) is bigger than the number | |
4563 | of elements that we can fit in a vectype (nunits), we have to | |
4564 | generate more than one vector stmt - i.e - we need to "unroll" | |
4565 | the vector stmt by a factor VF/nunits. */ | |
4566 | for (j = 0; j < ncopies; j++) | |
4567 | { | |
4a00c761 | 4568 | /* Handle uses. */ |
ebfd146a | 4569 | if (j == 0) |
4a00c761 JJ |
4570 | { |
4571 | if (slp_node) | |
4572 | { | |
4573 | if (code == WIDEN_LSHIFT_EXPR) | |
4574 | { | |
4575 | unsigned int k; | |
ebfd146a | 4576 | |
4a00c761 JJ |
4577 | vec_oprnd1 = op1; |
4578 | /* Store vec_oprnd1 for every vector stmt to be created | |
4579 | for SLP_NODE. We check during the analysis that all | |
4580 | the shift arguments are the same. */ | |
4581 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4582 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4583 | |
4584 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 4585 | slp_node); |
4a00c761 JJ |
4586 | } |
4587 | else | |
4588 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
306b0c92 | 4589 | &vec_oprnds1, slp_node); |
4a00c761 JJ |
4590 | } |
4591 | else | |
4592 | { | |
81c40241 | 4593 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 4594 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
4595 | if (op_type == binary_op) |
4596 | { | |
4597 | if (code == WIDEN_LSHIFT_EXPR) | |
4598 | vec_oprnd1 = op1; | |
4599 | else | |
81c40241 | 4600 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 4601 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4602 | } |
4603 | } | |
4604 | } | |
ebfd146a | 4605 | else |
4a00c761 JJ |
4606 | { |
4607 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
4608 | vec_oprnds0.truncate (0); |
4609 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
4610 | if (op_type == binary_op) |
4611 | { | |
4612 | if (code == WIDEN_LSHIFT_EXPR) | |
4613 | vec_oprnd1 = op1; | |
4614 | else | |
4615 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
4616 | vec_oprnd1); | |
9771b263 DN |
4617 | vec_oprnds1.truncate (0); |
4618 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
4619 | } |
4620 | } | |
ebfd146a | 4621 | |
4a00c761 JJ |
4622 | /* Arguments are ready. Create the new vector stmts. */ |
4623 | for (i = multi_step_cvt; i >= 0; i--) | |
4624 | { | |
9771b263 | 4625 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
4626 | enum tree_code c1 = code1, c2 = code2; |
4627 | if (i == 0 && codecvt2 != ERROR_MARK) | |
4628 | { | |
4629 | c1 = codecvt1; | |
4630 | c2 = codecvt2; | |
4631 | } | |
4632 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
4633 | &vec_oprnds1, | |
4634 | stmt, this_dest, gsi, | |
4635 | c1, c2, decl1, decl2, | |
4636 | op_type); | |
4637 | } | |
4638 | ||
9771b263 | 4639 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4640 | { |
4641 | if (cvt_type) | |
4642 | { | |
4643 | if (codecvt1 == CALL_EXPR) | |
4644 | { | |
4645 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4646 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4647 | gimple_call_set_lhs (new_stmt, new_temp); | |
4648 | } | |
4649 | else | |
4650 | { | |
4651 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4652 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4653 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4654 | vop0); | |
4a00c761 JJ |
4655 | } |
4656 | ||
4657 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4658 | } | |
4659 | else | |
4660 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
4661 | ||
4662 | if (slp_node) | |
9771b263 | 4663 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 4664 | else |
c689ce1e RB |
4665 | { |
4666 | if (!prev_stmt_info) | |
4667 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
4668 | else | |
4669 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4670 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4671 | } | |
4a00c761 | 4672 | } |
ebfd146a | 4673 | } |
4a00c761 JJ |
4674 | |
4675 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
4676 | break; |
4677 | ||
4678 | case NARROW: | |
4679 | /* In case the vectorization factor (VF) is bigger than the number | |
4680 | of elements that we can fit in a vectype (nunits), we have to | |
4681 | generate more than one vector stmt - i.e - we need to "unroll" | |
4682 | the vector stmt by a factor VF/nunits. */ | |
4683 | for (j = 0; j < ncopies; j++) | |
4684 | { | |
4685 | /* Handle uses. */ | |
4a00c761 JJ |
4686 | if (slp_node) |
4687 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 4688 | slp_node); |
ebfd146a IR |
4689 | else |
4690 | { | |
9771b263 | 4691 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
4692 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
4693 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
4694 | } |
4695 | ||
4a00c761 JJ |
4696 | /* Arguments are ready. Create the new vector stmts. */ |
4697 | if (cvt_type) | |
9771b263 | 4698 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4699 | { |
4700 | if (codecvt1 == CALL_EXPR) | |
4701 | { | |
4702 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4703 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4704 | gimple_call_set_lhs (new_stmt, new_temp); | |
4705 | } | |
4706 | else | |
4707 | { | |
4708 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4709 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4710 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4711 | vop0); | |
4a00c761 | 4712 | } |
ebfd146a | 4713 | |
4a00c761 | 4714 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 4715 | vec_oprnds0[i] = new_temp; |
4a00c761 | 4716 | } |
ebfd146a | 4717 | |
4a00c761 JJ |
4718 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
4719 | stmt, vec_dsts, gsi, | |
4720 | slp_node, code1, | |
4721 | &prev_stmt_info); | |
ebfd146a IR |
4722 | } |
4723 | ||
4724 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 4725 | break; |
ebfd146a IR |
4726 | } |
4727 | ||
9771b263 DN |
4728 | vec_oprnds0.release (); |
4729 | vec_oprnds1.release (); | |
9771b263 | 4730 | interm_types.release (); |
ebfd146a IR |
4731 | |
4732 | return true; | |
4733 | } | |
ff802fa1 IR |
4734 | |
4735 | ||
ebfd146a IR |
4736 | /* Function vectorizable_assignment. |
4737 | ||
b8698a0f L |
4738 | Check if STMT performs an assignment (copy) that can be vectorized. |
4739 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4740 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4741 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4742 | ||
4743 | static bool | |
355fe088 TS |
4744 | vectorizable_assignment (gimple *stmt, gimple_stmt_iterator *gsi, |
4745 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
4746 | { |
4747 | tree vec_dest; | |
4748 | tree scalar_dest; | |
4749 | tree op; | |
4750 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a IR |
4751 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4752 | tree new_temp; | |
355fe088 | 4753 | gimple *def_stmt; |
4fc5ebf1 JG |
4754 | enum vect_def_type dt[1] = {vect_unknown_def_type}; |
4755 | int ndts = 1; | |
ebfd146a | 4756 | int ncopies; |
f18b55bd | 4757 | int i, j; |
6e1aa848 | 4758 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 4759 | tree vop; |
a70d6342 | 4760 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4761 | vec_info *vinfo = stmt_info->vinfo; |
355fe088 | 4762 | gimple *new_stmt = NULL; |
f18b55bd | 4763 | stmt_vec_info prev_stmt_info = NULL; |
fde9c428 RG |
4764 | enum tree_code code; |
4765 | tree vectype_in; | |
ebfd146a | 4766 | |
a70d6342 | 4767 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4768 | return false; |
4769 | ||
66c16fd9 RB |
4770 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4771 | && ! vec_stmt) | |
ebfd146a IR |
4772 | return false; |
4773 | ||
4774 | /* Is vectorizable assignment? */ | |
4775 | if (!is_gimple_assign (stmt)) | |
4776 | return false; | |
4777 | ||
4778 | scalar_dest = gimple_assign_lhs (stmt); | |
4779 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4780 | return false; | |
4781 | ||
fde9c428 | 4782 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 4783 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
4784 | || code == PAREN_EXPR |
4785 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
4786 | op = gimple_assign_rhs1 (stmt); |
4787 | else | |
4788 | return false; | |
4789 | ||
7b7ec6c5 RG |
4790 | if (code == VIEW_CONVERT_EXPR) |
4791 | op = TREE_OPERAND (op, 0); | |
4792 | ||
465c8c19 | 4793 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
928686b1 | 4794 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 JJ |
4795 | |
4796 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4797 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4798 | case of SLP. */ | |
fce57248 | 4799 | if (slp_node) |
465c8c19 JJ |
4800 | ncopies = 1; |
4801 | else | |
e8f142e2 | 4802 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
4803 | |
4804 | gcc_assert (ncopies >= 1); | |
4805 | ||
81c40241 | 4806 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[0], &vectype_in)) |
ebfd146a | 4807 | { |
73fbfcad | 4808 | if (dump_enabled_p ()) |
78c60e3d | 4809 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4810 | "use not simple.\n"); |
ebfd146a IR |
4811 | return false; |
4812 | } | |
4813 | ||
fde9c428 RG |
4814 | /* We can handle NOP_EXPR conversions that do not change the number |
4815 | of elements or the vector size. */ | |
7b7ec6c5 RG |
4816 | if ((CONVERT_EXPR_CODE_P (code) |
4817 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 | 4818 | && (!vectype_in |
928686b1 | 4819 | || maybe_ne (TYPE_VECTOR_SUBPARTS (vectype_in), nunits) |
cf098191 RS |
4820 | || maybe_ne (GET_MODE_SIZE (TYPE_MODE (vectype)), |
4821 | GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
fde9c428 RG |
4822 | return false; |
4823 | ||
7b7b1813 RG |
4824 | /* We do not handle bit-precision changes. */ |
4825 | if ((CONVERT_EXPR_CODE_P (code) | |
4826 | || code == VIEW_CONVERT_EXPR) | |
4827 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
2be65d9e RS |
4828 | && (!type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
4829 | || !type_has_mode_precision_p (TREE_TYPE (op))) | |
7b7b1813 RG |
4830 | /* But a conversion that does not change the bit-pattern is ok. */ |
4831 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4832 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2dab46d5 IE |
4833 | && TYPE_UNSIGNED (TREE_TYPE (op))) |
4834 | /* Conversion between boolean types of different sizes is | |
4835 | a simple assignment in case their vectypes are same | |
4836 | boolean vectors. */ | |
4837 | && (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
4838 | || !VECTOR_BOOLEAN_TYPE_P (vectype_in))) | |
7b7b1813 | 4839 | { |
73fbfcad | 4840 | if (dump_enabled_p ()) |
78c60e3d SS |
4841 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4842 | "type conversion to/from bit-precision " | |
e645e942 | 4843 | "unsupported.\n"); |
7b7b1813 RG |
4844 | return false; |
4845 | } | |
4846 | ||
ebfd146a IR |
4847 | if (!vec_stmt) /* transformation not required. */ |
4848 | { | |
4849 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 4850 | if (dump_enabled_p ()) |
78c60e3d | 4851 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4852 | "=== vectorizable_assignment ===\n"); |
4fc5ebf1 | 4853 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); |
ebfd146a IR |
4854 | return true; |
4855 | } | |
4856 | ||
67b8dbac | 4857 | /* Transform. */ |
73fbfcad | 4858 | if (dump_enabled_p ()) |
e645e942 | 4859 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
4860 | |
4861 | /* Handle def. */ | |
4862 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4863 | ||
4864 | /* Handle use. */ | |
f18b55bd | 4865 | for (j = 0; j < ncopies; j++) |
ebfd146a | 4866 | { |
f18b55bd IR |
4867 | /* Handle uses. */ |
4868 | if (j == 0) | |
306b0c92 | 4869 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); |
f18b55bd IR |
4870 | else |
4871 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
4872 | ||
4873 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 4874 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 4875 | { |
7b7ec6c5 RG |
4876 | if (CONVERT_EXPR_CODE_P (code) |
4877 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 4878 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
4879 | new_stmt = gimple_build_assign (vec_dest, vop); |
4880 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4881 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4882 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4883 | if (slp_node) | |
9771b263 | 4884 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 4885 | } |
ebfd146a IR |
4886 | |
4887 | if (slp_node) | |
f18b55bd IR |
4888 | continue; |
4889 | ||
4890 | if (j == 0) | |
4891 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4892 | else | |
4893 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4894 | ||
4895 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4896 | } | |
b8698a0f | 4897 | |
9771b263 | 4898 | vec_oprnds.release (); |
ebfd146a IR |
4899 | return true; |
4900 | } | |
4901 | ||
9dc3f7de | 4902 | |
1107f3ae IR |
4903 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
4904 | either as shift by a scalar or by a vector. */ | |
4905 | ||
4906 | bool | |
4907 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
4908 | { | |
4909 | ||
ef4bddc2 | 4910 | machine_mode vec_mode; |
1107f3ae IR |
4911 | optab optab; |
4912 | int icode; | |
4913 | tree vectype; | |
4914 | ||
4915 | vectype = get_vectype_for_scalar_type (scalar_type); | |
4916 | if (!vectype) | |
4917 | return false; | |
4918 | ||
4919 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4920 | if (!optab | |
4921 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
4922 | { | |
4923 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4924 | if (!optab | |
4925 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
4926 | == CODE_FOR_nothing)) | |
4927 | return false; | |
4928 | } | |
4929 | ||
4930 | vec_mode = TYPE_MODE (vectype); | |
4931 | icode = (int) optab_handler (optab, vec_mode); | |
4932 | if (icode == CODE_FOR_nothing) | |
4933 | return false; | |
4934 | ||
4935 | return true; | |
4936 | } | |
4937 | ||
4938 | ||
9dc3f7de IR |
4939 | /* Function vectorizable_shift. |
4940 | ||
4941 | Check if STMT performs a shift operation that can be vectorized. | |
4942 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4943 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
4944 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4945 | ||
4946 | static bool | |
355fe088 TS |
4947 | vectorizable_shift (gimple *stmt, gimple_stmt_iterator *gsi, |
4948 | gimple **vec_stmt, slp_tree slp_node) | |
9dc3f7de IR |
4949 | { |
4950 | tree vec_dest; | |
4951 | tree scalar_dest; | |
4952 | tree op0, op1 = NULL; | |
4953 | tree vec_oprnd1 = NULL_TREE; | |
4954 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4955 | tree vectype; | |
4956 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4957 | enum tree_code code; | |
ef4bddc2 | 4958 | machine_mode vec_mode; |
9dc3f7de IR |
4959 | tree new_temp; |
4960 | optab optab; | |
4961 | int icode; | |
ef4bddc2 | 4962 | machine_mode optab_op2_mode; |
355fe088 | 4963 | gimple *def_stmt; |
9dc3f7de | 4964 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 4965 | int ndts = 2; |
355fe088 | 4966 | gimple *new_stmt = NULL; |
9dc3f7de | 4967 | stmt_vec_info prev_stmt_info; |
928686b1 RS |
4968 | poly_uint64 nunits_in; |
4969 | poly_uint64 nunits_out; | |
9dc3f7de | 4970 | tree vectype_out; |
cede2577 | 4971 | tree op1_vectype; |
9dc3f7de IR |
4972 | int ncopies; |
4973 | int j, i; | |
6e1aa848 DN |
4974 | vec<tree> vec_oprnds0 = vNULL; |
4975 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
4976 | tree vop0, vop1; |
4977 | unsigned int k; | |
49eab32e | 4978 | bool scalar_shift_arg = true; |
9dc3f7de | 4979 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4980 | vec_info *vinfo = stmt_info->vinfo; |
9dc3f7de IR |
4981 | |
4982 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
4983 | return false; | |
4984 | ||
66c16fd9 RB |
4985 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4986 | && ! vec_stmt) | |
9dc3f7de IR |
4987 | return false; |
4988 | ||
4989 | /* Is STMT a vectorizable binary/unary operation? */ | |
4990 | if (!is_gimple_assign (stmt)) | |
4991 | return false; | |
4992 | ||
4993 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4994 | return false; | |
4995 | ||
4996 | code = gimple_assign_rhs_code (stmt); | |
4997 | ||
4998 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
4999 | || code == RROTATE_EXPR)) | |
5000 | return false; | |
5001 | ||
5002 | scalar_dest = gimple_assign_lhs (stmt); | |
5003 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2be65d9e | 5004 | if (!type_has_mode_precision_p (TREE_TYPE (scalar_dest))) |
7b7b1813 | 5005 | { |
73fbfcad | 5006 | if (dump_enabled_p ()) |
78c60e3d | 5007 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5008 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
5009 | return false; |
5010 | } | |
9dc3f7de IR |
5011 | |
5012 | op0 = gimple_assign_rhs1 (stmt); | |
81c40241 | 5013 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
9dc3f7de | 5014 | { |
73fbfcad | 5015 | if (dump_enabled_p ()) |
78c60e3d | 5016 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5017 | "use not simple.\n"); |
9dc3f7de IR |
5018 | return false; |
5019 | } | |
5020 | /* If op0 is an external or constant def use a vector type with | |
5021 | the same size as the output vector type. */ | |
5022 | if (!vectype) | |
5023 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
5024 | if (vec_stmt) | |
5025 | gcc_assert (vectype); | |
5026 | if (!vectype) | |
5027 | { | |
73fbfcad | 5028 | if (dump_enabled_p ()) |
78c60e3d | 5029 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5030 | "no vectype for scalar type\n"); |
9dc3f7de IR |
5031 | return false; |
5032 | } | |
5033 | ||
5034 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
5035 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
928686b1 | 5036 | if (maybe_ne (nunits_out, nunits_in)) |
9dc3f7de IR |
5037 | return false; |
5038 | ||
5039 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 5040 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &op1_vectype)) |
9dc3f7de | 5041 | { |
73fbfcad | 5042 | if (dump_enabled_p ()) |
78c60e3d | 5043 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5044 | "use not simple.\n"); |
9dc3f7de IR |
5045 | return false; |
5046 | } | |
5047 | ||
9dc3f7de IR |
5048 | /* Multiple types in SLP are handled by creating the appropriate number of |
5049 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5050 | case of SLP. */ | |
fce57248 | 5051 | if (slp_node) |
9dc3f7de IR |
5052 | ncopies = 1; |
5053 | else | |
e8f142e2 | 5054 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
9dc3f7de IR |
5055 | |
5056 | gcc_assert (ncopies >= 1); | |
5057 | ||
5058 | /* Determine whether the shift amount is a vector, or scalar. If the | |
5059 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
5060 | ||
dbfa87aa YR |
5061 | if ((dt[1] == vect_internal_def |
5062 | || dt[1] == vect_induction_def) | |
5063 | && !slp_node) | |
49eab32e JJ |
5064 | scalar_shift_arg = false; |
5065 | else if (dt[1] == vect_constant_def | |
5066 | || dt[1] == vect_external_def | |
5067 | || dt[1] == vect_internal_def) | |
5068 | { | |
5069 | /* In SLP, need to check whether the shift count is the same, | |
5070 | in loops if it is a constant or invariant, it is always | |
5071 | a scalar shift. */ | |
5072 | if (slp_node) | |
5073 | { | |
355fe088 TS |
5074 | vec<gimple *> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
5075 | gimple *slpstmt; | |
49eab32e | 5076 | |
9771b263 | 5077 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
5078 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
5079 | scalar_shift_arg = false; | |
5080 | } | |
60d393e8 RB |
5081 | |
5082 | /* If the shift amount is computed by a pattern stmt we cannot | |
5083 | use the scalar amount directly thus give up and use a vector | |
5084 | shift. */ | |
5085 | if (dt[1] == vect_internal_def) | |
5086 | { | |
5087 | gimple *def = SSA_NAME_DEF_STMT (op1); | |
5088 | if (is_pattern_stmt_p (vinfo_for_stmt (def))) | |
5089 | scalar_shift_arg = false; | |
5090 | } | |
49eab32e JJ |
5091 | } |
5092 | else | |
5093 | { | |
73fbfcad | 5094 | if (dump_enabled_p ()) |
78c60e3d | 5095 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5096 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
5097 | return false; |
5098 | } | |
5099 | ||
9dc3f7de | 5100 | /* Vector shifted by vector. */ |
49eab32e | 5101 | if (!scalar_shift_arg) |
9dc3f7de IR |
5102 | { |
5103 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 5104 | if (dump_enabled_p ()) |
78c60e3d | 5105 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5106 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 5107 | |
aa948027 JJ |
5108 | if (!op1_vectype) |
5109 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
5110 | if (op1_vectype == NULL_TREE | |
5111 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 5112 | { |
73fbfcad | 5113 | if (dump_enabled_p ()) |
78c60e3d SS |
5114 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5115 | "unusable type for last operand in" | |
e645e942 | 5116 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
5117 | return false; |
5118 | } | |
9dc3f7de IR |
5119 | } |
5120 | /* See if the machine has a vector shifted by scalar insn and if not | |
5121 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 5122 | else |
9dc3f7de IR |
5123 | { |
5124 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
5125 | if (optab | |
5126 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
5127 | { | |
73fbfcad | 5128 | if (dump_enabled_p ()) |
78c60e3d | 5129 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5130 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
5131 | } |
5132 | else | |
5133 | { | |
5134 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
5135 | if (optab | |
5136 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
5137 | != CODE_FOR_nothing)) | |
5138 | { | |
49eab32e JJ |
5139 | scalar_shift_arg = false; |
5140 | ||
73fbfcad | 5141 | if (dump_enabled_p ()) |
78c60e3d | 5142 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5143 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
5144 | |
5145 | /* Unlike the other binary operators, shifts/rotates have | |
5146 | the rhs being int, instead of the same type as the lhs, | |
5147 | so make sure the scalar is the right type if we are | |
aa948027 | 5148 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
5149 | if (dt[1] == vect_constant_def) |
5150 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
5151 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
5152 | TREE_TYPE (op1))) | |
5153 | { | |
5154 | if (slp_node | |
5155 | && TYPE_MODE (TREE_TYPE (vectype)) | |
5156 | != TYPE_MODE (TREE_TYPE (op1))) | |
5157 | { | |
73fbfcad | 5158 | if (dump_enabled_p ()) |
78c60e3d SS |
5159 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5160 | "unusable type for last operand in" | |
e645e942 | 5161 | " vector/vector shift/rotate.\n"); |
21c0a521 | 5162 | return false; |
aa948027 JJ |
5163 | } |
5164 | if (vec_stmt && !slp_node) | |
5165 | { | |
5166 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
5167 | op1 = vect_init_vector (stmt, op1, | |
5168 | TREE_TYPE (vectype), NULL); | |
5169 | } | |
5170 | } | |
9dc3f7de IR |
5171 | } |
5172 | } | |
5173 | } | |
9dc3f7de IR |
5174 | |
5175 | /* Supportable by target? */ | |
5176 | if (!optab) | |
5177 | { | |
73fbfcad | 5178 | if (dump_enabled_p ()) |
78c60e3d | 5179 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5180 | "no optab.\n"); |
9dc3f7de IR |
5181 | return false; |
5182 | } | |
5183 | vec_mode = TYPE_MODE (vectype); | |
5184 | icode = (int) optab_handler (optab, vec_mode); | |
5185 | if (icode == CODE_FOR_nothing) | |
5186 | { | |
73fbfcad | 5187 | if (dump_enabled_p ()) |
78c60e3d | 5188 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5189 | "op not supported by target.\n"); |
9dc3f7de | 5190 | /* Check only during analysis. */ |
cf098191 | 5191 | if (maybe_ne (GET_MODE_SIZE (vec_mode), UNITS_PER_WORD) |
ca09abcb RS |
5192 | || (!vec_stmt |
5193 | && !vect_worthwhile_without_simd_p (vinfo, code))) | |
9dc3f7de | 5194 | return false; |
73fbfcad | 5195 | if (dump_enabled_p ()) |
e645e942 TJ |
5196 | dump_printf_loc (MSG_NOTE, vect_location, |
5197 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
5198 | } |
5199 | ||
5200 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
ca09abcb RS |
5201 | if (!vec_stmt |
5202 | && !VECTOR_MODE_P (TYPE_MODE (vectype)) | |
5203 | && !vect_worthwhile_without_simd_p (vinfo, code)) | |
9dc3f7de | 5204 | { |
73fbfcad | 5205 | if (dump_enabled_p ()) |
78c60e3d | 5206 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5207 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
5208 | return false; |
5209 | } | |
5210 | ||
5211 | if (!vec_stmt) /* transformation not required. */ | |
5212 | { | |
5213 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 5214 | if (dump_enabled_p ()) |
e645e942 TJ |
5215 | dump_printf_loc (MSG_NOTE, vect_location, |
5216 | "=== vectorizable_shift ===\n"); | |
4fc5ebf1 | 5217 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); |
9dc3f7de IR |
5218 | return true; |
5219 | } | |
5220 | ||
67b8dbac | 5221 | /* Transform. */ |
9dc3f7de | 5222 | |
73fbfcad | 5223 | if (dump_enabled_p ()) |
78c60e3d | 5224 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5225 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
5226 | |
5227 | /* Handle def. */ | |
5228 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5229 | ||
9dc3f7de IR |
5230 | prev_stmt_info = NULL; |
5231 | for (j = 0; j < ncopies; j++) | |
5232 | { | |
5233 | /* Handle uses. */ | |
5234 | if (j == 0) | |
5235 | { | |
5236 | if (scalar_shift_arg) | |
5237 | { | |
5238 | /* Vector shl and shr insn patterns can be defined with scalar | |
5239 | operand 2 (shift operand). In this case, use constant or loop | |
5240 | invariant op1 directly, without extending it to vector mode | |
5241 | first. */ | |
5242 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
5243 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
5244 | { | |
73fbfcad | 5245 | if (dump_enabled_p ()) |
78c60e3d | 5246 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5247 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 5248 | vec_oprnd1 = op1; |
8930f723 | 5249 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 5250 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5251 | if (slp_node) |
5252 | { | |
5253 | /* Store vec_oprnd1 for every vector stmt to be created | |
5254 | for SLP_NODE. We check during the analysis that all | |
5255 | the shift arguments are the same. | |
5256 | TODO: Allow different constants for different vector | |
5257 | stmts generated for an SLP instance. */ | |
5258 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 5259 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5260 | } |
5261 | } | |
5262 | } | |
5263 | ||
5264 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
5265 | (a special case for certain kind of vector shifts); otherwise, | |
5266 | operand 1 should be of a vector type (the usual case). */ | |
5267 | if (vec_oprnd1) | |
5268 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5269 | slp_node); |
9dc3f7de IR |
5270 | else |
5271 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
306b0c92 | 5272 | slp_node); |
9dc3f7de IR |
5273 | } |
5274 | else | |
5275 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5276 | ||
5277 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 5278 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 5279 | { |
9771b263 | 5280 | vop1 = vec_oprnds1[i]; |
0d0e4a03 | 5281 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
9dc3f7de IR |
5282 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5283 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5284 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5285 | if (slp_node) | |
9771b263 | 5286 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
5287 | } |
5288 | ||
5289 | if (slp_node) | |
5290 | continue; | |
5291 | ||
5292 | if (j == 0) | |
5293 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5294 | else | |
5295 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5296 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5297 | } | |
5298 | ||
9771b263 DN |
5299 | vec_oprnds0.release (); |
5300 | vec_oprnds1.release (); | |
9dc3f7de IR |
5301 | |
5302 | return true; | |
5303 | } | |
5304 | ||
5305 | ||
ebfd146a IR |
5306 | /* Function vectorizable_operation. |
5307 | ||
16949072 RG |
5308 | Check if STMT performs a binary, unary or ternary operation that can |
5309 | be vectorized. | |
b8698a0f | 5310 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
5311 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5312 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5313 | ||
5314 | static bool | |
355fe088 TS |
5315 | vectorizable_operation (gimple *stmt, gimple_stmt_iterator *gsi, |
5316 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a | 5317 | { |
00f07b86 | 5318 | tree vec_dest; |
ebfd146a | 5319 | tree scalar_dest; |
16949072 | 5320 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 5321 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 5322 | tree vectype; |
ebfd146a | 5323 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
0eb952ea | 5324 | enum tree_code code, orig_code; |
ef4bddc2 | 5325 | machine_mode vec_mode; |
ebfd146a IR |
5326 | tree new_temp; |
5327 | int op_type; | |
00f07b86 | 5328 | optab optab; |
523ba738 | 5329 | bool target_support_p; |
355fe088 | 5330 | gimple *def_stmt; |
16949072 RG |
5331 | enum vect_def_type dt[3] |
5332 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 5333 | int ndts = 3; |
355fe088 | 5334 | gimple *new_stmt = NULL; |
ebfd146a | 5335 | stmt_vec_info prev_stmt_info; |
928686b1 RS |
5336 | poly_uint64 nunits_in; |
5337 | poly_uint64 nunits_out; | |
ebfd146a IR |
5338 | tree vectype_out; |
5339 | int ncopies; | |
5340 | int j, i; | |
6e1aa848 DN |
5341 | vec<tree> vec_oprnds0 = vNULL; |
5342 | vec<tree> vec_oprnds1 = vNULL; | |
5343 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 5344 | tree vop0, vop1, vop2; |
a70d6342 | 5345 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5346 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 5347 | |
a70d6342 | 5348 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5349 | return false; |
5350 | ||
66c16fd9 RB |
5351 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5352 | && ! vec_stmt) | |
ebfd146a IR |
5353 | return false; |
5354 | ||
5355 | /* Is STMT a vectorizable binary/unary operation? */ | |
5356 | if (!is_gimple_assign (stmt)) | |
5357 | return false; | |
5358 | ||
5359 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
5360 | return false; | |
5361 | ||
0eb952ea | 5362 | orig_code = code = gimple_assign_rhs_code (stmt); |
ebfd146a | 5363 | |
1af4ebf5 MG |
5364 | /* For pointer addition and subtraction, we should use the normal |
5365 | plus and minus for the vector operation. */ | |
ebfd146a IR |
5366 | if (code == POINTER_PLUS_EXPR) |
5367 | code = PLUS_EXPR; | |
1af4ebf5 MG |
5368 | if (code == POINTER_DIFF_EXPR) |
5369 | code = MINUS_EXPR; | |
ebfd146a IR |
5370 | |
5371 | /* Support only unary or binary operations. */ | |
5372 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 5373 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 5374 | { |
73fbfcad | 5375 | if (dump_enabled_p ()) |
78c60e3d | 5376 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5377 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 5378 | op_type); |
ebfd146a IR |
5379 | return false; |
5380 | } | |
5381 | ||
b690cc0f RG |
5382 | scalar_dest = gimple_assign_lhs (stmt); |
5383 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
5384 | ||
7b7b1813 RG |
5385 | /* Most operations cannot handle bit-precision types without extra |
5386 | truncations. */ | |
045c1278 | 5387 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
2be65d9e | 5388 | && !type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
7b7b1813 RG |
5389 | /* Exception are bitwise binary operations. */ |
5390 | && code != BIT_IOR_EXPR | |
5391 | && code != BIT_XOR_EXPR | |
5392 | && code != BIT_AND_EXPR) | |
5393 | { | |
73fbfcad | 5394 | if (dump_enabled_p ()) |
78c60e3d | 5395 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5396 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
5397 | return false; |
5398 | } | |
5399 | ||
ebfd146a | 5400 | op0 = gimple_assign_rhs1 (stmt); |
81c40241 | 5401 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
ebfd146a | 5402 | { |
73fbfcad | 5403 | if (dump_enabled_p ()) |
78c60e3d | 5404 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5405 | "use not simple.\n"); |
ebfd146a IR |
5406 | return false; |
5407 | } | |
b690cc0f RG |
5408 | /* If op0 is an external or constant def use a vector type with |
5409 | the same size as the output vector type. */ | |
5410 | if (!vectype) | |
b036c6c5 IE |
5411 | { |
5412 | /* For boolean type we cannot determine vectype by | |
5413 | invariant value (don't know whether it is a vector | |
5414 | of booleans or vector of integers). We use output | |
5415 | vectype because operations on boolean don't change | |
5416 | type. */ | |
2568d8a1 | 5417 | if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op0))) |
b036c6c5 | 5418 | { |
2568d8a1 | 5419 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (scalar_dest))) |
b036c6c5 IE |
5420 | { |
5421 | if (dump_enabled_p ()) | |
5422 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5423 | "not supported operation on bool value.\n"); | |
5424 | return false; | |
5425 | } | |
5426 | vectype = vectype_out; | |
5427 | } | |
5428 | else | |
5429 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
5430 | } | |
7d8930a0 IR |
5431 | if (vec_stmt) |
5432 | gcc_assert (vectype); | |
5433 | if (!vectype) | |
5434 | { | |
73fbfcad | 5435 | if (dump_enabled_p ()) |
7d8930a0 | 5436 | { |
78c60e3d SS |
5437 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5438 | "no vectype for scalar type "); | |
5439 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
5440 | TREE_TYPE (op0)); | |
e645e942 | 5441 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
5442 | } |
5443 | ||
5444 | return false; | |
5445 | } | |
b690cc0f RG |
5446 | |
5447 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
5448 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
928686b1 | 5449 | if (maybe_ne (nunits_out, nunits_in)) |
b690cc0f | 5450 | return false; |
ebfd146a | 5451 | |
16949072 | 5452 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
5453 | { |
5454 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 5455 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1])) |
ebfd146a | 5456 | { |
73fbfcad | 5457 | if (dump_enabled_p ()) |
78c60e3d | 5458 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5459 | "use not simple.\n"); |
ebfd146a IR |
5460 | return false; |
5461 | } | |
5462 | } | |
16949072 RG |
5463 | if (op_type == ternary_op) |
5464 | { | |
5465 | op2 = gimple_assign_rhs3 (stmt); | |
81c40241 | 5466 | if (!vect_is_simple_use (op2, vinfo, &def_stmt, &dt[2])) |
16949072 | 5467 | { |
73fbfcad | 5468 | if (dump_enabled_p ()) |
78c60e3d | 5469 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5470 | "use not simple.\n"); |
16949072 RG |
5471 | return false; |
5472 | } | |
5473 | } | |
ebfd146a | 5474 | |
b690cc0f | 5475 | /* Multiple types in SLP are handled by creating the appropriate number of |
ff802fa1 | 5476 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 5477 | case of SLP. */ |
fce57248 | 5478 | if (slp_node) |
b690cc0f RG |
5479 | ncopies = 1; |
5480 | else | |
e8f142e2 | 5481 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
b690cc0f RG |
5482 | |
5483 | gcc_assert (ncopies >= 1); | |
5484 | ||
9dc3f7de | 5485 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
5486 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
5487 | || code == RROTATE_EXPR) | |
9dc3f7de | 5488 | return false; |
ebfd146a | 5489 | |
ebfd146a | 5490 | /* Supportable by target? */ |
00f07b86 RH |
5491 | |
5492 | vec_mode = TYPE_MODE (vectype); | |
5493 | if (code == MULT_HIGHPART_EXPR) | |
523ba738 | 5494 | target_support_p = can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype)); |
00f07b86 RH |
5495 | else |
5496 | { | |
5497 | optab = optab_for_tree_code (code, vectype, optab_default); | |
5498 | if (!optab) | |
5deb57cb | 5499 | { |
73fbfcad | 5500 | if (dump_enabled_p ()) |
78c60e3d | 5501 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5502 | "no optab.\n"); |
00f07b86 | 5503 | return false; |
5deb57cb | 5504 | } |
523ba738 RS |
5505 | target_support_p = (optab_handler (optab, vec_mode) |
5506 | != CODE_FOR_nothing); | |
5deb57cb JJ |
5507 | } |
5508 | ||
523ba738 | 5509 | if (!target_support_p) |
ebfd146a | 5510 | { |
73fbfcad | 5511 | if (dump_enabled_p ()) |
78c60e3d | 5512 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5513 | "op not supported by target.\n"); |
ebfd146a | 5514 | /* Check only during analysis. */ |
cf098191 | 5515 | if (maybe_ne (GET_MODE_SIZE (vec_mode), UNITS_PER_WORD) |
ca09abcb | 5516 | || (!vec_stmt && !vect_worthwhile_without_simd_p (vinfo, code))) |
ebfd146a | 5517 | return false; |
73fbfcad | 5518 | if (dump_enabled_p ()) |
e645e942 TJ |
5519 | dump_printf_loc (MSG_NOTE, vect_location, |
5520 | "proceeding using word mode.\n"); | |
383d9c83 IR |
5521 | } |
5522 | ||
4a00c761 | 5523 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
5524 | if (!VECTOR_MODE_P (vec_mode) |
5525 | && !vec_stmt | |
ca09abcb | 5526 | && !vect_worthwhile_without_simd_p (vinfo, code)) |
7d8930a0 | 5527 | { |
73fbfcad | 5528 | if (dump_enabled_p ()) |
78c60e3d | 5529 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5530 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 5531 | return false; |
7d8930a0 | 5532 | } |
ebfd146a | 5533 | |
ebfd146a IR |
5534 | if (!vec_stmt) /* transformation not required. */ |
5535 | { | |
4a00c761 | 5536 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 5537 | if (dump_enabled_p ()) |
78c60e3d | 5538 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5539 | "=== vectorizable_operation ===\n"); |
4fc5ebf1 | 5540 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); |
ebfd146a IR |
5541 | return true; |
5542 | } | |
5543 | ||
67b8dbac | 5544 | /* Transform. */ |
ebfd146a | 5545 | |
73fbfcad | 5546 | if (dump_enabled_p ()) |
78c60e3d | 5547 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5548 | "transform binary/unary operation.\n"); |
383d9c83 | 5549 | |
ebfd146a | 5550 | /* Handle def. */ |
00f07b86 | 5551 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 5552 | |
0eb952ea JJ |
5553 | /* POINTER_DIFF_EXPR has pointer arguments which are vectorized as |
5554 | vectors with unsigned elements, but the result is signed. So, we | |
5555 | need to compute the MINUS_EXPR into vectype temporary and | |
5556 | VIEW_CONVERT_EXPR it into the final vectype_out result. */ | |
5557 | tree vec_cvt_dest = NULL_TREE; | |
5558 | if (orig_code == POINTER_DIFF_EXPR) | |
5559 | vec_cvt_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
5560 | ||
ebfd146a IR |
5561 | /* In case the vectorization factor (VF) is bigger than the number |
5562 | of elements that we can fit in a vectype (nunits), we have to generate | |
5563 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
5564 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
5565 | from one copy of the vector stmt to the next, in the field | |
5566 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
5567 | stages to find the correct vector defs to be used when vectorizing | |
5568 | stmts that use the defs of the current stmt. The example below | |
5569 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
5570 | we need to create 4 vectorized stmts): | |
5571 | ||
5572 | before vectorization: | |
5573 | RELATED_STMT VEC_STMT | |
5574 | S1: x = memref - - | |
5575 | S2: z = x + 1 - - | |
5576 | ||
5577 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
5578 | there): | |
5579 | RELATED_STMT VEC_STMT | |
5580 | VS1_0: vx0 = memref0 VS1_1 - | |
5581 | VS1_1: vx1 = memref1 VS1_2 - | |
5582 | VS1_2: vx2 = memref2 VS1_3 - | |
5583 | VS1_3: vx3 = memref3 - - | |
5584 | S1: x = load - VS1_0 | |
5585 | S2: z = x + 1 - - | |
5586 | ||
5587 | step2: vectorize stmt S2 (done here): | |
5588 | To vectorize stmt S2 we first need to find the relevant vector | |
5589 | def for the first operand 'x'. This is, as usual, obtained from | |
5590 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
5591 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
5592 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
5593 | the vector stmt VS2_0, and as usual, record it in the | |
5594 | STMT_VINFO_VEC_STMT of stmt S2. | |
5595 | When creating the second copy (VS2_1), we obtain the relevant vector | |
5596 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
5597 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
5598 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
5599 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
5600 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
5601 | chain of stmts and pointers: | |
5602 | RELATED_STMT VEC_STMT | |
5603 | VS1_0: vx0 = memref0 VS1_1 - | |
5604 | VS1_1: vx1 = memref1 VS1_2 - | |
5605 | VS1_2: vx2 = memref2 VS1_3 - | |
5606 | VS1_3: vx3 = memref3 - - | |
5607 | S1: x = load - VS1_0 | |
5608 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
5609 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
5610 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
5611 | VS2_3: vz3 = vx3 + v1 - - | |
5612 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
5613 | |
5614 | prev_stmt_info = NULL; | |
5615 | for (j = 0; j < ncopies; j++) | |
5616 | { | |
5617 | /* Handle uses. */ | |
5618 | if (j == 0) | |
4a00c761 JJ |
5619 | { |
5620 | if (op_type == binary_op || op_type == ternary_op) | |
5621 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
306b0c92 | 5622 | slp_node); |
4a00c761 JJ |
5623 | else |
5624 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5625 | slp_node); |
4a00c761 | 5626 | if (op_type == ternary_op) |
c392943c | 5627 | vect_get_vec_defs (op2, NULL_TREE, stmt, &vec_oprnds2, NULL, |
306b0c92 | 5628 | slp_node); |
4a00c761 | 5629 | } |
ebfd146a | 5630 | else |
4a00c761 JJ |
5631 | { |
5632 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5633 | if (op_type == ternary_op) | |
5634 | { | |
9771b263 DN |
5635 | tree vec_oprnd = vec_oprnds2.pop (); |
5636 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
5637 | vec_oprnd)); | |
4a00c761 JJ |
5638 | } |
5639 | } | |
5640 | ||
5641 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 5642 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 5643 | { |
4a00c761 | 5644 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 5645 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 5646 | vop2 = ((op_type == ternary_op) |
9771b263 | 5647 | ? vec_oprnds2[i] : NULL_TREE); |
0d0e4a03 | 5648 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1, vop2); |
4a00c761 JJ |
5649 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5650 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5651 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
0eb952ea JJ |
5652 | if (vec_cvt_dest) |
5653 | { | |
5654 | new_temp = build1 (VIEW_CONVERT_EXPR, vectype_out, new_temp); | |
5655 | new_stmt = gimple_build_assign (vec_cvt_dest, VIEW_CONVERT_EXPR, | |
5656 | new_temp); | |
5657 | new_temp = make_ssa_name (vec_cvt_dest, new_stmt); | |
5658 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5659 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5660 | } | |
4a00c761 | 5661 | if (slp_node) |
9771b263 | 5662 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
5663 | } |
5664 | ||
4a00c761 JJ |
5665 | if (slp_node) |
5666 | continue; | |
5667 | ||
5668 | if (j == 0) | |
5669 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5670 | else | |
5671 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5672 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
5673 | } |
5674 | ||
9771b263 DN |
5675 | vec_oprnds0.release (); |
5676 | vec_oprnds1.release (); | |
5677 | vec_oprnds2.release (); | |
ebfd146a | 5678 | |
ebfd146a IR |
5679 | return true; |
5680 | } | |
5681 | ||
f702e7d4 | 5682 | /* A helper function to ensure data reference DR's base alignment. */ |
c716e67f XDL |
5683 | |
5684 | static void | |
f702e7d4 | 5685 | ensure_base_align (struct data_reference *dr) |
c716e67f XDL |
5686 | { |
5687 | if (!dr->aux) | |
5688 | return; | |
5689 | ||
52639a61 | 5690 | if (DR_VECT_AUX (dr)->base_misaligned) |
c716e67f | 5691 | { |
52639a61 | 5692 | tree base_decl = DR_VECT_AUX (dr)->base_decl; |
c716e67f | 5693 | |
f702e7d4 RS |
5694 | unsigned int align_base_to = DR_TARGET_ALIGNMENT (dr) * BITS_PER_UNIT; |
5695 | ||
428f0c67 | 5696 | if (decl_in_symtab_p (base_decl)) |
f702e7d4 | 5697 | symtab_node::get (base_decl)->increase_alignment (align_base_to); |
428f0c67 JH |
5698 | else |
5699 | { | |
f702e7d4 | 5700 | SET_DECL_ALIGN (base_decl, align_base_to); |
428f0c67 JH |
5701 | DECL_USER_ALIGN (base_decl) = 1; |
5702 | } | |
52639a61 | 5703 | DR_VECT_AUX (dr)->base_misaligned = false; |
c716e67f XDL |
5704 | } |
5705 | } | |
5706 | ||
ebfd146a | 5707 | |
44fc7854 BE |
5708 | /* Function get_group_alias_ptr_type. |
5709 | ||
5710 | Return the alias type for the group starting at FIRST_STMT. */ | |
5711 | ||
5712 | static tree | |
5713 | get_group_alias_ptr_type (gimple *first_stmt) | |
5714 | { | |
5715 | struct data_reference *first_dr, *next_dr; | |
5716 | gimple *next_stmt; | |
5717 | ||
5718 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
5719 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (first_stmt)); | |
5720 | while (next_stmt) | |
5721 | { | |
5722 | next_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (next_stmt)); | |
5723 | if (get_alias_set (DR_REF (first_dr)) | |
5724 | != get_alias_set (DR_REF (next_dr))) | |
5725 | { | |
5726 | if (dump_enabled_p ()) | |
5727 | dump_printf_loc (MSG_NOTE, vect_location, | |
5728 | "conflicting alias set types.\n"); | |
5729 | return ptr_type_node; | |
5730 | } | |
5731 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); | |
5732 | } | |
5733 | return reference_alias_ptr_type (DR_REF (first_dr)); | |
5734 | } | |
5735 | ||
5736 | ||
ebfd146a IR |
5737 | /* Function vectorizable_store. |
5738 | ||
b8698a0f L |
5739 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
5740 | can be vectorized. | |
5741 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
5742 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5743 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5744 | ||
5745 | static bool | |
355fe088 | 5746 | vectorizable_store (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 5747 | slp_tree slp_node) |
ebfd146a | 5748 | { |
ebfd146a IR |
5749 | tree data_ref; |
5750 | tree op; | |
5751 | tree vec_oprnd = NULL_TREE; | |
5752 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5753 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
272c6793 | 5754 | tree elem_type; |
ebfd146a | 5755 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 5756 | struct loop *loop = NULL; |
ef4bddc2 | 5757 | machine_mode vec_mode; |
ebfd146a IR |
5758 | tree dummy; |
5759 | enum dr_alignment_support alignment_support_scheme; | |
355fe088 | 5760 | gimple *def_stmt; |
ebfd146a IR |
5761 | enum vect_def_type dt; |
5762 | stmt_vec_info prev_stmt_info = NULL; | |
5763 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 5764 | tree dataref_offset = NULL_TREE; |
355fe088 | 5765 | gimple *ptr_incr = NULL; |
ebfd146a IR |
5766 | int ncopies; |
5767 | int j; | |
2de001ee RS |
5768 | gimple *next_stmt, *first_stmt; |
5769 | bool grouped_store; | |
ebfd146a | 5770 | unsigned int group_size, i; |
6e1aa848 DN |
5771 | vec<tree> oprnds = vNULL; |
5772 | vec<tree> result_chain = vNULL; | |
ebfd146a | 5773 | bool inv_p; |
09dfa495 | 5774 | tree offset = NULL_TREE; |
6e1aa848 | 5775 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 5776 | bool slp = (slp_node != NULL); |
ebfd146a | 5777 | unsigned int vec_num; |
a70d6342 | 5778 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5779 | vec_info *vinfo = stmt_info->vinfo; |
272c6793 | 5780 | tree aggr_type; |
134c85ca | 5781 | gather_scatter_info gs_info; |
3bab6342 | 5782 | enum vect_def_type scatter_src_dt = vect_unknown_def_type; |
355fe088 | 5783 | gimple *new_stmt; |
d9f21f6a | 5784 | poly_uint64 vf; |
2de001ee | 5785 | vec_load_store_type vls_type; |
44fc7854 | 5786 | tree ref_type; |
a70d6342 | 5787 | |
a70d6342 | 5788 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5789 | return false; |
5790 | ||
66c16fd9 RB |
5791 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5792 | && ! vec_stmt) | |
ebfd146a IR |
5793 | return false; |
5794 | ||
5795 | /* Is vectorizable store? */ | |
5796 | ||
c3a8f964 RS |
5797 | tree mask = NULL_TREE, mask_vectype = NULL_TREE; |
5798 | if (is_gimple_assign (stmt)) | |
5799 | { | |
5800 | tree scalar_dest = gimple_assign_lhs (stmt); | |
5801 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR | |
5802 | && is_pattern_stmt_p (stmt_info)) | |
5803 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
5804 | if (TREE_CODE (scalar_dest) != ARRAY_REF | |
5805 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF | |
5806 | && TREE_CODE (scalar_dest) != INDIRECT_REF | |
5807 | && TREE_CODE (scalar_dest) != COMPONENT_REF | |
5808 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
5809 | && TREE_CODE (scalar_dest) != REALPART_EXPR | |
5810 | && TREE_CODE (scalar_dest) != MEM_REF) | |
5811 | return false; | |
5812 | } | |
5813 | else | |
5814 | { | |
5815 | gcall *call = dyn_cast <gcall *> (stmt); | |
5816 | if (!call || !gimple_call_internal_p (call, IFN_MASK_STORE)) | |
5817 | return false; | |
ebfd146a | 5818 | |
c3a8f964 RS |
5819 | if (slp_node != NULL) |
5820 | { | |
5821 | if (dump_enabled_p ()) | |
5822 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5823 | "SLP of masked stores not supported.\n"); | |
5824 | return false; | |
5825 | } | |
5826 | ||
5827 | ref_type = TREE_TYPE (gimple_call_arg (call, 1)); | |
5828 | mask = gimple_call_arg (call, 2); | |
5829 | if (!vect_check_load_store_mask (stmt, mask, &mask_vectype)) | |
5830 | return false; | |
5831 | } | |
5832 | ||
5833 | op = vect_get_store_rhs (stmt); | |
ebfd146a | 5834 | |
fce57248 RS |
5835 | /* Cannot have hybrid store SLP -- that would mean storing to the |
5836 | same location twice. */ | |
5837 | gcc_assert (slp == PURE_SLP_STMT (stmt_info)); | |
5838 | ||
f4d09712 | 5839 | tree vectype = STMT_VINFO_VECTYPE (stmt_info), rhs_vectype = NULL_TREE; |
4d694b27 | 5840 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 JJ |
5841 | |
5842 | if (loop_vinfo) | |
b17dc4d4 RB |
5843 | { |
5844 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
5845 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
5846 | } | |
5847 | else | |
5848 | vf = 1; | |
465c8c19 JJ |
5849 | |
5850 | /* Multiple types in SLP are handled by creating the appropriate number of | |
5851 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5852 | case of SLP. */ | |
fce57248 | 5853 | if (slp) |
465c8c19 JJ |
5854 | ncopies = 1; |
5855 | else | |
e8f142e2 | 5856 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
5857 | |
5858 | gcc_assert (ncopies >= 1); | |
5859 | ||
5860 | /* FORNOW. This restriction should be relaxed. */ | |
5861 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) | |
5862 | { | |
5863 | if (dump_enabled_p ()) | |
5864 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5865 | "multiple types in nested loop.\n"); | |
5866 | return false; | |
5867 | } | |
5868 | ||
3133c3b6 | 5869 | if (!vect_check_store_rhs (stmt, op, &rhs_vectype, &vls_type)) |
f4d09712 KY |
5870 | return false; |
5871 | ||
272c6793 | 5872 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 5873 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 5874 | |
ebfd146a IR |
5875 | if (!STMT_VINFO_DATA_REF (stmt_info)) |
5876 | return false; | |
5877 | ||
2de001ee | 5878 | vect_memory_access_type memory_access_type; |
7e11fc7f | 5879 | if (!get_load_store_type (stmt, vectype, slp, mask, vls_type, ncopies, |
2de001ee RS |
5880 | &memory_access_type, &gs_info)) |
5881 | return false; | |
3bab6342 | 5882 | |
c3a8f964 RS |
5883 | if (mask) |
5884 | { | |
7e11fc7f RS |
5885 | if (memory_access_type == VMAT_CONTIGUOUS) |
5886 | { | |
5887 | if (!VECTOR_MODE_P (vec_mode) | |
5888 | || !can_vec_mask_load_store_p (vec_mode, | |
5889 | TYPE_MODE (mask_vectype), false)) | |
5890 | return false; | |
5891 | } | |
5892 | else if (memory_access_type != VMAT_LOAD_STORE_LANES) | |
c3a8f964 RS |
5893 | { |
5894 | if (dump_enabled_p ()) | |
5895 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5896 | "unsupported access type for masked store.\n"); | |
5897 | return false; | |
5898 | } | |
c3a8f964 RS |
5899 | } |
5900 | else | |
5901 | { | |
5902 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
5903 | (e.g. - array initialization with 0). */ | |
5904 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) | |
5905 | return false; | |
5906 | } | |
5907 | ||
7cfb4d93 RS |
5908 | grouped_store = STMT_VINFO_GROUPED_ACCESS (stmt_info); |
5909 | if (grouped_store) | |
5910 | { | |
5911 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); | |
5912 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
5913 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
5914 | } | |
5915 | else | |
5916 | { | |
5917 | first_stmt = stmt; | |
5918 | first_dr = dr; | |
5919 | group_size = vec_num = 1; | |
5920 | } | |
5921 | ||
ebfd146a IR |
5922 | if (!vec_stmt) /* transformation not required. */ |
5923 | { | |
2de001ee | 5924 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; |
7cfb4d93 RS |
5925 | |
5926 | if (loop_vinfo | |
5927 | && LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo)) | |
5928 | check_load_store_masking (loop_vinfo, vectype, vls_type, group_size, | |
5929 | memory_access_type); | |
5930 | ||
ebfd146a | 5931 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; |
2e8ab70c RB |
5932 | /* The SLP costs are calculated during SLP analysis. */ |
5933 | if (!PURE_SLP_STMT (stmt_info)) | |
9ce4345a RS |
5934 | vect_model_store_cost (stmt_info, ncopies, memory_access_type, |
5935 | vls_type, NULL, NULL, NULL); | |
ebfd146a IR |
5936 | return true; |
5937 | } | |
2de001ee | 5938 | gcc_assert (memory_access_type == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); |
ebfd146a | 5939 | |
67b8dbac | 5940 | /* Transform. */ |
ebfd146a | 5941 | |
f702e7d4 | 5942 | ensure_base_align (dr); |
c716e67f | 5943 | |
2de001ee | 5944 | if (memory_access_type == VMAT_GATHER_SCATTER) |
3bab6342 | 5945 | { |
c3a8f964 | 5946 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE, src; |
134c85ca | 5947 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
5948 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; |
5949 | tree ptr, mask, var, scale, perm_mask = NULL_TREE; | |
5950 | edge pe = loop_preheader_edge (loop); | |
5951 | gimple_seq seq; | |
5952 | basic_block new_bb; | |
5953 | enum { NARROW, NONE, WIDEN } modifier; | |
4d694b27 RS |
5954 | poly_uint64 scatter_off_nunits |
5955 | = TYPE_VECTOR_SUBPARTS (gs_info.offset_vectype); | |
3bab6342 | 5956 | |
4d694b27 | 5957 | if (known_eq (nunits, scatter_off_nunits)) |
3bab6342 | 5958 | modifier = NONE; |
4d694b27 | 5959 | else if (known_eq (nunits * 2, scatter_off_nunits)) |
3bab6342 | 5960 | { |
3bab6342 AT |
5961 | modifier = WIDEN; |
5962 | ||
4d694b27 RS |
5963 | /* Currently gathers and scatters are only supported for |
5964 | fixed-length vectors. */ | |
5965 | unsigned int count = scatter_off_nunits.to_constant (); | |
5966 | vec_perm_builder sel (count, count, 1); | |
5967 | for (i = 0; i < (unsigned int) count; ++i) | |
5968 | sel.quick_push (i | (count / 2)); | |
3bab6342 | 5969 | |
4d694b27 | 5970 | vec_perm_indices indices (sel, 1, count); |
e3342de4 RS |
5971 | perm_mask = vect_gen_perm_mask_checked (gs_info.offset_vectype, |
5972 | indices); | |
3bab6342 AT |
5973 | gcc_assert (perm_mask != NULL_TREE); |
5974 | } | |
4d694b27 | 5975 | else if (known_eq (nunits, scatter_off_nunits * 2)) |
3bab6342 | 5976 | { |
3bab6342 AT |
5977 | modifier = NARROW; |
5978 | ||
4d694b27 RS |
5979 | /* Currently gathers and scatters are only supported for |
5980 | fixed-length vectors. */ | |
5981 | unsigned int count = nunits.to_constant (); | |
5982 | vec_perm_builder sel (count, count, 1); | |
5983 | for (i = 0; i < (unsigned int) count; ++i) | |
5984 | sel.quick_push (i | (count / 2)); | |
3bab6342 | 5985 | |
4d694b27 | 5986 | vec_perm_indices indices (sel, 2, count); |
e3342de4 | 5987 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); |
3bab6342 AT |
5988 | gcc_assert (perm_mask != NULL_TREE); |
5989 | ncopies *= 2; | |
5990 | } | |
5991 | else | |
5992 | gcc_unreachable (); | |
5993 | ||
134c85ca | 5994 | rettype = TREE_TYPE (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
5995 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
5996 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5997 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5998 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5999 | scaletype = TREE_VALUE (arglist); | |
6000 | ||
6001 | gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE | |
6002 | && TREE_CODE (rettype) == VOID_TYPE); | |
6003 | ||
134c85ca | 6004 | ptr = fold_convert (ptrtype, gs_info.base); |
3bab6342 AT |
6005 | if (!is_gimple_min_invariant (ptr)) |
6006 | { | |
6007 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
6008 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
6009 | gcc_assert (!new_bb); | |
6010 | } | |
6011 | ||
6012 | /* Currently we support only unconditional scatter stores, | |
6013 | so mask should be all ones. */ | |
6014 | mask = build_int_cst (masktype, -1); | |
6015 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
6016 | ||
134c85ca | 6017 | scale = build_int_cst (scaletype, gs_info.scale); |
3bab6342 AT |
6018 | |
6019 | prev_stmt_info = NULL; | |
6020 | for (j = 0; j < ncopies; ++j) | |
6021 | { | |
6022 | if (j == 0) | |
6023 | { | |
6024 | src = vec_oprnd1 | |
c3a8f964 | 6025 | = vect_get_vec_def_for_operand (op, stmt); |
3bab6342 | 6026 | op = vec_oprnd0 |
134c85ca | 6027 | = vect_get_vec_def_for_operand (gs_info.offset, stmt); |
3bab6342 AT |
6028 | } |
6029 | else if (modifier != NONE && (j & 1)) | |
6030 | { | |
6031 | if (modifier == WIDEN) | |
6032 | { | |
6033 | src = vec_oprnd1 | |
6034 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
6035 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, perm_mask, | |
6036 | stmt, gsi); | |
6037 | } | |
6038 | else if (modifier == NARROW) | |
6039 | { | |
6040 | src = permute_vec_elements (vec_oprnd1, vec_oprnd1, perm_mask, | |
6041 | stmt, gsi); | |
6042 | op = vec_oprnd0 | |
134c85ca RS |
6043 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
6044 | vec_oprnd0); | |
3bab6342 AT |
6045 | } |
6046 | else | |
6047 | gcc_unreachable (); | |
6048 | } | |
6049 | else | |
6050 | { | |
6051 | src = vec_oprnd1 | |
6052 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
6053 | op = vec_oprnd0 | |
134c85ca RS |
6054 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
6055 | vec_oprnd0); | |
3bab6342 AT |
6056 | } |
6057 | ||
6058 | if (!useless_type_conversion_p (srctype, TREE_TYPE (src))) | |
6059 | { | |
928686b1 RS |
6060 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)), |
6061 | TYPE_VECTOR_SUBPARTS (srctype))); | |
0e22bb5a | 6062 | var = vect_get_new_ssa_name (srctype, vect_simple_var); |
3bab6342 AT |
6063 | src = build1 (VIEW_CONVERT_EXPR, srctype, src); |
6064 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, src); | |
6065 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6066 | src = var; | |
6067 | } | |
6068 | ||
6069 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
6070 | { | |
928686b1 RS |
6071 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)), |
6072 | TYPE_VECTOR_SUBPARTS (idxtype))); | |
0e22bb5a | 6073 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
3bab6342 AT |
6074 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
6075 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
6076 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6077 | op = var; | |
6078 | } | |
6079 | ||
6080 | new_stmt | |
134c85ca | 6081 | = gimple_build_call (gs_info.decl, 5, ptr, mask, op, src, scale); |
3bab6342 AT |
6082 | |
6083 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6084 | ||
6085 | if (prev_stmt_info == NULL) | |
6086 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6087 | else | |
6088 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6089 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6090 | } | |
6091 | return true; | |
6092 | } | |
6093 | ||
0d0293ac | 6094 | if (grouped_store) |
ebfd146a | 6095 | { |
e14c1050 | 6096 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
6097 | |
6098 | /* FORNOW */ | |
a70d6342 | 6099 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
6100 | |
6101 | /* We vectorize all the stmts of the interleaving group when we | |
6102 | reach the last stmt in the group. */ | |
e14c1050 IR |
6103 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
6104 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
6105 | && !slp) |
6106 | { | |
6107 | *vec_stmt = NULL; | |
6108 | return true; | |
6109 | } | |
6110 | ||
6111 | if (slp) | |
4b5caab7 | 6112 | { |
0d0293ac | 6113 | grouped_store = false; |
4b5caab7 IR |
6114 | /* VEC_NUM is the number of vect stmts to be created for this |
6115 | group. */ | |
6116 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 6117 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
52eab378 | 6118 | gcc_assert (GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_stmt)) == first_stmt); |
4b5caab7 | 6119 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
c3a8f964 | 6120 | op = vect_get_store_rhs (first_stmt); |
4b5caab7 | 6121 | } |
ebfd146a | 6122 | else |
4b5caab7 IR |
6123 | /* VEC_NUM is the number of vect stmts to be created for this |
6124 | group. */ | |
ebfd146a | 6125 | vec_num = group_size; |
44fc7854 BE |
6126 | |
6127 | ref_type = get_group_alias_ptr_type (first_stmt); | |
ebfd146a | 6128 | } |
b8698a0f | 6129 | else |
7cfb4d93 | 6130 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); |
b8698a0f | 6131 | |
73fbfcad | 6132 | if (dump_enabled_p ()) |
78c60e3d | 6133 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6134 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 6135 | |
2de001ee RS |
6136 | if (memory_access_type == VMAT_ELEMENTWISE |
6137 | || memory_access_type == VMAT_STRIDED_SLP) | |
f2e2a985 MM |
6138 | { |
6139 | gimple_stmt_iterator incr_gsi; | |
6140 | bool insert_after; | |
355fe088 | 6141 | gimple *incr; |
f2e2a985 MM |
6142 | tree offvar; |
6143 | tree ivstep; | |
6144 | tree running_off; | |
6145 | gimple_seq stmts = NULL; | |
6146 | tree stride_base, stride_step, alias_off; | |
6147 | tree vec_oprnd; | |
f502d50e | 6148 | unsigned int g; |
4d694b27 RS |
6149 | /* Checked by get_load_store_type. */ |
6150 | unsigned int const_nunits = nunits.to_constant (); | |
f2e2a985 | 6151 | |
7cfb4d93 | 6152 | gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); |
f2e2a985 MM |
6153 | gcc_assert (!nested_in_vect_loop_p (loop, stmt)); |
6154 | ||
6155 | stride_base | |
6156 | = fold_build_pointer_plus | |
f502d50e | 6157 | (unshare_expr (DR_BASE_ADDRESS (first_dr)), |
f2e2a985 | 6158 | size_binop (PLUS_EXPR, |
f502d50e | 6159 | convert_to_ptrofftype (unshare_expr (DR_OFFSET (first_dr))), |
44fc7854 | 6160 | convert_to_ptrofftype (DR_INIT (first_dr)))); |
f502d50e | 6161 | stride_step = fold_convert (sizetype, unshare_expr (DR_STEP (first_dr))); |
f2e2a985 MM |
6162 | |
6163 | /* For a store with loop-invariant (but other than power-of-2) | |
6164 | stride (i.e. not a grouped access) like so: | |
6165 | ||
6166 | for (i = 0; i < n; i += stride) | |
6167 | array[i] = ...; | |
6168 | ||
6169 | we generate a new induction variable and new stores from | |
6170 | the components of the (vectorized) rhs: | |
6171 | ||
6172 | for (j = 0; ; j += VF*stride) | |
6173 | vectemp = ...; | |
6174 | tmp1 = vectemp[0]; | |
6175 | array[j] = tmp1; | |
6176 | tmp2 = vectemp[1]; | |
6177 | array[j + stride] = tmp2; | |
6178 | ... | |
6179 | */ | |
6180 | ||
4d694b27 | 6181 | unsigned nstores = const_nunits; |
b17dc4d4 | 6182 | unsigned lnel = 1; |
cee62fee | 6183 | tree ltype = elem_type; |
04199738 | 6184 | tree lvectype = vectype; |
cee62fee MM |
6185 | if (slp) |
6186 | { | |
4d694b27 RS |
6187 | if (group_size < const_nunits |
6188 | && const_nunits % group_size == 0) | |
b17dc4d4 | 6189 | { |
4d694b27 | 6190 | nstores = const_nunits / group_size; |
b17dc4d4 RB |
6191 | lnel = group_size; |
6192 | ltype = build_vector_type (elem_type, group_size); | |
04199738 RB |
6193 | lvectype = vectype; |
6194 | ||
6195 | /* First check if vec_extract optab doesn't support extraction | |
6196 | of vector elts directly. */ | |
b397965c | 6197 | scalar_mode elmode = SCALAR_TYPE_MODE (elem_type); |
9da15d40 RS |
6198 | machine_mode vmode; |
6199 | if (!mode_for_vector (elmode, group_size).exists (&vmode) | |
6200 | || !VECTOR_MODE_P (vmode) | |
04199738 RB |
6201 | || (convert_optab_handler (vec_extract_optab, |
6202 | TYPE_MODE (vectype), vmode) | |
6203 | == CODE_FOR_nothing)) | |
6204 | { | |
6205 | /* Try to avoid emitting an extract of vector elements | |
6206 | by performing the extracts using an integer type of the | |
6207 | same size, extracting from a vector of those and then | |
6208 | re-interpreting it as the original vector type if | |
6209 | supported. */ | |
6210 | unsigned lsize | |
6211 | = group_size * GET_MODE_BITSIZE (elmode); | |
fffbab82 | 6212 | elmode = int_mode_for_size (lsize, 0).require (); |
4d694b27 | 6213 | unsigned int lnunits = const_nunits / group_size; |
04199738 RB |
6214 | /* If we can't construct such a vector fall back to |
6215 | element extracts from the original vector type and | |
6216 | element size stores. */ | |
4d694b27 | 6217 | if (mode_for_vector (elmode, lnunits).exists (&vmode) |
9da15d40 | 6218 | && VECTOR_MODE_P (vmode) |
04199738 RB |
6219 | && (convert_optab_handler (vec_extract_optab, |
6220 | vmode, elmode) | |
6221 | != CODE_FOR_nothing)) | |
6222 | { | |
4d694b27 | 6223 | nstores = lnunits; |
04199738 RB |
6224 | lnel = group_size; |
6225 | ltype = build_nonstandard_integer_type (lsize, 1); | |
6226 | lvectype = build_vector_type (ltype, nstores); | |
6227 | } | |
6228 | /* Else fall back to vector extraction anyway. | |
6229 | Fewer stores are more important than avoiding spilling | |
6230 | of the vector we extract from. Compared to the | |
6231 | construction case in vectorizable_load no store-forwarding | |
6232 | issue exists here for reasonable archs. */ | |
6233 | } | |
b17dc4d4 | 6234 | } |
4d694b27 RS |
6235 | else if (group_size >= const_nunits |
6236 | && group_size % const_nunits == 0) | |
b17dc4d4 RB |
6237 | { |
6238 | nstores = 1; | |
4d694b27 | 6239 | lnel = const_nunits; |
b17dc4d4 | 6240 | ltype = vectype; |
04199738 | 6241 | lvectype = vectype; |
b17dc4d4 | 6242 | } |
cee62fee MM |
6243 | ltype = build_aligned_type (ltype, TYPE_ALIGN (elem_type)); |
6244 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
6245 | } | |
6246 | ||
f2e2a985 MM |
6247 | ivstep = stride_step; |
6248 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
b17dc4d4 | 6249 | build_int_cst (TREE_TYPE (ivstep), vf)); |
f2e2a985 MM |
6250 | |
6251 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
6252 | ||
6253 | create_iv (stride_base, ivstep, NULL, | |
6254 | loop, &incr_gsi, insert_after, | |
6255 | &offvar, NULL); | |
6256 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 6257 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
f2e2a985 MM |
6258 | |
6259 | stride_step = force_gimple_operand (stride_step, &stmts, true, NULL_TREE); | |
6260 | if (stmts) | |
6261 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
6262 | ||
6263 | prev_stmt_info = NULL; | |
44fc7854 | 6264 | alias_off = build_int_cst (ref_type, 0); |
f502d50e MM |
6265 | next_stmt = first_stmt; |
6266 | for (g = 0; g < group_size; g++) | |
f2e2a985 | 6267 | { |
f502d50e MM |
6268 | running_off = offvar; |
6269 | if (g) | |
f2e2a985 | 6270 | { |
f502d50e MM |
6271 | tree size = TYPE_SIZE_UNIT (ltype); |
6272 | tree pos = fold_build2 (MULT_EXPR, sizetype, size_int (g), | |
f2e2a985 | 6273 | size); |
f502d50e | 6274 | tree newoff = copy_ssa_name (running_off, NULL); |
f2e2a985 | 6275 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
f502d50e | 6276 | running_off, pos); |
f2e2a985 | 6277 | vect_finish_stmt_generation (stmt, incr, gsi); |
f2e2a985 | 6278 | running_off = newoff; |
f502d50e | 6279 | } |
b17dc4d4 RB |
6280 | unsigned int group_el = 0; |
6281 | unsigned HOST_WIDE_INT | |
6282 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
f502d50e MM |
6283 | for (j = 0; j < ncopies; j++) |
6284 | { | |
c3a8f964 | 6285 | /* We've set op and dt above, from vect_get_store_rhs, |
f502d50e MM |
6286 | and first_stmt == stmt. */ |
6287 | if (j == 0) | |
6288 | { | |
6289 | if (slp) | |
6290 | { | |
6291 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, NULL, | |
306b0c92 | 6292 | slp_node); |
f502d50e MM |
6293 | vec_oprnd = vec_oprnds[0]; |
6294 | } | |
6295 | else | |
6296 | { | |
c3a8f964 | 6297 | op = vect_get_store_rhs (next_stmt); |
81c40241 | 6298 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
f502d50e MM |
6299 | } |
6300 | } | |
f2e2a985 | 6301 | else |
f502d50e MM |
6302 | { |
6303 | if (slp) | |
6304 | vec_oprnd = vec_oprnds[j]; | |
6305 | else | |
c079cbac | 6306 | { |
81c40241 | 6307 | vect_is_simple_use (vec_oprnd, vinfo, &def_stmt, &dt); |
c079cbac RB |
6308 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); |
6309 | } | |
f502d50e | 6310 | } |
04199738 RB |
6311 | /* Pun the vector to extract from if necessary. */ |
6312 | if (lvectype != vectype) | |
6313 | { | |
6314 | tree tem = make_ssa_name (lvectype); | |
6315 | gimple *pun | |
6316 | = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
6317 | lvectype, vec_oprnd)); | |
6318 | vect_finish_stmt_generation (stmt, pun, gsi); | |
6319 | vec_oprnd = tem; | |
6320 | } | |
f502d50e MM |
6321 | for (i = 0; i < nstores; i++) |
6322 | { | |
6323 | tree newref, newoff; | |
355fe088 | 6324 | gimple *incr, *assign; |
f502d50e MM |
6325 | tree size = TYPE_SIZE (ltype); |
6326 | /* Extract the i'th component. */ | |
6327 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, | |
6328 | bitsize_int (i), size); | |
6329 | tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd, | |
6330 | size, pos); | |
6331 | ||
6332 | elem = force_gimple_operand_gsi (gsi, elem, true, | |
6333 | NULL_TREE, true, | |
6334 | GSI_SAME_STMT); | |
6335 | ||
b17dc4d4 RB |
6336 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
6337 | group_el * elsz); | |
f502d50e | 6338 | newref = build2 (MEM_REF, ltype, |
b17dc4d4 | 6339 | running_off, this_off); |
f502d50e MM |
6340 | |
6341 | /* And store it to *running_off. */ | |
6342 | assign = gimple_build_assign (newref, elem); | |
6343 | vect_finish_stmt_generation (stmt, assign, gsi); | |
6344 | ||
b17dc4d4 RB |
6345 | group_el += lnel; |
6346 | if (! slp | |
6347 | || group_el == group_size) | |
6348 | { | |
6349 | newoff = copy_ssa_name (running_off, NULL); | |
6350 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
6351 | running_off, stride_step); | |
6352 | vect_finish_stmt_generation (stmt, incr, gsi); | |
f502d50e | 6353 | |
b17dc4d4 RB |
6354 | running_off = newoff; |
6355 | group_el = 0; | |
6356 | } | |
225ce44b RB |
6357 | if (g == group_size - 1 |
6358 | && !slp) | |
f502d50e MM |
6359 | { |
6360 | if (j == 0 && i == 0) | |
225ce44b RB |
6361 | STMT_VINFO_VEC_STMT (stmt_info) |
6362 | = *vec_stmt = assign; | |
f502d50e MM |
6363 | else |
6364 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = assign; | |
6365 | prev_stmt_info = vinfo_for_stmt (assign); | |
6366 | } | |
6367 | } | |
f2e2a985 | 6368 | } |
f502d50e | 6369 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
b17dc4d4 RB |
6370 | if (slp) |
6371 | break; | |
f2e2a985 | 6372 | } |
778dd3b6 RB |
6373 | |
6374 | vec_oprnds.release (); | |
f2e2a985 MM |
6375 | return true; |
6376 | } | |
6377 | ||
8c681247 | 6378 | auto_vec<tree> dr_chain (group_size); |
9771b263 | 6379 | oprnds.create (group_size); |
ebfd146a | 6380 | |
720f5239 | 6381 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 6382 | gcc_assert (alignment_support_scheme); |
7cfb4d93 | 6383 | bool masked_loop_p = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); |
272c6793 | 6384 | /* Targets with store-lane instructions must not require explicit |
c3a8f964 RS |
6385 | realignment. vect_supportable_dr_alignment always returns either |
6386 | dr_aligned or dr_unaligned_supported for masked operations. */ | |
7cfb4d93 RS |
6387 | gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES |
6388 | && !mask | |
6389 | && !masked_loop_p) | |
272c6793 RS |
6390 | || alignment_support_scheme == dr_aligned |
6391 | || alignment_support_scheme == dr_unaligned_supported); | |
6392 | ||
62da9e14 RS |
6393 | if (memory_access_type == VMAT_CONTIGUOUS_DOWN |
6394 | || memory_access_type == VMAT_CONTIGUOUS_REVERSE) | |
09dfa495 BM |
6395 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
6396 | ||
2de001ee | 6397 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
272c6793 RS |
6398 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); |
6399 | else | |
6400 | aggr_type = vectype; | |
ebfd146a | 6401 | |
c3a8f964 RS |
6402 | if (mask) |
6403 | LOOP_VINFO_HAS_MASK_STORE (loop_vinfo) = true; | |
6404 | ||
ebfd146a IR |
6405 | /* In case the vectorization factor (VF) is bigger than the number |
6406 | of elements that we can fit in a vectype (nunits), we have to generate | |
6407 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 6408 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
6409 | vect_get_vec_def_for_copy_stmt. */ |
6410 | ||
0d0293ac | 6411 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
6412 | |
6413 | S1: &base + 2 = x2 | |
6414 | S2: &base = x0 | |
6415 | S3: &base + 1 = x1 | |
6416 | S4: &base + 3 = x3 | |
6417 | ||
6418 | We create vectorized stores starting from base address (the access of the | |
6419 | first stmt in the chain (S2 in the above example), when the last store stmt | |
6420 | of the chain (S4) is reached: | |
6421 | ||
6422 | VS1: &base = vx2 | |
6423 | VS2: &base + vec_size*1 = vx0 | |
6424 | VS3: &base + vec_size*2 = vx1 | |
6425 | VS4: &base + vec_size*3 = vx3 | |
6426 | ||
6427 | Then permutation statements are generated: | |
6428 | ||
3fcc1b55 JJ |
6429 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
6430 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 6431 | ... |
b8698a0f | 6432 | |
ebfd146a IR |
6433 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
6434 | (the order of the data-refs in the output of vect_permute_store_chain | |
6435 | corresponds to the order of scalar stmts in the interleaving chain - see | |
6436 | the documentation of vect_permute_store_chain()). | |
6437 | ||
6438 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 6439 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 6440 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 6441 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
6442 | */ |
6443 | ||
6444 | prev_stmt_info = NULL; | |
c3a8f964 | 6445 | tree vec_mask = NULL_TREE; |
7cfb4d93 | 6446 | vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); |
ebfd146a IR |
6447 | for (j = 0; j < ncopies; j++) |
6448 | { | |
ebfd146a IR |
6449 | |
6450 | if (j == 0) | |
6451 | { | |
6452 | if (slp) | |
6453 | { | |
6454 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c | 6455 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
306b0c92 | 6456 | NULL, slp_node); |
ebfd146a | 6457 | |
9771b263 | 6458 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
6459 | } |
6460 | else | |
6461 | { | |
b8698a0f L |
6462 | /* For interleaved stores we collect vectorized defs for all the |
6463 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
6464 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
6465 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
6466 | ||
0d0293ac | 6467 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 6468 | OPRNDS are of size 1. */ |
b8698a0f | 6469 | next_stmt = first_stmt; |
ebfd146a IR |
6470 | for (i = 0; i < group_size; i++) |
6471 | { | |
b8698a0f L |
6472 | /* Since gaps are not supported for interleaved stores, |
6473 | GROUP_SIZE is the exact number of stmts in the chain. | |
6474 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
6475 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a | 6476 | iteration of the loop will be executed. */ |
c3a8f964 | 6477 | op = vect_get_store_rhs (next_stmt); |
81c40241 | 6478 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
9771b263 DN |
6479 | dr_chain.quick_push (vec_oprnd); |
6480 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 6481 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a | 6482 | } |
c3a8f964 RS |
6483 | if (mask) |
6484 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, | |
6485 | mask_vectype); | |
ebfd146a IR |
6486 | } |
6487 | ||
6488 | /* We should have catched mismatched types earlier. */ | |
6489 | gcc_assert (useless_type_conversion_p (vectype, | |
6490 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
6491 | bool simd_lane_access_p |
6492 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
6493 | if (simd_lane_access_p | |
6494 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
6495 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
6496 | && integer_zerop (DR_OFFSET (first_dr)) | |
6497 | && integer_zerop (DR_INIT (first_dr)) | |
6498 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
44fc7854 | 6499 | get_alias_set (TREE_TYPE (ref_type)))) |
74bf76ed JJ |
6500 | { |
6501 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
44fc7854 | 6502 | dataref_offset = build_int_cst (ref_type, 0); |
8928eff3 | 6503 | inv_p = false; |
74bf76ed JJ |
6504 | } |
6505 | else | |
6506 | dataref_ptr | |
6507 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
6508 | simd_lane_access_p ? loop : NULL, | |
09dfa495 | 6509 | offset, &dummy, gsi, &ptr_incr, |
74bf76ed | 6510 | simd_lane_access_p, &inv_p); |
a70d6342 | 6511 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 6512 | } |
b8698a0f | 6513 | else |
ebfd146a | 6514 | { |
b8698a0f L |
6515 | /* For interleaved stores we created vectorized defs for all the |
6516 | defs stored in OPRNDS in the previous iteration (previous copy). | |
6517 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
6518 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
6519 | next copy. | |
0d0293ac | 6520 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
6521 | OPRNDS are of size 1. */ |
6522 | for (i = 0; i < group_size; i++) | |
6523 | { | |
9771b263 | 6524 | op = oprnds[i]; |
81c40241 | 6525 | vect_is_simple_use (op, vinfo, &def_stmt, &dt); |
b8698a0f | 6526 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
9771b263 DN |
6527 | dr_chain[i] = vec_oprnd; |
6528 | oprnds[i] = vec_oprnd; | |
ebfd146a | 6529 | } |
c3a8f964 RS |
6530 | if (mask) |
6531 | { | |
6532 | vect_is_simple_use (vec_mask, vinfo, &def_stmt, &dt); | |
6533 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); | |
6534 | } | |
74bf76ed JJ |
6535 | if (dataref_offset) |
6536 | dataref_offset | |
6537 | = int_const_binop (PLUS_EXPR, dataref_offset, | |
6538 | TYPE_SIZE_UNIT (aggr_type)); | |
6539 | else | |
6540 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
6541 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
6542 | } |
6543 | ||
2de001ee | 6544 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
ebfd146a | 6545 | { |
272c6793 | 6546 | tree vec_array; |
267d3070 | 6547 | |
272c6793 RS |
6548 | /* Combine all the vectors into an array. */ |
6549 | vec_array = create_vector_array (vectype, vec_num); | |
6550 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 6551 | { |
9771b263 | 6552 | vec_oprnd = dr_chain[i]; |
272c6793 | 6553 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 6554 | } |
b8698a0f | 6555 | |
7cfb4d93 RS |
6556 | tree final_mask = NULL; |
6557 | if (masked_loop_p) | |
6558 | final_mask = vect_get_loop_mask (gsi, masks, ncopies, vectype, j); | |
6559 | if (vec_mask) | |
6560 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
6561 | vec_mask, gsi); | |
6562 | ||
7e11fc7f | 6563 | gcall *call; |
7cfb4d93 | 6564 | if (final_mask) |
7e11fc7f RS |
6565 | { |
6566 | /* Emit: | |
6567 | MASK_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK, | |
6568 | VEC_ARRAY). */ | |
6569 | unsigned int align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype)); | |
6570 | tree alias_ptr = build_int_cst (ref_type, align); | |
6571 | call = gimple_build_call_internal (IFN_MASK_STORE_LANES, 4, | |
6572 | dataref_ptr, alias_ptr, | |
7cfb4d93 | 6573 | final_mask, vec_array); |
7e11fc7f RS |
6574 | } |
6575 | else | |
6576 | { | |
6577 | /* Emit: | |
6578 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
6579 | data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type); | |
6580 | call = gimple_build_call_internal (IFN_STORE_LANES, 1, | |
6581 | vec_array); | |
6582 | gimple_call_set_lhs (call, data_ref); | |
6583 | } | |
a844293d RS |
6584 | gimple_call_set_nothrow (call, true); |
6585 | new_stmt = call; | |
267d3070 | 6586 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
6587 | } |
6588 | else | |
6589 | { | |
6590 | new_stmt = NULL; | |
0d0293ac | 6591 | if (grouped_store) |
272c6793 | 6592 | { |
b6b9227d JJ |
6593 | if (j == 0) |
6594 | result_chain.create (group_size); | |
272c6793 RS |
6595 | /* Permute. */ |
6596 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
6597 | &result_chain); | |
6598 | } | |
c2d7ab2a | 6599 | |
272c6793 RS |
6600 | next_stmt = first_stmt; |
6601 | for (i = 0; i < vec_num; i++) | |
6602 | { | |
644ffefd | 6603 | unsigned align, misalign; |
272c6793 | 6604 | |
7cfb4d93 RS |
6605 | tree final_mask = NULL_TREE; |
6606 | if (masked_loop_p) | |
6607 | final_mask = vect_get_loop_mask (gsi, masks, vec_num * ncopies, | |
6608 | vectype, vec_num * j + i); | |
6609 | if (vec_mask) | |
6610 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
6611 | vec_mask, gsi); | |
6612 | ||
272c6793 RS |
6613 | if (i > 0) |
6614 | /* Bump the vector pointer. */ | |
6615 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
6616 | stmt, NULL_TREE); | |
6617 | ||
6618 | if (slp) | |
9771b263 | 6619 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
6620 | else if (grouped_store) |
6621 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 6622 | vect_permute_store_chain(). */ |
9771b263 | 6623 | vec_oprnd = result_chain[i]; |
272c6793 | 6624 | |
f702e7d4 | 6625 | align = DR_TARGET_ALIGNMENT (first_dr); |
272c6793 | 6626 | if (aligned_access_p (first_dr)) |
644ffefd | 6627 | misalign = 0; |
272c6793 RS |
6628 | else if (DR_MISALIGNMENT (first_dr) == -1) |
6629 | { | |
25f68d90 | 6630 | align = dr_alignment (vect_dr_behavior (first_dr)); |
52639a61 | 6631 | misalign = 0; |
272c6793 RS |
6632 | } |
6633 | else | |
c3a8f964 | 6634 | misalign = DR_MISALIGNMENT (first_dr); |
aed93b23 RB |
6635 | if (dataref_offset == NULL_TREE |
6636 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
6637 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
6638 | misalign); | |
c2d7ab2a | 6639 | |
62da9e14 | 6640 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
09dfa495 BM |
6641 | { |
6642 | tree perm_mask = perm_mask_for_reverse (vectype); | |
6643 | tree perm_dest | |
c3a8f964 | 6644 | = vect_create_destination_var (vect_get_store_rhs (stmt), |
09dfa495 | 6645 | vectype); |
b731b390 | 6646 | tree new_temp = make_ssa_name (perm_dest); |
09dfa495 BM |
6647 | |
6648 | /* Generate the permute statement. */ | |
355fe088 | 6649 | gimple *perm_stmt |
0d0e4a03 JJ |
6650 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
6651 | vec_oprnd, perm_mask); | |
09dfa495 BM |
6652 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
6653 | ||
6654 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); | |
6655 | vec_oprnd = new_temp; | |
6656 | } | |
6657 | ||
272c6793 | 6658 | /* Arguments are ready. Create the new vector stmt. */ |
7cfb4d93 | 6659 | if (final_mask) |
c3a8f964 RS |
6660 | { |
6661 | align = least_bit_hwi (misalign | align); | |
6662 | tree ptr = build_int_cst (ref_type, align); | |
6663 | gcall *call | |
6664 | = gimple_build_call_internal (IFN_MASK_STORE, 4, | |
6665 | dataref_ptr, ptr, | |
7cfb4d93 | 6666 | final_mask, vec_oprnd); |
c3a8f964 RS |
6667 | gimple_call_set_nothrow (call, true); |
6668 | new_stmt = call; | |
6669 | } | |
6670 | else | |
6671 | { | |
6672 | data_ref = fold_build2 (MEM_REF, vectype, | |
6673 | dataref_ptr, | |
6674 | dataref_offset | |
6675 | ? dataref_offset | |
6676 | : build_int_cst (ref_type, 0)); | |
6677 | if (aligned_access_p (first_dr)) | |
6678 | ; | |
6679 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
6680 | TREE_TYPE (data_ref) | |
6681 | = build_aligned_type (TREE_TYPE (data_ref), | |
6682 | align * BITS_PER_UNIT); | |
6683 | else | |
6684 | TREE_TYPE (data_ref) | |
6685 | = build_aligned_type (TREE_TYPE (data_ref), | |
6686 | TYPE_ALIGN (elem_type)); | |
6687 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
6688 | } | |
272c6793 | 6689 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
6690 | |
6691 | if (slp) | |
6692 | continue; | |
6693 | ||
e14c1050 | 6694 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
6695 | if (!next_stmt) |
6696 | break; | |
6697 | } | |
ebfd146a | 6698 | } |
1da0876c RS |
6699 | if (!slp) |
6700 | { | |
6701 | if (j == 0) | |
6702 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6703 | else | |
6704 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6705 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6706 | } | |
ebfd146a IR |
6707 | } |
6708 | ||
9771b263 DN |
6709 | oprnds.release (); |
6710 | result_chain.release (); | |
6711 | vec_oprnds.release (); | |
ebfd146a IR |
6712 | |
6713 | return true; | |
6714 | } | |
6715 | ||
557be5a8 AL |
6716 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
6717 | VECTOR_CST mask. No checks are made that the target platform supports the | |
7ac7e286 | 6718 | mask, so callers may wish to test can_vec_perm_const_p separately, or use |
557be5a8 | 6719 | vect_gen_perm_mask_checked. */ |
a1e53f3f | 6720 | |
3fcc1b55 | 6721 | tree |
4aae3cb3 | 6722 | vect_gen_perm_mask_any (tree vectype, const vec_perm_indices &sel) |
a1e53f3f | 6723 | { |
b00cb3bf | 6724 | tree mask_type; |
a1e53f3f | 6725 | |
0ecc2b7d RS |
6726 | poly_uint64 nunits = sel.length (); |
6727 | gcc_assert (known_eq (nunits, TYPE_VECTOR_SUBPARTS (vectype))); | |
b00cb3bf RS |
6728 | |
6729 | mask_type = build_vector_type (ssizetype, nunits); | |
736d0f28 | 6730 | return vec_perm_indices_to_tree (mask_type, sel); |
a1e53f3f L |
6731 | } |
6732 | ||
7ac7e286 | 6733 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_const_p, |
cf7aa6a3 | 6734 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ |
557be5a8 AL |
6735 | |
6736 | tree | |
4aae3cb3 | 6737 | vect_gen_perm_mask_checked (tree vectype, const vec_perm_indices &sel) |
557be5a8 | 6738 | { |
7ac7e286 | 6739 | gcc_assert (can_vec_perm_const_p (TYPE_MODE (vectype), sel)); |
557be5a8 AL |
6740 | return vect_gen_perm_mask_any (vectype, sel); |
6741 | } | |
6742 | ||
aec7ae7d JJ |
6743 | /* Given a vector variable X and Y, that was generated for the scalar |
6744 | STMT, generate instructions to permute the vector elements of X and Y | |
6745 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
6746 | permuted vector variable. */ | |
a1e53f3f L |
6747 | |
6748 | static tree | |
355fe088 | 6749 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple *stmt, |
aec7ae7d | 6750 | gimple_stmt_iterator *gsi) |
a1e53f3f L |
6751 | { |
6752 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 6753 | tree perm_dest, data_ref; |
355fe088 | 6754 | gimple *perm_stmt; |
a1e53f3f | 6755 | |
7ad429a4 RS |
6756 | tree scalar_dest = gimple_get_lhs (stmt); |
6757 | if (TREE_CODE (scalar_dest) == SSA_NAME) | |
6758 | perm_dest = vect_create_destination_var (scalar_dest, vectype); | |
6759 | else | |
6760 | perm_dest = vect_get_new_vect_var (vectype, vect_simple_var, NULL); | |
b731b390 | 6761 | data_ref = make_ssa_name (perm_dest); |
a1e53f3f L |
6762 | |
6763 | /* Generate the permute statement. */ | |
0d0e4a03 | 6764 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
a1e53f3f L |
6765 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
6766 | ||
6767 | return data_ref; | |
6768 | } | |
6769 | ||
6b916b36 RB |
6770 | /* Hoist the definitions of all SSA uses on STMT out of the loop LOOP, |
6771 | inserting them on the loops preheader edge. Returns true if we | |
6772 | were successful in doing so (and thus STMT can be moved then), | |
6773 | otherwise returns false. */ | |
6774 | ||
6775 | static bool | |
355fe088 | 6776 | hoist_defs_of_uses (gimple *stmt, struct loop *loop) |
6b916b36 RB |
6777 | { |
6778 | ssa_op_iter i; | |
6779 | tree op; | |
6780 | bool any = false; | |
6781 | ||
6782 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6783 | { | |
355fe088 | 6784 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6785 | if (!gimple_nop_p (def_stmt) |
6786 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6787 | { | |
6788 | /* Make sure we don't need to recurse. While we could do | |
6789 | so in simple cases when there are more complex use webs | |
6790 | we don't have an easy way to preserve stmt order to fulfil | |
6791 | dependencies within them. */ | |
6792 | tree op2; | |
6793 | ssa_op_iter i2; | |
d1417442 JJ |
6794 | if (gimple_code (def_stmt) == GIMPLE_PHI) |
6795 | return false; | |
6b916b36 RB |
6796 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
6797 | { | |
355fe088 | 6798 | gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2); |
6b916b36 RB |
6799 | if (!gimple_nop_p (def_stmt2) |
6800 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2))) | |
6801 | return false; | |
6802 | } | |
6803 | any = true; | |
6804 | } | |
6805 | } | |
6806 | ||
6807 | if (!any) | |
6808 | return true; | |
6809 | ||
6810 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6811 | { | |
355fe088 | 6812 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6813 | if (!gimple_nop_p (def_stmt) |
6814 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6815 | { | |
6816 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); | |
6817 | gsi_remove (&gsi, false); | |
6818 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); | |
6819 | } | |
6820 | } | |
6821 | ||
6822 | return true; | |
6823 | } | |
6824 | ||
ebfd146a IR |
6825 | /* vectorizable_load. |
6826 | ||
b8698a0f L |
6827 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
6828 | can be vectorized. | |
6829 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
6830 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
6831 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
6832 | ||
6833 | static bool | |
355fe088 | 6834 | vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 6835 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a IR |
6836 | { |
6837 | tree scalar_dest; | |
6838 | tree vec_dest = NULL; | |
6839 | tree data_ref = NULL; | |
6840 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 6841 | stmt_vec_info prev_stmt_info; |
ebfd146a | 6842 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 6843 | struct loop *loop = NULL; |
ebfd146a | 6844 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 6845 | bool nested_in_vect_loop = false; |
c716e67f | 6846 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
272c6793 | 6847 | tree elem_type; |
ebfd146a | 6848 | tree new_temp; |
ef4bddc2 | 6849 | machine_mode mode; |
355fe088 | 6850 | gimple *new_stmt = NULL; |
ebfd146a IR |
6851 | tree dummy; |
6852 | enum dr_alignment_support alignment_support_scheme; | |
6853 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 6854 | tree dataref_offset = NULL_TREE; |
355fe088 | 6855 | gimple *ptr_incr = NULL; |
ebfd146a | 6856 | int ncopies; |
4d694b27 RS |
6857 | int i, j; |
6858 | unsigned int group_size; | |
6859 | poly_uint64 group_gap_adj; | |
ebfd146a IR |
6860 | tree msq = NULL_TREE, lsq; |
6861 | tree offset = NULL_TREE; | |
356bbc4c | 6862 | tree byte_offset = NULL_TREE; |
ebfd146a | 6863 | tree realignment_token = NULL_TREE; |
538dd0b7 | 6864 | gphi *phi = NULL; |
6e1aa848 | 6865 | vec<tree> dr_chain = vNULL; |
0d0293ac | 6866 | bool grouped_load = false; |
355fe088 | 6867 | gimple *first_stmt; |
4f0a0218 | 6868 | gimple *first_stmt_for_drptr = NULL; |
ebfd146a IR |
6869 | bool inv_p; |
6870 | bool compute_in_loop = false; | |
6871 | struct loop *at_loop; | |
6872 | int vec_num; | |
6873 | bool slp = (slp_node != NULL); | |
6874 | bool slp_perm = false; | |
a70d6342 | 6875 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
d9f21f6a | 6876 | poly_uint64 vf; |
272c6793 | 6877 | tree aggr_type; |
134c85ca | 6878 | gather_scatter_info gs_info; |
310213d4 | 6879 | vec_info *vinfo = stmt_info->vinfo; |
44fc7854 | 6880 | tree ref_type; |
a70d6342 | 6881 | |
465c8c19 JJ |
6882 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
6883 | return false; | |
6884 | ||
66c16fd9 RB |
6885 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
6886 | && ! vec_stmt) | |
465c8c19 JJ |
6887 | return false; |
6888 | ||
c3a8f964 RS |
6889 | tree mask = NULL_TREE, mask_vectype = NULL_TREE; |
6890 | if (is_gimple_assign (stmt)) | |
6891 | { | |
6892 | scalar_dest = gimple_assign_lhs (stmt); | |
6893 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
6894 | return false; | |
465c8c19 | 6895 | |
c3a8f964 RS |
6896 | tree_code code = gimple_assign_rhs_code (stmt); |
6897 | if (code != ARRAY_REF | |
6898 | && code != BIT_FIELD_REF | |
6899 | && code != INDIRECT_REF | |
6900 | && code != COMPONENT_REF | |
6901 | && code != IMAGPART_EXPR | |
6902 | && code != REALPART_EXPR | |
6903 | && code != MEM_REF | |
6904 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
6905 | return false; | |
6906 | } | |
6907 | else | |
6908 | { | |
6909 | gcall *call = dyn_cast <gcall *> (stmt); | |
6910 | if (!call || !gimple_call_internal_p (call, IFN_MASK_LOAD)) | |
6911 | return false; | |
465c8c19 | 6912 | |
c3a8f964 RS |
6913 | scalar_dest = gimple_call_lhs (call); |
6914 | if (!scalar_dest) | |
6915 | return false; | |
6916 | ||
6917 | if (slp_node != NULL) | |
6918 | { | |
6919 | if (dump_enabled_p ()) | |
6920 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6921 | "SLP of masked loads not supported.\n"); | |
6922 | return false; | |
6923 | } | |
6924 | ||
6925 | mask = gimple_call_arg (call, 2); | |
6926 | if (!vect_check_load_store_mask (stmt, mask, &mask_vectype)) | |
6927 | return false; | |
6928 | } | |
465c8c19 JJ |
6929 | |
6930 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
6931 | return false; | |
6932 | ||
6933 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
4d694b27 | 6934 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 | 6935 | |
a70d6342 IR |
6936 | if (loop_vinfo) |
6937 | { | |
6938 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
6939 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
6940 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
6941 | } | |
6942 | else | |
3533e503 | 6943 | vf = 1; |
ebfd146a IR |
6944 | |
6945 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 6946 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 6947 | case of SLP. */ |
fce57248 | 6948 | if (slp) |
ebfd146a IR |
6949 | ncopies = 1; |
6950 | else | |
e8f142e2 | 6951 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
ebfd146a IR |
6952 | |
6953 | gcc_assert (ncopies >= 1); | |
6954 | ||
6955 | /* FORNOW. This restriction should be relaxed. */ | |
6956 | if (nested_in_vect_loop && ncopies > 1) | |
6957 | { | |
73fbfcad | 6958 | if (dump_enabled_p ()) |
78c60e3d | 6959 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6960 | "multiple types in nested loop.\n"); |
ebfd146a IR |
6961 | return false; |
6962 | } | |
6963 | ||
f2556b68 RB |
6964 | /* Invalidate assumptions made by dependence analysis when vectorization |
6965 | on the unrolled body effectively re-orders stmts. */ | |
6966 | if (ncopies > 1 | |
6967 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
d9f21f6a RS |
6968 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
6969 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
f2556b68 RB |
6970 | { |
6971 | if (dump_enabled_p ()) | |
6972 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6973 | "cannot perform implicit CSE when unrolling " | |
6974 | "with negative dependence distance\n"); | |
6975 | return false; | |
6976 | } | |
6977 | ||
7b7b1813 | 6978 | elem_type = TREE_TYPE (vectype); |
947131ba | 6979 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
6980 | |
6981 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
6982 | (e.g. - data copies). */ | |
947131ba | 6983 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 6984 | { |
73fbfcad | 6985 | if (dump_enabled_p ()) |
78c60e3d | 6986 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6987 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
6988 | return false; |
6989 | } | |
6990 | ||
ebfd146a | 6991 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 6992 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 6993 | { |
0d0293ac | 6994 | grouped_load = true; |
ebfd146a | 6995 | /* FORNOW */ |
2de001ee RS |
6996 | gcc_assert (!nested_in_vect_loop); |
6997 | gcc_assert (!STMT_VINFO_GATHER_SCATTER_P (stmt_info)); | |
ebfd146a | 6998 | |
e14c1050 | 6999 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
d3465d72 | 7000 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
d5f035ea | 7001 | |
b1af7da6 RB |
7002 | if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
7003 | slp_perm = true; | |
7004 | ||
f2556b68 RB |
7005 | /* Invalidate assumptions made by dependence analysis when vectorization |
7006 | on the unrolled body effectively re-orders stmts. */ | |
7007 | if (!PURE_SLP_STMT (stmt_info) | |
7008 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
d9f21f6a RS |
7009 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
7010 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
f2556b68 RB |
7011 | { |
7012 | if (dump_enabled_p ()) | |
7013 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7014 | "cannot perform implicit CSE when performing " | |
7015 | "group loads with negative dependence distance\n"); | |
7016 | return false; | |
7017 | } | |
96bb56b2 RB |
7018 | |
7019 | /* Similarly when the stmt is a load that is both part of a SLP | |
7020 | instance and a loop vectorized stmt via the same-dr mechanism | |
7021 | we have to give up. */ | |
7022 | if (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info) | |
7023 | && (STMT_SLP_TYPE (stmt_info) | |
7024 | != STMT_SLP_TYPE (vinfo_for_stmt | |
7025 | (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info))))) | |
7026 | { | |
7027 | if (dump_enabled_p ()) | |
7028 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7029 | "conflicting SLP types for CSEd load\n"); | |
7030 | return false; | |
7031 | } | |
ebfd146a | 7032 | } |
7cfb4d93 RS |
7033 | else |
7034 | group_size = 1; | |
ebfd146a | 7035 | |
2de001ee | 7036 | vect_memory_access_type memory_access_type; |
7e11fc7f | 7037 | if (!get_load_store_type (stmt, vectype, slp, mask, VLS_LOAD, ncopies, |
2de001ee RS |
7038 | &memory_access_type, &gs_info)) |
7039 | return false; | |
a1e53f3f | 7040 | |
c3a8f964 RS |
7041 | if (mask) |
7042 | { | |
7043 | if (memory_access_type == VMAT_CONTIGUOUS) | |
7044 | { | |
7e11fc7f RS |
7045 | machine_mode vec_mode = TYPE_MODE (vectype); |
7046 | if (!VECTOR_MODE_P (vec_mode) | |
7047 | || !can_vec_mask_load_store_p (vec_mode, | |
c3a8f964 RS |
7048 | TYPE_MODE (mask_vectype), true)) |
7049 | return false; | |
7050 | } | |
7051 | else if (memory_access_type == VMAT_GATHER_SCATTER) | |
7052 | { | |
7053 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); | |
7054 | tree masktype | |
7055 | = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (TREE_CHAIN (arglist)))); | |
7056 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
7057 | { | |
7058 | if (dump_enabled_p ()) | |
7059 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7060 | "masked gather with integer mask not" | |
7061 | " supported."); | |
7062 | return false; | |
7063 | } | |
7064 | } | |
7e11fc7f | 7065 | else if (memory_access_type != VMAT_LOAD_STORE_LANES) |
c3a8f964 RS |
7066 | { |
7067 | if (dump_enabled_p ()) | |
7068 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7069 | "unsupported access type for masked load.\n"); | |
7070 | return false; | |
7071 | } | |
7072 | } | |
7073 | ||
ebfd146a IR |
7074 | if (!vec_stmt) /* transformation not required. */ |
7075 | { | |
2de001ee RS |
7076 | if (!slp) |
7077 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; | |
7cfb4d93 RS |
7078 | |
7079 | if (loop_vinfo | |
7080 | && LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo)) | |
7081 | check_load_store_masking (loop_vinfo, vectype, VLS_LOAD, group_size, | |
7082 | memory_access_type); | |
7083 | ||
ebfd146a | 7084 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; |
2e8ab70c RB |
7085 | /* The SLP costs are calculated during SLP analysis. */ |
7086 | if (!PURE_SLP_STMT (stmt_info)) | |
2de001ee | 7087 | vect_model_load_cost (stmt_info, ncopies, memory_access_type, |
2e8ab70c | 7088 | NULL, NULL, NULL); |
ebfd146a IR |
7089 | return true; |
7090 | } | |
7091 | ||
2de001ee RS |
7092 | if (!slp) |
7093 | gcc_assert (memory_access_type | |
7094 | == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); | |
7095 | ||
73fbfcad | 7096 | if (dump_enabled_p ()) |
78c60e3d | 7097 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 7098 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a | 7099 | |
67b8dbac | 7100 | /* Transform. */ |
ebfd146a | 7101 | |
f702e7d4 | 7102 | ensure_base_align (dr); |
c716e67f | 7103 | |
2de001ee | 7104 | if (memory_access_type == VMAT_GATHER_SCATTER) |
aec7ae7d | 7105 | { |
c3a8f964 | 7106 | vect_build_gather_load_calls (stmt, gsi, vec_stmt, &gs_info, mask); |
aec7ae7d JJ |
7107 | return true; |
7108 | } | |
2de001ee RS |
7109 | |
7110 | if (memory_access_type == VMAT_ELEMENTWISE | |
7111 | || memory_access_type == VMAT_STRIDED_SLP) | |
7d75abc8 MM |
7112 | { |
7113 | gimple_stmt_iterator incr_gsi; | |
7114 | bool insert_after; | |
355fe088 | 7115 | gimple *incr; |
7d75abc8 | 7116 | tree offvar; |
7d75abc8 MM |
7117 | tree ivstep; |
7118 | tree running_off; | |
9771b263 | 7119 | vec<constructor_elt, va_gc> *v = NULL; |
7d75abc8 | 7120 | gimple_seq stmts = NULL; |
14ac6aa2 | 7121 | tree stride_base, stride_step, alias_off; |
4d694b27 RS |
7122 | /* Checked by get_load_store_type. */ |
7123 | unsigned int const_nunits = nunits.to_constant (); | |
14ac6aa2 | 7124 | |
7cfb4d93 | 7125 | gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); |
14ac6aa2 | 7126 | gcc_assert (!nested_in_vect_loop); |
7d75abc8 | 7127 | |
f502d50e | 7128 | if (slp && grouped_load) |
44fc7854 BE |
7129 | { |
7130 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); | |
7131 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
7132 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
7133 | ref_type = get_group_alias_ptr_type (first_stmt); | |
7134 | } | |
ab313a8c | 7135 | else |
44fc7854 BE |
7136 | { |
7137 | first_stmt = stmt; | |
7138 | first_dr = dr; | |
7139 | group_size = 1; | |
7140 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); | |
7141 | } | |
ab313a8c | 7142 | |
14ac6aa2 RB |
7143 | stride_base |
7144 | = fold_build_pointer_plus | |
ab313a8c | 7145 | (DR_BASE_ADDRESS (first_dr), |
14ac6aa2 | 7146 | size_binop (PLUS_EXPR, |
ab313a8c RB |
7147 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
7148 | convert_to_ptrofftype (DR_INIT (first_dr)))); | |
7149 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); | |
7d75abc8 MM |
7150 | |
7151 | /* For a load with loop-invariant (but other than power-of-2) | |
7152 | stride (i.e. not a grouped access) like so: | |
7153 | ||
7154 | for (i = 0; i < n; i += stride) | |
7155 | ... = array[i]; | |
7156 | ||
7157 | we generate a new induction variable and new accesses to | |
7158 | form a new vector (or vectors, depending on ncopies): | |
7159 | ||
7160 | for (j = 0; ; j += VF*stride) | |
7161 | tmp1 = array[j]; | |
7162 | tmp2 = array[j + stride]; | |
7163 | ... | |
7164 | vectemp = {tmp1, tmp2, ...} | |
7165 | */ | |
7166 | ||
ab313a8c RB |
7167 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step, |
7168 | build_int_cst (TREE_TYPE (stride_step), vf)); | |
7d75abc8 MM |
7169 | |
7170 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
7171 | ||
ab313a8c | 7172 | create_iv (unshare_expr (stride_base), unshare_expr (ivstep), NULL, |
7d75abc8 MM |
7173 | loop, &incr_gsi, insert_after, |
7174 | &offvar, NULL); | |
7175 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 7176 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
7d75abc8 | 7177 | |
ab313a8c RB |
7178 | stride_step = force_gimple_operand (unshare_expr (stride_step), |
7179 | &stmts, true, NULL_TREE); | |
7d75abc8 MM |
7180 | if (stmts) |
7181 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
7182 | ||
7183 | prev_stmt_info = NULL; | |
7184 | running_off = offvar; | |
44fc7854 | 7185 | alias_off = build_int_cst (ref_type, 0); |
4d694b27 | 7186 | int nloads = const_nunits; |
e09b4c37 | 7187 | int lnel = 1; |
7b5fc413 | 7188 | tree ltype = TREE_TYPE (vectype); |
ea60dd34 | 7189 | tree lvectype = vectype; |
b266b968 | 7190 | auto_vec<tree> dr_chain; |
2de001ee | 7191 | if (memory_access_type == VMAT_STRIDED_SLP) |
7b5fc413 | 7192 | { |
4d694b27 | 7193 | if (group_size < const_nunits) |
e09b4c37 | 7194 | { |
ff03930a JJ |
7195 | /* First check if vec_init optab supports construction from |
7196 | vector elts directly. */ | |
b397965c | 7197 | scalar_mode elmode = SCALAR_TYPE_MODE (TREE_TYPE (vectype)); |
9da15d40 RS |
7198 | machine_mode vmode; |
7199 | if (mode_for_vector (elmode, group_size).exists (&vmode) | |
7200 | && VECTOR_MODE_P (vmode) | |
ff03930a JJ |
7201 | && (convert_optab_handler (vec_init_optab, |
7202 | TYPE_MODE (vectype), vmode) | |
7203 | != CODE_FOR_nothing)) | |
ea60dd34 | 7204 | { |
4d694b27 | 7205 | nloads = const_nunits / group_size; |
ea60dd34 | 7206 | lnel = group_size; |
ff03930a JJ |
7207 | ltype = build_vector_type (TREE_TYPE (vectype), group_size); |
7208 | } | |
7209 | else | |
7210 | { | |
7211 | /* Otherwise avoid emitting a constructor of vector elements | |
7212 | by performing the loads using an integer type of the same | |
7213 | size, constructing a vector of those and then | |
7214 | re-interpreting it as the original vector type. | |
7215 | This avoids a huge runtime penalty due to the general | |
7216 | inability to perform store forwarding from smaller stores | |
7217 | to a larger load. */ | |
7218 | unsigned lsize | |
7219 | = group_size * TYPE_PRECISION (TREE_TYPE (vectype)); | |
fffbab82 | 7220 | elmode = int_mode_for_size (lsize, 0).require (); |
4d694b27 | 7221 | unsigned int lnunits = const_nunits / group_size; |
ff03930a JJ |
7222 | /* If we can't construct such a vector fall back to |
7223 | element loads of the original vector type. */ | |
4d694b27 | 7224 | if (mode_for_vector (elmode, lnunits).exists (&vmode) |
9da15d40 | 7225 | && VECTOR_MODE_P (vmode) |
ff03930a JJ |
7226 | && (convert_optab_handler (vec_init_optab, vmode, elmode) |
7227 | != CODE_FOR_nothing)) | |
7228 | { | |
4d694b27 | 7229 | nloads = lnunits; |
ff03930a JJ |
7230 | lnel = group_size; |
7231 | ltype = build_nonstandard_integer_type (lsize, 1); | |
7232 | lvectype = build_vector_type (ltype, nloads); | |
7233 | } | |
ea60dd34 | 7234 | } |
e09b4c37 | 7235 | } |
2de001ee | 7236 | else |
e09b4c37 | 7237 | { |
ea60dd34 | 7238 | nloads = 1; |
4d694b27 | 7239 | lnel = const_nunits; |
e09b4c37 | 7240 | ltype = vectype; |
e09b4c37 | 7241 | } |
2de001ee RS |
7242 | ltype = build_aligned_type (ltype, TYPE_ALIGN (TREE_TYPE (vectype))); |
7243 | } | |
7244 | if (slp) | |
7245 | { | |
66c16fd9 RB |
7246 | /* For SLP permutation support we need to load the whole group, |
7247 | not only the number of vector stmts the permutation result | |
7248 | fits in. */ | |
b266b968 | 7249 | if (slp_perm) |
66c16fd9 | 7250 | { |
d9f21f6a RS |
7251 | /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for |
7252 | variable VF. */ | |
7253 | unsigned int const_vf = vf.to_constant (); | |
4d694b27 | 7254 | ncopies = CEIL (group_size * const_vf, const_nunits); |
66c16fd9 RB |
7255 | dr_chain.create (ncopies); |
7256 | } | |
7257 | else | |
7258 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
7b5fc413 | 7259 | } |
4d694b27 | 7260 | unsigned int group_el = 0; |
e09b4c37 RB |
7261 | unsigned HOST_WIDE_INT |
7262 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
7d75abc8 MM |
7263 | for (j = 0; j < ncopies; j++) |
7264 | { | |
7b5fc413 | 7265 | if (nloads > 1) |
e09b4c37 RB |
7266 | vec_alloc (v, nloads); |
7267 | for (i = 0; i < nloads; i++) | |
7b5fc413 | 7268 | { |
e09b4c37 RB |
7269 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
7270 | group_el * elsz); | |
7271 | new_stmt = gimple_build_assign (make_ssa_name (ltype), | |
7272 | build2 (MEM_REF, ltype, | |
7273 | running_off, this_off)); | |
7274 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7275 | if (nloads > 1) | |
7276 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, | |
7277 | gimple_assign_lhs (new_stmt)); | |
7278 | ||
7279 | group_el += lnel; | |
7280 | if (! slp | |
7281 | || group_el == group_size) | |
7b5fc413 | 7282 | { |
e09b4c37 RB |
7283 | tree newoff = copy_ssa_name (running_off); |
7284 | gimple *incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
7285 | running_off, stride_step); | |
7b5fc413 RB |
7286 | vect_finish_stmt_generation (stmt, incr, gsi); |
7287 | ||
7288 | running_off = newoff; | |
e09b4c37 | 7289 | group_el = 0; |
7b5fc413 | 7290 | } |
7b5fc413 | 7291 | } |
e09b4c37 | 7292 | if (nloads > 1) |
7d75abc8 | 7293 | { |
ea60dd34 RB |
7294 | tree vec_inv = build_constructor (lvectype, v); |
7295 | new_temp = vect_init_vector (stmt, vec_inv, lvectype, gsi); | |
e09b4c37 | 7296 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
ea60dd34 RB |
7297 | if (lvectype != vectype) |
7298 | { | |
7299 | new_stmt = gimple_build_assign (make_ssa_name (vectype), | |
7300 | VIEW_CONVERT_EXPR, | |
7301 | build1 (VIEW_CONVERT_EXPR, | |
7302 | vectype, new_temp)); | |
7303 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7304 | } | |
7d75abc8 MM |
7305 | } |
7306 | ||
7b5fc413 | 7307 | if (slp) |
b266b968 | 7308 | { |
b266b968 RB |
7309 | if (slp_perm) |
7310 | dr_chain.quick_push (gimple_assign_lhs (new_stmt)); | |
66c16fd9 RB |
7311 | else |
7312 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
b266b968 | 7313 | } |
7d75abc8 | 7314 | else |
225ce44b RB |
7315 | { |
7316 | if (j == 0) | |
7317 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7318 | else | |
7319 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7320 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7321 | } | |
7d75abc8 | 7322 | } |
b266b968 | 7323 | if (slp_perm) |
29afecdf RB |
7324 | { |
7325 | unsigned n_perms; | |
7326 | vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, | |
7327 | slp_node_instance, false, &n_perms); | |
7328 | } | |
7d75abc8 MM |
7329 | return true; |
7330 | } | |
aec7ae7d | 7331 | |
0d0293ac | 7332 | if (grouped_load) |
ebfd146a | 7333 | { |
e14c1050 | 7334 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
44fc7854 | 7335 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
4f0a0218 | 7336 | /* For SLP vectorization we directly vectorize a subchain |
52eab378 RB |
7337 | without permutation. */ |
7338 | if (slp && ! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
4f0a0218 RB |
7339 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
7340 | /* For BB vectorization always use the first stmt to base | |
7341 | the data ref pointer on. */ | |
7342 | if (bb_vinfo) | |
7343 | first_stmt_for_drptr = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 7344 | |
ebfd146a | 7345 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
7346 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
7347 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
7348 | ??? But we can only do so if there is exactly one | |
7349 | as we have no way to get at the rest. Leave the CSE | |
7350 | opportunity alone. | |
7351 | ??? With the group load eventually participating | |
7352 | in multiple different permutations (having multiple | |
7353 | slp nodes which refer to the same group) the CSE | |
7354 | is even wrong code. See PR56270. */ | |
7355 | && !slp) | |
ebfd146a IR |
7356 | { |
7357 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
7358 | return true; | |
7359 | } | |
7360 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
9b999e8c | 7361 | group_gap_adj = 0; |
ebfd146a IR |
7362 | |
7363 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
7364 | if (slp) | |
7365 | { | |
0d0293ac | 7366 | grouped_load = false; |
91ff1504 RB |
7367 | /* For SLP permutation support we need to load the whole group, |
7368 | not only the number of vector stmts the permutation result | |
7369 | fits in. */ | |
7370 | if (slp_perm) | |
b267968e | 7371 | { |
d9f21f6a RS |
7372 | /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for |
7373 | variable VF. */ | |
7374 | unsigned int const_vf = vf.to_constant (); | |
4d694b27 RS |
7375 | unsigned int const_nunits = nunits.to_constant (); |
7376 | vec_num = CEIL (group_size * const_vf, const_nunits); | |
b267968e RB |
7377 | group_gap_adj = vf * group_size - nunits * vec_num; |
7378 | } | |
91ff1504 | 7379 | else |
b267968e RB |
7380 | { |
7381 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
796bd467 RB |
7382 | group_gap_adj |
7383 | = group_size - SLP_INSTANCE_GROUP_SIZE (slp_node_instance); | |
b267968e | 7384 | } |
a70d6342 | 7385 | } |
ebfd146a | 7386 | else |
9b999e8c | 7387 | vec_num = group_size; |
44fc7854 BE |
7388 | |
7389 | ref_type = get_group_alias_ptr_type (first_stmt); | |
ebfd146a IR |
7390 | } |
7391 | else | |
7392 | { | |
7393 | first_stmt = stmt; | |
7394 | first_dr = dr; | |
7395 | group_size = vec_num = 1; | |
9b999e8c | 7396 | group_gap_adj = 0; |
44fc7854 | 7397 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); |
ebfd146a IR |
7398 | } |
7399 | ||
720f5239 | 7400 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 7401 | gcc_assert (alignment_support_scheme); |
7cfb4d93 RS |
7402 | bool masked_loop_p = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); |
7403 | /* Targets with store-lane instructions must not require explicit | |
7404 | realignment. vect_supportable_dr_alignment always returns either | |
7405 | dr_aligned or dr_unaligned_supported for masked operations. */ | |
7406 | gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES | |
7407 | && !mask | |
7408 | && !masked_loop_p) | |
272c6793 RS |
7409 | || alignment_support_scheme == dr_aligned |
7410 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
7411 | |
7412 | /* In case the vectorization factor (VF) is bigger than the number | |
7413 | of elements that we can fit in a vectype (nunits), we have to generate | |
7414 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 7415 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 7416 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 7417 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 7418 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
7419 | stmts that use the defs of the current stmt. The example below |
7420 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
7421 | need to create 4 vectorized stmts): | |
ebfd146a IR |
7422 | |
7423 | before vectorization: | |
7424 | RELATED_STMT VEC_STMT | |
7425 | S1: x = memref - - | |
7426 | S2: z = x + 1 - - | |
7427 | ||
7428 | step 1: vectorize stmt S1: | |
7429 | We first create the vector stmt VS1_0, and, as usual, record a | |
7430 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
7431 | Next, we create the vector stmt VS1_1, and record a pointer to | |
7432 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 7433 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
7434 | stmts and pointers: |
7435 | RELATED_STMT VEC_STMT | |
7436 | VS1_0: vx0 = memref0 VS1_1 - | |
7437 | VS1_1: vx1 = memref1 VS1_2 - | |
7438 | VS1_2: vx2 = memref2 VS1_3 - | |
7439 | VS1_3: vx3 = memref3 - - | |
7440 | S1: x = load - VS1_0 | |
7441 | S2: z = x + 1 - - | |
7442 | ||
b8698a0f L |
7443 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
7444 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
7445 | stmt S2. */ |
7446 | ||
0d0293ac | 7447 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
7448 | |
7449 | S1: x2 = &base + 2 | |
7450 | S2: x0 = &base | |
7451 | S3: x1 = &base + 1 | |
7452 | S4: x3 = &base + 3 | |
7453 | ||
b8698a0f | 7454 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
7455 | starting from the access of the first stmt of the chain: |
7456 | ||
7457 | VS1: vx0 = &base | |
7458 | VS2: vx1 = &base + vec_size*1 | |
7459 | VS3: vx3 = &base + vec_size*2 | |
7460 | VS4: vx4 = &base + vec_size*3 | |
7461 | ||
7462 | Then permutation statements are generated: | |
7463 | ||
e2c83630 RH |
7464 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
7465 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
7466 | ... |
7467 | ||
7468 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
7469 | (the order of the data-refs in the output of vect_permute_load_chain | |
7470 | corresponds to the order of scalar stmts in the interleaving chain - see | |
7471 | the documentation of vect_permute_load_chain()). | |
7472 | The generation of permutation stmts and recording them in | |
0d0293ac | 7473 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 7474 | |
b8698a0f | 7475 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
7476 | permutation stmts above are created for every copy. The result vector |
7477 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
7478 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
7479 | |
7480 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
7481 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
7482 | we generate the following code: | |
7483 | p = initial_addr; | |
7484 | indx = 0; | |
7485 | loop { | |
7486 | p = p + indx * vectype_size; | |
7487 | vec_dest = *(p); | |
7488 | indx = indx + 1; | |
7489 | } | |
7490 | ||
7491 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 7492 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
7493 | then generate the following code, in which the data in each iteration is |
7494 | obtained by two vector loads, one from the previous iteration, and one | |
7495 | from the current iteration: | |
7496 | p1 = initial_addr; | |
7497 | msq_init = *(floor(p1)) | |
7498 | p2 = initial_addr + VS - 1; | |
7499 | realignment_token = call target_builtin; | |
7500 | indx = 0; | |
7501 | loop { | |
7502 | p2 = p2 + indx * vectype_size | |
7503 | lsq = *(floor(p2)) | |
7504 | vec_dest = realign_load (msq, lsq, realignment_token) | |
7505 | indx = indx + 1; | |
7506 | msq = lsq; | |
7507 | } */ | |
7508 | ||
7509 | /* If the misalignment remains the same throughout the execution of the | |
7510 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 7511 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
7512 | This can only occur when vectorizing memory accesses in the inner-loop |
7513 | nested within an outer-loop that is being vectorized. */ | |
7514 | ||
d1e4b493 | 7515 | if (nested_in_vect_loop |
cf098191 RS |
7516 | && !multiple_p (DR_STEP_ALIGNMENT (dr), |
7517 | GET_MODE_SIZE (TYPE_MODE (vectype)))) | |
ebfd146a IR |
7518 | { |
7519 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
7520 | compute_in_loop = true; | |
7521 | } | |
7522 | ||
7523 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
7524 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 7525 | && !compute_in_loop) |
ebfd146a IR |
7526 | { |
7527 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
7528 | alignment_support_scheme, NULL_TREE, | |
7529 | &at_loop); | |
7530 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7531 | { | |
538dd0b7 | 7532 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (msq)); |
356bbc4c JJ |
7533 | byte_offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
7534 | size_one_node); | |
ebfd146a IR |
7535 | } |
7536 | } | |
7537 | else | |
7538 | at_loop = loop; | |
7539 | ||
62da9e14 | 7540 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
a1e53f3f L |
7541 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
7542 | ||
2de001ee | 7543 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
272c6793 RS |
7544 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); |
7545 | else | |
7546 | aggr_type = vectype; | |
7547 | ||
c3a8f964 | 7548 | tree vec_mask = NULL_TREE; |
ebfd146a | 7549 | prev_stmt_info = NULL; |
4d694b27 | 7550 | poly_uint64 group_elt = 0; |
7cfb4d93 | 7551 | vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); |
ebfd146a | 7552 | for (j = 0; j < ncopies; j++) |
b8698a0f | 7553 | { |
272c6793 | 7554 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 7555 | if (j == 0) |
74bf76ed JJ |
7556 | { |
7557 | bool simd_lane_access_p | |
7558 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
7559 | if (simd_lane_access_p | |
7560 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
7561 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
7562 | && integer_zerop (DR_OFFSET (first_dr)) | |
7563 | && integer_zerop (DR_INIT (first_dr)) | |
7564 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
44fc7854 | 7565 | get_alias_set (TREE_TYPE (ref_type))) |
74bf76ed JJ |
7566 | && (alignment_support_scheme == dr_aligned |
7567 | || alignment_support_scheme == dr_unaligned_supported)) | |
7568 | { | |
7569 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
44fc7854 | 7570 | dataref_offset = build_int_cst (ref_type, 0); |
8928eff3 | 7571 | inv_p = false; |
74bf76ed | 7572 | } |
4f0a0218 RB |
7573 | else if (first_stmt_for_drptr |
7574 | && first_stmt != first_stmt_for_drptr) | |
7575 | { | |
7576 | dataref_ptr | |
7577 | = vect_create_data_ref_ptr (first_stmt_for_drptr, aggr_type, | |
7578 | at_loop, offset, &dummy, gsi, | |
7579 | &ptr_incr, simd_lane_access_p, | |
7580 | &inv_p, byte_offset); | |
7581 | /* Adjust the pointer by the difference to first_stmt. */ | |
7582 | data_reference_p ptrdr | |
7583 | = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt_for_drptr)); | |
7584 | tree diff = fold_convert (sizetype, | |
7585 | size_binop (MINUS_EXPR, | |
7586 | DR_INIT (first_dr), | |
7587 | DR_INIT (ptrdr))); | |
7588 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
7589 | stmt, diff); | |
7590 | } | |
74bf76ed JJ |
7591 | else |
7592 | dataref_ptr | |
7593 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
7594 | offset, &dummy, gsi, &ptr_incr, | |
356bbc4c JJ |
7595 | simd_lane_access_p, &inv_p, |
7596 | byte_offset); | |
c3a8f964 RS |
7597 | if (mask) |
7598 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, | |
7599 | mask_vectype); | |
74bf76ed | 7600 | } |
ebfd146a | 7601 | else |
c3a8f964 RS |
7602 | { |
7603 | if (dataref_offset) | |
7604 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
7605 | TYPE_SIZE_UNIT (aggr_type)); | |
7606 | else | |
7607 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
7608 | TYPE_SIZE_UNIT (aggr_type)); | |
7609 | if (mask) | |
7610 | { | |
7611 | gimple *def_stmt; | |
7612 | vect_def_type dt; | |
7613 | vect_is_simple_use (vec_mask, vinfo, &def_stmt, &dt); | |
7614 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); | |
7615 | } | |
7616 | } | |
ebfd146a | 7617 | |
0d0293ac | 7618 | if (grouped_load || slp_perm) |
9771b263 | 7619 | dr_chain.create (vec_num); |
5ce1ee7f | 7620 | |
2de001ee | 7621 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
ebfd146a | 7622 | { |
272c6793 RS |
7623 | tree vec_array; |
7624 | ||
7625 | vec_array = create_vector_array (vectype, vec_num); | |
7626 | ||
7cfb4d93 RS |
7627 | tree final_mask = NULL_TREE; |
7628 | if (masked_loop_p) | |
7629 | final_mask = vect_get_loop_mask (gsi, masks, ncopies, vectype, j); | |
7630 | if (vec_mask) | |
7631 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
7632 | vec_mask, gsi); | |
7633 | ||
7e11fc7f | 7634 | gcall *call; |
7cfb4d93 | 7635 | if (final_mask) |
7e11fc7f RS |
7636 | { |
7637 | /* Emit: | |
7638 | VEC_ARRAY = MASK_LOAD_LANES (DATAREF_PTR, ALIAS_PTR, | |
7639 | VEC_MASK). */ | |
7640 | unsigned int align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype)); | |
7641 | tree alias_ptr = build_int_cst (ref_type, align); | |
7642 | call = gimple_build_call_internal (IFN_MASK_LOAD_LANES, 3, | |
7643 | dataref_ptr, alias_ptr, | |
7cfb4d93 | 7644 | final_mask); |
7e11fc7f RS |
7645 | } |
7646 | else | |
7647 | { | |
7648 | /* Emit: | |
7649 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
7650 | data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type); | |
7651 | call = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
7652 | } | |
a844293d RS |
7653 | gimple_call_set_lhs (call, vec_array); |
7654 | gimple_call_set_nothrow (call, true); | |
7655 | new_stmt = call; | |
272c6793 | 7656 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a | 7657 | |
272c6793 RS |
7658 | /* Extract each vector into an SSA_NAME. */ |
7659 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 7660 | { |
272c6793 RS |
7661 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
7662 | vec_array, i); | |
9771b263 | 7663 | dr_chain.quick_push (new_temp); |
272c6793 RS |
7664 | } |
7665 | ||
7666 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 7667 | vect_record_grouped_load_vectors (stmt, dr_chain); |
272c6793 RS |
7668 | } |
7669 | else | |
7670 | { | |
7671 | for (i = 0; i < vec_num; i++) | |
7672 | { | |
7cfb4d93 RS |
7673 | tree final_mask = NULL_TREE; |
7674 | if (masked_loop_p | |
7675 | && memory_access_type != VMAT_INVARIANT) | |
7676 | final_mask = vect_get_loop_mask (gsi, masks, vec_num * ncopies, | |
7677 | vectype, vec_num * j + i); | |
7678 | if (vec_mask) | |
7679 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
7680 | vec_mask, gsi); | |
7681 | ||
272c6793 RS |
7682 | if (i > 0) |
7683 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
7684 | stmt, NULL_TREE); | |
7685 | ||
7686 | /* 2. Create the vector-load in the loop. */ | |
7687 | switch (alignment_support_scheme) | |
7688 | { | |
7689 | case dr_aligned: | |
7690 | case dr_unaligned_supported: | |
be1ac4ec | 7691 | { |
644ffefd MJ |
7692 | unsigned int align, misalign; |
7693 | ||
f702e7d4 | 7694 | align = DR_TARGET_ALIGNMENT (dr); |
272c6793 RS |
7695 | if (alignment_support_scheme == dr_aligned) |
7696 | { | |
7697 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 7698 | misalign = 0; |
272c6793 RS |
7699 | } |
7700 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
7701 | { | |
25f68d90 | 7702 | align = dr_alignment (vect_dr_behavior (first_dr)); |
52639a61 | 7703 | misalign = 0; |
272c6793 RS |
7704 | } |
7705 | else | |
c3a8f964 | 7706 | misalign = DR_MISALIGNMENT (first_dr); |
aed93b23 RB |
7707 | if (dataref_offset == NULL_TREE |
7708 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
7709 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), |
7710 | align, misalign); | |
c3a8f964 | 7711 | |
7cfb4d93 | 7712 | if (final_mask) |
c3a8f964 RS |
7713 | { |
7714 | align = least_bit_hwi (misalign | align); | |
7715 | tree ptr = build_int_cst (ref_type, align); | |
7716 | gcall *call | |
7717 | = gimple_build_call_internal (IFN_MASK_LOAD, 3, | |
7718 | dataref_ptr, ptr, | |
7cfb4d93 | 7719 | final_mask); |
c3a8f964 RS |
7720 | gimple_call_set_nothrow (call, true); |
7721 | new_stmt = call; | |
7722 | data_ref = NULL_TREE; | |
7723 | } | |
7724 | else | |
7725 | { | |
7726 | data_ref | |
7727 | = fold_build2 (MEM_REF, vectype, dataref_ptr, | |
7728 | dataref_offset | |
7729 | ? dataref_offset | |
7730 | : build_int_cst (ref_type, 0)); | |
7731 | if (alignment_support_scheme == dr_aligned) | |
7732 | ; | |
7733 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
7734 | TREE_TYPE (data_ref) | |
7735 | = build_aligned_type (TREE_TYPE (data_ref), | |
7736 | align * BITS_PER_UNIT); | |
7737 | else | |
7738 | TREE_TYPE (data_ref) | |
7739 | = build_aligned_type (TREE_TYPE (data_ref), | |
7740 | TYPE_ALIGN (elem_type)); | |
7741 | } | |
272c6793 | 7742 | break; |
be1ac4ec | 7743 | } |
272c6793 | 7744 | case dr_explicit_realign: |
267d3070 | 7745 | { |
272c6793 | 7746 | tree ptr, bump; |
272c6793 | 7747 | |
d88981fc | 7748 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
272c6793 RS |
7749 | |
7750 | if (compute_in_loop) | |
7751 | msq = vect_setup_realignment (first_stmt, gsi, | |
7752 | &realignment_token, | |
7753 | dr_explicit_realign, | |
7754 | dataref_ptr, NULL); | |
7755 | ||
aed93b23 RB |
7756 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
7757 | ptr = copy_ssa_name (dataref_ptr); | |
7758 | else | |
7759 | ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
f702e7d4 | 7760 | unsigned int align = DR_TARGET_ALIGNMENT (first_dr); |
0d0e4a03 JJ |
7761 | new_stmt = gimple_build_assign |
7762 | (ptr, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
7763 | build_int_cst |
7764 | (TREE_TYPE (dataref_ptr), | |
f702e7d4 | 7765 | -(HOST_WIDE_INT) align)); |
272c6793 RS |
7766 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7767 | data_ref | |
7768 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 7769 | build_int_cst (ref_type, 0)); |
272c6793 RS |
7770 | vec_dest = vect_create_destination_var (scalar_dest, |
7771 | vectype); | |
7772 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
7773 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
7774 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7775 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
7776 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
7777 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7778 | msq = new_temp; | |
7779 | ||
d88981fc | 7780 | bump = size_binop (MULT_EXPR, vs, |
7b7b1813 | 7781 | TYPE_SIZE_UNIT (elem_type)); |
d88981fc | 7782 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
272c6793 | 7783 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
0d0e4a03 JJ |
7784 | new_stmt = gimple_build_assign |
7785 | (NULL_TREE, BIT_AND_EXPR, ptr, | |
272c6793 | 7786 | build_int_cst |
f702e7d4 | 7787 | (TREE_TYPE (ptr), -(HOST_WIDE_INT) align)); |
aed93b23 | 7788 | ptr = copy_ssa_name (ptr, new_stmt); |
272c6793 RS |
7789 | gimple_assign_set_lhs (new_stmt, ptr); |
7790 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7791 | data_ref | |
7792 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 7793 | build_int_cst (ref_type, 0)); |
272c6793 | 7794 | break; |
267d3070 | 7795 | } |
272c6793 | 7796 | case dr_explicit_realign_optimized: |
f702e7d4 RS |
7797 | { |
7798 | if (TREE_CODE (dataref_ptr) == SSA_NAME) | |
7799 | new_temp = copy_ssa_name (dataref_ptr); | |
7800 | else | |
7801 | new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
7802 | unsigned int align = DR_TARGET_ALIGNMENT (first_dr); | |
7803 | new_stmt = gimple_build_assign | |
7804 | (new_temp, BIT_AND_EXPR, dataref_ptr, | |
7805 | build_int_cst (TREE_TYPE (dataref_ptr), | |
7806 | -(HOST_WIDE_INT) align)); | |
7807 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7808 | data_ref | |
7809 | = build2 (MEM_REF, vectype, new_temp, | |
7810 | build_int_cst (ref_type, 0)); | |
7811 | break; | |
7812 | } | |
272c6793 RS |
7813 | default: |
7814 | gcc_unreachable (); | |
7815 | } | |
ebfd146a | 7816 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
c3a8f964 RS |
7817 | /* DATA_REF is null if we've already built the statement. */ |
7818 | if (data_ref) | |
7819 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
ebfd146a | 7820 | new_temp = make_ssa_name (vec_dest, new_stmt); |
c3a8f964 | 7821 | gimple_set_lhs (new_stmt, new_temp); |
ebfd146a IR |
7822 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7823 | ||
272c6793 RS |
7824 | /* 3. Handle explicit realignment if necessary/supported. |
7825 | Create in loop: | |
7826 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
7827 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
7828 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 7829 | { |
272c6793 RS |
7830 | lsq = gimple_assign_lhs (new_stmt); |
7831 | if (!realignment_token) | |
7832 | realignment_token = dataref_ptr; | |
7833 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
0d0e4a03 JJ |
7834 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, |
7835 | msq, lsq, realignment_token); | |
272c6793 RS |
7836 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7837 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7838 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7839 | ||
7840 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7841 | { | |
7842 | gcc_assert (phi); | |
7843 | if (i == vec_num - 1 && j == ncopies - 1) | |
7844 | add_phi_arg (phi, lsq, | |
7845 | loop_latch_edge (containing_loop), | |
9e227d60 | 7846 | UNKNOWN_LOCATION); |
272c6793 RS |
7847 | msq = lsq; |
7848 | } | |
ebfd146a | 7849 | } |
ebfd146a | 7850 | |
59fd17e3 RB |
7851 | /* 4. Handle invariant-load. */ |
7852 | if (inv_p && !bb_vinfo) | |
7853 | { | |
59fd17e3 | 7854 | gcc_assert (!grouped_load); |
d1417442 JJ |
7855 | /* If we have versioned for aliasing or the loop doesn't |
7856 | have any data dependencies that would preclude this, | |
7857 | then we are sure this is a loop invariant load and | |
7858 | thus we can insert it on the preheader edge. */ | |
7859 | if (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) | |
7860 | && !nested_in_vect_loop | |
6b916b36 | 7861 | && hoist_defs_of_uses (stmt, loop)) |
a0e35eb0 RB |
7862 | { |
7863 | if (dump_enabled_p ()) | |
7864 | { | |
7865 | dump_printf_loc (MSG_NOTE, vect_location, | |
7866 | "hoisting out of the vectorized " | |
7867 | "loop: "); | |
7868 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
a0e35eb0 | 7869 | } |
b731b390 | 7870 | tree tem = copy_ssa_name (scalar_dest); |
a0e35eb0 RB |
7871 | gsi_insert_on_edge_immediate |
7872 | (loop_preheader_edge (loop), | |
7873 | gimple_build_assign (tem, | |
7874 | unshare_expr | |
7875 | (gimple_assign_rhs1 (stmt)))); | |
7876 | new_temp = vect_init_vector (stmt, tem, vectype, NULL); | |
34cd48e5 RB |
7877 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
7878 | set_vinfo_for_stmt (new_stmt, | |
7879 | new_stmt_vec_info (new_stmt, vinfo)); | |
a0e35eb0 RB |
7880 | } |
7881 | else | |
7882 | { | |
7883 | gimple_stmt_iterator gsi2 = *gsi; | |
7884 | gsi_next (&gsi2); | |
7885 | new_temp = vect_init_vector (stmt, scalar_dest, | |
7886 | vectype, &gsi2); | |
34cd48e5 | 7887 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
a0e35eb0 | 7888 | } |
59fd17e3 RB |
7889 | } |
7890 | ||
62da9e14 | 7891 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
272c6793 | 7892 | { |
aec7ae7d JJ |
7893 | tree perm_mask = perm_mask_for_reverse (vectype); |
7894 | new_temp = permute_vec_elements (new_temp, new_temp, | |
7895 | perm_mask, stmt, gsi); | |
ebfd146a IR |
7896 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
7897 | } | |
267d3070 | 7898 | |
272c6793 | 7899 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
7900 | vect_transform_grouped_load (). */ |
7901 | if (grouped_load || slp_perm) | |
9771b263 | 7902 | dr_chain.quick_push (new_temp); |
267d3070 | 7903 | |
272c6793 RS |
7904 | /* Store vector loads in the corresponding SLP_NODE. */ |
7905 | if (slp && !slp_perm) | |
9771b263 | 7906 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
b267968e RB |
7907 | |
7908 | /* With SLP permutation we load the gaps as well, without | |
7909 | we need to skip the gaps after we manage to fully load | |
7910 | all elements. group_gap_adj is GROUP_SIZE here. */ | |
7911 | group_elt += nunits; | |
d9f21f6a RS |
7912 | if (maybe_ne (group_gap_adj, 0U) |
7913 | && !slp_perm | |
7914 | && known_eq (group_elt, group_size - group_gap_adj)) | |
b267968e | 7915 | { |
d9f21f6a RS |
7916 | poly_wide_int bump_val |
7917 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) | |
7918 | * group_gap_adj); | |
8e6cdc90 | 7919 | tree bump = wide_int_to_tree (sizetype, bump_val); |
b267968e RB |
7920 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
7921 | stmt, bump); | |
7922 | group_elt = 0; | |
7923 | } | |
272c6793 | 7924 | } |
9b999e8c RB |
7925 | /* Bump the vector pointer to account for a gap or for excess |
7926 | elements loaded for a permuted SLP load. */ | |
d9f21f6a | 7927 | if (maybe_ne (group_gap_adj, 0U) && slp_perm) |
a64b9c26 | 7928 | { |
d9f21f6a RS |
7929 | poly_wide_int bump_val |
7930 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) | |
7931 | * group_gap_adj); | |
8e6cdc90 | 7932 | tree bump = wide_int_to_tree (sizetype, bump_val); |
a64b9c26 RB |
7933 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
7934 | stmt, bump); | |
7935 | } | |
ebfd146a IR |
7936 | } |
7937 | ||
7938 | if (slp && !slp_perm) | |
7939 | continue; | |
7940 | ||
7941 | if (slp_perm) | |
7942 | { | |
29afecdf | 7943 | unsigned n_perms; |
01d8bf07 | 7944 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
29afecdf RB |
7945 | slp_node_instance, false, |
7946 | &n_perms)) | |
ebfd146a | 7947 | { |
9771b263 | 7948 | dr_chain.release (); |
ebfd146a IR |
7949 | return false; |
7950 | } | |
7951 | } | |
7952 | else | |
7953 | { | |
0d0293ac | 7954 | if (grouped_load) |
ebfd146a | 7955 | { |
2de001ee | 7956 | if (memory_access_type != VMAT_LOAD_STORE_LANES) |
0d0293ac | 7957 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 7958 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
7959 | } |
7960 | else | |
7961 | { | |
7962 | if (j == 0) | |
7963 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7964 | else | |
7965 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7966 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7967 | } | |
7968 | } | |
9771b263 | 7969 | dr_chain.release (); |
ebfd146a IR |
7970 | } |
7971 | ||
ebfd146a IR |
7972 | return true; |
7973 | } | |
7974 | ||
7975 | /* Function vect_is_simple_cond. | |
b8698a0f | 7976 | |
ebfd146a IR |
7977 | Input: |
7978 | LOOP - the loop that is being vectorized. | |
7979 | COND - Condition that is checked for simple use. | |
7980 | ||
e9e1d143 RG |
7981 | Output: |
7982 | *COMP_VECTYPE - the vector type for the comparison. | |
4fc5ebf1 | 7983 | *DTS - The def types for the arguments of the comparison |
e9e1d143 | 7984 | |
ebfd146a IR |
7985 | Returns whether a COND can be vectorized. Checks whether |
7986 | condition operands are supportable using vec_is_simple_use. */ | |
7987 | ||
87aab9b2 | 7988 | static bool |
4fc5ebf1 | 7989 | vect_is_simple_cond (tree cond, vec_info *vinfo, |
8da4c8d8 RB |
7990 | tree *comp_vectype, enum vect_def_type *dts, |
7991 | tree vectype) | |
ebfd146a IR |
7992 | { |
7993 | tree lhs, rhs; | |
e9e1d143 | 7994 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a | 7995 | |
a414c77f IE |
7996 | /* Mask case. */ |
7997 | if (TREE_CODE (cond) == SSA_NAME | |
2568d8a1 | 7998 | && VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (cond))) |
a414c77f IE |
7999 | { |
8000 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (cond); | |
8001 | if (!vect_is_simple_use (cond, vinfo, &lhs_def_stmt, | |
4fc5ebf1 | 8002 | &dts[0], comp_vectype) |
a414c77f IE |
8003 | || !*comp_vectype |
8004 | || !VECTOR_BOOLEAN_TYPE_P (*comp_vectype)) | |
8005 | return false; | |
8006 | return true; | |
8007 | } | |
8008 | ||
ebfd146a IR |
8009 | if (!COMPARISON_CLASS_P (cond)) |
8010 | return false; | |
8011 | ||
8012 | lhs = TREE_OPERAND (cond, 0); | |
8013 | rhs = TREE_OPERAND (cond, 1); | |
8014 | ||
8015 | if (TREE_CODE (lhs) == SSA_NAME) | |
8016 | { | |
355fe088 | 8017 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); |
4fc5ebf1 | 8018 | if (!vect_is_simple_use (lhs, vinfo, &lhs_def_stmt, &dts[0], &vectype1)) |
ebfd146a IR |
8019 | return false; |
8020 | } | |
4fc5ebf1 JG |
8021 | else if (TREE_CODE (lhs) == INTEGER_CST || TREE_CODE (lhs) == REAL_CST |
8022 | || TREE_CODE (lhs) == FIXED_CST) | |
8023 | dts[0] = vect_constant_def; | |
8024 | else | |
ebfd146a IR |
8025 | return false; |
8026 | ||
8027 | if (TREE_CODE (rhs) == SSA_NAME) | |
8028 | { | |
355fe088 | 8029 | gimple *rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); |
4fc5ebf1 | 8030 | if (!vect_is_simple_use (rhs, vinfo, &rhs_def_stmt, &dts[1], &vectype2)) |
ebfd146a IR |
8031 | return false; |
8032 | } | |
4fc5ebf1 JG |
8033 | else if (TREE_CODE (rhs) == INTEGER_CST || TREE_CODE (rhs) == REAL_CST |
8034 | || TREE_CODE (rhs) == FIXED_CST) | |
8035 | dts[1] = vect_constant_def; | |
8036 | else | |
ebfd146a IR |
8037 | return false; |
8038 | ||
28b33016 | 8039 | if (vectype1 && vectype2 |
928686b1 RS |
8040 | && maybe_ne (TYPE_VECTOR_SUBPARTS (vectype1), |
8041 | TYPE_VECTOR_SUBPARTS (vectype2))) | |
28b33016 IE |
8042 | return false; |
8043 | ||
e9e1d143 | 8044 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
8da4c8d8 RB |
8045 | /* Invariant comparison. */ |
8046 | if (! *comp_vectype) | |
8047 | { | |
8048 | tree scalar_type = TREE_TYPE (lhs); | |
8049 | /* If we can widen the comparison to match vectype do so. */ | |
8050 | if (INTEGRAL_TYPE_P (scalar_type) | |
8051 | && tree_int_cst_lt (TYPE_SIZE (scalar_type), | |
8052 | TYPE_SIZE (TREE_TYPE (vectype)))) | |
8053 | scalar_type = build_nonstandard_integer_type | |
8054 | (tree_to_uhwi (TYPE_SIZE (TREE_TYPE (vectype))), | |
8055 | TYPE_UNSIGNED (scalar_type)); | |
8056 | *comp_vectype = get_vectype_for_scalar_type (scalar_type); | |
8057 | } | |
8058 | ||
ebfd146a IR |
8059 | return true; |
8060 | } | |
8061 | ||
8062 | /* vectorizable_condition. | |
8063 | ||
b8698a0f L |
8064 | Check if STMT is conditional modify expression that can be vectorized. |
8065 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
8066 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
8067 | at GSI. |
8068 | ||
8069 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
8070 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
0ad23163 | 8071 | else clause if it is 2). |
ebfd146a IR |
8072 | |
8073 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
8074 | ||
4bbe8262 | 8075 | bool |
355fe088 TS |
8076 | vectorizable_condition (gimple *stmt, gimple_stmt_iterator *gsi, |
8077 | gimple **vec_stmt, tree reduc_def, int reduc_index, | |
f7e531cf | 8078 | slp_tree slp_node) |
ebfd146a IR |
8079 | { |
8080 | tree scalar_dest = NULL_TREE; | |
8081 | tree vec_dest = NULL_TREE; | |
01216d27 JJ |
8082 | tree cond_expr, cond_expr0 = NULL_TREE, cond_expr1 = NULL_TREE; |
8083 | tree then_clause, else_clause; | |
ebfd146a | 8084 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
df11cc78 | 8085 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
8086 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
8087 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
5958f9e2 | 8088 | tree vec_compare; |
ebfd146a IR |
8089 | tree new_temp; |
8090 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4fc5ebf1 JG |
8091 | enum vect_def_type dts[4] |
8092 | = {vect_unknown_def_type, vect_unknown_def_type, | |
8093 | vect_unknown_def_type, vect_unknown_def_type}; | |
8094 | int ndts = 4; | |
f7e531cf | 8095 | int ncopies; |
01216d27 | 8096 | enum tree_code code, cond_code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
a855b1b1 | 8097 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
8098 | int i, j; |
8099 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
8100 | vec<tree> vec_oprnds0 = vNULL; |
8101 | vec<tree> vec_oprnds1 = vNULL; | |
8102 | vec<tree> vec_oprnds2 = vNULL; | |
8103 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 8104 | tree vec_cmp_type; |
a414c77f | 8105 | bool masked = false; |
b8698a0f | 8106 | |
f7e531cf IR |
8107 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
8108 | return false; | |
8109 | ||
af29617a AH |
8110 | if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) == TREE_CODE_REDUCTION) |
8111 | { | |
8112 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
8113 | return false; | |
ebfd146a | 8114 | |
af29617a AH |
8115 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
8116 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
8117 | && reduc_def)) | |
8118 | return false; | |
ebfd146a | 8119 | |
af29617a AH |
8120 | /* FORNOW: not yet supported. */ |
8121 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
8122 | { | |
8123 | if (dump_enabled_p ()) | |
8124 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
8125 | "value used after loop.\n"); | |
8126 | return false; | |
8127 | } | |
ebfd146a IR |
8128 | } |
8129 | ||
8130 | /* Is vectorizable conditional operation? */ | |
8131 | if (!is_gimple_assign (stmt)) | |
8132 | return false; | |
8133 | ||
8134 | code = gimple_assign_rhs_code (stmt); | |
8135 | ||
8136 | if (code != COND_EXPR) | |
8137 | return false; | |
8138 | ||
465c8c19 | 8139 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
2947d3b2 | 8140 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
465c8c19 | 8141 | |
fce57248 | 8142 | if (slp_node) |
465c8c19 JJ |
8143 | ncopies = 1; |
8144 | else | |
e8f142e2 | 8145 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
8146 | |
8147 | gcc_assert (ncopies >= 1); | |
8148 | if (reduc_index && ncopies > 1) | |
8149 | return false; /* FORNOW */ | |
8150 | ||
4e71066d RG |
8151 | cond_expr = gimple_assign_rhs1 (stmt); |
8152 | then_clause = gimple_assign_rhs2 (stmt); | |
8153 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 8154 | |
4fc5ebf1 | 8155 | if (!vect_is_simple_cond (cond_expr, stmt_info->vinfo, |
8da4c8d8 | 8156 | &comp_vectype, &dts[0], vectype) |
e9e1d143 | 8157 | || !comp_vectype) |
ebfd146a IR |
8158 | return false; |
8159 | ||
81c40241 | 8160 | gimple *def_stmt; |
4fc5ebf1 | 8161 | if (!vect_is_simple_use (then_clause, stmt_info->vinfo, &def_stmt, &dts[2], |
2947d3b2 IE |
8162 | &vectype1)) |
8163 | return false; | |
4fc5ebf1 | 8164 | if (!vect_is_simple_use (else_clause, stmt_info->vinfo, &def_stmt, &dts[3], |
2947d3b2 | 8165 | &vectype2)) |
ebfd146a | 8166 | return false; |
2947d3b2 IE |
8167 | |
8168 | if (vectype1 && !useless_type_conversion_p (vectype, vectype1)) | |
8169 | return false; | |
8170 | ||
8171 | if (vectype2 && !useless_type_conversion_p (vectype, vectype2)) | |
ebfd146a IR |
8172 | return false; |
8173 | ||
28b33016 IE |
8174 | masked = !COMPARISON_CLASS_P (cond_expr); |
8175 | vec_cmp_type = build_same_sized_truth_vector_type (comp_vectype); | |
8176 | ||
74946978 MP |
8177 | if (vec_cmp_type == NULL_TREE) |
8178 | return false; | |
784fb9b3 | 8179 | |
01216d27 JJ |
8180 | cond_code = TREE_CODE (cond_expr); |
8181 | if (!masked) | |
8182 | { | |
8183 | cond_expr0 = TREE_OPERAND (cond_expr, 0); | |
8184 | cond_expr1 = TREE_OPERAND (cond_expr, 1); | |
8185 | } | |
8186 | ||
8187 | if (!masked && VECTOR_BOOLEAN_TYPE_P (comp_vectype)) | |
8188 | { | |
8189 | /* Boolean values may have another representation in vectors | |
8190 | and therefore we prefer bit operations over comparison for | |
8191 | them (which also works for scalar masks). We store opcodes | |
8192 | to use in bitop1 and bitop2. Statement is vectorized as | |
8193 | BITOP2 (rhs1 BITOP1 rhs2) or rhs1 BITOP2 (BITOP1 rhs2) | |
8194 | depending on bitop1 and bitop2 arity. */ | |
8195 | switch (cond_code) | |
8196 | { | |
8197 | case GT_EXPR: | |
8198 | bitop1 = BIT_NOT_EXPR; | |
8199 | bitop2 = BIT_AND_EXPR; | |
8200 | break; | |
8201 | case GE_EXPR: | |
8202 | bitop1 = BIT_NOT_EXPR; | |
8203 | bitop2 = BIT_IOR_EXPR; | |
8204 | break; | |
8205 | case LT_EXPR: | |
8206 | bitop1 = BIT_NOT_EXPR; | |
8207 | bitop2 = BIT_AND_EXPR; | |
8208 | std::swap (cond_expr0, cond_expr1); | |
8209 | break; | |
8210 | case LE_EXPR: | |
8211 | bitop1 = BIT_NOT_EXPR; | |
8212 | bitop2 = BIT_IOR_EXPR; | |
8213 | std::swap (cond_expr0, cond_expr1); | |
8214 | break; | |
8215 | case NE_EXPR: | |
8216 | bitop1 = BIT_XOR_EXPR; | |
8217 | break; | |
8218 | case EQ_EXPR: | |
8219 | bitop1 = BIT_XOR_EXPR; | |
8220 | bitop2 = BIT_NOT_EXPR; | |
8221 | break; | |
8222 | default: | |
8223 | return false; | |
8224 | } | |
8225 | cond_code = SSA_NAME; | |
8226 | } | |
8227 | ||
b8698a0f | 8228 | if (!vec_stmt) |
ebfd146a IR |
8229 | { |
8230 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
01216d27 JJ |
8231 | if (bitop1 != NOP_EXPR) |
8232 | { | |
8233 | machine_mode mode = TYPE_MODE (comp_vectype); | |
8234 | optab optab; | |
8235 | ||
8236 | optab = optab_for_tree_code (bitop1, comp_vectype, optab_default); | |
8237 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8238 | return false; | |
8239 | ||
8240 | if (bitop2 != NOP_EXPR) | |
8241 | { | |
8242 | optab = optab_for_tree_code (bitop2, comp_vectype, | |
8243 | optab_default); | |
8244 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8245 | return false; | |
8246 | } | |
8247 | } | |
4fc5ebf1 JG |
8248 | if (expand_vec_cond_expr_p (vectype, comp_vectype, |
8249 | cond_code)) | |
8250 | { | |
8251 | vect_model_simple_cost (stmt_info, ncopies, dts, ndts, NULL, NULL); | |
8252 | return true; | |
8253 | } | |
8254 | return false; | |
ebfd146a IR |
8255 | } |
8256 | ||
f7e531cf IR |
8257 | /* Transform. */ |
8258 | ||
8259 | if (!slp_node) | |
8260 | { | |
9771b263 DN |
8261 | vec_oprnds0.create (1); |
8262 | vec_oprnds1.create (1); | |
8263 | vec_oprnds2.create (1); | |
8264 | vec_oprnds3.create (1); | |
f7e531cf | 8265 | } |
ebfd146a IR |
8266 | |
8267 | /* Handle def. */ | |
8268 | scalar_dest = gimple_assign_lhs (stmt); | |
8269 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
8270 | ||
8271 | /* Handle cond expr. */ | |
a855b1b1 MM |
8272 | for (j = 0; j < ncopies; j++) |
8273 | { | |
538dd0b7 | 8274 | gassign *new_stmt = NULL; |
a855b1b1 MM |
8275 | if (j == 0) |
8276 | { | |
f7e531cf IR |
8277 | if (slp_node) |
8278 | { | |
00f96dc9 TS |
8279 | auto_vec<tree, 4> ops; |
8280 | auto_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 8281 | |
a414c77f | 8282 | if (masked) |
01216d27 | 8283 | ops.safe_push (cond_expr); |
a414c77f IE |
8284 | else |
8285 | { | |
01216d27 JJ |
8286 | ops.safe_push (cond_expr0); |
8287 | ops.safe_push (cond_expr1); | |
a414c77f | 8288 | } |
9771b263 DN |
8289 | ops.safe_push (then_clause); |
8290 | ops.safe_push (else_clause); | |
306b0c92 | 8291 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
37b5ec8f JJ |
8292 | vec_oprnds3 = vec_defs.pop (); |
8293 | vec_oprnds2 = vec_defs.pop (); | |
a414c77f IE |
8294 | if (!masked) |
8295 | vec_oprnds1 = vec_defs.pop (); | |
37b5ec8f | 8296 | vec_oprnds0 = vec_defs.pop (); |
f7e531cf IR |
8297 | } |
8298 | else | |
8299 | { | |
355fe088 | 8300 | gimple *gtemp; |
a414c77f IE |
8301 | if (masked) |
8302 | { | |
8303 | vec_cond_lhs | |
8304 | = vect_get_vec_def_for_operand (cond_expr, stmt, | |
8305 | comp_vectype); | |
8306 | vect_is_simple_use (cond_expr, stmt_info->vinfo, | |
8307 | >emp, &dts[0]); | |
8308 | } | |
8309 | else | |
8310 | { | |
01216d27 JJ |
8311 | vec_cond_lhs |
8312 | = vect_get_vec_def_for_operand (cond_expr0, | |
8313 | stmt, comp_vectype); | |
8314 | vect_is_simple_use (cond_expr0, loop_vinfo, >emp, &dts[0]); | |
8315 | ||
8316 | vec_cond_rhs | |
8317 | = vect_get_vec_def_for_operand (cond_expr1, | |
8318 | stmt, comp_vectype); | |
8319 | vect_is_simple_use (cond_expr1, loop_vinfo, >emp, &dts[1]); | |
a414c77f | 8320 | } |
f7e531cf IR |
8321 | if (reduc_index == 1) |
8322 | vec_then_clause = reduc_def; | |
8323 | else | |
8324 | { | |
8325 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
81c40241 RB |
8326 | stmt); |
8327 | vect_is_simple_use (then_clause, loop_vinfo, | |
8328 | >emp, &dts[2]); | |
f7e531cf IR |
8329 | } |
8330 | if (reduc_index == 2) | |
8331 | vec_else_clause = reduc_def; | |
8332 | else | |
8333 | { | |
8334 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
81c40241 RB |
8335 | stmt); |
8336 | vect_is_simple_use (else_clause, loop_vinfo, >emp, &dts[3]); | |
f7e531cf | 8337 | } |
a855b1b1 MM |
8338 | } |
8339 | } | |
8340 | else | |
8341 | { | |
a414c77f IE |
8342 | vec_cond_lhs |
8343 | = vect_get_vec_def_for_stmt_copy (dts[0], | |
8344 | vec_oprnds0.pop ()); | |
8345 | if (!masked) | |
8346 | vec_cond_rhs | |
8347 | = vect_get_vec_def_for_stmt_copy (dts[1], | |
8348 | vec_oprnds1.pop ()); | |
8349 | ||
a855b1b1 | 8350 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 8351 | vec_oprnds2.pop ()); |
a855b1b1 | 8352 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 8353 | vec_oprnds3.pop ()); |
f7e531cf IR |
8354 | } |
8355 | ||
8356 | if (!slp_node) | |
8357 | { | |
9771b263 | 8358 | vec_oprnds0.quick_push (vec_cond_lhs); |
a414c77f IE |
8359 | if (!masked) |
8360 | vec_oprnds1.quick_push (vec_cond_rhs); | |
9771b263 DN |
8361 | vec_oprnds2.quick_push (vec_then_clause); |
8362 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
8363 | } |
8364 | ||
9dc3f7de | 8365 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 8366 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 8367 | { |
9771b263 DN |
8368 | vec_then_clause = vec_oprnds2[i]; |
8369 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 8370 | |
a414c77f IE |
8371 | if (masked) |
8372 | vec_compare = vec_cond_lhs; | |
8373 | else | |
8374 | { | |
8375 | vec_cond_rhs = vec_oprnds1[i]; | |
01216d27 JJ |
8376 | if (bitop1 == NOP_EXPR) |
8377 | vec_compare = build2 (cond_code, vec_cmp_type, | |
8378 | vec_cond_lhs, vec_cond_rhs); | |
8379 | else | |
8380 | { | |
8381 | new_temp = make_ssa_name (vec_cmp_type); | |
8382 | if (bitop1 == BIT_NOT_EXPR) | |
8383 | new_stmt = gimple_build_assign (new_temp, bitop1, | |
8384 | vec_cond_rhs); | |
8385 | else | |
8386 | new_stmt | |
8387 | = gimple_build_assign (new_temp, bitop1, vec_cond_lhs, | |
8388 | vec_cond_rhs); | |
8389 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8390 | if (bitop2 == NOP_EXPR) | |
8391 | vec_compare = new_temp; | |
8392 | else if (bitop2 == BIT_NOT_EXPR) | |
8393 | { | |
8394 | /* Instead of doing ~x ? y : z do x ? z : y. */ | |
8395 | vec_compare = new_temp; | |
8396 | std::swap (vec_then_clause, vec_else_clause); | |
8397 | } | |
8398 | else | |
8399 | { | |
8400 | vec_compare = make_ssa_name (vec_cmp_type); | |
8401 | new_stmt | |
8402 | = gimple_build_assign (vec_compare, bitop2, | |
8403 | vec_cond_lhs, new_temp); | |
8404 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8405 | } | |
8406 | } | |
a414c77f | 8407 | } |
5958f9e2 JJ |
8408 | new_temp = make_ssa_name (vec_dest); |
8409 | new_stmt = gimple_build_assign (new_temp, VEC_COND_EXPR, | |
8410 | vec_compare, vec_then_clause, | |
8411 | vec_else_clause); | |
f7e531cf IR |
8412 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
8413 | if (slp_node) | |
9771b263 | 8414 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
8415 | } |
8416 | ||
8417 | if (slp_node) | |
8418 | continue; | |
8419 | ||
8420 | if (j == 0) | |
8421 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
8422 | else | |
8423 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
8424 | ||
8425 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 8426 | } |
b8698a0f | 8427 | |
9771b263 DN |
8428 | vec_oprnds0.release (); |
8429 | vec_oprnds1.release (); | |
8430 | vec_oprnds2.release (); | |
8431 | vec_oprnds3.release (); | |
f7e531cf | 8432 | |
ebfd146a IR |
8433 | return true; |
8434 | } | |
8435 | ||
42fd8198 IE |
8436 | /* vectorizable_comparison. |
8437 | ||
8438 | Check if STMT is comparison expression that can be vectorized. | |
8439 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
8440 | comparison, put it in VEC_STMT, and insert it at GSI. | |
8441 | ||
8442 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
8443 | ||
fce57248 | 8444 | static bool |
42fd8198 IE |
8445 | vectorizable_comparison (gimple *stmt, gimple_stmt_iterator *gsi, |
8446 | gimple **vec_stmt, tree reduc_def, | |
8447 | slp_tree slp_node) | |
8448 | { | |
8449 | tree lhs, rhs1, rhs2; | |
8450 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8451 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; | |
8452 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
8453 | tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE; | |
8454 | tree new_temp; | |
8455 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
8456 | enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 8457 | int ndts = 2; |
928686b1 | 8458 | poly_uint64 nunits; |
42fd8198 | 8459 | int ncopies; |
49e76ff1 | 8460 | enum tree_code code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
42fd8198 IE |
8461 | stmt_vec_info prev_stmt_info = NULL; |
8462 | int i, j; | |
8463 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
8464 | vec<tree> vec_oprnds0 = vNULL; | |
8465 | vec<tree> vec_oprnds1 = vNULL; | |
8466 | gimple *def_stmt; | |
8467 | tree mask_type; | |
8468 | tree mask; | |
8469 | ||
c245362b IE |
8470 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
8471 | return false; | |
8472 | ||
30480bcd | 8473 | if (!vectype || !VECTOR_BOOLEAN_TYPE_P (vectype)) |
42fd8198 IE |
8474 | return false; |
8475 | ||
8476 | mask_type = vectype; | |
8477 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
8478 | ||
fce57248 | 8479 | if (slp_node) |
42fd8198 IE |
8480 | ncopies = 1; |
8481 | else | |
e8f142e2 | 8482 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
42fd8198 IE |
8483 | |
8484 | gcc_assert (ncopies >= 1); | |
42fd8198 IE |
8485 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
8486 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
8487 | && reduc_def)) | |
8488 | return false; | |
8489 | ||
8490 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
8491 | { | |
8492 | if (dump_enabled_p ()) | |
8493 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
8494 | "value used after loop.\n"); | |
8495 | return false; | |
8496 | } | |
8497 | ||
8498 | if (!is_gimple_assign (stmt)) | |
8499 | return false; | |
8500 | ||
8501 | code = gimple_assign_rhs_code (stmt); | |
8502 | ||
8503 | if (TREE_CODE_CLASS (code) != tcc_comparison) | |
8504 | return false; | |
8505 | ||
8506 | rhs1 = gimple_assign_rhs1 (stmt); | |
8507 | rhs2 = gimple_assign_rhs2 (stmt); | |
8508 | ||
8509 | if (!vect_is_simple_use (rhs1, stmt_info->vinfo, &def_stmt, | |
8510 | &dts[0], &vectype1)) | |
8511 | return false; | |
8512 | ||
8513 | if (!vect_is_simple_use (rhs2, stmt_info->vinfo, &def_stmt, | |
8514 | &dts[1], &vectype2)) | |
8515 | return false; | |
8516 | ||
8517 | if (vectype1 && vectype2 | |
928686b1 RS |
8518 | && maybe_ne (TYPE_VECTOR_SUBPARTS (vectype1), |
8519 | TYPE_VECTOR_SUBPARTS (vectype2))) | |
42fd8198 IE |
8520 | return false; |
8521 | ||
8522 | vectype = vectype1 ? vectype1 : vectype2; | |
8523 | ||
8524 | /* Invariant comparison. */ | |
8525 | if (!vectype) | |
8526 | { | |
69a9a66f | 8527 | vectype = get_vectype_for_scalar_type (TREE_TYPE (rhs1)); |
928686b1 | 8528 | if (maybe_ne (TYPE_VECTOR_SUBPARTS (vectype), nunits)) |
42fd8198 IE |
8529 | return false; |
8530 | } | |
928686b1 | 8531 | else if (maybe_ne (nunits, TYPE_VECTOR_SUBPARTS (vectype))) |
42fd8198 IE |
8532 | return false; |
8533 | ||
49e76ff1 IE |
8534 | /* Can't compare mask and non-mask types. */ |
8535 | if (vectype1 && vectype2 | |
8536 | && (VECTOR_BOOLEAN_TYPE_P (vectype1) ^ VECTOR_BOOLEAN_TYPE_P (vectype2))) | |
8537 | return false; | |
8538 | ||
8539 | /* Boolean values may have another representation in vectors | |
8540 | and therefore we prefer bit operations over comparison for | |
8541 | them (which also works for scalar masks). We store opcodes | |
8542 | to use in bitop1 and bitop2. Statement is vectorized as | |
8543 | BITOP2 (rhs1 BITOP1 rhs2) or | |
8544 | rhs1 BITOP2 (BITOP1 rhs2) | |
8545 | depending on bitop1 and bitop2 arity. */ | |
8546 | if (VECTOR_BOOLEAN_TYPE_P (vectype)) | |
8547 | { | |
8548 | if (code == GT_EXPR) | |
8549 | { | |
8550 | bitop1 = BIT_NOT_EXPR; | |
8551 | bitop2 = BIT_AND_EXPR; | |
8552 | } | |
8553 | else if (code == GE_EXPR) | |
8554 | { | |
8555 | bitop1 = BIT_NOT_EXPR; | |
8556 | bitop2 = BIT_IOR_EXPR; | |
8557 | } | |
8558 | else if (code == LT_EXPR) | |
8559 | { | |
8560 | bitop1 = BIT_NOT_EXPR; | |
8561 | bitop2 = BIT_AND_EXPR; | |
8562 | std::swap (rhs1, rhs2); | |
264d951a | 8563 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
8564 | } |
8565 | else if (code == LE_EXPR) | |
8566 | { | |
8567 | bitop1 = BIT_NOT_EXPR; | |
8568 | bitop2 = BIT_IOR_EXPR; | |
8569 | std::swap (rhs1, rhs2); | |
264d951a | 8570 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
8571 | } |
8572 | else | |
8573 | { | |
8574 | bitop1 = BIT_XOR_EXPR; | |
8575 | if (code == EQ_EXPR) | |
8576 | bitop2 = BIT_NOT_EXPR; | |
8577 | } | |
8578 | } | |
8579 | ||
42fd8198 IE |
8580 | if (!vec_stmt) |
8581 | { | |
8582 | STMT_VINFO_TYPE (stmt_info) = comparison_vec_info_type; | |
49e76ff1 | 8583 | vect_model_simple_cost (stmt_info, ncopies * (1 + (bitop2 != NOP_EXPR)), |
4fc5ebf1 | 8584 | dts, ndts, NULL, NULL); |
49e76ff1 | 8585 | if (bitop1 == NOP_EXPR) |
96592eed | 8586 | return expand_vec_cmp_expr_p (vectype, mask_type, code); |
49e76ff1 IE |
8587 | else |
8588 | { | |
8589 | machine_mode mode = TYPE_MODE (vectype); | |
8590 | optab optab; | |
8591 | ||
8592 | optab = optab_for_tree_code (bitop1, vectype, optab_default); | |
8593 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8594 | return false; | |
8595 | ||
8596 | if (bitop2 != NOP_EXPR) | |
8597 | { | |
8598 | optab = optab_for_tree_code (bitop2, vectype, optab_default); | |
8599 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8600 | return false; | |
8601 | } | |
8602 | return true; | |
8603 | } | |
42fd8198 IE |
8604 | } |
8605 | ||
8606 | /* Transform. */ | |
8607 | if (!slp_node) | |
8608 | { | |
8609 | vec_oprnds0.create (1); | |
8610 | vec_oprnds1.create (1); | |
8611 | } | |
8612 | ||
8613 | /* Handle def. */ | |
8614 | lhs = gimple_assign_lhs (stmt); | |
8615 | mask = vect_create_destination_var (lhs, mask_type); | |
8616 | ||
8617 | /* Handle cmp expr. */ | |
8618 | for (j = 0; j < ncopies; j++) | |
8619 | { | |
8620 | gassign *new_stmt = NULL; | |
8621 | if (j == 0) | |
8622 | { | |
8623 | if (slp_node) | |
8624 | { | |
8625 | auto_vec<tree, 2> ops; | |
8626 | auto_vec<vec<tree>, 2> vec_defs; | |
8627 | ||
8628 | ops.safe_push (rhs1); | |
8629 | ops.safe_push (rhs2); | |
306b0c92 | 8630 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
42fd8198 IE |
8631 | vec_oprnds1 = vec_defs.pop (); |
8632 | vec_oprnds0 = vec_defs.pop (); | |
8633 | } | |
8634 | else | |
8635 | { | |
e4af0bc4 IE |
8636 | vec_rhs1 = vect_get_vec_def_for_operand (rhs1, stmt, vectype); |
8637 | vec_rhs2 = vect_get_vec_def_for_operand (rhs2, stmt, vectype); | |
42fd8198 IE |
8638 | } |
8639 | } | |
8640 | else | |
8641 | { | |
8642 | vec_rhs1 = vect_get_vec_def_for_stmt_copy (dts[0], | |
8643 | vec_oprnds0.pop ()); | |
8644 | vec_rhs2 = vect_get_vec_def_for_stmt_copy (dts[1], | |
8645 | vec_oprnds1.pop ()); | |
8646 | } | |
8647 | ||
8648 | if (!slp_node) | |
8649 | { | |
8650 | vec_oprnds0.quick_push (vec_rhs1); | |
8651 | vec_oprnds1.quick_push (vec_rhs2); | |
8652 | } | |
8653 | ||
8654 | /* Arguments are ready. Create the new vector stmt. */ | |
8655 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1) | |
8656 | { | |
8657 | vec_rhs2 = vec_oprnds1[i]; | |
8658 | ||
8659 | new_temp = make_ssa_name (mask); | |
49e76ff1 IE |
8660 | if (bitop1 == NOP_EXPR) |
8661 | { | |
8662 | new_stmt = gimple_build_assign (new_temp, code, | |
8663 | vec_rhs1, vec_rhs2); | |
8664 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8665 | } | |
8666 | else | |
8667 | { | |
8668 | if (bitop1 == BIT_NOT_EXPR) | |
8669 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs2); | |
8670 | else | |
8671 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs1, | |
8672 | vec_rhs2); | |
8673 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8674 | if (bitop2 != NOP_EXPR) | |
8675 | { | |
8676 | tree res = make_ssa_name (mask); | |
8677 | if (bitop2 == BIT_NOT_EXPR) | |
8678 | new_stmt = gimple_build_assign (res, bitop2, new_temp); | |
8679 | else | |
8680 | new_stmt = gimple_build_assign (res, bitop2, vec_rhs1, | |
8681 | new_temp); | |
8682 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8683 | } | |
8684 | } | |
42fd8198 IE |
8685 | if (slp_node) |
8686 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
8687 | } | |
8688 | ||
8689 | if (slp_node) | |
8690 | continue; | |
8691 | ||
8692 | if (j == 0) | |
8693 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
8694 | else | |
8695 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
8696 | ||
8697 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
8698 | } | |
8699 | ||
8700 | vec_oprnds0.release (); | |
8701 | vec_oprnds1.release (); | |
8702 | ||
8703 | return true; | |
8704 | } | |
ebfd146a | 8705 | |
68a0f2ff RS |
8706 | /* If SLP_NODE is nonnull, return true if vectorizable_live_operation |
8707 | can handle all live statements in the node. Otherwise return true | |
8708 | if STMT is not live or if vectorizable_live_operation can handle it. | |
8709 | GSI and VEC_STMT are as for vectorizable_live_operation. */ | |
8710 | ||
8711 | static bool | |
8712 | can_vectorize_live_stmts (gimple *stmt, gimple_stmt_iterator *gsi, | |
8713 | slp_tree slp_node, gimple **vec_stmt) | |
8714 | { | |
8715 | if (slp_node) | |
8716 | { | |
8717 | gimple *slp_stmt; | |
8718 | unsigned int i; | |
8719 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (slp_node), i, slp_stmt) | |
8720 | { | |
8721 | stmt_vec_info slp_stmt_info = vinfo_for_stmt (slp_stmt); | |
8722 | if (STMT_VINFO_LIVE_P (slp_stmt_info) | |
8723 | && !vectorizable_live_operation (slp_stmt, gsi, slp_node, i, | |
8724 | vec_stmt)) | |
8725 | return false; | |
8726 | } | |
8727 | } | |
8728 | else if (STMT_VINFO_LIVE_P (vinfo_for_stmt (stmt)) | |
8729 | && !vectorizable_live_operation (stmt, gsi, slp_node, -1, vec_stmt)) | |
8730 | return false; | |
8731 | ||
8732 | return true; | |
8733 | } | |
8734 | ||
8644a673 | 8735 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
8736 | |
8737 | bool | |
891ad31c RB |
8738 | vect_analyze_stmt (gimple *stmt, bool *need_to_vectorize, slp_tree node, |
8739 | slp_instance node_instance) | |
ebfd146a | 8740 | { |
8644a673 | 8741 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 8742 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 8743 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 8744 | bool ok; |
355fe088 | 8745 | gimple *pattern_stmt; |
363477c0 | 8746 | gimple_seq pattern_def_seq; |
ebfd146a | 8747 | |
73fbfcad | 8748 | if (dump_enabled_p ()) |
ebfd146a | 8749 | { |
78c60e3d SS |
8750 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
8751 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
8644a673 | 8752 | } |
ebfd146a | 8753 | |
1825a1f3 | 8754 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 8755 | { |
73fbfcad | 8756 | if (dump_enabled_p ()) |
78c60e3d | 8757 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8758 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
8759 | |
8760 | return false; | |
8761 | } | |
b8698a0f L |
8762 | |
8763 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
8764 | to include: |
8765 | - the COND_EXPR which is the loop exit condition | |
8766 | - any LABEL_EXPRs in the loop | |
b8698a0f | 8767 | - computations that are used only for array indexing or loop control. |
8644a673 | 8768 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 8769 | instance, therefore, all the statements are relevant. |
ebfd146a | 8770 | |
d092494c | 8771 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 8772 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
8773 | statements. In basic blocks we are called from some SLP instance |
8774 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
8775 | already will be part of SLP instance. */ | |
83197f37 IR |
8776 | |
8777 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 8778 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 8779 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 8780 | { |
9d5e7640 | 8781 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 8782 | && pattern_stmt |
9d5e7640 IR |
8783 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
8784 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
8785 | { | |
83197f37 | 8786 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
8787 | stmt = pattern_stmt; |
8788 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 8789 | if (dump_enabled_p ()) |
9d5e7640 | 8790 | { |
78c60e3d SS |
8791 | dump_printf_loc (MSG_NOTE, vect_location, |
8792 | "==> examining pattern statement: "); | |
8793 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
9d5e7640 IR |
8794 | } |
8795 | } | |
8796 | else | |
8797 | { | |
73fbfcad | 8798 | if (dump_enabled_p ()) |
e645e942 | 8799 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 8800 | |
9d5e7640 IR |
8801 | return true; |
8802 | } | |
8644a673 | 8803 | } |
83197f37 | 8804 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 8805 | && node == NULL |
83197f37 IR |
8806 | && pattern_stmt |
8807 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
8808 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
8809 | { | |
8810 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 8811 | if (dump_enabled_p ()) |
83197f37 | 8812 | { |
78c60e3d SS |
8813 | dump_printf_loc (MSG_NOTE, vect_location, |
8814 | "==> examining pattern statement: "); | |
8815 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
83197f37 IR |
8816 | } |
8817 | ||
891ad31c RB |
8818 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node, |
8819 | node_instance)) | |
83197f37 IR |
8820 | return false; |
8821 | } | |
ebfd146a | 8822 | |
1107f3ae | 8823 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 8824 | && node == NULL |
363477c0 | 8825 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 8826 | { |
363477c0 | 8827 | gimple_stmt_iterator si; |
1107f3ae | 8828 | |
363477c0 JJ |
8829 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
8830 | { | |
355fe088 | 8831 | gimple *pattern_def_stmt = gsi_stmt (si); |
363477c0 JJ |
8832 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) |
8833 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
8834 | { | |
8835 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 8836 | if (dump_enabled_p ()) |
363477c0 | 8837 | { |
78c60e3d SS |
8838 | dump_printf_loc (MSG_NOTE, vect_location, |
8839 | "==> examining pattern def statement: "); | |
8840 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
363477c0 | 8841 | } |
1107f3ae | 8842 | |
363477c0 | 8843 | if (!vect_analyze_stmt (pattern_def_stmt, |
891ad31c | 8844 | need_to_vectorize, node, node_instance)) |
363477c0 JJ |
8845 | return false; |
8846 | } | |
8847 | } | |
8848 | } | |
1107f3ae | 8849 | |
8644a673 IR |
8850 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
8851 | { | |
8852 | case vect_internal_def: | |
8853 | break; | |
ebfd146a | 8854 | |
8644a673 | 8855 | case vect_reduction_def: |
7c5222ff | 8856 | case vect_nested_cycle: |
14a61437 RB |
8857 | gcc_assert (!bb_vinfo |
8858 | && (relevance == vect_used_in_outer | |
8859 | || relevance == vect_used_in_outer_by_reduction | |
8860 | || relevance == vect_used_by_reduction | |
b28ead45 AH |
8861 | || relevance == vect_unused_in_scope |
8862 | || relevance == vect_used_only_live)); | |
8644a673 IR |
8863 | break; |
8864 | ||
8865 | case vect_induction_def: | |
e7baeb39 RB |
8866 | gcc_assert (!bb_vinfo); |
8867 | break; | |
8868 | ||
8644a673 IR |
8869 | case vect_constant_def: |
8870 | case vect_external_def: | |
8871 | case vect_unknown_def_type: | |
8872 | default: | |
8873 | gcc_unreachable (); | |
8874 | } | |
ebfd146a | 8875 | |
8644a673 | 8876 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 8877 | { |
8644a673 | 8878 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
0136f8f0 AH |
8879 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
8880 | || (is_gimple_call (stmt) | |
8881 | && gimple_call_lhs (stmt) == NULL_TREE)); | |
8644a673 | 8882 | *need_to_vectorize = true; |
ebfd146a IR |
8883 | } |
8884 | ||
b1af7da6 RB |
8885 | if (PURE_SLP_STMT (stmt_info) && !node) |
8886 | { | |
8887 | dump_printf_loc (MSG_NOTE, vect_location, | |
8888 | "handled only by SLP analysis\n"); | |
8889 | return true; | |
8890 | } | |
8891 | ||
8892 | ok = true; | |
8893 | if (!bb_vinfo | |
8894 | && (STMT_VINFO_RELEVANT_P (stmt_info) | |
8895 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
8896 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
8897 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
8898 | || vectorizable_shift (stmt, NULL, NULL, node) | |
8899 | || vectorizable_operation (stmt, NULL, NULL, node) | |
8900 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
8901 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
8902 | || vectorizable_call (stmt, NULL, NULL, node) | |
8903 | || vectorizable_store (stmt, NULL, NULL, node) | |
891ad31c | 8904 | || vectorizable_reduction (stmt, NULL, NULL, node, node_instance) |
e7baeb39 | 8905 | || vectorizable_induction (stmt, NULL, NULL, node) |
42fd8198 IE |
8906 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
8907 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 RB |
8908 | else |
8909 | { | |
8910 | if (bb_vinfo) | |
8911 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
8912 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
8913 | || vectorizable_shift (stmt, NULL, NULL, node) | |
8914 | || vectorizable_operation (stmt, NULL, NULL, node) | |
8915 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
8916 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
8917 | || vectorizable_call (stmt, NULL, NULL, node) | |
8918 | || vectorizable_store (stmt, NULL, NULL, node) | |
42fd8198 IE |
8919 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
8920 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 | 8921 | } |
8644a673 IR |
8922 | |
8923 | if (!ok) | |
ebfd146a | 8924 | { |
73fbfcad | 8925 | if (dump_enabled_p ()) |
8644a673 | 8926 | { |
78c60e3d SS |
8927 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8928 | "not vectorized: relevant stmt not "); | |
8929 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
8930 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 8931 | } |
b8698a0f | 8932 | |
ebfd146a IR |
8933 | return false; |
8934 | } | |
8935 | ||
a70d6342 IR |
8936 | if (bb_vinfo) |
8937 | return true; | |
8938 | ||
8644a673 IR |
8939 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
8940 | need extra handling, except for vectorizable reductions. */ | |
68a0f2ff RS |
8941 | if (STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
8942 | && !can_vectorize_live_stmts (stmt, NULL, node, NULL)) | |
ebfd146a | 8943 | { |
73fbfcad | 8944 | if (dump_enabled_p ()) |
8644a673 | 8945 | { |
78c60e3d | 8946 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
68a0f2ff | 8947 | "not vectorized: live stmt not supported: "); |
78c60e3d | 8948 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); |
8644a673 | 8949 | } |
b8698a0f | 8950 | |
8644a673 | 8951 | return false; |
ebfd146a IR |
8952 | } |
8953 | ||
ebfd146a IR |
8954 | return true; |
8955 | } | |
8956 | ||
8957 | ||
8958 | /* Function vect_transform_stmt. | |
8959 | ||
8960 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
8961 | ||
8962 | bool | |
355fe088 | 8963 | vect_transform_stmt (gimple *stmt, gimple_stmt_iterator *gsi, |
0d0293ac | 8964 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
8965 | slp_instance slp_node_instance) |
8966 | { | |
8967 | bool is_store = false; | |
355fe088 | 8968 | gimple *vec_stmt = NULL; |
ebfd146a | 8969 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
ebfd146a | 8970 | bool done; |
ebfd146a | 8971 | |
fce57248 | 8972 | gcc_assert (slp_node || !PURE_SLP_STMT (stmt_info)); |
355fe088 | 8973 | gimple *old_vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
225ce44b | 8974 | |
ebfd146a IR |
8975 | switch (STMT_VINFO_TYPE (stmt_info)) |
8976 | { | |
8977 | case type_demotion_vec_info_type: | |
ebfd146a | 8978 | case type_promotion_vec_info_type: |
ebfd146a IR |
8979 | case type_conversion_vec_info_type: |
8980 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
8981 | gcc_assert (done); | |
8982 | break; | |
8983 | ||
8984 | case induc_vec_info_type: | |
e7baeb39 | 8985 | done = vectorizable_induction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
8986 | gcc_assert (done); |
8987 | break; | |
8988 | ||
9dc3f7de IR |
8989 | case shift_vec_info_type: |
8990 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
8991 | gcc_assert (done); | |
8992 | break; | |
8993 | ||
ebfd146a IR |
8994 | case op_vec_info_type: |
8995 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
8996 | gcc_assert (done); | |
8997 | break; | |
8998 | ||
8999 | case assignment_vec_info_type: | |
9000 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
9001 | gcc_assert (done); | |
9002 | break; | |
9003 | ||
9004 | case load_vec_info_type: | |
b8698a0f | 9005 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
9006 | slp_node_instance); |
9007 | gcc_assert (done); | |
9008 | break; | |
9009 | ||
9010 | case store_vec_info_type: | |
9011 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
9012 | gcc_assert (done); | |
0d0293ac | 9013 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
9014 | { |
9015 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 9016 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
9017 | one are skipped, and there vec_stmt_info shouldn't be freed |
9018 | meanwhile. */ | |
0d0293ac | 9019 | *grouped_store = true; |
ebfd146a IR |
9020 | if (STMT_VINFO_VEC_STMT (stmt_info)) |
9021 | is_store = true; | |
9022 | } | |
9023 | else | |
9024 | is_store = true; | |
9025 | break; | |
9026 | ||
9027 | case condition_vec_info_type: | |
f7e531cf | 9028 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
9029 | gcc_assert (done); |
9030 | break; | |
9031 | ||
42fd8198 IE |
9032 | case comparison_vec_info_type: |
9033 | done = vectorizable_comparison (stmt, gsi, &vec_stmt, NULL, slp_node); | |
9034 | gcc_assert (done); | |
9035 | break; | |
9036 | ||
ebfd146a | 9037 | case call_vec_info_type: |
190c2236 | 9038 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 9039 | stmt = gsi_stmt (*gsi); |
ebfd146a IR |
9040 | break; |
9041 | ||
0136f8f0 AH |
9042 | case call_simd_clone_vec_info_type: |
9043 | done = vectorizable_simd_clone_call (stmt, gsi, &vec_stmt, slp_node); | |
9044 | stmt = gsi_stmt (*gsi); | |
9045 | break; | |
9046 | ||
ebfd146a | 9047 | case reduc_vec_info_type: |
891ad31c RB |
9048 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node, |
9049 | slp_node_instance); | |
ebfd146a IR |
9050 | gcc_assert (done); |
9051 | break; | |
9052 | ||
9053 | default: | |
9054 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
9055 | { | |
73fbfcad | 9056 | if (dump_enabled_p ()) |
78c60e3d | 9057 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9058 | "stmt not supported.\n"); |
ebfd146a IR |
9059 | gcc_unreachable (); |
9060 | } | |
9061 | } | |
9062 | ||
225ce44b RB |
9063 | /* Verify SLP vectorization doesn't mess with STMT_VINFO_VEC_STMT. |
9064 | This would break hybrid SLP vectorization. */ | |
9065 | if (slp_node) | |
d90f8440 RB |
9066 | gcc_assert (!vec_stmt |
9067 | && STMT_VINFO_VEC_STMT (stmt_info) == old_vec_stmt); | |
225ce44b | 9068 | |
ebfd146a IR |
9069 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that |
9070 | is being vectorized, but outside the immediately enclosing loop. */ | |
9071 | if (vec_stmt | |
a70d6342 IR |
9072 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
9073 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
9074 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
9075 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
9076 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 9077 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 9078 | vect_used_in_outer_by_reduction)) |
ebfd146a | 9079 | { |
a70d6342 IR |
9080 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
9081 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
9082 | imm_use_iterator imm_iter; |
9083 | use_operand_p use_p; | |
9084 | tree scalar_dest; | |
355fe088 | 9085 | gimple *exit_phi; |
ebfd146a | 9086 | |
73fbfcad | 9087 | if (dump_enabled_p ()) |
78c60e3d | 9088 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 9089 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
9090 | |
9091 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
9092 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
9093 | STMT). */ | |
9094 | if (gimple_code (stmt) == GIMPLE_PHI) | |
9095 | scalar_dest = PHI_RESULT (stmt); | |
9096 | else | |
9097 | scalar_dest = gimple_assign_lhs (stmt); | |
9098 | ||
9099 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
9100 | { | |
9101 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
9102 | { | |
9103 | exit_phi = USE_STMT (use_p); | |
9104 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
9105 | } | |
9106 | } | |
9107 | } | |
9108 | ||
9109 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
9110 | being vectorized. */ | |
68a0f2ff | 9111 | if (STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) |
ebfd146a | 9112 | { |
68a0f2ff | 9113 | done = can_vectorize_live_stmts (stmt, gsi, slp_node, &vec_stmt); |
ebfd146a IR |
9114 | gcc_assert (done); |
9115 | } | |
9116 | ||
9117 | if (vec_stmt) | |
83197f37 | 9118 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 9119 | |
b8698a0f | 9120 | return is_store; |
ebfd146a IR |
9121 | } |
9122 | ||
9123 | ||
b8698a0f | 9124 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
9125 | stmt_vec_info. */ |
9126 | ||
9127 | void | |
355fe088 | 9128 | vect_remove_stores (gimple *first_stmt) |
ebfd146a | 9129 | { |
355fe088 TS |
9130 | gimple *next = first_stmt; |
9131 | gimple *tmp; | |
ebfd146a IR |
9132 | gimple_stmt_iterator next_si; |
9133 | ||
9134 | while (next) | |
9135 | { | |
78048b1c JJ |
9136 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
9137 | ||
9138 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
9139 | if (is_pattern_stmt_p (stmt_info)) | |
9140 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
9141 | /* Free the attached stmt_vec_info and remove the stmt. */ |
9142 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 9143 | unlink_stmt_vdef (next); |
ebfd146a | 9144 | gsi_remove (&next_si, true); |
3d3f2249 | 9145 | release_defs (next); |
ebfd146a IR |
9146 | free_stmt_vec_info (next); |
9147 | next = tmp; | |
9148 | } | |
9149 | } | |
9150 | ||
9151 | ||
9152 | /* Function new_stmt_vec_info. | |
9153 | ||
9154 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
9155 | ||
9156 | stmt_vec_info | |
310213d4 | 9157 | new_stmt_vec_info (gimple *stmt, vec_info *vinfo) |
ebfd146a IR |
9158 | { |
9159 | stmt_vec_info res; | |
9160 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
9161 | ||
9162 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
9163 | STMT_VINFO_STMT (res) = stmt; | |
310213d4 | 9164 | res->vinfo = vinfo; |
8644a673 | 9165 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
9166 | STMT_VINFO_LIVE_P (res) = false; |
9167 | STMT_VINFO_VECTYPE (res) = NULL; | |
9168 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 9169 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
9170 | STMT_VINFO_IN_PATTERN_P (res) = false; |
9171 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 9172 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a | 9173 | STMT_VINFO_DATA_REF (res) = NULL; |
af29617a | 9174 | STMT_VINFO_VEC_REDUCTION_TYPE (res) = TREE_CODE_REDUCTION; |
7e16ce79 | 9175 | STMT_VINFO_VEC_CONST_COND_REDUC_CODE (res) = ERROR_MARK; |
ebfd146a | 9176 | |
ebfd146a IR |
9177 | if (gimple_code (stmt) == GIMPLE_PHI |
9178 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
9179 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
9180 | else | |
8644a673 IR |
9181 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
9182 | ||
9771b263 | 9183 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 9184 | STMT_SLP_TYPE (res) = loop_vect; |
78810bd3 RB |
9185 | STMT_VINFO_NUM_SLP_USES (res) = 0; |
9186 | ||
e14c1050 IR |
9187 | GROUP_FIRST_ELEMENT (res) = NULL; |
9188 | GROUP_NEXT_ELEMENT (res) = NULL; | |
9189 | GROUP_SIZE (res) = 0; | |
9190 | GROUP_STORE_COUNT (res) = 0; | |
9191 | GROUP_GAP (res) = 0; | |
9192 | GROUP_SAME_DR_STMT (res) = NULL; | |
ebfd146a IR |
9193 | |
9194 | return res; | |
9195 | } | |
9196 | ||
9197 | ||
9198 | /* Create a hash table for stmt_vec_info. */ | |
9199 | ||
9200 | void | |
9201 | init_stmt_vec_info_vec (void) | |
9202 | { | |
9771b263 DN |
9203 | gcc_assert (!stmt_vec_info_vec.exists ()); |
9204 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
9205 | } |
9206 | ||
9207 | ||
9208 | /* Free hash table for stmt_vec_info. */ | |
9209 | ||
9210 | void | |
9211 | free_stmt_vec_info_vec (void) | |
9212 | { | |
93675444 | 9213 | unsigned int i; |
3161455c | 9214 | stmt_vec_info info; |
93675444 JJ |
9215 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) |
9216 | if (info != NULL) | |
3161455c | 9217 | free_stmt_vec_info (STMT_VINFO_STMT (info)); |
9771b263 DN |
9218 | gcc_assert (stmt_vec_info_vec.exists ()); |
9219 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
9220 | } |
9221 | ||
9222 | ||
9223 | /* Free stmt vectorization related info. */ | |
9224 | ||
9225 | void | |
355fe088 | 9226 | free_stmt_vec_info (gimple *stmt) |
ebfd146a IR |
9227 | { |
9228 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
9229 | ||
9230 | if (!stmt_info) | |
9231 | return; | |
9232 | ||
78048b1c JJ |
9233 | /* Check if this statement has a related "pattern stmt" |
9234 | (introduced by the vectorizer during the pattern recognition | |
9235 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
9236 | too. */ | |
9237 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
9238 | { | |
9239 | stmt_vec_info patt_info | |
9240 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
9241 | if (patt_info) | |
9242 | { | |
363477c0 | 9243 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
355fe088 | 9244 | gimple *patt_stmt = STMT_VINFO_STMT (patt_info); |
f0281fde RB |
9245 | gimple_set_bb (patt_stmt, NULL); |
9246 | tree lhs = gimple_get_lhs (patt_stmt); | |
e6f5c25d | 9247 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde | 9248 | release_ssa_name (lhs); |
363477c0 JJ |
9249 | if (seq) |
9250 | { | |
9251 | gimple_stmt_iterator si; | |
9252 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
f0281fde | 9253 | { |
355fe088 | 9254 | gimple *seq_stmt = gsi_stmt (si); |
f0281fde | 9255 | gimple_set_bb (seq_stmt, NULL); |
7532abf2 | 9256 | lhs = gimple_get_lhs (seq_stmt); |
e6f5c25d | 9257 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde RB |
9258 | release_ssa_name (lhs); |
9259 | free_stmt_vec_info (seq_stmt); | |
9260 | } | |
363477c0 | 9261 | } |
f0281fde | 9262 | free_stmt_vec_info (patt_stmt); |
78048b1c JJ |
9263 | } |
9264 | } | |
9265 | ||
9771b263 | 9266 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
6c9e85fb | 9267 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).release (); |
ebfd146a IR |
9268 | set_vinfo_for_stmt (stmt, NULL); |
9269 | free (stmt_info); | |
9270 | } | |
9271 | ||
9272 | ||
bb67d9c7 | 9273 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 9274 | |
bb67d9c7 | 9275 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
9276 | by the target. */ |
9277 | ||
c803b2a9 | 9278 | tree |
86e36728 | 9279 | get_vectype_for_scalar_type_and_size (tree scalar_type, poly_uint64 size) |
ebfd146a | 9280 | { |
c7d97b28 | 9281 | tree orig_scalar_type = scalar_type; |
3bd8f481 | 9282 | scalar_mode inner_mode; |
ef4bddc2 | 9283 | machine_mode simd_mode; |
86e36728 | 9284 | poly_uint64 nunits; |
ebfd146a IR |
9285 | tree vectype; |
9286 | ||
3bd8f481 RS |
9287 | if (!is_int_mode (TYPE_MODE (scalar_type), &inner_mode) |
9288 | && !is_float_mode (TYPE_MODE (scalar_type), &inner_mode)) | |
ebfd146a IR |
9289 | return NULL_TREE; |
9290 | ||
3bd8f481 | 9291 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
48f2e373 | 9292 | |
7b7b1813 RG |
9293 | /* For vector types of elements whose mode precision doesn't |
9294 | match their types precision we use a element type of mode | |
9295 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
9296 | they support the proper result truncation/extension. |
9297 | We also make sure to build vector types with INTEGER_TYPE | |
9298 | component type only. */ | |
6d7971b8 | 9299 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
9300 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
9301 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
9302 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
9303 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 9304 | |
ccbf5bb4 RG |
9305 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
9306 | When the component mode passes the above test simply use a type | |
9307 | corresponding to that mode. The theory is that any use that | |
9308 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 9309 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 9310 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
9311 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
9312 | ||
9313 | /* We can't build a vector type of elements with alignment bigger than | |
9314 | their size. */ | |
dfc2e2ac | 9315 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
9316 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
9317 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 9318 | |
dfc2e2ac RB |
9319 | /* If we felt back to using the mode fail if there was |
9320 | no scalar type for it. */ | |
9321 | if (scalar_type == NULL_TREE) | |
9322 | return NULL_TREE; | |
9323 | ||
bb67d9c7 RG |
9324 | /* If no size was supplied use the mode the target prefers. Otherwise |
9325 | lookup a vector mode of the specified size. */ | |
86e36728 | 9326 | if (known_eq (size, 0U)) |
bb67d9c7 | 9327 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); |
86e36728 RS |
9328 | else if (!multiple_p (size, nbytes, &nunits) |
9329 | || !mode_for_vector (inner_mode, nunits).exists (&simd_mode)) | |
9da15d40 | 9330 | return NULL_TREE; |
4c8fd8ac | 9331 | /* NOTE: nunits == 1 is allowed to support single element vector types. */ |
86e36728 | 9332 | if (!multiple_p (GET_MODE_SIZE (simd_mode), nbytes, &nunits)) |
cc4b5170 | 9333 | return NULL_TREE; |
ebfd146a IR |
9334 | |
9335 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
9336 | |
9337 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
9338 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 9339 | return NULL_TREE; |
ebfd146a | 9340 | |
c7d97b28 RB |
9341 | /* Re-attach the address-space qualifier if we canonicalized the scalar |
9342 | type. */ | |
9343 | if (TYPE_ADDR_SPACE (orig_scalar_type) != TYPE_ADDR_SPACE (vectype)) | |
9344 | return build_qualified_type | |
9345 | (vectype, KEEP_QUAL_ADDR_SPACE (TYPE_QUALS (orig_scalar_type))); | |
9346 | ||
ebfd146a IR |
9347 | return vectype; |
9348 | } | |
9349 | ||
86e36728 | 9350 | poly_uint64 current_vector_size; |
bb67d9c7 RG |
9351 | |
9352 | /* Function get_vectype_for_scalar_type. | |
9353 | ||
9354 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
9355 | by the target. */ | |
9356 | ||
9357 | tree | |
9358 | get_vectype_for_scalar_type (tree scalar_type) | |
9359 | { | |
9360 | tree vectype; | |
9361 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
9362 | current_vector_size); | |
9363 | if (vectype | |
86e36728 | 9364 | && known_eq (current_vector_size, 0U)) |
bb67d9c7 RG |
9365 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); |
9366 | return vectype; | |
9367 | } | |
9368 | ||
42fd8198 IE |
9369 | /* Function get_mask_type_for_scalar_type. |
9370 | ||
9371 | Returns the mask type corresponding to a result of comparison | |
9372 | of vectors of specified SCALAR_TYPE as supported by target. */ | |
9373 | ||
9374 | tree | |
9375 | get_mask_type_for_scalar_type (tree scalar_type) | |
9376 | { | |
9377 | tree vectype = get_vectype_for_scalar_type (scalar_type); | |
9378 | ||
9379 | if (!vectype) | |
9380 | return NULL; | |
9381 | ||
9382 | return build_truth_vector_type (TYPE_VECTOR_SUBPARTS (vectype), | |
9383 | current_vector_size); | |
9384 | } | |
9385 | ||
b690cc0f RG |
9386 | /* Function get_same_sized_vectype |
9387 | ||
9388 | Returns a vector type corresponding to SCALAR_TYPE of size | |
9389 | VECTOR_TYPE if supported by the target. */ | |
9390 | ||
9391 | tree | |
bb67d9c7 | 9392 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 9393 | { |
2568d8a1 | 9394 | if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type)) |
9f47c7e5 IE |
9395 | return build_same_sized_truth_vector_type (vector_type); |
9396 | ||
bb67d9c7 RG |
9397 | return get_vectype_for_scalar_type_and_size |
9398 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
9399 | } |
9400 | ||
ebfd146a IR |
9401 | /* Function vect_is_simple_use. |
9402 | ||
9403 | Input: | |
81c40241 RB |
9404 | VINFO - the vect info of the loop or basic block that is being vectorized. |
9405 | OPERAND - operand in the loop or bb. | |
9406 | Output: | |
9407 | DEF_STMT - the defining stmt in case OPERAND is an SSA_NAME. | |
9408 | DT - the type of definition | |
ebfd146a IR |
9409 | |
9410 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 9411 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 9412 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 9413 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
9414 | is the case in reduction/induction computations). |
9415 | For basic blocks, supportable operands are constants and bb invariants. | |
9416 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
9417 | |
9418 | bool | |
81c40241 RB |
9419 | vect_is_simple_use (tree operand, vec_info *vinfo, |
9420 | gimple **def_stmt, enum vect_def_type *dt) | |
b8698a0f | 9421 | { |
ebfd146a | 9422 | *def_stmt = NULL; |
3fc356dc | 9423 | *dt = vect_unknown_def_type; |
b8698a0f | 9424 | |
73fbfcad | 9425 | if (dump_enabled_p ()) |
ebfd146a | 9426 | { |
78c60e3d SS |
9427 | dump_printf_loc (MSG_NOTE, vect_location, |
9428 | "vect_is_simple_use: operand "); | |
9429 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
e645e942 | 9430 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 9431 | } |
b8698a0f | 9432 | |
b758f602 | 9433 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
9434 | { |
9435 | *dt = vect_constant_def; | |
9436 | return true; | |
9437 | } | |
b8698a0f | 9438 | |
ebfd146a IR |
9439 | if (is_gimple_min_invariant (operand)) |
9440 | { | |
8644a673 | 9441 | *dt = vect_external_def; |
ebfd146a IR |
9442 | return true; |
9443 | } | |
9444 | ||
ebfd146a IR |
9445 | if (TREE_CODE (operand) != SSA_NAME) |
9446 | { | |
73fbfcad | 9447 | if (dump_enabled_p ()) |
af29617a AH |
9448 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9449 | "not ssa-name.\n"); | |
ebfd146a IR |
9450 | return false; |
9451 | } | |
b8698a0f | 9452 | |
3fc356dc | 9453 | if (SSA_NAME_IS_DEFAULT_DEF (operand)) |
ebfd146a | 9454 | { |
3fc356dc RB |
9455 | *dt = vect_external_def; |
9456 | return true; | |
ebfd146a IR |
9457 | } |
9458 | ||
3fc356dc | 9459 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
73fbfcad | 9460 | if (dump_enabled_p ()) |
ebfd146a | 9461 | { |
78c60e3d SS |
9462 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
9463 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
ebfd146a IR |
9464 | } |
9465 | ||
61d371eb | 9466 | if (! vect_stmt_in_region_p (vinfo, *def_stmt)) |
8644a673 | 9467 | *dt = vect_external_def; |
ebfd146a IR |
9468 | else |
9469 | { | |
3fc356dc | 9470 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (*def_stmt); |
603cca93 | 9471 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
ebfd146a IR |
9472 | } |
9473 | ||
2e8ab70c RB |
9474 | if (dump_enabled_p ()) |
9475 | { | |
9476 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: "); | |
9477 | switch (*dt) | |
9478 | { | |
9479 | case vect_uninitialized_def: | |
9480 | dump_printf (MSG_NOTE, "uninitialized\n"); | |
9481 | break; | |
9482 | case vect_constant_def: | |
9483 | dump_printf (MSG_NOTE, "constant\n"); | |
9484 | break; | |
9485 | case vect_external_def: | |
9486 | dump_printf (MSG_NOTE, "external\n"); | |
9487 | break; | |
9488 | case vect_internal_def: | |
9489 | dump_printf (MSG_NOTE, "internal\n"); | |
9490 | break; | |
9491 | case vect_induction_def: | |
9492 | dump_printf (MSG_NOTE, "induction\n"); | |
9493 | break; | |
9494 | case vect_reduction_def: | |
9495 | dump_printf (MSG_NOTE, "reduction\n"); | |
9496 | break; | |
9497 | case vect_double_reduction_def: | |
9498 | dump_printf (MSG_NOTE, "double reduction\n"); | |
9499 | break; | |
9500 | case vect_nested_cycle: | |
9501 | dump_printf (MSG_NOTE, "nested cycle\n"); | |
9502 | break; | |
9503 | case vect_unknown_def_type: | |
9504 | dump_printf (MSG_NOTE, "unknown\n"); | |
9505 | break; | |
9506 | } | |
9507 | } | |
9508 | ||
81c40241 | 9509 | if (*dt == vect_unknown_def_type) |
ebfd146a | 9510 | { |
73fbfcad | 9511 | if (dump_enabled_p ()) |
78c60e3d | 9512 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9513 | "Unsupported pattern.\n"); |
ebfd146a IR |
9514 | return false; |
9515 | } | |
9516 | ||
ebfd146a IR |
9517 | switch (gimple_code (*def_stmt)) |
9518 | { | |
9519 | case GIMPLE_PHI: | |
ebfd146a | 9520 | case GIMPLE_ASSIGN: |
ebfd146a | 9521 | case GIMPLE_CALL: |
81c40241 | 9522 | break; |
ebfd146a | 9523 | default: |
73fbfcad | 9524 | if (dump_enabled_p ()) |
78c60e3d | 9525 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9526 | "unsupported defining stmt:\n"); |
ebfd146a IR |
9527 | return false; |
9528 | } | |
9529 | ||
9530 | return true; | |
9531 | } | |
9532 | ||
81c40241 | 9533 | /* Function vect_is_simple_use. |
b690cc0f | 9534 | |
81c40241 | 9535 | Same as vect_is_simple_use but also determines the vector operand |
b690cc0f RG |
9536 | type of OPERAND and stores it to *VECTYPE. If the definition of |
9537 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
9538 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
9539 | is responsible to compute the best suited vector type for the | |
9540 | scalar operand. */ | |
9541 | ||
9542 | bool | |
81c40241 RB |
9543 | vect_is_simple_use (tree operand, vec_info *vinfo, |
9544 | gimple **def_stmt, enum vect_def_type *dt, tree *vectype) | |
b690cc0f | 9545 | { |
81c40241 | 9546 | if (!vect_is_simple_use (operand, vinfo, def_stmt, dt)) |
b690cc0f RG |
9547 | return false; |
9548 | ||
9549 | /* Now get a vector type if the def is internal, otherwise supply | |
9550 | NULL_TREE and leave it up to the caller to figure out a proper | |
9551 | type for the use stmt. */ | |
9552 | if (*dt == vect_internal_def | |
9553 | || *dt == vect_induction_def | |
9554 | || *dt == vect_reduction_def | |
9555 | || *dt == vect_double_reduction_def | |
9556 | || *dt == vect_nested_cycle) | |
9557 | { | |
9558 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
9559 | |
9560 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
9561 | && !STMT_VINFO_RELEVANT (stmt_info) | |
9562 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 9563 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 9564 | |
b690cc0f RG |
9565 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
9566 | gcc_assert (*vectype != NULL_TREE); | |
9567 | } | |
9568 | else if (*dt == vect_uninitialized_def | |
9569 | || *dt == vect_constant_def | |
9570 | || *dt == vect_external_def) | |
9571 | *vectype = NULL_TREE; | |
9572 | else | |
9573 | gcc_unreachable (); | |
9574 | ||
9575 | return true; | |
9576 | } | |
9577 | ||
ebfd146a IR |
9578 | |
9579 | /* Function supportable_widening_operation | |
9580 | ||
b8698a0f L |
9581 | Check whether an operation represented by the code CODE is a |
9582 | widening operation that is supported by the target platform in | |
b690cc0f RG |
9583 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
9584 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 9585 | |
ebfd146a IR |
9586 | Widening operations we currently support are NOP (CONVERT), FLOAT |
9587 | and WIDEN_MULT. This function checks if these operations are supported | |
9588 | by the target platform either directly (via vector tree-codes), or via | |
9589 | target builtins. | |
9590 | ||
9591 | Output: | |
b8698a0f L |
9592 | - CODE1 and CODE2 are codes of vector operations to be used when |
9593 | vectorizing the operation, if available. | |
ebfd146a IR |
9594 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
9595 | case of multi-step conversion (like char->short->int - in that case | |
9596 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
9597 | - INTERM_TYPES contains the intermediate type required to perform the |
9598 | widening operation (short in the above example). */ | |
ebfd146a IR |
9599 | |
9600 | bool | |
355fe088 | 9601 | supportable_widening_operation (enum tree_code code, gimple *stmt, |
b690cc0f | 9602 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
9603 | enum tree_code *code1, enum tree_code *code2, |
9604 | int *multi_step_cvt, | |
9771b263 | 9605 | vec<tree> *interm_types) |
ebfd146a IR |
9606 | { |
9607 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
9608 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 9609 | struct loop *vect_loop = NULL; |
ef4bddc2 | 9610 | machine_mode vec_mode; |
81f40b79 | 9611 | enum insn_code icode1, icode2; |
ebfd146a | 9612 | optab optab1, optab2; |
b690cc0f RG |
9613 | tree vectype = vectype_in; |
9614 | tree wide_vectype = vectype_out; | |
ebfd146a | 9615 | enum tree_code c1, c2; |
4a00c761 JJ |
9616 | int i; |
9617 | tree prev_type, intermediate_type; | |
ef4bddc2 | 9618 | machine_mode intermediate_mode, prev_mode; |
4a00c761 | 9619 | optab optab3, optab4; |
ebfd146a | 9620 | |
4a00c761 | 9621 | *multi_step_cvt = 0; |
4ef69dfc IR |
9622 | if (loop_info) |
9623 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
9624 | ||
ebfd146a IR |
9625 | switch (code) |
9626 | { | |
9627 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
9628 | /* The result of a vectorized widening operation usually requires |
9629 | two vectors (because the widened results do not fit into one vector). | |
9630 | The generated vector results would normally be expected to be | |
9631 | generated in the same order as in the original scalar computation, | |
9632 | i.e. if 8 results are generated in each vector iteration, they are | |
9633 | to be organized as follows: | |
9634 | vect1: [res1,res2,res3,res4], | |
9635 | vect2: [res5,res6,res7,res8]. | |
9636 | ||
9637 | However, in the special case that the result of the widening | |
9638 | operation is used in a reduction computation only, the order doesn't | |
9639 | matter (because when vectorizing a reduction we change the order of | |
9640 | the computation). Some targets can take advantage of this and | |
9641 | generate more efficient code. For example, targets like Altivec, | |
9642 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
9643 | generate the following vectors: | |
9644 | vect1: [res1,res3,res5,res7], | |
9645 | vect2: [res2,res4,res6,res8]. | |
9646 | ||
9647 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
9648 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
9649 | iterations in parallel). We therefore don't allow to change the | |
9650 | order of the computation in the inner-loop during outer-loop | |
9651 | vectorization. */ | |
9652 | /* TODO: Another case in which order doesn't *really* matter is when we | |
9653 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
9654 | Normally, pack_trunc performs an even/odd permute, whereas the | |
9655 | repack from an even/odd expansion would be an interleave, which | |
9656 | would be significantly simpler for e.g. AVX2. */ | |
9657 | /* In any case, in order to avoid duplicating the code below, recurse | |
9658 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
9659 | are properly set up for the caller. If we fail, we'll continue with | |
9660 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
9661 | if (vect_loop | |
9662 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
9663 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
9664 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
9665 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
9666 | code1, code2, multi_step_cvt, |
9667 | interm_types)) | |
ebc047a2 CH |
9668 | { |
9669 | /* Elements in a vector with vect_used_by_reduction property cannot | |
9670 | be reordered if the use chain with this property does not have the | |
9671 | same operation. One such an example is s += a * b, where elements | |
9672 | in a and b cannot be reordered. Here we check if the vector defined | |
9673 | by STMT is only directly used in the reduction statement. */ | |
9674 | tree lhs = gimple_assign_lhs (stmt); | |
9675 | use_operand_p dummy; | |
355fe088 | 9676 | gimple *use_stmt; |
ebc047a2 CH |
9677 | stmt_vec_info use_stmt_info = NULL; |
9678 | if (single_imm_use (lhs, &dummy, &use_stmt) | |
9679 | && (use_stmt_info = vinfo_for_stmt (use_stmt)) | |
9680 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) | |
9681 | return true; | |
9682 | } | |
4a00c761 JJ |
9683 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
9684 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
9685 | break; |
9686 | ||
81c40241 RB |
9687 | case DOT_PROD_EXPR: |
9688 | c1 = DOT_PROD_EXPR; | |
9689 | c2 = DOT_PROD_EXPR; | |
9690 | break; | |
9691 | ||
9692 | case SAD_EXPR: | |
9693 | c1 = SAD_EXPR; | |
9694 | c2 = SAD_EXPR; | |
9695 | break; | |
9696 | ||
6ae6116f RH |
9697 | case VEC_WIDEN_MULT_EVEN_EXPR: |
9698 | /* Support the recursion induced just above. */ | |
9699 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
9700 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
9701 | break; | |
9702 | ||
36ba4aae | 9703 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
9704 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
9705 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
9706 | break; |
9707 | ||
ebfd146a | 9708 | CASE_CONVERT: |
4a00c761 JJ |
9709 | c1 = VEC_UNPACK_LO_EXPR; |
9710 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
9711 | break; |
9712 | ||
9713 | case FLOAT_EXPR: | |
4a00c761 JJ |
9714 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
9715 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
9716 | break; |
9717 | ||
9718 | case FIX_TRUNC_EXPR: | |
9719 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
9720 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
9721 | computing the operation. */ | |
9722 | return false; | |
9723 | ||
9724 | default: | |
9725 | gcc_unreachable (); | |
9726 | } | |
9727 | ||
6ae6116f | 9728 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
6b4db501 | 9729 | std::swap (c1, c2); |
4a00c761 | 9730 | |
ebfd146a IR |
9731 | if (code == FIX_TRUNC_EXPR) |
9732 | { | |
9733 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
9734 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
9735 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
9736 | } |
9737 | else | |
9738 | { | |
9739 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
9740 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
9741 | } | |
9742 | ||
9743 | if (!optab1 || !optab2) | |
9744 | return false; | |
9745 | ||
9746 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
9747 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
9748 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
9749 | return false; |
9750 | ||
4a00c761 JJ |
9751 | *code1 = c1; |
9752 | *code2 = c2; | |
9753 | ||
9754 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
9755 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff IE |
9756 | /* For scalar masks we may have different boolean |
9757 | vector types having the same QImode. Thus we | |
9758 | add additional check for elements number. */ | |
9759 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
928686b1 RS |
9760 | || known_eq (TYPE_VECTOR_SUBPARTS (vectype), |
9761 | TYPE_VECTOR_SUBPARTS (wide_vectype) * 2)); | |
4a00c761 | 9762 | |
b8698a0f | 9763 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 9764 | types. */ |
ebfd146a | 9765 | |
4a00c761 JJ |
9766 | prev_type = vectype; |
9767 | prev_mode = vec_mode; | |
b8698a0f | 9768 | |
4a00c761 JJ |
9769 | if (!CONVERT_EXPR_CODE_P (code)) |
9770 | return false; | |
b8698a0f | 9771 | |
4a00c761 JJ |
9772 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
9773 | intermediate steps in promotion sequence. We try | |
9774 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
9775 | not. */ | |
9771b263 | 9776 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
9777 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
9778 | { | |
9779 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
9780 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
9781 | { | |
7cfb4d93 | 9782 | intermediate_type = vect_halve_mask_nunits (prev_type); |
3ae0661a IE |
9783 | if (intermediate_mode != TYPE_MODE (intermediate_type)) |
9784 | return false; | |
9785 | } | |
9786 | else | |
9787 | intermediate_type | |
9788 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
9789 | TYPE_UNSIGNED (prev_type)); | |
9790 | ||
4a00c761 JJ |
9791 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); |
9792 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
9793 | ||
9794 | if (!optab3 || !optab4 | |
9795 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
9796 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
9797 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
9798 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
9799 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
9800 | == CODE_FOR_nothing) | |
9801 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
9802 | == CODE_FOR_nothing)) | |
9803 | break; | |
ebfd146a | 9804 | |
9771b263 | 9805 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
9806 | (*multi_step_cvt)++; |
9807 | ||
9808 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
9809 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff | 9810 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
928686b1 RS |
9811 | || known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type), |
9812 | TYPE_VECTOR_SUBPARTS (wide_vectype) * 2)); | |
4a00c761 JJ |
9813 | |
9814 | prev_type = intermediate_type; | |
9815 | prev_mode = intermediate_mode; | |
ebfd146a IR |
9816 | } |
9817 | ||
9771b263 | 9818 | interm_types->release (); |
4a00c761 | 9819 | return false; |
ebfd146a IR |
9820 | } |
9821 | ||
9822 | ||
9823 | /* Function supportable_narrowing_operation | |
9824 | ||
b8698a0f L |
9825 | Check whether an operation represented by the code CODE is a |
9826 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
9827 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
9828 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 9829 | |
ebfd146a | 9830 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 9831 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
9832 | the target platform directly via vector tree-codes. |
9833 | ||
9834 | Output: | |
b8698a0f L |
9835 | - CODE1 is the code of a vector operation to be used when |
9836 | vectorizing the operation, if available. | |
ebfd146a IR |
9837 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
9838 | case of multi-step conversion (like int->short->char - in that case | |
9839 | MULTI_STEP_CVT will be 1). | |
9840 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 9841 | narrowing operation (short in the above example). */ |
ebfd146a IR |
9842 | |
9843 | bool | |
9844 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 9845 | tree vectype_out, tree vectype_in, |
ebfd146a | 9846 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 9847 | vec<tree> *interm_types) |
ebfd146a | 9848 | { |
ef4bddc2 | 9849 | machine_mode vec_mode; |
ebfd146a IR |
9850 | enum insn_code icode1; |
9851 | optab optab1, interm_optab; | |
b690cc0f RG |
9852 | tree vectype = vectype_in; |
9853 | tree narrow_vectype = vectype_out; | |
ebfd146a | 9854 | enum tree_code c1; |
3ae0661a | 9855 | tree intermediate_type, prev_type; |
ef4bddc2 | 9856 | machine_mode intermediate_mode, prev_mode; |
ebfd146a | 9857 | int i; |
4a00c761 | 9858 | bool uns; |
ebfd146a | 9859 | |
4a00c761 | 9860 | *multi_step_cvt = 0; |
ebfd146a IR |
9861 | switch (code) |
9862 | { | |
9863 | CASE_CONVERT: | |
9864 | c1 = VEC_PACK_TRUNC_EXPR; | |
9865 | break; | |
9866 | ||
9867 | case FIX_TRUNC_EXPR: | |
9868 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
9869 | break; | |
9870 | ||
9871 | case FLOAT_EXPR: | |
9872 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
9873 | tree code and optabs used for computing the operation. */ | |
9874 | return false; | |
9875 | ||
9876 | default: | |
9877 | gcc_unreachable (); | |
9878 | } | |
9879 | ||
9880 | if (code == FIX_TRUNC_EXPR) | |
9881 | /* The signedness is determined from output operand. */ | |
b690cc0f | 9882 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
9883 | else |
9884 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
9885 | ||
9886 | if (!optab1) | |
9887 | return false; | |
9888 | ||
9889 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 9890 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
9891 | return false; |
9892 | ||
4a00c761 JJ |
9893 | *code1 = c1; |
9894 | ||
9895 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff IE |
9896 | /* For scalar masks we may have different boolean |
9897 | vector types having the same QImode. Thus we | |
9898 | add additional check for elements number. */ | |
9899 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
928686b1 RS |
9900 | || known_eq (TYPE_VECTOR_SUBPARTS (vectype) * 2, |
9901 | TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 | 9902 | |
ebfd146a IR |
9903 | /* Check if it's a multi-step conversion that can be done using intermediate |
9904 | types. */ | |
4a00c761 | 9905 | prev_mode = vec_mode; |
3ae0661a | 9906 | prev_type = vectype; |
4a00c761 JJ |
9907 | if (code == FIX_TRUNC_EXPR) |
9908 | uns = TYPE_UNSIGNED (vectype_out); | |
9909 | else | |
9910 | uns = TYPE_UNSIGNED (vectype); | |
9911 | ||
9912 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
9913 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
9914 | costly than signed. */ | |
9915 | if (code == FIX_TRUNC_EXPR && uns) | |
9916 | { | |
9917 | enum insn_code icode2; | |
9918 | ||
9919 | intermediate_type | |
9920 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
9921 | interm_optab | |
9922 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 9923 | if (interm_optab != unknown_optab |
4a00c761 JJ |
9924 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
9925 | && insn_data[icode1].operand[0].mode | |
9926 | == insn_data[icode2].operand[0].mode) | |
9927 | { | |
9928 | uns = false; | |
9929 | optab1 = interm_optab; | |
9930 | icode1 = icode2; | |
9931 | } | |
9932 | } | |
ebfd146a | 9933 | |
4a00c761 JJ |
9934 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
9935 | intermediate steps in promotion sequence. We try | |
9936 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 9937 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
9938 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
9939 | { | |
9940 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
9941 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
9942 | { | |
7cfb4d93 | 9943 | intermediate_type = vect_double_mask_nunits (prev_type); |
3ae0661a | 9944 | if (intermediate_mode != TYPE_MODE (intermediate_type)) |
7cfb4d93 | 9945 | return false; |
3ae0661a IE |
9946 | } |
9947 | else | |
9948 | intermediate_type | |
9949 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
4a00c761 JJ |
9950 | interm_optab |
9951 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
9952 | optab_default); | |
9953 | if (!interm_optab | |
9954 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
9955 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
9956 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
9957 | == CODE_FOR_nothing)) | |
9958 | break; | |
9959 | ||
9771b263 | 9960 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
9961 | (*multi_step_cvt)++; |
9962 | ||
9963 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff | 9964 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
928686b1 RS |
9965 | || known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type) * 2, |
9966 | TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 JJ |
9967 | |
9968 | prev_mode = intermediate_mode; | |
3ae0661a | 9969 | prev_type = intermediate_type; |
4a00c761 | 9970 | optab1 = interm_optab; |
ebfd146a IR |
9971 | } |
9972 | ||
9771b263 | 9973 | interm_types->release (); |
4a00c761 | 9974 | return false; |
ebfd146a | 9975 | } |
7cfb4d93 RS |
9976 | |
9977 | /* Generate and return a statement that sets vector mask MASK such that | |
9978 | MASK[I] is true iff J + START_INDEX < END_INDEX for all J <= I. */ | |
9979 | ||
9980 | gcall * | |
9981 | vect_gen_while (tree mask, tree start_index, tree end_index) | |
9982 | { | |
9983 | tree cmp_type = TREE_TYPE (start_index); | |
9984 | tree mask_type = TREE_TYPE (mask); | |
9985 | gcc_checking_assert (direct_internal_fn_supported_p (IFN_WHILE_ULT, | |
9986 | cmp_type, mask_type, | |
9987 | OPTIMIZE_FOR_SPEED)); | |
9988 | gcall *call = gimple_build_call_internal (IFN_WHILE_ULT, 3, | |
9989 | start_index, end_index, | |
9990 | build_zero_cst (mask_type)); | |
9991 | gimple_call_set_lhs (call, mask); | |
9992 | return call; | |
9993 | } | |
535e7c11 RS |
9994 | |
9995 | /* Generate a vector mask of type MASK_TYPE for which index I is false iff | |
9996 | J + START_INDEX < END_INDEX for all J <= I. Add the statements to SEQ. */ | |
9997 | ||
9998 | tree | |
9999 | vect_gen_while_not (gimple_seq *seq, tree mask_type, tree start_index, | |
10000 | tree end_index) | |
10001 | { | |
10002 | tree tmp = make_ssa_name (mask_type); | |
10003 | gcall *call = vect_gen_while (tmp, start_index, end_index); | |
10004 | gimple_seq_add_stmt (seq, call); | |
10005 | return gimple_build (seq, BIT_NOT_EXPR, mask_type, tmp); | |
10006 | } |