]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
cbe34bb5 | 2 | Copyright (C) 2003-2017 Free Software Foundation, Inc. |
b8698a0f | 3 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
4 | and Ira Rosen <irar@il.ibm.com> |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 3, or (at your option) any later | |
11 | version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
c7131fb2 | 25 | #include "backend.h" |
957060b5 AM |
26 | #include "target.h" |
27 | #include "rtl.h" | |
ebfd146a | 28 | #include "tree.h" |
c7131fb2 | 29 | #include "gimple.h" |
c7131fb2 | 30 | #include "ssa.h" |
957060b5 AM |
31 | #include "optabs-tree.h" |
32 | #include "insn-config.h" | |
33 | #include "recog.h" /* FIXME: for insn_data */ | |
34 | #include "cgraph.h" | |
957060b5 | 35 | #include "dumpfile.h" |
c7131fb2 | 36 | #include "alias.h" |
40e23961 | 37 | #include "fold-const.h" |
d8a2d370 | 38 | #include "stor-layout.h" |
2fb9a547 | 39 | #include "tree-eh.h" |
45b0be94 | 40 | #include "gimplify.h" |
5be5c238 | 41 | #include "gimple-iterator.h" |
18f429e2 | 42 | #include "gimplify-me.h" |
442b4905 | 43 | #include "tree-cfg.h" |
e28030cf | 44 | #include "tree-ssa-loop-manip.h" |
ebfd146a | 45 | #include "cfgloop.h" |
0136f8f0 AH |
46 | #include "tree-ssa-loop.h" |
47 | #include "tree-scalar-evolution.h" | |
ebfd146a | 48 | #include "tree-vectorizer.h" |
9b2b7279 | 49 | #include "builtins.h" |
70439f0d | 50 | #include "internal-fn.h" |
5ebaa477 | 51 | #include "tree-vector-builder.h" |
f151c9e1 | 52 | #include "vec-perm-indices.h" |
ebfd146a | 53 | |
7ee2468b SB |
54 | /* For lang_hooks.types.type_for_mode. */ |
55 | #include "langhooks.h" | |
ebfd146a | 56 | |
2de001ee RS |
57 | /* Says whether a statement is a load, a store of a vectorized statement |
58 | result, or a store of an invariant value. */ | |
59 | enum vec_load_store_type { | |
60 | VLS_LOAD, | |
61 | VLS_STORE, | |
62 | VLS_STORE_INVARIANT | |
63 | }; | |
64 | ||
c3e7ee41 BS |
65 | /* Return the vectorized type for the given statement. */ |
66 | ||
67 | tree | |
68 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
69 | { | |
70 | return STMT_VINFO_VECTYPE (stmt_info); | |
71 | } | |
72 | ||
73 | /* Return TRUE iff the given statement is in an inner loop relative to | |
74 | the loop being vectorized. */ | |
75 | bool | |
76 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
77 | { | |
355fe088 | 78 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
c3e7ee41 BS |
79 | basic_block bb = gimple_bb (stmt); |
80 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
81 | struct loop* loop; | |
82 | ||
83 | if (!loop_vinfo) | |
84 | return false; | |
85 | ||
86 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
87 | ||
88 | return (bb->loop_father == loop->inner); | |
89 | } | |
90 | ||
91 | /* Record the cost of a statement, either by directly informing the | |
92 | target model or by saving it in a vector for later processing. | |
93 | Return a preliminary estimate of the statement's cost. */ | |
94 | ||
95 | unsigned | |
92345349 | 96 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 97 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 98 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 99 | { |
cc9fe6bb JH |
100 | if ((kind == vector_load || kind == unaligned_load) |
101 | && STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
102 | kind = vector_gather_load; | |
103 | if ((kind == vector_store || kind == unaligned_store) | |
104 | && STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
105 | kind = vector_scatter_store; | |
92345349 | 106 | if (body_cost_vec) |
c3e7ee41 | 107 | { |
92345349 | 108 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
ddf56386 RB |
109 | stmt_info_for_cost si = { count, kind, |
110 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
111 | misalign }; | |
112 | body_cost_vec->safe_push (si); | |
c3e7ee41 | 113 | return (unsigned) |
92345349 | 114 | (builtin_vectorization_cost (kind, vectype, misalign) * count); |
c3e7ee41 BS |
115 | } |
116 | else | |
310213d4 RB |
117 | return add_stmt_cost (stmt_info->vinfo->target_cost_data, |
118 | count, kind, stmt_info, misalign, where); | |
c3e7ee41 BS |
119 | } |
120 | ||
272c6793 RS |
121 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
122 | ||
123 | static tree | |
124 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
125 | { | |
126 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
127 | "vect_array"); | |
128 | } | |
129 | ||
130 | /* ARRAY is an array of vectors created by create_vector_array. | |
131 | Return an SSA_NAME for the vector in index N. The reference | |
132 | is part of the vectorization of STMT and the vector is associated | |
133 | with scalar destination SCALAR_DEST. */ | |
134 | ||
135 | static tree | |
355fe088 | 136 | read_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree scalar_dest, |
272c6793 RS |
137 | tree array, unsigned HOST_WIDE_INT n) |
138 | { | |
139 | tree vect_type, vect, vect_name, array_ref; | |
355fe088 | 140 | gimple *new_stmt; |
272c6793 RS |
141 | |
142 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
143 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
144 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
145 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
146 | build_int_cst (size_type_node, n), | |
147 | NULL_TREE, NULL_TREE); | |
148 | ||
149 | new_stmt = gimple_build_assign (vect, array_ref); | |
150 | vect_name = make_ssa_name (vect, new_stmt); | |
151 | gimple_assign_set_lhs (new_stmt, vect_name); | |
152 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
153 | |
154 | return vect_name; | |
155 | } | |
156 | ||
157 | /* ARRAY is an array of vectors created by create_vector_array. | |
158 | Emit code to store SSA_NAME VECT in index N of the array. | |
159 | The store is part of the vectorization of STMT. */ | |
160 | ||
161 | static void | |
355fe088 | 162 | write_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree vect, |
272c6793 RS |
163 | tree array, unsigned HOST_WIDE_INT n) |
164 | { | |
165 | tree array_ref; | |
355fe088 | 166 | gimple *new_stmt; |
272c6793 RS |
167 | |
168 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
169 | build_int_cst (size_type_node, n), | |
170 | NULL_TREE, NULL_TREE); | |
171 | ||
172 | new_stmt = gimple_build_assign (array_ref, vect); | |
173 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
174 | } |
175 | ||
176 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
177 | of *PTR. The memory reference replaces those in FIRST_DR | |
178 | (and its group). */ | |
179 | ||
180 | static tree | |
44fc7854 | 181 | create_array_ref (tree type, tree ptr, tree alias_ptr_type) |
272c6793 | 182 | { |
44fc7854 | 183 | tree mem_ref; |
272c6793 | 184 | |
272c6793 RS |
185 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); |
186 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 187 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
188 | return mem_ref; |
189 | } | |
190 | ||
ebfd146a IR |
191 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
192 | ||
193 | /* Function vect_mark_relevant. | |
194 | ||
195 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
196 | ||
197 | static void | |
355fe088 | 198 | vect_mark_relevant (vec<gimple *> *worklist, gimple *stmt, |
97ecdb46 | 199 | enum vect_relevant relevant, bool live_p) |
ebfd146a IR |
200 | { |
201 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
202 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
203 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
355fe088 | 204 | gimple *pattern_stmt; |
ebfd146a | 205 | |
73fbfcad | 206 | if (dump_enabled_p ()) |
66c16fd9 RB |
207 | { |
208 | dump_printf_loc (MSG_NOTE, vect_location, | |
209 | "mark relevant %d, live %d: ", relevant, live_p); | |
210 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
211 | } | |
ebfd146a | 212 | |
83197f37 IR |
213 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
214 | related pattern stmt instead of the original stmt. However, such stmts | |
215 | may have their own uses that are not in any pattern, in such cases the | |
216 | stmt itself should be marked. */ | |
ebfd146a IR |
217 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
218 | { | |
97ecdb46 JJ |
219 | /* This is the last stmt in a sequence that was detected as a |
220 | pattern that can potentially be vectorized. Don't mark the stmt | |
221 | as relevant/live because it's not going to be vectorized. | |
222 | Instead mark the pattern-stmt that replaces it. */ | |
83197f37 | 223 | |
97ecdb46 JJ |
224 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); |
225 | ||
226 | if (dump_enabled_p ()) | |
227 | dump_printf_loc (MSG_NOTE, vect_location, | |
228 | "last stmt in pattern. don't mark" | |
229 | " relevant/live.\n"); | |
230 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
231 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
232 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
233 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
234 | stmt = pattern_stmt; | |
ebfd146a IR |
235 | } |
236 | ||
237 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
238 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
239 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
240 | ||
241 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
242 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
243 | { | |
73fbfcad | 244 | if (dump_enabled_p ()) |
78c60e3d | 245 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 246 | "already marked relevant/live.\n"); |
ebfd146a IR |
247 | return; |
248 | } | |
249 | ||
9771b263 | 250 | worklist->safe_push (stmt); |
ebfd146a IR |
251 | } |
252 | ||
253 | ||
b28ead45 AH |
254 | /* Function is_simple_and_all_uses_invariant |
255 | ||
256 | Return true if STMT is simple and all uses of it are invariant. */ | |
257 | ||
258 | bool | |
259 | is_simple_and_all_uses_invariant (gimple *stmt, loop_vec_info loop_vinfo) | |
260 | { | |
261 | tree op; | |
262 | gimple *def_stmt; | |
263 | ssa_op_iter iter; | |
264 | ||
265 | if (!is_gimple_assign (stmt)) | |
266 | return false; | |
267 | ||
268 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) | |
269 | { | |
270 | enum vect_def_type dt = vect_uninitialized_def; | |
271 | ||
272 | if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt)) | |
273 | { | |
274 | if (dump_enabled_p ()) | |
275 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
276 | "use not simple.\n"); | |
277 | return false; | |
278 | } | |
279 | ||
280 | if (dt != vect_external_def && dt != vect_constant_def) | |
281 | return false; | |
282 | } | |
283 | return true; | |
284 | } | |
285 | ||
ebfd146a IR |
286 | /* Function vect_stmt_relevant_p. |
287 | ||
288 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
289 | "relevant for vectorization". | |
290 | ||
291 | A stmt is considered "relevant for vectorization" if: | |
292 | - it has uses outside the loop. | |
293 | - it has vdefs (it alters memory). | |
294 | - control stmts in the loop (except for the exit condition). | |
295 | ||
296 | CHECKME: what other side effects would the vectorizer allow? */ | |
297 | ||
298 | static bool | |
355fe088 | 299 | vect_stmt_relevant_p (gimple *stmt, loop_vec_info loop_vinfo, |
ebfd146a IR |
300 | enum vect_relevant *relevant, bool *live_p) |
301 | { | |
302 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
303 | ssa_op_iter op_iter; | |
304 | imm_use_iterator imm_iter; | |
305 | use_operand_p use_p; | |
306 | def_operand_p def_p; | |
307 | ||
8644a673 | 308 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
309 | *live_p = false; |
310 | ||
311 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
312 | if (is_ctrl_stmt (stmt) |
313 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
314 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 315 | *relevant = vect_used_in_scope; |
ebfd146a IR |
316 | |
317 | /* changing memory. */ | |
318 | if (gimple_code (stmt) != GIMPLE_PHI) | |
ac6aeab4 RB |
319 | if (gimple_vdef (stmt) |
320 | && !gimple_clobber_p (stmt)) | |
ebfd146a | 321 | { |
73fbfcad | 322 | if (dump_enabled_p ()) |
78c60e3d | 323 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 324 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 325 | *relevant = vect_used_in_scope; |
ebfd146a IR |
326 | } |
327 | ||
328 | /* uses outside the loop. */ | |
329 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
330 | { | |
331 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
332 | { | |
333 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
334 | if (!flow_bb_inside_loop_p (loop, bb)) | |
335 | { | |
73fbfcad | 336 | if (dump_enabled_p ()) |
78c60e3d | 337 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 338 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 339 | |
3157b0c2 AO |
340 | if (is_gimple_debug (USE_STMT (use_p))) |
341 | continue; | |
342 | ||
ebfd146a IR |
343 | /* We expect all such uses to be in the loop exit phis |
344 | (because of loop closed form) */ | |
345 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
346 | gcc_assert (bb == single_exit (loop)->dest); | |
347 | ||
348 | *live_p = true; | |
349 | } | |
350 | } | |
351 | } | |
352 | ||
3a2edf4c AH |
353 | if (*live_p && *relevant == vect_unused_in_scope |
354 | && !is_simple_and_all_uses_invariant (stmt, loop_vinfo)) | |
b28ead45 AH |
355 | { |
356 | if (dump_enabled_p ()) | |
357 | dump_printf_loc (MSG_NOTE, vect_location, | |
358 | "vec_stmt_relevant_p: stmt live but not relevant.\n"); | |
359 | *relevant = vect_used_only_live; | |
360 | } | |
361 | ||
ebfd146a IR |
362 | return (*live_p || *relevant); |
363 | } | |
364 | ||
365 | ||
b8698a0f | 366 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 367 | |
ff802fa1 | 368 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
369 | used in STMT for anything other than indexing an array. */ |
370 | ||
371 | static bool | |
355fe088 | 372 | exist_non_indexing_operands_for_use_p (tree use, gimple *stmt) |
ebfd146a IR |
373 | { |
374 | tree operand; | |
375 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 376 | |
ff802fa1 | 377 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
378 | reference in STMT, then any operand that corresponds to USE |
379 | is not indexing an array. */ | |
380 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
381 | return true; | |
59a05b0c | 382 | |
ebfd146a IR |
383 | /* STMT has a data_ref. FORNOW this means that its of one of |
384 | the following forms: | |
385 | -1- ARRAY_REF = var | |
386 | -2- var = ARRAY_REF | |
387 | (This should have been verified in analyze_data_refs). | |
388 | ||
389 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 390 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
391 | for array indexing. |
392 | ||
393 | Therefore, all we need to check is if STMT falls into the | |
394 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
395 | |
396 | if (!gimple_assign_copy_p (stmt)) | |
5ce9450f JJ |
397 | { |
398 | if (is_gimple_call (stmt) | |
399 | && gimple_call_internal_p (stmt)) | |
400 | switch (gimple_call_internal_fn (stmt)) | |
401 | { | |
402 | case IFN_MASK_STORE: | |
403 | operand = gimple_call_arg (stmt, 3); | |
404 | if (operand == use) | |
405 | return true; | |
406 | /* FALLTHRU */ | |
407 | case IFN_MASK_LOAD: | |
408 | operand = gimple_call_arg (stmt, 2); | |
409 | if (operand == use) | |
410 | return true; | |
411 | break; | |
412 | default: | |
413 | break; | |
414 | } | |
415 | return false; | |
416 | } | |
417 | ||
59a05b0c EB |
418 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
419 | return false; | |
ebfd146a | 420 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
421 | if (TREE_CODE (operand) != SSA_NAME) |
422 | return false; | |
423 | ||
424 | if (operand == use) | |
425 | return true; | |
426 | ||
427 | return false; | |
428 | } | |
429 | ||
430 | ||
b8698a0f | 431 | /* |
ebfd146a IR |
432 | Function process_use. |
433 | ||
434 | Inputs: | |
435 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b28ead45 | 436 | - RELEVANT - enum value to be set in the STMT_VINFO of the stmt |
ff802fa1 | 437 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 438 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
439 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
440 | be performed. | |
ebfd146a IR |
441 | |
442 | Outputs: | |
443 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
444 | relevance info of the DEF_STMT of this USE: | |
445 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
446 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
447 | Exceptions: | |
448 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 449 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 450 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
451 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
452 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
453 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
454 | be modified accordingly. | |
455 | ||
456 | Return true if everything is as expected. Return false otherwise. */ | |
457 | ||
458 | static bool | |
b28ead45 | 459 | process_use (gimple *stmt, tree use, loop_vec_info loop_vinfo, |
355fe088 | 460 | enum vect_relevant relevant, vec<gimple *> *worklist, |
aec7ae7d | 461 | bool force) |
ebfd146a IR |
462 | { |
463 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
464 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
465 | stmt_vec_info dstmt_vinfo; | |
466 | basic_block bb, def_bb; | |
355fe088 | 467 | gimple *def_stmt; |
ebfd146a IR |
468 | enum vect_def_type dt; |
469 | ||
b8698a0f | 470 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 471 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 472 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
473 | return true; |
474 | ||
81c40241 | 475 | if (!vect_is_simple_use (use, loop_vinfo, &def_stmt, &dt)) |
b8698a0f | 476 | { |
73fbfcad | 477 | if (dump_enabled_p ()) |
78c60e3d | 478 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 479 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
480 | return false; |
481 | } | |
482 | ||
483 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
484 | return true; | |
485 | ||
486 | def_bb = gimple_bb (def_stmt); | |
487 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
488 | { | |
73fbfcad | 489 | if (dump_enabled_p ()) |
e645e942 | 490 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n"); |
ebfd146a IR |
491 | return true; |
492 | } | |
493 | ||
b8698a0f L |
494 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
495 | DEF_STMT must have already been processed, because this should be the | |
496 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
497 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
498 | check that everything is as expected, and we are done. */ |
499 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
500 | bb = gimple_bb (stmt); | |
501 | if (gimple_code (stmt) == GIMPLE_PHI | |
502 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
503 | && gimple_code (def_stmt) != GIMPLE_PHI | |
504 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
505 | && bb->loop_father == def_bb->loop_father) | |
506 | { | |
73fbfcad | 507 | if (dump_enabled_p ()) |
78c60e3d | 508 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 509 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a IR |
510 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
511 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
512 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 513 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 514 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
515 | return true; |
516 | } | |
517 | ||
518 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
519 | outer-loop-header-bb: | |
520 | d = def_stmt | |
521 | inner-loop: | |
522 | stmt # use (d) | |
523 | outer-loop-tail-bb: | |
524 | ... */ | |
525 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
526 | { | |
73fbfcad | 527 | if (dump_enabled_p ()) |
78c60e3d | 528 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 529 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 530 | |
ebfd146a IR |
531 | switch (relevant) |
532 | { | |
8644a673 | 533 | case vect_unused_in_scope: |
7c5222ff IR |
534 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
535 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 536 | break; |
7c5222ff | 537 | |
ebfd146a | 538 | case vect_used_in_outer_by_reduction: |
7c5222ff | 539 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
540 | relevant = vect_used_by_reduction; |
541 | break; | |
7c5222ff | 542 | |
ebfd146a | 543 | case vect_used_in_outer: |
7c5222ff | 544 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 545 | relevant = vect_used_in_scope; |
ebfd146a | 546 | break; |
7c5222ff | 547 | |
8644a673 | 548 | case vect_used_in_scope: |
ebfd146a IR |
549 | break; |
550 | ||
551 | default: | |
552 | gcc_unreachable (); | |
b8698a0f | 553 | } |
ebfd146a IR |
554 | } |
555 | ||
556 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
557 | outer-loop-header-bb: | |
558 | ... | |
559 | inner-loop: | |
560 | d = def_stmt | |
06066f92 | 561 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
562 | stmt # use (d) */ |
563 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
564 | { | |
73fbfcad | 565 | if (dump_enabled_p ()) |
78c60e3d | 566 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 567 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 568 | |
ebfd146a IR |
569 | switch (relevant) |
570 | { | |
8644a673 | 571 | case vect_unused_in_scope: |
b8698a0f | 572 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 573 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 574 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
575 | break; |
576 | ||
ebfd146a | 577 | case vect_used_by_reduction: |
b28ead45 | 578 | case vect_used_only_live: |
ebfd146a IR |
579 | relevant = vect_used_in_outer_by_reduction; |
580 | break; | |
581 | ||
8644a673 | 582 | case vect_used_in_scope: |
ebfd146a IR |
583 | relevant = vect_used_in_outer; |
584 | break; | |
585 | ||
586 | default: | |
587 | gcc_unreachable (); | |
588 | } | |
589 | } | |
643a9684 RB |
590 | /* We are also not interested in uses on loop PHI backedges that are |
591 | inductions. Otherwise we'll needlessly vectorize the IV increment | |
e294f495 RB |
592 | and cause hybrid SLP for SLP inductions. Unless the PHI is live |
593 | of course. */ | |
643a9684 RB |
594 | else if (gimple_code (stmt) == GIMPLE_PHI |
595 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_induction_def | |
e294f495 | 596 | && ! STMT_VINFO_LIVE_P (stmt_vinfo) |
643a9684 RB |
597 | && (PHI_ARG_DEF_FROM_EDGE (stmt, loop_latch_edge (bb->loop_father)) |
598 | == use)) | |
599 | { | |
600 | if (dump_enabled_p ()) | |
601 | dump_printf_loc (MSG_NOTE, vect_location, | |
602 | "induction value on backedge.\n"); | |
603 | return true; | |
604 | } | |
605 | ||
ebfd146a | 606 | |
b28ead45 | 607 | vect_mark_relevant (worklist, def_stmt, relevant, false); |
ebfd146a IR |
608 | return true; |
609 | } | |
610 | ||
611 | ||
612 | /* Function vect_mark_stmts_to_be_vectorized. | |
613 | ||
614 | Not all stmts in the loop need to be vectorized. For example: | |
615 | ||
616 | for i... | |
617 | for j... | |
618 | 1. T0 = i + j | |
619 | 2. T1 = a[T0] | |
620 | ||
621 | 3. j = j + 1 | |
622 | ||
623 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
624 | addressing of vectorized data-refs are handled differently. | |
625 | ||
626 | This pass detects such stmts. */ | |
627 | ||
628 | bool | |
629 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
630 | { | |
ebfd146a IR |
631 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
632 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
633 | unsigned int nbbs = loop->num_nodes; | |
634 | gimple_stmt_iterator si; | |
355fe088 | 635 | gimple *stmt; |
ebfd146a IR |
636 | unsigned int i; |
637 | stmt_vec_info stmt_vinfo; | |
638 | basic_block bb; | |
355fe088 | 639 | gimple *phi; |
ebfd146a | 640 | bool live_p; |
b28ead45 | 641 | enum vect_relevant relevant; |
ebfd146a | 642 | |
73fbfcad | 643 | if (dump_enabled_p ()) |
78c60e3d | 644 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 645 | "=== vect_mark_stmts_to_be_vectorized ===\n"); |
ebfd146a | 646 | |
355fe088 | 647 | auto_vec<gimple *, 64> worklist; |
ebfd146a IR |
648 | |
649 | /* 1. Init worklist. */ | |
650 | for (i = 0; i < nbbs; i++) | |
651 | { | |
652 | bb = bbs[i]; | |
653 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 654 | { |
ebfd146a | 655 | phi = gsi_stmt (si); |
73fbfcad | 656 | if (dump_enabled_p ()) |
ebfd146a | 657 | { |
78c60e3d SS |
658 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
659 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
660 | } |
661 | ||
662 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 663 | vect_mark_relevant (&worklist, phi, relevant, live_p); |
ebfd146a IR |
664 | } |
665 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
666 | { | |
667 | stmt = gsi_stmt (si); | |
73fbfcad | 668 | if (dump_enabled_p ()) |
ebfd146a | 669 | { |
78c60e3d SS |
670 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
671 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
b8698a0f | 672 | } |
ebfd146a IR |
673 | |
674 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 675 | vect_mark_relevant (&worklist, stmt, relevant, live_p); |
ebfd146a IR |
676 | } |
677 | } | |
678 | ||
679 | /* 2. Process_worklist */ | |
9771b263 | 680 | while (worklist.length () > 0) |
ebfd146a IR |
681 | { |
682 | use_operand_p use_p; | |
683 | ssa_op_iter iter; | |
684 | ||
9771b263 | 685 | stmt = worklist.pop (); |
73fbfcad | 686 | if (dump_enabled_p ()) |
ebfd146a | 687 | { |
78c60e3d SS |
688 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
689 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
ebfd146a IR |
690 | } |
691 | ||
b8698a0f | 692 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
b28ead45 AH |
693 | (DEF_STMT) as relevant/irrelevant according to the relevance property |
694 | of STMT. */ | |
ebfd146a IR |
695 | stmt_vinfo = vinfo_for_stmt (stmt); |
696 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
ebfd146a | 697 | |
b28ead45 AH |
698 | /* Generally, the relevance property of STMT (in STMT_VINFO_RELEVANT) is |
699 | propagated as is to the DEF_STMTs of its USEs. | |
ebfd146a IR |
700 | |
701 | One exception is when STMT has been identified as defining a reduction | |
b28ead45 | 702 | variable; in this case we set the relevance to vect_used_by_reduction. |
ebfd146a | 703 | This is because we distinguish between two kinds of relevant stmts - |
b8698a0f | 704 | those that are used by a reduction computation, and those that are |
ff802fa1 | 705 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 706 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 707 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 708 | |
b28ead45 | 709 | switch (STMT_VINFO_DEF_TYPE (stmt_vinfo)) |
ebfd146a | 710 | { |
06066f92 | 711 | case vect_reduction_def: |
b28ead45 AH |
712 | gcc_assert (relevant != vect_unused_in_scope); |
713 | if (relevant != vect_unused_in_scope | |
714 | && relevant != vect_used_in_scope | |
715 | && relevant != vect_used_by_reduction | |
716 | && relevant != vect_used_only_live) | |
06066f92 | 717 | { |
b28ead45 AH |
718 | if (dump_enabled_p ()) |
719 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
720 | "unsupported use of reduction.\n"); | |
721 | return false; | |
06066f92 | 722 | } |
06066f92 | 723 | break; |
b8698a0f | 724 | |
06066f92 | 725 | case vect_nested_cycle: |
b28ead45 AH |
726 | if (relevant != vect_unused_in_scope |
727 | && relevant != vect_used_in_outer_by_reduction | |
728 | && relevant != vect_used_in_outer) | |
06066f92 | 729 | { |
73fbfcad | 730 | if (dump_enabled_p ()) |
78c60e3d | 731 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 732 | "unsupported use of nested cycle.\n"); |
7c5222ff | 733 | |
06066f92 IR |
734 | return false; |
735 | } | |
b8698a0f L |
736 | break; |
737 | ||
06066f92 | 738 | case vect_double_reduction_def: |
b28ead45 AH |
739 | if (relevant != vect_unused_in_scope |
740 | && relevant != vect_used_by_reduction | |
741 | && relevant != vect_used_only_live) | |
06066f92 | 742 | { |
73fbfcad | 743 | if (dump_enabled_p ()) |
78c60e3d | 744 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 745 | "unsupported use of double reduction.\n"); |
7c5222ff | 746 | |
7c5222ff | 747 | return false; |
06066f92 | 748 | } |
b8698a0f | 749 | break; |
7c5222ff | 750 | |
06066f92 IR |
751 | default: |
752 | break; | |
7c5222ff | 753 | } |
b8698a0f | 754 | |
aec7ae7d | 755 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
756 | { |
757 | /* Pattern statements are not inserted into the code, so | |
758 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
759 | have to scan the RHS or function arguments instead. */ | |
760 | if (is_gimple_assign (stmt)) | |
761 | { | |
69d2aade JJ |
762 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
763 | tree op = gimple_assign_rhs1 (stmt); | |
764 | ||
765 | i = 1; | |
766 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
767 | { | |
768 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
b28ead45 | 769 | relevant, &worklist, false) |
69d2aade | 770 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
b28ead45 | 771 | relevant, &worklist, false)) |
566d377a | 772 | return false; |
69d2aade JJ |
773 | i = 2; |
774 | } | |
775 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 776 | { |
69d2aade | 777 | op = gimple_op (stmt, i); |
afbe6325 | 778 | if (TREE_CODE (op) == SSA_NAME |
b28ead45 | 779 | && !process_use (stmt, op, loop_vinfo, relevant, |
afbe6325 | 780 | &worklist, false)) |
07687835 | 781 | return false; |
9d5e7640 IR |
782 | } |
783 | } | |
784 | else if (is_gimple_call (stmt)) | |
785 | { | |
786 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
787 | { | |
788 | tree arg = gimple_call_arg (stmt, i); | |
b28ead45 | 789 | if (!process_use (stmt, arg, loop_vinfo, relevant, |
aec7ae7d | 790 | &worklist, false)) |
07687835 | 791 | return false; |
9d5e7640 IR |
792 | } |
793 | } | |
794 | } | |
795 | else | |
796 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
797 | { | |
798 | tree op = USE_FROM_PTR (use_p); | |
b28ead45 | 799 | if (!process_use (stmt, op, loop_vinfo, relevant, |
aec7ae7d | 800 | &worklist, false)) |
07687835 | 801 | return false; |
9d5e7640 | 802 | } |
aec7ae7d | 803 | |
3bab6342 | 804 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo)) |
aec7ae7d | 805 | { |
134c85ca RS |
806 | gather_scatter_info gs_info; |
807 | if (!vect_check_gather_scatter (stmt, loop_vinfo, &gs_info)) | |
808 | gcc_unreachable (); | |
809 | if (!process_use (stmt, gs_info.offset, loop_vinfo, relevant, | |
810 | &worklist, true)) | |
566d377a | 811 | return false; |
aec7ae7d | 812 | } |
ebfd146a IR |
813 | } /* while worklist */ |
814 | ||
ebfd146a IR |
815 | return true; |
816 | } | |
817 | ||
818 | ||
b8698a0f | 819 | /* Function vect_model_simple_cost. |
ebfd146a | 820 | |
b8698a0f | 821 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
822 | single op. Right now, this does not account for multiple insns that could |
823 | be generated for the single vector op. We will handle that shortly. */ | |
824 | ||
825 | void | |
b8698a0f | 826 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 | 827 | enum vect_def_type *dt, |
4fc5ebf1 | 828 | int ndts, |
92345349 BS |
829 | stmt_vector_for_cost *prologue_cost_vec, |
830 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
831 | { |
832 | int i; | |
92345349 | 833 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
834 | |
835 | /* The SLP costs were already calculated during SLP tree build. */ | |
836 | if (PURE_SLP_STMT (stmt_info)) | |
837 | return; | |
838 | ||
4fc5ebf1 JG |
839 | /* Cost the "broadcast" of a scalar operand in to a vector operand. |
840 | Use scalar_to_vec to cost the broadcast, as elsewhere in the vector | |
841 | cost model. */ | |
842 | for (i = 0; i < ndts; i++) | |
92345349 | 843 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
4fc5ebf1 | 844 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
92345349 | 845 | stmt_info, 0, vect_prologue); |
c3e7ee41 BS |
846 | |
847 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
92345349 BS |
848 | inside_cost = record_stmt_cost (body_cost_vec, ncopies, vector_stmt, |
849 | stmt_info, 0, vect_body); | |
c3e7ee41 | 850 | |
73fbfcad | 851 | if (dump_enabled_p ()) |
78c60e3d SS |
852 | dump_printf_loc (MSG_NOTE, vect_location, |
853 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 854 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
855 | } |
856 | ||
857 | ||
8bd37302 BS |
858 | /* Model cost for type demotion and promotion operations. PWR is normally |
859 | zero for single-step promotions and demotions. It will be one if | |
860 | two-step promotion/demotion is required, and so on. Each additional | |
861 | step doubles the number of instructions required. */ | |
862 | ||
863 | static void | |
864 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
865 | enum vect_def_type *dt, int pwr) | |
866 | { | |
867 | int i, tmp; | |
92345349 | 868 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 BS |
869 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
870 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
871 | void *target_cost_data; | |
8bd37302 BS |
872 | |
873 | /* The SLP costs were already calculated during SLP tree build. */ | |
874 | if (PURE_SLP_STMT (stmt_info)) | |
875 | return; | |
876 | ||
c3e7ee41 BS |
877 | if (loop_vinfo) |
878 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
879 | else | |
880 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
881 | ||
8bd37302 BS |
882 | for (i = 0; i < pwr + 1; i++) |
883 | { | |
884 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
885 | (i + 1) : i; | |
c3e7ee41 | 886 | inside_cost += add_stmt_cost (target_cost_data, vect_pow2 (tmp), |
92345349 BS |
887 | vec_promote_demote, stmt_info, 0, |
888 | vect_body); | |
8bd37302 BS |
889 | } |
890 | ||
891 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
892 | for (i = 0; i < 2; i++) | |
92345349 BS |
893 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
894 | prologue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt, | |
895 | stmt_info, 0, vect_prologue); | |
8bd37302 | 896 | |
73fbfcad | 897 | if (dump_enabled_p ()) |
78c60e3d SS |
898 | dump_printf_loc (MSG_NOTE, vect_location, |
899 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 900 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
901 | } |
902 | ||
ebfd146a IR |
903 | /* Function vect_model_store_cost |
904 | ||
0d0293ac MM |
905 | Models cost for stores. In the case of grouped accesses, one access |
906 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a IR |
907 | |
908 | void | |
b8698a0f | 909 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
2de001ee RS |
910 | vect_memory_access_type memory_access_type, |
911 | enum vect_def_type dt, slp_tree slp_node, | |
92345349 BS |
912 | stmt_vector_for_cost *prologue_cost_vec, |
913 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a | 914 | { |
92345349 | 915 | unsigned int inside_cost = 0, prologue_cost = 0; |
892a981f RS |
916 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
917 | gimple *first_stmt = STMT_VINFO_STMT (stmt_info); | |
918 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); | |
ebfd146a | 919 | |
8644a673 | 920 | if (dt == vect_constant_def || dt == vect_external_def) |
92345349 BS |
921 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
922 | stmt_info, 0, vect_prologue); | |
ebfd146a | 923 | |
892a981f RS |
924 | /* Grouped stores update all elements in the group at once, |
925 | so we want the DR for the first statement. */ | |
926 | if (!slp_node && grouped_access_p) | |
720f5239 | 927 | { |
892a981f RS |
928 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
929 | dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
720f5239 | 930 | } |
ebfd146a | 931 | |
892a981f RS |
932 | /* True if we should include any once-per-group costs as well as |
933 | the cost of the statement itself. For SLP we only get called | |
934 | once per group anyhow. */ | |
935 | bool first_stmt_p = (first_stmt == STMT_VINFO_STMT (stmt_info)); | |
936 | ||
272c6793 | 937 | /* We assume that the cost of a single store-lanes instruction is |
0d0293ac | 938 | equivalent to the cost of GROUP_SIZE separate stores. If a grouped |
272c6793 | 939 | access is instead being provided by a permute-and-store operation, |
2de001ee RS |
940 | include the cost of the permutes. */ |
941 | if (first_stmt_p | |
942 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 943 | { |
e1377713 ES |
944 | /* Uses a high and low interleave or shuffle operations for each |
945 | needed permute. */ | |
892a981f | 946 | int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
e1377713 | 947 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
92345349 BS |
948 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
949 | stmt_info, 0, vect_body); | |
ebfd146a | 950 | |
73fbfcad | 951 | if (dump_enabled_p ()) |
78c60e3d | 952 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 953 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 954 | group_size); |
ebfd146a IR |
955 | } |
956 | ||
cee62fee | 957 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a | 958 | /* Costs of the stores. */ |
067bc855 RB |
959 | if (memory_access_type == VMAT_ELEMENTWISE |
960 | || memory_access_type == VMAT_GATHER_SCATTER) | |
c5126ce8 RS |
961 | { |
962 | /* N scalar stores plus extracting the elements. */ | |
963 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); | |
964 | inside_cost += record_stmt_cost (body_cost_vec, | |
965 | ncopies * assumed_nunits, | |
966 | scalar_store, stmt_info, 0, vect_body); | |
967 | } | |
f2e2a985 | 968 | else |
892a981f | 969 | vect_get_store_cost (dr, ncopies, &inside_cost, body_cost_vec); |
ebfd146a | 970 | |
2de001ee RS |
971 | if (memory_access_type == VMAT_ELEMENTWISE |
972 | || memory_access_type == VMAT_STRIDED_SLP) | |
c5126ce8 RS |
973 | { |
974 | /* N scalar stores plus extracting the elements. */ | |
975 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); | |
976 | inside_cost += record_stmt_cost (body_cost_vec, | |
977 | ncopies * assumed_nunits, | |
978 | vec_to_scalar, stmt_info, 0, vect_body); | |
979 | } | |
cee62fee | 980 | |
73fbfcad | 981 | if (dump_enabled_p ()) |
78c60e3d SS |
982 | dump_printf_loc (MSG_NOTE, vect_location, |
983 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 984 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
985 | } |
986 | ||
987 | ||
720f5239 IR |
988 | /* Calculate cost of DR's memory access. */ |
989 | void | |
990 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 991 | unsigned int *inside_cost, |
92345349 | 992 | stmt_vector_for_cost *body_cost_vec) |
720f5239 IR |
993 | { |
994 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 995 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 996 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
997 | |
998 | switch (alignment_support_scheme) | |
999 | { | |
1000 | case dr_aligned: | |
1001 | { | |
92345349 BS |
1002 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
1003 | vector_store, stmt_info, 0, | |
1004 | vect_body); | |
720f5239 | 1005 | |
73fbfcad | 1006 | if (dump_enabled_p ()) |
78c60e3d | 1007 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1008 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
1009 | break; |
1010 | } | |
1011 | ||
1012 | case dr_unaligned_supported: | |
1013 | { | |
720f5239 | 1014 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1015 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1016 | unaligned_store, stmt_info, |
92345349 | 1017 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1018 | if (dump_enabled_p ()) |
78c60e3d SS |
1019 | dump_printf_loc (MSG_NOTE, vect_location, |
1020 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1021 | "hardware.\n"); |
720f5239 IR |
1022 | break; |
1023 | } | |
1024 | ||
38eec4c6 UW |
1025 | case dr_unaligned_unsupported: |
1026 | { | |
1027 | *inside_cost = VECT_MAX_COST; | |
1028 | ||
73fbfcad | 1029 | if (dump_enabled_p ()) |
78c60e3d | 1030 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1031 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1032 | break; |
1033 | } | |
1034 | ||
720f5239 IR |
1035 | default: |
1036 | gcc_unreachable (); | |
1037 | } | |
1038 | } | |
1039 | ||
1040 | ||
ebfd146a IR |
1041 | /* Function vect_model_load_cost |
1042 | ||
892a981f RS |
1043 | Models cost for loads. In the case of grouped accesses, one access has |
1044 | the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1045 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1046 | access scheme chosen. */ |
1047 | ||
1048 | void | |
92345349 | 1049 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, |
2de001ee RS |
1050 | vect_memory_access_type memory_access_type, |
1051 | slp_tree slp_node, | |
92345349 BS |
1052 | stmt_vector_for_cost *prologue_cost_vec, |
1053 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a | 1054 | { |
892a981f RS |
1055 | gimple *first_stmt = STMT_VINFO_STMT (stmt_info); |
1056 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
92345349 | 1057 | unsigned int inside_cost = 0, prologue_cost = 0; |
892a981f | 1058 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); |
ebfd146a | 1059 | |
892a981f RS |
1060 | /* Grouped loads read all elements in the group at once, |
1061 | so we want the DR for the first statement. */ | |
1062 | if (!slp_node && grouped_access_p) | |
ebfd146a | 1063 | { |
892a981f RS |
1064 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
1065 | dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
ebfd146a IR |
1066 | } |
1067 | ||
892a981f RS |
1068 | /* True if we should include any once-per-group costs as well as |
1069 | the cost of the statement itself. For SLP we only get called | |
1070 | once per group anyhow. */ | |
1071 | bool first_stmt_p = (first_stmt == STMT_VINFO_STMT (stmt_info)); | |
1072 | ||
272c6793 | 1073 | /* We assume that the cost of a single load-lanes instruction is |
0d0293ac | 1074 | equivalent to the cost of GROUP_SIZE separate loads. If a grouped |
272c6793 | 1075 | access is instead being provided by a load-and-permute operation, |
2de001ee RS |
1076 | include the cost of the permutes. */ |
1077 | if (first_stmt_p | |
1078 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 1079 | { |
2c23db6d ES |
1080 | /* Uses an even and odd extract operations or shuffle operations |
1081 | for each needed permute. */ | |
892a981f | 1082 | int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
2c23db6d ES |
1083 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
1084 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, | |
1085 | stmt_info, 0, vect_body); | |
ebfd146a | 1086 | |
73fbfcad | 1087 | if (dump_enabled_p ()) |
e645e942 TJ |
1088 | dump_printf_loc (MSG_NOTE, vect_location, |
1089 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1090 | group_size); |
ebfd146a IR |
1091 | } |
1092 | ||
1093 | /* The loads themselves. */ | |
067bc855 RB |
1094 | if (memory_access_type == VMAT_ELEMENTWISE |
1095 | || memory_access_type == VMAT_GATHER_SCATTER) | |
a82960aa | 1096 | { |
a21892ad BS |
1097 | /* N scalar loads plus gathering them into a vector. */ |
1098 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
c5126ce8 | 1099 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); |
92345349 | 1100 | inside_cost += record_stmt_cost (body_cost_vec, |
c5126ce8 | 1101 | ncopies * assumed_nunits, |
92345349 | 1102 | scalar_load, stmt_info, 0, vect_body); |
a82960aa RG |
1103 | } |
1104 | else | |
892a981f | 1105 | vect_get_load_cost (dr, ncopies, first_stmt_p, |
92345349 BS |
1106 | &inside_cost, &prologue_cost, |
1107 | prologue_cost_vec, body_cost_vec, true); | |
2de001ee RS |
1108 | if (memory_access_type == VMAT_ELEMENTWISE |
1109 | || memory_access_type == VMAT_STRIDED_SLP) | |
892a981f RS |
1110 | inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct, |
1111 | stmt_info, 0, vect_body); | |
720f5239 | 1112 | |
73fbfcad | 1113 | if (dump_enabled_p ()) |
78c60e3d SS |
1114 | dump_printf_loc (MSG_NOTE, vect_location, |
1115 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1116 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1117 | } |
1118 | ||
1119 | ||
1120 | /* Calculate cost of DR's memory access. */ | |
1121 | void | |
1122 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1123 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1124 | unsigned int *prologue_cost, |
1125 | stmt_vector_for_cost *prologue_cost_vec, | |
1126 | stmt_vector_for_cost *body_cost_vec, | |
1127 | bool record_prologue_costs) | |
720f5239 IR |
1128 | { |
1129 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 1130 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 1131 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
1132 | |
1133 | switch (alignment_support_scheme) | |
ebfd146a IR |
1134 | { |
1135 | case dr_aligned: | |
1136 | { | |
92345349 BS |
1137 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1138 | stmt_info, 0, vect_body); | |
ebfd146a | 1139 | |
73fbfcad | 1140 | if (dump_enabled_p ()) |
78c60e3d | 1141 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1142 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1143 | |
1144 | break; | |
1145 | } | |
1146 | case dr_unaligned_supported: | |
1147 | { | |
720f5239 | 1148 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1149 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1150 | unaligned_load, stmt_info, |
92345349 | 1151 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1152 | |
73fbfcad | 1153 | if (dump_enabled_p ()) |
78c60e3d SS |
1154 | dump_printf_loc (MSG_NOTE, vect_location, |
1155 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1156 | "hardware.\n"); |
ebfd146a IR |
1157 | |
1158 | break; | |
1159 | } | |
1160 | case dr_explicit_realign: | |
1161 | { | |
92345349 BS |
1162 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1163 | vector_load, stmt_info, 0, vect_body); | |
1164 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1165 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1166 | |
1167 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1168 | the containing loop, the following cost should be added to the | |
92345349 | 1169 | prologue costs. */ |
ebfd146a | 1170 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1171 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1172 | stmt_info, 0, vect_body); | |
ebfd146a | 1173 | |
73fbfcad | 1174 | if (dump_enabled_p ()) |
e645e942 TJ |
1175 | dump_printf_loc (MSG_NOTE, vect_location, |
1176 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1177 | |
ebfd146a IR |
1178 | break; |
1179 | } | |
1180 | case dr_explicit_realign_optimized: | |
1181 | { | |
73fbfcad | 1182 | if (dump_enabled_p ()) |
e645e942 | 1183 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1184 | "vect_model_load_cost: unaligned software " |
e645e942 | 1185 | "pipelined.\n"); |
ebfd146a IR |
1186 | |
1187 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1188 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1189 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1190 | access, then the above cost should only be considered for one |
ff802fa1 | 1191 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1192 | and a realignment op. */ |
1193 | ||
92345349 | 1194 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1195 | { |
92345349 BS |
1196 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1197 | vector_stmt, stmt_info, | |
1198 | 0, vect_prologue); | |
ebfd146a | 1199 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1200 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1201 | vector_stmt, stmt_info, | |
1202 | 0, vect_prologue); | |
ebfd146a IR |
1203 | } |
1204 | ||
92345349 BS |
1205 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1206 | stmt_info, 0, vect_body); | |
1207 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1208 | stmt_info, 0, vect_body); | |
8bd37302 | 1209 | |
73fbfcad | 1210 | if (dump_enabled_p ()) |
78c60e3d | 1211 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1212 | "vect_model_load_cost: explicit realign optimized" |
1213 | "\n"); | |
8bd37302 | 1214 | |
ebfd146a IR |
1215 | break; |
1216 | } | |
1217 | ||
38eec4c6 UW |
1218 | case dr_unaligned_unsupported: |
1219 | { | |
1220 | *inside_cost = VECT_MAX_COST; | |
1221 | ||
73fbfcad | 1222 | if (dump_enabled_p ()) |
78c60e3d | 1223 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1224 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1225 | break; |
1226 | } | |
1227 | ||
ebfd146a IR |
1228 | default: |
1229 | gcc_unreachable (); | |
1230 | } | |
ebfd146a IR |
1231 | } |
1232 | ||
418b7df3 RG |
1233 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1234 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1235 | |
418b7df3 | 1236 | static void |
355fe088 | 1237 | vect_init_vector_1 (gimple *stmt, gimple *new_stmt, gimple_stmt_iterator *gsi) |
ebfd146a | 1238 | { |
ebfd146a | 1239 | if (gsi) |
418b7df3 | 1240 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1241 | else |
1242 | { | |
418b7df3 | 1243 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1244 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1245 | |
a70d6342 IR |
1246 | if (loop_vinfo) |
1247 | { | |
1248 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1249 | basic_block new_bb; |
1250 | edge pe; | |
a70d6342 IR |
1251 | |
1252 | if (nested_in_vect_loop_p (loop, stmt)) | |
1253 | loop = loop->inner; | |
b8698a0f | 1254 | |
a70d6342 | 1255 | pe = loop_preheader_edge (loop); |
418b7df3 | 1256 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1257 | gcc_assert (!new_bb); |
1258 | } | |
1259 | else | |
1260 | { | |
1261 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1262 | basic_block bb; | |
1263 | gimple_stmt_iterator gsi_bb_start; | |
1264 | ||
1265 | gcc_assert (bb_vinfo); | |
1266 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1267 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1268 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1269 | } |
ebfd146a IR |
1270 | } |
1271 | ||
73fbfcad | 1272 | if (dump_enabled_p ()) |
ebfd146a | 1273 | { |
78c60e3d SS |
1274 | dump_printf_loc (MSG_NOTE, vect_location, |
1275 | "created new init_stmt: "); | |
1276 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
ebfd146a | 1277 | } |
418b7df3 RG |
1278 | } |
1279 | ||
1280 | /* Function vect_init_vector. | |
ebfd146a | 1281 | |
5467ee52 RG |
1282 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1283 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1284 | vector type a vector with all elements equal to VAL is created first. | |
1285 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1286 | initialization at the loop preheader. | |
418b7df3 RG |
1287 | Return the DEF of INIT_STMT. |
1288 | It will be used in the vectorization of STMT. */ | |
1289 | ||
1290 | tree | |
355fe088 | 1291 | vect_init_vector (gimple *stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 | 1292 | { |
355fe088 | 1293 | gimple *init_stmt; |
418b7df3 RG |
1294 | tree new_temp; |
1295 | ||
e412ece4 RB |
1296 | /* We abuse this function to push sth to a SSA name with initial 'val'. */ |
1297 | if (! useless_type_conversion_p (type, TREE_TYPE (val))) | |
418b7df3 | 1298 | { |
e412ece4 RB |
1299 | gcc_assert (TREE_CODE (type) == VECTOR_TYPE); |
1300 | if (! types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) | |
418b7df3 | 1301 | { |
5a308cf1 IE |
1302 | /* Scalar boolean value should be transformed into |
1303 | all zeros or all ones value before building a vector. */ | |
1304 | if (VECTOR_BOOLEAN_TYPE_P (type)) | |
1305 | { | |
b3d51f23 IE |
1306 | tree true_val = build_all_ones_cst (TREE_TYPE (type)); |
1307 | tree false_val = build_zero_cst (TREE_TYPE (type)); | |
5a308cf1 IE |
1308 | |
1309 | if (CONSTANT_CLASS_P (val)) | |
1310 | val = integer_zerop (val) ? false_val : true_val; | |
1311 | else | |
1312 | { | |
1313 | new_temp = make_ssa_name (TREE_TYPE (type)); | |
1314 | init_stmt = gimple_build_assign (new_temp, COND_EXPR, | |
1315 | val, true_val, false_val); | |
1316 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
1317 | val = new_temp; | |
1318 | } | |
1319 | } | |
1320 | else if (CONSTANT_CLASS_P (val)) | |
42fd8198 | 1321 | val = fold_convert (TREE_TYPE (type), val); |
418b7df3 RG |
1322 | else |
1323 | { | |
b731b390 | 1324 | new_temp = make_ssa_name (TREE_TYPE (type)); |
e412ece4 RB |
1325 | if (! INTEGRAL_TYPE_P (TREE_TYPE (val))) |
1326 | init_stmt = gimple_build_assign (new_temp, | |
1327 | fold_build1 (VIEW_CONVERT_EXPR, | |
1328 | TREE_TYPE (type), | |
1329 | val)); | |
1330 | else | |
1331 | init_stmt = gimple_build_assign (new_temp, NOP_EXPR, val); | |
418b7df3 | 1332 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1333 | val = new_temp; |
418b7df3 RG |
1334 | } |
1335 | } | |
5467ee52 | 1336 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1337 | } |
1338 | ||
0e22bb5a RB |
1339 | new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_"); |
1340 | init_stmt = gimple_build_assign (new_temp, val); | |
418b7df3 | 1341 | vect_init_vector_1 (stmt, init_stmt, gsi); |
0e22bb5a | 1342 | return new_temp; |
ebfd146a IR |
1343 | } |
1344 | ||
c83a894c | 1345 | /* Function vect_get_vec_def_for_operand_1. |
a70d6342 | 1346 | |
c83a894c AH |
1347 | For a defining stmt DEF_STMT of a scalar stmt, return a vector def with type |
1348 | DT that will be used in the vectorized stmt. */ | |
ebfd146a IR |
1349 | |
1350 | tree | |
c83a894c | 1351 | vect_get_vec_def_for_operand_1 (gimple *def_stmt, enum vect_def_type dt) |
ebfd146a IR |
1352 | { |
1353 | tree vec_oprnd; | |
355fe088 | 1354 | gimple *vec_stmt; |
ebfd146a | 1355 | stmt_vec_info def_stmt_info = NULL; |
ebfd146a IR |
1356 | |
1357 | switch (dt) | |
1358 | { | |
81c40241 | 1359 | /* operand is a constant or a loop invariant. */ |
ebfd146a | 1360 | case vect_constant_def: |
81c40241 | 1361 | case vect_external_def: |
c83a894c AH |
1362 | /* Code should use vect_get_vec_def_for_operand. */ |
1363 | gcc_unreachable (); | |
ebfd146a | 1364 | |
81c40241 | 1365 | /* operand is defined inside the loop. */ |
8644a673 | 1366 | case vect_internal_def: |
ebfd146a | 1367 | { |
ebfd146a IR |
1368 | /* Get the def from the vectorized stmt. */ |
1369 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1370 | |
ebfd146a | 1371 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1372 | /* Get vectorized pattern statement. */ |
1373 | if (!vec_stmt | |
1374 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1375 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1376 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1377 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1378 | gcc_assert (vec_stmt); |
1379 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1380 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1381 | else if (is_gimple_call (vec_stmt)) | |
1382 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1383 | else | |
1384 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1385 | return vec_oprnd; | |
1386 | } | |
1387 | ||
c78e3652 | 1388 | /* operand is defined by a loop header phi. */ |
ebfd146a | 1389 | case vect_reduction_def: |
06066f92 | 1390 | case vect_double_reduction_def: |
7c5222ff | 1391 | case vect_nested_cycle: |
ebfd146a IR |
1392 | case vect_induction_def: |
1393 | { | |
1394 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1395 | ||
1396 | /* Get the def from the vectorized stmt. */ | |
1397 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1398 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1399 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1400 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1401 | else | |
1402 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1403 | return vec_oprnd; |
1404 | } | |
1405 | ||
1406 | default: | |
1407 | gcc_unreachable (); | |
1408 | } | |
1409 | } | |
1410 | ||
1411 | ||
c83a894c AH |
1412 | /* Function vect_get_vec_def_for_operand. |
1413 | ||
1414 | OP is an operand in STMT. This function returns a (vector) def that will be | |
1415 | used in the vectorized stmt for STMT. | |
1416 | ||
1417 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1418 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1419 | ||
1420 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1421 | needs to be introduced. VECTYPE may be used to specify a required type for | |
1422 | vector invariant. */ | |
1423 | ||
1424 | tree | |
1425 | vect_get_vec_def_for_operand (tree op, gimple *stmt, tree vectype) | |
1426 | { | |
1427 | gimple *def_stmt; | |
1428 | enum vect_def_type dt; | |
1429 | bool is_simple_use; | |
1430 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
1431 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
1432 | ||
1433 | if (dump_enabled_p ()) | |
1434 | { | |
1435 | dump_printf_loc (MSG_NOTE, vect_location, | |
1436 | "vect_get_vec_def_for_operand: "); | |
1437 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
1438 | dump_printf (MSG_NOTE, "\n"); | |
1439 | } | |
1440 | ||
1441 | is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt); | |
1442 | gcc_assert (is_simple_use); | |
1443 | if (def_stmt && dump_enabled_p ()) | |
1444 | { | |
1445 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1446 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
1447 | } | |
1448 | ||
1449 | if (dt == vect_constant_def || dt == vect_external_def) | |
1450 | { | |
1451 | tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo); | |
1452 | tree vector_type; | |
1453 | ||
1454 | if (vectype) | |
1455 | vector_type = vectype; | |
2568d8a1 | 1456 | else if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op)) |
c83a894c AH |
1457 | && VECTOR_BOOLEAN_TYPE_P (stmt_vectype)) |
1458 | vector_type = build_same_sized_truth_vector_type (stmt_vectype); | |
1459 | else | |
1460 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
1461 | ||
1462 | gcc_assert (vector_type); | |
1463 | return vect_init_vector (stmt, op, vector_type, NULL); | |
1464 | } | |
1465 | else | |
1466 | return vect_get_vec_def_for_operand_1 (def_stmt, dt); | |
1467 | } | |
1468 | ||
1469 | ||
ebfd146a IR |
1470 | /* Function vect_get_vec_def_for_stmt_copy |
1471 | ||
ff802fa1 | 1472 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1473 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1474 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1475 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1476 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1477 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1478 | DT is the type of the vector def VEC_OPRND. |
1479 | ||
1480 | Context: | |
1481 | In case the vectorization factor (VF) is bigger than the number | |
1482 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1483 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1484 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1485 | smallest data-type determines the VF, and as a result, when vectorizing |
1486 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1487 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1488 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1489 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1490 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1491 | ||
1492 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1493 | |
ebfd146a IR |
1494 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1495 | VS1.1: vx.1 = memref1 VS1.2 | |
1496 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1497 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1498 | |
1499 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1500 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1501 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1502 | VSnew.3: vz3 = vx.3 + ... | |
1503 | ||
1504 | The vectorization of S1 is explained in vectorizable_load. | |
1505 | The vectorization of S2: | |
b8698a0f L |
1506 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1507 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1508 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1509 | returns the vector-def 'vx.0'. |
1510 | ||
b8698a0f L |
1511 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1512 | function is called to get the relevant vector-def for each operand. It is | |
1513 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1514 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1515 | ||
b8698a0f L |
1516 | For example, to obtain the vector-def 'vx.1' in order to create the |
1517 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1518 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1519 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1520 | and return its def ('vx.1'). | |
1521 | Overall, to create the above sequence this function will be called 3 times: | |
1522 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1523 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1524 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1525 | ||
1526 | tree | |
1527 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1528 | { | |
355fe088 | 1529 | gimple *vec_stmt_for_operand; |
ebfd146a IR |
1530 | stmt_vec_info def_stmt_info; |
1531 | ||
1532 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1533 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1534 | return vec_oprnd; |
1535 | ||
1536 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1537 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1538 | gcc_assert (def_stmt_info); | |
1539 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1540 | gcc_assert (vec_stmt_for_operand); | |
ebfd146a IR |
1541 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) |
1542 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1543 | else | |
1544 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1545 | return vec_oprnd; | |
1546 | } | |
1547 | ||
1548 | ||
1549 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1550 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a | 1551 | |
c78e3652 | 1552 | void |
b8698a0f | 1553 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1554 | vec<tree> *vec_oprnds0, |
1555 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1556 | { |
9771b263 | 1557 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1558 | |
1559 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1560 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1561 | |
9771b263 | 1562 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1563 | { |
9771b263 | 1564 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1565 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1566 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1567 | } |
1568 | } | |
1569 | ||
1570 | ||
c78e3652 | 1571 | /* Get vectorized definitions for OP0 and OP1. */ |
ebfd146a | 1572 | |
c78e3652 | 1573 | void |
355fe088 | 1574 | vect_get_vec_defs (tree op0, tree op1, gimple *stmt, |
9771b263 DN |
1575 | vec<tree> *vec_oprnds0, |
1576 | vec<tree> *vec_oprnds1, | |
306b0c92 | 1577 | slp_tree slp_node) |
ebfd146a IR |
1578 | { |
1579 | if (slp_node) | |
d092494c IR |
1580 | { |
1581 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1582 | auto_vec<tree> ops (nops); |
1583 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1584 | |
9771b263 | 1585 | ops.quick_push (op0); |
d092494c | 1586 | if (op1) |
9771b263 | 1587 | ops.quick_push (op1); |
d092494c | 1588 | |
306b0c92 | 1589 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
d092494c | 1590 | |
37b5ec8f | 1591 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1592 | if (op1) |
37b5ec8f | 1593 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1594 | } |
ebfd146a IR |
1595 | else |
1596 | { | |
1597 | tree vec_oprnd; | |
1598 | ||
9771b263 | 1599 | vec_oprnds0->create (1); |
81c40241 | 1600 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 1601 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1602 | |
1603 | if (op1) | |
1604 | { | |
9771b263 | 1605 | vec_oprnds1->create (1); |
81c40241 | 1606 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 1607 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1608 | } |
1609 | } | |
1610 | } | |
1611 | ||
1612 | ||
1613 | /* Function vect_finish_stmt_generation. | |
1614 | ||
1615 | Insert a new stmt. */ | |
1616 | ||
1617 | void | |
355fe088 | 1618 | vect_finish_stmt_generation (gimple *stmt, gimple *vec_stmt, |
ebfd146a IR |
1619 | gimple_stmt_iterator *gsi) |
1620 | { | |
1621 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
310213d4 | 1622 | vec_info *vinfo = stmt_info->vinfo; |
ebfd146a IR |
1623 | |
1624 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1625 | ||
54e8e2c3 RG |
1626 | if (!gsi_end_p (*gsi) |
1627 | && gimple_has_mem_ops (vec_stmt)) | |
1628 | { | |
355fe088 | 1629 | gimple *at_stmt = gsi_stmt (*gsi); |
54e8e2c3 RG |
1630 | tree vuse = gimple_vuse (at_stmt); |
1631 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1632 | { | |
1633 | tree vdef = gimple_vdef (at_stmt); | |
1634 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1635 | /* If we have an SSA vuse and insert a store, update virtual | |
1636 | SSA form to avoid triggering the renamer. Do so only | |
1637 | if we can easily see all uses - which is what almost always | |
1638 | happens with the way vectorized stmts are inserted. */ | |
1639 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1640 | && ((is_gimple_assign (vec_stmt) | |
1641 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1642 | || (is_gimple_call (vec_stmt) | |
1643 | && !(gimple_call_flags (vec_stmt) | |
1644 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1645 | { | |
1646 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1647 | gimple_set_vdef (vec_stmt, new_vdef); | |
1648 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1649 | } | |
1650 | } | |
1651 | } | |
ebfd146a IR |
1652 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1653 | ||
310213d4 | 1654 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, vinfo)); |
ebfd146a | 1655 | |
73fbfcad | 1656 | if (dump_enabled_p ()) |
ebfd146a | 1657 | { |
78c60e3d SS |
1658 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); |
1659 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
ebfd146a IR |
1660 | } |
1661 | ||
ad885386 | 1662 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
8e91d222 JJ |
1663 | |
1664 | /* While EH edges will generally prevent vectorization, stmt might | |
1665 | e.g. be in a must-not-throw region. Ensure newly created stmts | |
1666 | that could throw are part of the same region. */ | |
1667 | int lp_nr = lookup_stmt_eh_lp (stmt); | |
1668 | if (lp_nr != 0 && stmt_could_throw_p (vec_stmt)) | |
1669 | add_stmt_to_eh_lp (vec_stmt, lp_nr); | |
ebfd146a IR |
1670 | } |
1671 | ||
70439f0d RS |
1672 | /* We want to vectorize a call to combined function CFN with function |
1673 | decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN | |
1674 | as the types of all inputs. Check whether this is possible using | |
1675 | an internal function, returning its code if so or IFN_LAST if not. */ | |
ebfd146a | 1676 | |
70439f0d RS |
1677 | static internal_fn |
1678 | vectorizable_internal_function (combined_fn cfn, tree fndecl, | |
1679 | tree vectype_out, tree vectype_in) | |
ebfd146a | 1680 | { |
70439f0d RS |
1681 | internal_fn ifn; |
1682 | if (internal_fn_p (cfn)) | |
1683 | ifn = as_internal_fn (cfn); | |
1684 | else | |
1685 | ifn = associated_internal_fn (fndecl); | |
1686 | if (ifn != IFN_LAST && direct_internal_fn_p (ifn)) | |
1687 | { | |
1688 | const direct_internal_fn_info &info = direct_internal_fn (ifn); | |
1689 | if (info.vectorizable) | |
1690 | { | |
1691 | tree type0 = (info.type0 < 0 ? vectype_out : vectype_in); | |
1692 | tree type1 = (info.type1 < 0 ? vectype_out : vectype_in); | |
d95ab70a RS |
1693 | if (direct_internal_fn_supported_p (ifn, tree_pair (type0, type1), |
1694 | OPTIMIZE_FOR_SPEED)) | |
70439f0d RS |
1695 | return ifn; |
1696 | } | |
1697 | } | |
1698 | return IFN_LAST; | |
ebfd146a IR |
1699 | } |
1700 | ||
5ce9450f | 1701 | |
355fe088 | 1702 | static tree permute_vec_elements (tree, tree, tree, gimple *, |
5ce9450f JJ |
1703 | gimple_stmt_iterator *); |
1704 | ||
62da9e14 RS |
1705 | /* STMT is a non-strided load or store, meaning that it accesses |
1706 | elements with a known constant step. Return -1 if that step | |
1707 | is negative, 0 if it is zero, and 1 if it is greater than zero. */ | |
1708 | ||
1709 | static int | |
1710 | compare_step_with_zero (gimple *stmt) | |
1711 | { | |
1712 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3f5e8a76 RS |
1713 | data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
1714 | return tree_int_cst_compare (vect_dr_behavior (dr)->step, | |
1715 | size_zero_node); | |
62da9e14 RS |
1716 | } |
1717 | ||
1718 | /* If the target supports a permute mask that reverses the elements in | |
1719 | a vector of type VECTYPE, return that mask, otherwise return null. */ | |
1720 | ||
1721 | static tree | |
1722 | perm_mask_for_reverse (tree vectype) | |
1723 | { | |
1724 | int i, nunits; | |
62da9e14 RS |
1725 | |
1726 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
62da9e14 | 1727 | |
d980067b RS |
1728 | /* The encoding has a single stepped pattern. */ |
1729 | vec_perm_builder sel (nunits, 1, 3); | |
1730 | for (i = 0; i < 3; ++i) | |
908a1a16 | 1731 | sel.quick_push (nunits - 1 - i); |
62da9e14 | 1732 | |
e3342de4 RS |
1733 | vec_perm_indices indices (sel, 1, nunits); |
1734 | if (!can_vec_perm_const_p (TYPE_MODE (vectype), indices)) | |
62da9e14 | 1735 | return NULL_TREE; |
e3342de4 | 1736 | return vect_gen_perm_mask_checked (vectype, indices); |
62da9e14 | 1737 | } |
5ce9450f | 1738 | |
2de001ee RS |
1739 | /* A subroutine of get_load_store_type, with a subset of the same |
1740 | arguments. Handle the case where STMT is part of a grouped load | |
1741 | or store. | |
1742 | ||
1743 | For stores, the statements in the group are all consecutive | |
1744 | and there is no gap at the end. For loads, the statements in the | |
1745 | group might not be consecutive; there can be gaps between statements | |
1746 | as well as at the end. */ | |
1747 | ||
1748 | static bool | |
1749 | get_group_load_store_type (gimple *stmt, tree vectype, bool slp, | |
1750 | vec_load_store_type vls_type, | |
1751 | vect_memory_access_type *memory_access_type) | |
1752 | { | |
1753 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1754 | vec_info *vinfo = stmt_info->vinfo; | |
1755 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1756 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; | |
1757 | gimple *first_stmt = GROUP_FIRST_ELEMENT (stmt_info); | |
f702e7d4 | 1758 | data_reference *first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
2de001ee RS |
1759 | unsigned int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
1760 | bool single_element_p = (stmt == first_stmt | |
1761 | && !GROUP_NEXT_ELEMENT (stmt_info)); | |
1762 | unsigned HOST_WIDE_INT gap = GROUP_GAP (vinfo_for_stmt (first_stmt)); | |
522fcdd7 | 1763 | unsigned nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2de001ee RS |
1764 | |
1765 | /* True if the vectorized statements would access beyond the last | |
1766 | statement in the group. */ | |
1767 | bool overrun_p = false; | |
1768 | ||
1769 | /* True if we can cope with such overrun by peeling for gaps, so that | |
1770 | there is at least one final scalar iteration after the vector loop. */ | |
1771 | bool can_overrun_p = (vls_type == VLS_LOAD && loop_vinfo && !loop->inner); | |
1772 | ||
1773 | /* There can only be a gap at the end of the group if the stride is | |
1774 | known at compile time. */ | |
1775 | gcc_assert (!STMT_VINFO_STRIDED_P (stmt_info) || gap == 0); | |
1776 | ||
1777 | /* Stores can't yet have gaps. */ | |
1778 | gcc_assert (slp || vls_type == VLS_LOAD || gap == 0); | |
1779 | ||
1780 | if (slp) | |
1781 | { | |
1782 | if (STMT_VINFO_STRIDED_P (stmt_info)) | |
1783 | { | |
1784 | /* Try to use consecutive accesses of GROUP_SIZE elements, | |
1785 | separated by the stride, until we have a complete vector. | |
1786 | Fall back to scalar accesses if that isn't possible. */ | |
1787 | if (nunits % group_size == 0) | |
1788 | *memory_access_type = VMAT_STRIDED_SLP; | |
1789 | else | |
1790 | *memory_access_type = VMAT_ELEMENTWISE; | |
1791 | } | |
1792 | else | |
1793 | { | |
1794 | overrun_p = loop_vinfo && gap != 0; | |
1795 | if (overrun_p && vls_type != VLS_LOAD) | |
1796 | { | |
1797 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1798 | "Grouped store with gaps requires" | |
1799 | " non-consecutive accesses\n"); | |
1800 | return false; | |
1801 | } | |
f702e7d4 RS |
1802 | /* An overrun is fine if the trailing elements are smaller |
1803 | than the alignment boundary B. Every vector access will | |
1804 | be a multiple of B and so we are guaranteed to access a | |
1805 | non-gap element in the same B-sized block. */ | |
f9ef2c76 | 1806 | if (overrun_p |
f702e7d4 RS |
1807 | && gap < (vect_known_alignment_in_bytes (first_dr) |
1808 | / vect_get_scalar_dr_size (first_dr))) | |
f9ef2c76 | 1809 | overrun_p = false; |
2de001ee RS |
1810 | if (overrun_p && !can_overrun_p) |
1811 | { | |
1812 | if (dump_enabled_p ()) | |
1813 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1814 | "Peeling for outer loop is not supported\n"); | |
1815 | return false; | |
1816 | } | |
1817 | *memory_access_type = VMAT_CONTIGUOUS; | |
1818 | } | |
1819 | } | |
1820 | else | |
1821 | { | |
1822 | /* We can always handle this case using elementwise accesses, | |
1823 | but see if something more efficient is available. */ | |
1824 | *memory_access_type = VMAT_ELEMENTWISE; | |
1825 | ||
1826 | /* If there is a gap at the end of the group then these optimizations | |
1827 | would access excess elements in the last iteration. */ | |
1828 | bool would_overrun_p = (gap != 0); | |
f702e7d4 RS |
1829 | /* An overrun is fine if the trailing elements are smaller than the |
1830 | alignment boundary B. Every vector access will be a multiple of B | |
1831 | and so we are guaranteed to access a non-gap element in the | |
1832 | same B-sized block. */ | |
f9ef2c76 | 1833 | if (would_overrun_p |
f702e7d4 RS |
1834 | && gap < (vect_known_alignment_in_bytes (first_dr) |
1835 | / vect_get_scalar_dr_size (first_dr))) | |
f9ef2c76 | 1836 | would_overrun_p = false; |
f702e7d4 | 1837 | |
2de001ee | 1838 | if (!STMT_VINFO_STRIDED_P (stmt_info) |
62da9e14 RS |
1839 | && (can_overrun_p || !would_overrun_p) |
1840 | && compare_step_with_zero (stmt) > 0) | |
2de001ee RS |
1841 | { |
1842 | /* First try using LOAD/STORE_LANES. */ | |
1843 | if (vls_type == VLS_LOAD | |
1844 | ? vect_load_lanes_supported (vectype, group_size) | |
1845 | : vect_store_lanes_supported (vectype, group_size)) | |
1846 | { | |
1847 | *memory_access_type = VMAT_LOAD_STORE_LANES; | |
1848 | overrun_p = would_overrun_p; | |
1849 | } | |
1850 | ||
1851 | /* If that fails, try using permuting loads. */ | |
1852 | if (*memory_access_type == VMAT_ELEMENTWISE | |
1853 | && (vls_type == VLS_LOAD | |
1854 | ? vect_grouped_load_supported (vectype, single_element_p, | |
1855 | group_size) | |
1856 | : vect_grouped_store_supported (vectype, group_size))) | |
1857 | { | |
1858 | *memory_access_type = VMAT_CONTIGUOUS_PERMUTE; | |
1859 | overrun_p = would_overrun_p; | |
1860 | } | |
1861 | } | |
1862 | } | |
1863 | ||
1864 | if (vls_type != VLS_LOAD && first_stmt == stmt) | |
1865 | { | |
1866 | /* STMT is the leader of the group. Check the operands of all the | |
1867 | stmts of the group. */ | |
1868 | gimple *next_stmt = GROUP_NEXT_ELEMENT (stmt_info); | |
1869 | while (next_stmt) | |
1870 | { | |
1871 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
1872 | tree op = gimple_assign_rhs1 (next_stmt); | |
1873 | gimple *def_stmt; | |
1874 | enum vect_def_type dt; | |
1875 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt)) | |
1876 | { | |
1877 | if (dump_enabled_p ()) | |
1878 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1879 | "use not simple.\n"); | |
1880 | return false; | |
1881 | } | |
1882 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); | |
1883 | } | |
1884 | } | |
1885 | ||
1886 | if (overrun_p) | |
1887 | { | |
1888 | gcc_assert (can_overrun_p); | |
1889 | if (dump_enabled_p ()) | |
1890 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1891 | "Data access with gaps requires scalar " | |
1892 | "epilogue loop\n"); | |
1893 | LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true; | |
1894 | } | |
1895 | ||
1896 | return true; | |
1897 | } | |
1898 | ||
62da9e14 RS |
1899 | /* A subroutine of get_load_store_type, with a subset of the same |
1900 | arguments. Handle the case where STMT is a load or store that | |
1901 | accesses consecutive elements with a negative step. */ | |
1902 | ||
1903 | static vect_memory_access_type | |
1904 | get_negative_load_store_type (gimple *stmt, tree vectype, | |
1905 | vec_load_store_type vls_type, | |
1906 | unsigned int ncopies) | |
1907 | { | |
1908 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1909 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
1910 | dr_alignment_support alignment_support_scheme; | |
1911 | ||
1912 | if (ncopies > 1) | |
1913 | { | |
1914 | if (dump_enabled_p ()) | |
1915 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1916 | "multiple types with negative step.\n"); | |
1917 | return VMAT_ELEMENTWISE; | |
1918 | } | |
1919 | ||
1920 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
1921 | if (alignment_support_scheme != dr_aligned | |
1922 | && alignment_support_scheme != dr_unaligned_supported) | |
1923 | { | |
1924 | if (dump_enabled_p ()) | |
1925 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1926 | "negative step but alignment required.\n"); | |
1927 | return VMAT_ELEMENTWISE; | |
1928 | } | |
1929 | ||
1930 | if (vls_type == VLS_STORE_INVARIANT) | |
1931 | { | |
1932 | if (dump_enabled_p ()) | |
1933 | dump_printf_loc (MSG_NOTE, vect_location, | |
1934 | "negative step with invariant source;" | |
1935 | " no permute needed.\n"); | |
1936 | return VMAT_CONTIGUOUS_DOWN; | |
1937 | } | |
1938 | ||
1939 | if (!perm_mask_for_reverse (vectype)) | |
1940 | { | |
1941 | if (dump_enabled_p ()) | |
1942 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1943 | "negative step and reversing not supported.\n"); | |
1944 | return VMAT_ELEMENTWISE; | |
1945 | } | |
1946 | ||
1947 | return VMAT_CONTIGUOUS_REVERSE; | |
1948 | } | |
1949 | ||
2de001ee RS |
1950 | /* Analyze load or store statement STMT of type VLS_TYPE. Return true |
1951 | if there is a memory access type that the vectorized form can use, | |
1952 | storing it in *MEMORY_ACCESS_TYPE if so. If we decide to use gathers | |
1953 | or scatters, fill in GS_INFO accordingly. | |
1954 | ||
1955 | SLP says whether we're performing SLP rather than loop vectorization. | |
62da9e14 RS |
1956 | VECTYPE is the vector type that the vectorized statements will use. |
1957 | NCOPIES is the number of vector statements that will be needed. */ | |
2de001ee RS |
1958 | |
1959 | static bool | |
1960 | get_load_store_type (gimple *stmt, tree vectype, bool slp, | |
62da9e14 | 1961 | vec_load_store_type vls_type, unsigned int ncopies, |
2de001ee RS |
1962 | vect_memory_access_type *memory_access_type, |
1963 | gather_scatter_info *gs_info) | |
1964 | { | |
1965 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1966 | vec_info *vinfo = stmt_info->vinfo; | |
1967 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4d694b27 | 1968 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2de001ee RS |
1969 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
1970 | { | |
1971 | *memory_access_type = VMAT_GATHER_SCATTER; | |
1972 | gimple *def_stmt; | |
1973 | if (!vect_check_gather_scatter (stmt, loop_vinfo, gs_info)) | |
1974 | gcc_unreachable (); | |
1975 | else if (!vect_is_simple_use (gs_info->offset, vinfo, &def_stmt, | |
1976 | &gs_info->offset_dt, | |
1977 | &gs_info->offset_vectype)) | |
1978 | { | |
1979 | if (dump_enabled_p ()) | |
1980 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1981 | "%s index use not simple.\n", | |
1982 | vls_type == VLS_LOAD ? "gather" : "scatter"); | |
1983 | return false; | |
1984 | } | |
1985 | } | |
1986 | else if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1987 | { | |
1988 | if (!get_group_load_store_type (stmt, vectype, slp, vls_type, | |
1989 | memory_access_type)) | |
1990 | return false; | |
1991 | } | |
1992 | else if (STMT_VINFO_STRIDED_P (stmt_info)) | |
1993 | { | |
1994 | gcc_assert (!slp); | |
1995 | *memory_access_type = VMAT_ELEMENTWISE; | |
1996 | } | |
1997 | else | |
62da9e14 RS |
1998 | { |
1999 | int cmp = compare_step_with_zero (stmt); | |
2000 | if (cmp < 0) | |
2001 | *memory_access_type = get_negative_load_store_type | |
2002 | (stmt, vectype, vls_type, ncopies); | |
2003 | else if (cmp == 0) | |
2004 | { | |
2005 | gcc_assert (vls_type == VLS_LOAD); | |
2006 | *memory_access_type = VMAT_INVARIANT; | |
2007 | } | |
2008 | else | |
2009 | *memory_access_type = VMAT_CONTIGUOUS; | |
2010 | } | |
2de001ee | 2011 | |
4d694b27 RS |
2012 | if ((*memory_access_type == VMAT_ELEMENTWISE |
2013 | || *memory_access_type == VMAT_STRIDED_SLP) | |
2014 | && !nunits.is_constant ()) | |
2015 | { | |
2016 | if (dump_enabled_p ()) | |
2017 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2018 | "Not using elementwise accesses due to variable " | |
2019 | "vectorization factor.\n"); | |
2020 | return false; | |
2021 | } | |
2022 | ||
2de001ee RS |
2023 | /* FIXME: At the moment the cost model seems to underestimate the |
2024 | cost of using elementwise accesses. This check preserves the | |
2025 | traditional behavior until that can be fixed. */ | |
2026 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2027 | && !STMT_VINFO_STRIDED_P (stmt_info)) | |
2028 | { | |
2029 | if (dump_enabled_p ()) | |
2030 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2031 | "not falling back to elementwise accesses\n"); | |
2032 | return false; | |
2033 | } | |
2034 | return true; | |
2035 | } | |
2036 | ||
5ce9450f JJ |
2037 | /* Function vectorizable_mask_load_store. |
2038 | ||
2039 | Check if STMT performs a conditional load or store that can be vectorized. | |
2040 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2041 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. | |
2042 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2043 | ||
2044 | static bool | |
355fe088 TS |
2045 | vectorizable_mask_load_store (gimple *stmt, gimple_stmt_iterator *gsi, |
2046 | gimple **vec_stmt, slp_tree slp_node) | |
5ce9450f JJ |
2047 | { |
2048 | tree vec_dest = NULL; | |
2049 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2050 | stmt_vec_info prev_stmt_info; | |
2051 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2052 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2053 | bool nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
2054 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
2055 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
57e2f6ad | 2056 | tree rhs_vectype = NULL_TREE; |
045c1278 | 2057 | tree mask_vectype; |
5ce9450f | 2058 | tree elem_type; |
355fe088 | 2059 | gimple *new_stmt; |
5ce9450f JJ |
2060 | tree dummy; |
2061 | tree dataref_ptr = NULL_TREE; | |
355fe088 | 2062 | gimple *ptr_incr; |
4d694b27 | 2063 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
5ce9450f JJ |
2064 | int ncopies; |
2065 | int i, j; | |
2066 | bool inv_p; | |
134c85ca | 2067 | gather_scatter_info gs_info; |
2de001ee | 2068 | vec_load_store_type vls_type; |
5ce9450f | 2069 | tree mask; |
355fe088 | 2070 | gimple *def_stmt; |
5ce9450f JJ |
2071 | enum vect_def_type dt; |
2072 | ||
2073 | if (slp_node != NULL) | |
2074 | return false; | |
2075 | ||
e8f142e2 | 2076 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
5ce9450f JJ |
2077 | gcc_assert (ncopies >= 1); |
2078 | ||
5ce9450f | 2079 | mask = gimple_call_arg (stmt, 2); |
045c1278 | 2080 | |
2568d8a1 | 2081 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (mask))) |
5ce9450f JJ |
2082 | return false; |
2083 | ||
2084 | /* FORNOW. This restriction should be relaxed. */ | |
2085 | if (nested_in_vect_loop && ncopies > 1) | |
2086 | { | |
2087 | if (dump_enabled_p ()) | |
2088 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2089 | "multiple types in nested loop."); | |
2090 | return false; | |
2091 | } | |
2092 | ||
2093 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) | |
2094 | return false; | |
2095 | ||
66c16fd9 RB |
2096 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
2097 | && ! vec_stmt) | |
5ce9450f JJ |
2098 | return false; |
2099 | ||
2100 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
2101 | return false; | |
2102 | ||
2103 | elem_type = TREE_TYPE (vectype); | |
2104 | ||
045c1278 IE |
2105 | if (TREE_CODE (mask) != SSA_NAME) |
2106 | return false; | |
2107 | ||
2108 | if (!vect_is_simple_use (mask, loop_vinfo, &def_stmt, &dt, &mask_vectype)) | |
2109 | return false; | |
2110 | ||
2111 | if (!mask_vectype) | |
2112 | mask_vectype = get_mask_type_for_scalar_type (TREE_TYPE (vectype)); | |
2113 | ||
dc6a3147 IE |
2114 | if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype) |
2115 | || TYPE_VECTOR_SUBPARTS (mask_vectype) != TYPE_VECTOR_SUBPARTS (vectype)) | |
045c1278 IE |
2116 | return false; |
2117 | ||
2de001ee | 2118 | if (gimple_call_internal_fn (stmt) == IFN_MASK_STORE) |
57e2f6ad IE |
2119 | { |
2120 | tree rhs = gimple_call_arg (stmt, 3); | |
2121 | if (!vect_is_simple_use (rhs, loop_vinfo, &def_stmt, &dt, &rhs_vectype)) | |
2122 | return false; | |
2de001ee RS |
2123 | if (dt == vect_constant_def || dt == vect_external_def) |
2124 | vls_type = VLS_STORE_INVARIANT; | |
2125 | else | |
2126 | vls_type = VLS_STORE; | |
57e2f6ad | 2127 | } |
2de001ee RS |
2128 | else |
2129 | vls_type = VLS_LOAD; | |
57e2f6ad | 2130 | |
2de001ee | 2131 | vect_memory_access_type memory_access_type; |
62da9e14 | 2132 | if (!get_load_store_type (stmt, vectype, false, vls_type, ncopies, |
2de001ee RS |
2133 | &memory_access_type, &gs_info)) |
2134 | return false; | |
03b9e8e4 | 2135 | |
2de001ee RS |
2136 | if (memory_access_type == VMAT_GATHER_SCATTER) |
2137 | { | |
134c85ca | 2138 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); |
03b9e8e4 JJ |
2139 | tree masktype |
2140 | = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (TREE_CHAIN (arglist)))); | |
2141 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
2142 | { | |
2143 | if (dump_enabled_p ()) | |
2144 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2145 | "masked gather with integer mask not supported."); | |
2146 | return false; | |
2147 | } | |
5ce9450f | 2148 | } |
2de001ee RS |
2149 | else if (memory_access_type != VMAT_CONTIGUOUS) |
2150 | { | |
2151 | if (dump_enabled_p ()) | |
2152 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2153 | "unsupported access type for masked %s.\n", | |
2154 | vls_type == VLS_LOAD ? "load" : "store"); | |
2155 | return false; | |
2156 | } | |
5ce9450f | 2157 | else if (!VECTOR_MODE_P (TYPE_MODE (vectype)) |
045c1278 IE |
2158 | || !can_vec_mask_load_store_p (TYPE_MODE (vectype), |
2159 | TYPE_MODE (mask_vectype), | |
2de001ee | 2160 | vls_type == VLS_LOAD) |
57e2f6ad IE |
2161 | || (rhs_vectype |
2162 | && !useless_type_conversion_p (vectype, rhs_vectype))) | |
5ce9450f JJ |
2163 | return false; |
2164 | ||
5ce9450f JJ |
2165 | if (!vec_stmt) /* transformation not required. */ |
2166 | { | |
2de001ee | 2167 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; |
5ce9450f | 2168 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; |
2de001ee RS |
2169 | if (vls_type == VLS_LOAD) |
2170 | vect_model_load_cost (stmt_info, ncopies, memory_access_type, | |
2171 | NULL, NULL, NULL); | |
5ce9450f | 2172 | else |
2de001ee RS |
2173 | vect_model_store_cost (stmt_info, ncopies, memory_access_type, |
2174 | dt, NULL, NULL, NULL); | |
5ce9450f JJ |
2175 | return true; |
2176 | } | |
2de001ee | 2177 | gcc_assert (memory_access_type == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); |
5ce9450f | 2178 | |
67b8dbac | 2179 | /* Transform. */ |
5ce9450f | 2180 | |
2de001ee | 2181 | if (memory_access_type == VMAT_GATHER_SCATTER) |
5ce9450f JJ |
2182 | { |
2183 | tree vec_oprnd0 = NULL_TREE, op; | |
134c85ca | 2184 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); |
5ce9450f | 2185 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; |
acdcd61b | 2186 | tree ptr, vec_mask = NULL_TREE, mask_op = NULL_TREE, var, scale; |
5ce9450f | 2187 | tree perm_mask = NULL_TREE, prev_res = NULL_TREE; |
acdcd61b | 2188 | tree mask_perm_mask = NULL_TREE; |
5ce9450f JJ |
2189 | edge pe = loop_preheader_edge (loop); |
2190 | gimple_seq seq; | |
2191 | basic_block new_bb; | |
2192 | enum { NARROW, NONE, WIDEN } modifier; | |
4d694b27 RS |
2193 | poly_uint64 gather_off_nunits |
2194 | = TYPE_VECTOR_SUBPARTS (gs_info.offset_vectype); | |
5ce9450f | 2195 | |
134c85ca | 2196 | rettype = TREE_TYPE (TREE_TYPE (gs_info.decl)); |
acdcd61b JJ |
2197 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
2198 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2199 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2200 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2201 | scaletype = TREE_VALUE (arglist); | |
2202 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
2203 | && types_compatible_p (srctype, masktype)); | |
2204 | ||
4d694b27 | 2205 | if (known_eq (nunits, gather_off_nunits)) |
5ce9450f | 2206 | modifier = NONE; |
4d694b27 | 2207 | else if (known_eq (nunits * 2, gather_off_nunits)) |
5ce9450f | 2208 | { |
5ce9450f JJ |
2209 | modifier = WIDEN; |
2210 | ||
4d694b27 RS |
2211 | /* Currently widening gathers and scatters are only supported for |
2212 | fixed-length vectors. */ | |
2213 | int count = gather_off_nunits.to_constant (); | |
2214 | vec_perm_builder sel (count, count, 1); | |
2215 | for (i = 0; i < count; ++i) | |
2216 | sel.quick_push (i | (count / 2)); | |
5ce9450f | 2217 | |
4d694b27 | 2218 | vec_perm_indices indices (sel, 1, count); |
e3342de4 RS |
2219 | perm_mask = vect_gen_perm_mask_checked (gs_info.offset_vectype, |
2220 | indices); | |
5ce9450f | 2221 | } |
4d694b27 | 2222 | else if (known_eq (nunits, gather_off_nunits * 2)) |
5ce9450f | 2223 | { |
5ce9450f JJ |
2224 | modifier = NARROW; |
2225 | ||
4d694b27 RS |
2226 | /* Currently narrowing gathers and scatters are only supported for |
2227 | fixed-length vectors. */ | |
2228 | int count = nunits.to_constant (); | |
2229 | vec_perm_builder sel (count, count, 1); | |
2230 | sel.quick_grow (count); | |
2231 | for (i = 0; i < count; ++i) | |
2232 | sel[i] = i < count / 2 ? i : i + count / 2; | |
2233 | vec_perm_indices indices (sel, 2, count); | |
e3342de4 | 2234 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); |
5ce9450f | 2235 | |
5ce9450f | 2236 | ncopies *= 2; |
4d694b27 RS |
2237 | for (i = 0; i < count; ++i) |
2238 | sel[i] = i | (count / 2); | |
2239 | indices.new_vector (sel, 2, count); | |
e3342de4 | 2240 | mask_perm_mask = vect_gen_perm_mask_checked (masktype, indices); |
5ce9450f JJ |
2241 | } |
2242 | else | |
2243 | gcc_unreachable (); | |
2244 | ||
5ce9450f JJ |
2245 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); |
2246 | ||
134c85ca | 2247 | ptr = fold_convert (ptrtype, gs_info.base); |
5ce9450f JJ |
2248 | if (!is_gimple_min_invariant (ptr)) |
2249 | { | |
2250 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
2251 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
2252 | gcc_assert (!new_bb); | |
2253 | } | |
2254 | ||
134c85ca | 2255 | scale = build_int_cst (scaletype, gs_info.scale); |
5ce9450f JJ |
2256 | |
2257 | prev_stmt_info = NULL; | |
2258 | for (j = 0; j < ncopies; ++j) | |
2259 | { | |
2260 | if (modifier == WIDEN && (j & 1)) | |
2261 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
2262 | perm_mask, stmt, gsi); | |
2263 | else if (j == 0) | |
2264 | op = vec_oprnd0 | |
134c85ca | 2265 | = vect_get_vec_def_for_operand (gs_info.offset, stmt); |
5ce9450f JJ |
2266 | else |
2267 | op = vec_oprnd0 | |
134c85ca | 2268 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, vec_oprnd0); |
5ce9450f JJ |
2269 | |
2270 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
2271 | { | |
2272 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
2273 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 2274 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
5ce9450f JJ |
2275 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
2276 | new_stmt | |
0d0e4a03 | 2277 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
2278 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2279 | op = var; | |
2280 | } | |
2281 | ||
acdcd61b JJ |
2282 | if (mask_perm_mask && (j & 1)) |
2283 | mask_op = permute_vec_elements (mask_op, mask_op, | |
2284 | mask_perm_mask, stmt, gsi); | |
5ce9450f JJ |
2285 | else |
2286 | { | |
acdcd61b | 2287 | if (j == 0) |
81c40241 | 2288 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); |
acdcd61b JJ |
2289 | else |
2290 | { | |
81c40241 | 2291 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
acdcd61b JJ |
2292 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2293 | } | |
5ce9450f | 2294 | |
acdcd61b JJ |
2295 | mask_op = vec_mask; |
2296 | if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask))) | |
2297 | { | |
2298 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)) | |
2299 | == TYPE_VECTOR_SUBPARTS (masktype)); | |
0e22bb5a | 2300 | var = vect_get_new_ssa_name (masktype, vect_simple_var); |
acdcd61b JJ |
2301 | mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op); |
2302 | new_stmt | |
0d0e4a03 | 2303 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, mask_op); |
acdcd61b JJ |
2304 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2305 | mask_op = var; | |
2306 | } | |
5ce9450f JJ |
2307 | } |
2308 | ||
2309 | new_stmt | |
134c85ca | 2310 | = gimple_build_call (gs_info.decl, 5, mask_op, ptr, op, mask_op, |
5ce9450f JJ |
2311 | scale); |
2312 | ||
2313 | if (!useless_type_conversion_p (vectype, rettype)) | |
2314 | { | |
2315 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
2316 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 2317 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
5ce9450f JJ |
2318 | gimple_call_set_lhs (new_stmt, op); |
2319 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 2320 | var = make_ssa_name (vec_dest); |
5ce9450f | 2321 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
0d0e4a03 | 2322 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
2323 | } |
2324 | else | |
2325 | { | |
2326 | var = make_ssa_name (vec_dest, new_stmt); | |
2327 | gimple_call_set_lhs (new_stmt, var); | |
2328 | } | |
2329 | ||
2330 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2331 | ||
2332 | if (modifier == NARROW) | |
2333 | { | |
2334 | if ((j & 1) == 0) | |
2335 | { | |
2336 | prev_res = var; | |
2337 | continue; | |
2338 | } | |
2339 | var = permute_vec_elements (prev_res, var, | |
2340 | perm_mask, stmt, gsi); | |
2341 | new_stmt = SSA_NAME_DEF_STMT (var); | |
2342 | } | |
2343 | ||
2344 | if (prev_stmt_info == NULL) | |
2345 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2346 | else | |
2347 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2348 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2349 | } | |
3efe2e2c JJ |
2350 | |
2351 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
2352 | from the IL. */ | |
e6f5c25d IE |
2353 | if (STMT_VINFO_RELATED_STMT (stmt_info)) |
2354 | { | |
2355 | stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
2356 | stmt_info = vinfo_for_stmt (stmt); | |
2357 | } | |
3efe2e2c JJ |
2358 | tree lhs = gimple_call_lhs (stmt); |
2359 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
2360 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
2361 | set_vinfo_for_stmt (stmt, NULL); | |
2362 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2363 | gsi_replace (gsi, new_stmt, true); | |
5ce9450f JJ |
2364 | return true; |
2365 | } | |
2de001ee | 2366 | else if (vls_type != VLS_LOAD) |
5ce9450f JJ |
2367 | { |
2368 | tree vec_rhs = NULL_TREE, vec_mask = NULL_TREE; | |
2369 | prev_stmt_info = NULL; | |
2d4dc223 | 2370 | LOOP_VINFO_HAS_MASK_STORE (loop_vinfo) = true; |
5ce9450f JJ |
2371 | for (i = 0; i < ncopies; i++) |
2372 | { | |
2373 | unsigned align, misalign; | |
2374 | ||
2375 | if (i == 0) | |
2376 | { | |
2377 | tree rhs = gimple_call_arg (stmt, 3); | |
81c40241 | 2378 | vec_rhs = vect_get_vec_def_for_operand (rhs, stmt); |
7251b0bf RS |
2379 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, |
2380 | mask_vectype); | |
5ce9450f JJ |
2381 | /* We should have catched mismatched types earlier. */ |
2382 | gcc_assert (useless_type_conversion_p (vectype, | |
2383 | TREE_TYPE (vec_rhs))); | |
2384 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, | |
2385 | NULL_TREE, &dummy, gsi, | |
2386 | &ptr_incr, false, &inv_p); | |
2387 | gcc_assert (!inv_p); | |
2388 | } | |
2389 | else | |
2390 | { | |
81c40241 | 2391 | vect_is_simple_use (vec_rhs, loop_vinfo, &def_stmt, &dt); |
5ce9450f | 2392 | vec_rhs = vect_get_vec_def_for_stmt_copy (dt, vec_rhs); |
81c40241 | 2393 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2394 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2395 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2396 | TYPE_SIZE_UNIT (vectype)); | |
2397 | } | |
2398 | ||
f702e7d4 | 2399 | align = DR_TARGET_ALIGNMENT (dr); |
5ce9450f JJ |
2400 | if (aligned_access_p (dr)) |
2401 | misalign = 0; | |
2402 | else if (DR_MISALIGNMENT (dr) == -1) | |
2403 | { | |
2404 | align = TYPE_ALIGN_UNIT (elem_type); | |
2405 | misalign = 0; | |
2406 | } | |
2407 | else | |
2408 | misalign = DR_MISALIGNMENT (dr); | |
2409 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2410 | misalign); | |
08554c26 | 2411 | tree ptr = build_int_cst (TREE_TYPE (gimple_call_arg (stmt, 1)), |
146ec50f | 2412 | misalign ? least_bit_hwi (misalign) : align); |
a844293d | 2413 | gcall *call |
5ce9450f | 2414 | = gimple_build_call_internal (IFN_MASK_STORE, 4, dataref_ptr, |
08554c26 | 2415 | ptr, vec_mask, vec_rhs); |
a844293d RS |
2416 | gimple_call_set_nothrow (call, true); |
2417 | new_stmt = call; | |
5ce9450f JJ |
2418 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2419 | if (i == 0) | |
2420 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2421 | else | |
2422 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2423 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2424 | } | |
2425 | } | |
2426 | else | |
2427 | { | |
2428 | tree vec_mask = NULL_TREE; | |
2429 | prev_stmt_info = NULL; | |
2430 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); | |
2431 | for (i = 0; i < ncopies; i++) | |
2432 | { | |
2433 | unsigned align, misalign; | |
2434 | ||
2435 | if (i == 0) | |
2436 | { | |
7251b0bf RS |
2437 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, |
2438 | mask_vectype); | |
5ce9450f JJ |
2439 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, |
2440 | NULL_TREE, &dummy, gsi, | |
2441 | &ptr_incr, false, &inv_p); | |
2442 | gcc_assert (!inv_p); | |
2443 | } | |
2444 | else | |
2445 | { | |
81c40241 | 2446 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2447 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2448 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2449 | TYPE_SIZE_UNIT (vectype)); | |
2450 | } | |
2451 | ||
f702e7d4 | 2452 | align = DR_TARGET_ALIGNMENT (dr); |
5ce9450f JJ |
2453 | if (aligned_access_p (dr)) |
2454 | misalign = 0; | |
2455 | else if (DR_MISALIGNMENT (dr) == -1) | |
2456 | { | |
2457 | align = TYPE_ALIGN_UNIT (elem_type); | |
2458 | misalign = 0; | |
2459 | } | |
2460 | else | |
2461 | misalign = DR_MISALIGNMENT (dr); | |
2462 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2463 | misalign); | |
08554c26 | 2464 | tree ptr = build_int_cst (TREE_TYPE (gimple_call_arg (stmt, 1)), |
146ec50f | 2465 | misalign ? least_bit_hwi (misalign) : align); |
a844293d | 2466 | gcall *call |
5ce9450f | 2467 | = gimple_build_call_internal (IFN_MASK_LOAD, 3, dataref_ptr, |
08554c26 | 2468 | ptr, vec_mask); |
a844293d RS |
2469 | gimple_call_set_lhs (call, make_ssa_name (vec_dest)); |
2470 | gimple_call_set_nothrow (call, true); | |
2471 | vect_finish_stmt_generation (stmt, call, gsi); | |
5ce9450f | 2472 | if (i == 0) |
a844293d | 2473 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = call; |
5ce9450f | 2474 | else |
a844293d RS |
2475 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = call; |
2476 | prev_stmt_info = vinfo_for_stmt (call); | |
5ce9450f JJ |
2477 | } |
2478 | } | |
2479 | ||
2de001ee | 2480 | if (vls_type == VLS_LOAD) |
3efe2e2c JJ |
2481 | { |
2482 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
2483 | from the IL. */ | |
e6f5c25d IE |
2484 | if (STMT_VINFO_RELATED_STMT (stmt_info)) |
2485 | { | |
2486 | stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
2487 | stmt_info = vinfo_for_stmt (stmt); | |
2488 | } | |
3efe2e2c JJ |
2489 | tree lhs = gimple_call_lhs (stmt); |
2490 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
2491 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
2492 | set_vinfo_for_stmt (stmt, NULL); | |
2493 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2494 | gsi_replace (gsi, new_stmt, true); | |
2495 | } | |
2496 | ||
5ce9450f JJ |
2497 | return true; |
2498 | } | |
2499 | ||
37b14185 RB |
2500 | /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64}. */ |
2501 | ||
2502 | static bool | |
2503 | vectorizable_bswap (gimple *stmt, gimple_stmt_iterator *gsi, | |
2504 | gimple **vec_stmt, slp_tree slp_node, | |
2505 | tree vectype_in, enum vect_def_type *dt) | |
2506 | { | |
2507 | tree op, vectype; | |
2508 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2509 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2510 | unsigned ncopies, nunits; | |
2511 | ||
2512 | op = gimple_call_arg (stmt, 0); | |
2513 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2514 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
2515 | ||
2516 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2517 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2518 | case of SLP. */ | |
2519 | if (slp_node) | |
2520 | ncopies = 1; | |
2521 | else | |
e8f142e2 | 2522 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
37b14185 RB |
2523 | |
2524 | gcc_assert (ncopies >= 1); | |
2525 | ||
2526 | tree char_vectype = get_same_sized_vectype (char_type_node, vectype_in); | |
2527 | if (! char_vectype) | |
2528 | return false; | |
2529 | ||
794e3180 | 2530 | unsigned int num_bytes = TYPE_VECTOR_SUBPARTS (char_vectype); |
794e3180 | 2531 | unsigned word_bytes = num_bytes / nunits; |
908a1a16 | 2532 | |
d980067b RS |
2533 | /* The encoding uses one stepped pattern for each byte in the word. */ |
2534 | vec_perm_builder elts (num_bytes, word_bytes, 3); | |
2535 | for (unsigned i = 0; i < 3; ++i) | |
37b14185 | 2536 | for (unsigned j = 0; j < word_bytes; ++j) |
908a1a16 | 2537 | elts.quick_push ((i + 1) * word_bytes - j - 1); |
37b14185 | 2538 | |
e3342de4 RS |
2539 | vec_perm_indices indices (elts, 1, num_bytes); |
2540 | if (!can_vec_perm_const_p (TYPE_MODE (char_vectype), indices)) | |
37b14185 RB |
2541 | return false; |
2542 | ||
2543 | if (! vec_stmt) | |
2544 | { | |
2545 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
2546 | if (dump_enabled_p ()) | |
2547 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_bswap ===" | |
2548 | "\n"); | |
2549 | if (! PURE_SLP_STMT (stmt_info)) | |
2550 | { | |
2551 | add_stmt_cost (stmt_info->vinfo->target_cost_data, | |
2552 | 1, vector_stmt, stmt_info, 0, vect_prologue); | |
2553 | add_stmt_cost (stmt_info->vinfo->target_cost_data, | |
2554 | ncopies, vec_perm, stmt_info, 0, vect_body); | |
2555 | } | |
2556 | return true; | |
2557 | } | |
2558 | ||
736d0f28 | 2559 | tree bswap_vconst = vec_perm_indices_to_tree (char_vectype, indices); |
37b14185 RB |
2560 | |
2561 | /* Transform. */ | |
2562 | vec<tree> vec_oprnds = vNULL; | |
2563 | gimple *new_stmt = NULL; | |
2564 | stmt_vec_info prev_stmt_info = NULL; | |
2565 | for (unsigned j = 0; j < ncopies; j++) | |
2566 | { | |
2567 | /* Handle uses. */ | |
2568 | if (j == 0) | |
306b0c92 | 2569 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); |
37b14185 RB |
2570 | else |
2571 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
2572 | ||
2573 | /* Arguments are ready. create the new vector stmt. */ | |
2574 | unsigned i; | |
2575 | tree vop; | |
2576 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) | |
2577 | { | |
2578 | tree tem = make_ssa_name (char_vectype); | |
2579 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
2580 | char_vectype, vop)); | |
2581 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2582 | tree tem2 = make_ssa_name (char_vectype); | |
2583 | new_stmt = gimple_build_assign (tem2, VEC_PERM_EXPR, | |
2584 | tem, tem, bswap_vconst); | |
2585 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2586 | tem = make_ssa_name (vectype); | |
2587 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
2588 | vectype, tem2)); | |
2589 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2590 | if (slp_node) | |
2591 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
2592 | } | |
2593 | ||
2594 | if (slp_node) | |
2595 | continue; | |
2596 | ||
2597 | if (j == 0) | |
2598 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2599 | else | |
2600 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2601 | ||
2602 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2603 | } | |
2604 | ||
2605 | vec_oprnds.release (); | |
2606 | return true; | |
2607 | } | |
2608 | ||
b1b6836e RS |
2609 | /* Return true if vector types VECTYPE_IN and VECTYPE_OUT have |
2610 | integer elements and if we can narrow VECTYPE_IN to VECTYPE_OUT | |
2611 | in a single step. On success, store the binary pack code in | |
2612 | *CONVERT_CODE. */ | |
2613 | ||
2614 | static bool | |
2615 | simple_integer_narrowing (tree vectype_out, tree vectype_in, | |
2616 | tree_code *convert_code) | |
2617 | { | |
2618 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_out)) | |
2619 | || !INTEGRAL_TYPE_P (TREE_TYPE (vectype_in))) | |
2620 | return false; | |
2621 | ||
2622 | tree_code code; | |
2623 | int multi_step_cvt = 0; | |
2624 | auto_vec <tree, 8> interm_types; | |
2625 | if (!supportable_narrowing_operation (NOP_EXPR, vectype_out, vectype_in, | |
2626 | &code, &multi_step_cvt, | |
2627 | &interm_types) | |
2628 | || multi_step_cvt) | |
2629 | return false; | |
2630 | ||
2631 | *convert_code = code; | |
2632 | return true; | |
2633 | } | |
5ce9450f | 2634 | |
ebfd146a IR |
2635 | /* Function vectorizable_call. |
2636 | ||
538dd0b7 | 2637 | Check if GS performs a function call that can be vectorized. |
b8698a0f | 2638 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
2639 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2640 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2641 | ||
2642 | static bool | |
355fe088 | 2643 | vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
190c2236 | 2644 | slp_tree slp_node) |
ebfd146a | 2645 | { |
538dd0b7 | 2646 | gcall *stmt; |
ebfd146a IR |
2647 | tree vec_dest; |
2648 | tree scalar_dest; | |
2649 | tree op, type; | |
2650 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
538dd0b7 | 2651 | stmt_vec_info stmt_info = vinfo_for_stmt (gs), prev_stmt_info; |
ebfd146a | 2652 | tree vectype_out, vectype_in; |
c7bda0f4 RS |
2653 | poly_uint64 nunits_in; |
2654 | poly_uint64 nunits_out; | |
ebfd146a | 2655 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
190c2236 | 2656 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 2657 | vec_info *vinfo = stmt_info->vinfo; |
81c40241 | 2658 | tree fndecl, new_temp, rhs_type; |
355fe088 | 2659 | gimple *def_stmt; |
0502fb85 UB |
2660 | enum vect_def_type dt[3] |
2661 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 2662 | int ndts = 3; |
355fe088 | 2663 | gimple *new_stmt = NULL; |
ebfd146a | 2664 | int ncopies, j; |
6e1aa848 | 2665 | vec<tree> vargs = vNULL; |
ebfd146a IR |
2666 | enum { NARROW, NONE, WIDEN } modifier; |
2667 | size_t i, nargs; | |
9d5e7640 | 2668 | tree lhs; |
ebfd146a | 2669 | |
190c2236 | 2670 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2671 | return false; |
2672 | ||
66c16fd9 RB |
2673 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
2674 | && ! vec_stmt) | |
ebfd146a IR |
2675 | return false; |
2676 | ||
538dd0b7 DM |
2677 | /* Is GS a vectorizable call? */ |
2678 | stmt = dyn_cast <gcall *> (gs); | |
2679 | if (!stmt) | |
ebfd146a IR |
2680 | return false; |
2681 | ||
5ce9450f JJ |
2682 | if (gimple_call_internal_p (stmt) |
2683 | && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD | |
2684 | || gimple_call_internal_fn (stmt) == IFN_MASK_STORE)) | |
2685 | return vectorizable_mask_load_store (stmt, gsi, vec_stmt, | |
2686 | slp_node); | |
2687 | ||
0136f8f0 AH |
2688 | if (gimple_call_lhs (stmt) == NULL_TREE |
2689 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
ebfd146a IR |
2690 | return false; |
2691 | ||
0136f8f0 | 2692 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); |
5a2c1986 | 2693 | |
b690cc0f RG |
2694 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
2695 | ||
ebfd146a IR |
2696 | /* Process function arguments. */ |
2697 | rhs_type = NULL_TREE; | |
b690cc0f | 2698 | vectype_in = NULL_TREE; |
ebfd146a IR |
2699 | nargs = gimple_call_num_args (stmt); |
2700 | ||
1b1562a5 MM |
2701 | /* Bail out if the function has more than three arguments, we do not have |
2702 | interesting builtin functions to vectorize with more than two arguments | |
2703 | except for fma. No arguments is also not good. */ | |
2704 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
2705 | return false; |
2706 | ||
74bf76ed JJ |
2707 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
2708 | if (gimple_call_internal_p (stmt) | |
2709 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2710 | { | |
2711 | nargs = 0; | |
2712 | rhs_type = unsigned_type_node; | |
2713 | } | |
2714 | ||
ebfd146a IR |
2715 | for (i = 0; i < nargs; i++) |
2716 | { | |
b690cc0f RG |
2717 | tree opvectype; |
2718 | ||
ebfd146a IR |
2719 | op = gimple_call_arg (stmt, i); |
2720 | ||
2721 | /* We can only handle calls with arguments of the same type. */ | |
2722 | if (rhs_type | |
8533c9d8 | 2723 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 2724 | { |
73fbfcad | 2725 | if (dump_enabled_p ()) |
78c60e3d | 2726 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2727 | "argument types differ.\n"); |
ebfd146a IR |
2728 | return false; |
2729 | } | |
b690cc0f RG |
2730 | if (!rhs_type) |
2731 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 2732 | |
81c40241 | 2733 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[i], &opvectype)) |
ebfd146a | 2734 | { |
73fbfcad | 2735 | if (dump_enabled_p ()) |
78c60e3d | 2736 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2737 | "use not simple.\n"); |
ebfd146a IR |
2738 | return false; |
2739 | } | |
ebfd146a | 2740 | |
b690cc0f RG |
2741 | if (!vectype_in) |
2742 | vectype_in = opvectype; | |
2743 | else if (opvectype | |
2744 | && opvectype != vectype_in) | |
2745 | { | |
73fbfcad | 2746 | if (dump_enabled_p ()) |
78c60e3d | 2747 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2748 | "argument vector types differ.\n"); |
b690cc0f RG |
2749 | return false; |
2750 | } | |
2751 | } | |
2752 | /* If all arguments are external or constant defs use a vector type with | |
2753 | the same size as the output vector type. */ | |
ebfd146a | 2754 | if (!vectype_in) |
b690cc0f | 2755 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
2756 | if (vec_stmt) |
2757 | gcc_assert (vectype_in); | |
2758 | if (!vectype_in) | |
2759 | { | |
73fbfcad | 2760 | if (dump_enabled_p ()) |
7d8930a0 | 2761 | { |
78c60e3d SS |
2762 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2763 | "no vectype for scalar type "); | |
2764 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 2765 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
2766 | } |
2767 | ||
2768 | return false; | |
2769 | } | |
ebfd146a IR |
2770 | |
2771 | /* FORNOW */ | |
b690cc0f RG |
2772 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
2773 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
c7bda0f4 | 2774 | if (known_eq (nunits_in * 2, nunits_out)) |
ebfd146a | 2775 | modifier = NARROW; |
c7bda0f4 | 2776 | else if (known_eq (nunits_out, nunits_in)) |
ebfd146a | 2777 | modifier = NONE; |
c7bda0f4 | 2778 | else if (known_eq (nunits_out * 2, nunits_in)) |
ebfd146a IR |
2779 | modifier = WIDEN; |
2780 | else | |
2781 | return false; | |
2782 | ||
70439f0d RS |
2783 | /* We only handle functions that do not read or clobber memory. */ |
2784 | if (gimple_vuse (stmt)) | |
2785 | { | |
2786 | if (dump_enabled_p ()) | |
2787 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2788 | "function reads from or writes to memory.\n"); | |
2789 | return false; | |
2790 | } | |
2791 | ||
ebfd146a IR |
2792 | /* For now, we only vectorize functions if a target specific builtin |
2793 | is available. TODO -- in some cases, it might be profitable to | |
2794 | insert the calls for pieces of the vector, in order to be able | |
2795 | to vectorize other operations in the loop. */ | |
70439f0d RS |
2796 | fndecl = NULL_TREE; |
2797 | internal_fn ifn = IFN_LAST; | |
2798 | combined_fn cfn = gimple_call_combined_fn (stmt); | |
2799 | tree callee = gimple_call_fndecl (stmt); | |
2800 | ||
2801 | /* First try using an internal function. */ | |
b1b6836e RS |
2802 | tree_code convert_code = ERROR_MARK; |
2803 | if (cfn != CFN_LAST | |
2804 | && (modifier == NONE | |
2805 | || (modifier == NARROW | |
2806 | && simple_integer_narrowing (vectype_out, vectype_in, | |
2807 | &convert_code)))) | |
70439f0d RS |
2808 | ifn = vectorizable_internal_function (cfn, callee, vectype_out, |
2809 | vectype_in); | |
2810 | ||
2811 | /* If that fails, try asking for a target-specific built-in function. */ | |
2812 | if (ifn == IFN_LAST) | |
2813 | { | |
2814 | if (cfn != CFN_LAST) | |
2815 | fndecl = targetm.vectorize.builtin_vectorized_function | |
2816 | (cfn, vectype_out, vectype_in); | |
2817 | else | |
2818 | fndecl = targetm.vectorize.builtin_md_vectorized_function | |
2819 | (callee, vectype_out, vectype_in); | |
2820 | } | |
2821 | ||
2822 | if (ifn == IFN_LAST && !fndecl) | |
ebfd146a | 2823 | { |
70439f0d | 2824 | if (cfn == CFN_GOMP_SIMD_LANE |
74bf76ed JJ |
2825 | && !slp_node |
2826 | && loop_vinfo | |
2827 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2828 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
2829 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2830 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
2831 | { | |
2832 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
2833 | { 0, 1, 2, ... vf - 1 } vector. */ | |
2834 | gcc_assert (nargs == 0); | |
2835 | } | |
37b14185 RB |
2836 | else if (modifier == NONE |
2837 | && (gimple_call_builtin_p (stmt, BUILT_IN_BSWAP16) | |
2838 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP32) | |
2839 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP64))) | |
2840 | return vectorizable_bswap (stmt, gsi, vec_stmt, slp_node, | |
2841 | vectype_in, dt); | |
74bf76ed JJ |
2842 | else |
2843 | { | |
2844 | if (dump_enabled_p ()) | |
2845 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 2846 | "function is not vectorizable.\n"); |
74bf76ed JJ |
2847 | return false; |
2848 | } | |
ebfd146a IR |
2849 | } |
2850 | ||
fce57248 | 2851 | if (slp_node) |
190c2236 | 2852 | ncopies = 1; |
b1b6836e | 2853 | else if (modifier == NARROW && ifn == IFN_LAST) |
e8f142e2 | 2854 | ncopies = vect_get_num_copies (loop_vinfo, vectype_out); |
ebfd146a | 2855 | else |
e8f142e2 | 2856 | ncopies = vect_get_num_copies (loop_vinfo, vectype_in); |
ebfd146a IR |
2857 | |
2858 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
2859 | needs to be generated. */ | |
2860 | gcc_assert (ncopies >= 1); | |
2861 | ||
2862 | if (!vec_stmt) /* transformation not required. */ | |
2863 | { | |
2864 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 2865 | if (dump_enabled_p ()) |
e645e942 TJ |
2866 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ===" |
2867 | "\n"); | |
4fc5ebf1 | 2868 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); |
b1b6836e RS |
2869 | if (ifn != IFN_LAST && modifier == NARROW && !slp_node) |
2870 | add_stmt_cost (stmt_info->vinfo->target_cost_data, ncopies / 2, | |
2871 | vec_promote_demote, stmt_info, 0, vect_body); | |
2872 | ||
ebfd146a IR |
2873 | return true; |
2874 | } | |
2875 | ||
67b8dbac | 2876 | /* Transform. */ |
ebfd146a | 2877 | |
73fbfcad | 2878 | if (dump_enabled_p ()) |
e645e942 | 2879 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
2880 | |
2881 | /* Handle def. */ | |
2882 | scalar_dest = gimple_call_lhs (stmt); | |
2883 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
2884 | ||
2885 | prev_stmt_info = NULL; | |
b1b6836e | 2886 | if (modifier == NONE || ifn != IFN_LAST) |
ebfd146a | 2887 | { |
b1b6836e | 2888 | tree prev_res = NULL_TREE; |
ebfd146a IR |
2889 | for (j = 0; j < ncopies; ++j) |
2890 | { | |
2891 | /* Build argument list for the vectorized call. */ | |
2892 | if (j == 0) | |
9771b263 | 2893 | vargs.create (nargs); |
ebfd146a | 2894 | else |
9771b263 | 2895 | vargs.truncate (0); |
ebfd146a | 2896 | |
190c2236 JJ |
2897 | if (slp_node) |
2898 | { | |
ef062b13 | 2899 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2900 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2901 | |
2902 | for (i = 0; i < nargs; i++) | |
9771b263 | 2903 | vargs.quick_push (gimple_call_arg (stmt, i)); |
306b0c92 | 2904 | vect_get_slp_defs (vargs, slp_node, &vec_defs); |
37b5ec8f | 2905 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2906 | |
2907 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2908 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
2909 | { |
2910 | size_t k; | |
2911 | for (k = 0; k < nargs; k++) | |
2912 | { | |
37b5ec8f | 2913 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 2914 | vargs[k] = vec_oprndsk[i]; |
190c2236 | 2915 | } |
b1b6836e RS |
2916 | if (modifier == NARROW) |
2917 | { | |
2918 | tree half_res = make_ssa_name (vectype_in); | |
a844293d RS |
2919 | gcall *call |
2920 | = gimple_build_call_internal_vec (ifn, vargs); | |
2921 | gimple_call_set_lhs (call, half_res); | |
2922 | gimple_call_set_nothrow (call, true); | |
2923 | new_stmt = call; | |
b1b6836e RS |
2924 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2925 | if ((i & 1) == 0) | |
2926 | { | |
2927 | prev_res = half_res; | |
2928 | continue; | |
2929 | } | |
2930 | new_temp = make_ssa_name (vec_dest); | |
2931 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
2932 | prev_res, half_res); | |
2933 | } | |
70439f0d | 2934 | else |
b1b6836e | 2935 | { |
a844293d | 2936 | gcall *call; |
b1b6836e | 2937 | if (ifn != IFN_LAST) |
a844293d | 2938 | call = gimple_build_call_internal_vec (ifn, vargs); |
b1b6836e | 2939 | else |
a844293d RS |
2940 | call = gimple_build_call_vec (fndecl, vargs); |
2941 | new_temp = make_ssa_name (vec_dest, call); | |
2942 | gimple_call_set_lhs (call, new_temp); | |
2943 | gimple_call_set_nothrow (call, true); | |
2944 | new_stmt = call; | |
b1b6836e | 2945 | } |
190c2236 | 2946 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 2947 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2948 | } |
2949 | ||
2950 | for (i = 0; i < nargs; i++) | |
2951 | { | |
37b5ec8f | 2952 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2953 | vec_oprndsi.release (); |
190c2236 | 2954 | } |
190c2236 JJ |
2955 | continue; |
2956 | } | |
2957 | ||
ebfd146a IR |
2958 | for (i = 0; i < nargs; i++) |
2959 | { | |
2960 | op = gimple_call_arg (stmt, i); | |
2961 | if (j == 0) | |
2962 | vec_oprnd0 | |
81c40241 | 2963 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 2964 | else |
63827fb8 IR |
2965 | { |
2966 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
2967 | vec_oprnd0 | |
2968 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
2969 | } | |
ebfd146a | 2970 | |
9771b263 | 2971 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
2972 | } |
2973 | ||
74bf76ed JJ |
2974 | if (gimple_call_internal_p (stmt) |
2975 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2976 | { | |
c7bda0f4 | 2977 | tree cst = build_index_vector (vectype_out, j * nunits_out, 1); |
74bf76ed | 2978 | tree new_var |
0e22bb5a | 2979 | = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_"); |
355fe088 | 2980 | gimple *init_stmt = gimple_build_assign (new_var, cst); |
74bf76ed | 2981 | vect_init_vector_1 (stmt, init_stmt, NULL); |
b731b390 | 2982 | new_temp = make_ssa_name (vec_dest); |
0e22bb5a | 2983 | new_stmt = gimple_build_assign (new_temp, new_var); |
74bf76ed | 2984 | } |
b1b6836e RS |
2985 | else if (modifier == NARROW) |
2986 | { | |
2987 | tree half_res = make_ssa_name (vectype_in); | |
a844293d RS |
2988 | gcall *call = gimple_build_call_internal_vec (ifn, vargs); |
2989 | gimple_call_set_lhs (call, half_res); | |
2990 | gimple_call_set_nothrow (call, true); | |
2991 | new_stmt = call; | |
b1b6836e RS |
2992 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2993 | if ((j & 1) == 0) | |
2994 | { | |
2995 | prev_res = half_res; | |
2996 | continue; | |
2997 | } | |
2998 | new_temp = make_ssa_name (vec_dest); | |
2999 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
3000 | prev_res, half_res); | |
3001 | } | |
74bf76ed JJ |
3002 | else |
3003 | { | |
a844293d | 3004 | gcall *call; |
70439f0d | 3005 | if (ifn != IFN_LAST) |
a844293d | 3006 | call = gimple_build_call_internal_vec (ifn, vargs); |
70439f0d | 3007 | else |
a844293d | 3008 | call = gimple_build_call_vec (fndecl, vargs); |
74bf76ed | 3009 | new_temp = make_ssa_name (vec_dest, new_stmt); |
a844293d RS |
3010 | gimple_call_set_lhs (call, new_temp); |
3011 | gimple_call_set_nothrow (call, true); | |
3012 | new_stmt = call; | |
74bf76ed | 3013 | } |
ebfd146a IR |
3014 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3015 | ||
b1b6836e | 3016 | if (j == (modifier == NARROW ? 1 : 0)) |
ebfd146a IR |
3017 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; |
3018 | else | |
3019 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3020 | ||
3021 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3022 | } | |
b1b6836e RS |
3023 | } |
3024 | else if (modifier == NARROW) | |
3025 | { | |
ebfd146a IR |
3026 | for (j = 0; j < ncopies; ++j) |
3027 | { | |
3028 | /* Build argument list for the vectorized call. */ | |
3029 | if (j == 0) | |
9771b263 | 3030 | vargs.create (nargs * 2); |
ebfd146a | 3031 | else |
9771b263 | 3032 | vargs.truncate (0); |
ebfd146a | 3033 | |
190c2236 JJ |
3034 | if (slp_node) |
3035 | { | |
ef062b13 | 3036 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 3037 | vec<tree> vec_oprnds0; |
190c2236 JJ |
3038 | |
3039 | for (i = 0; i < nargs; i++) | |
9771b263 | 3040 | vargs.quick_push (gimple_call_arg (stmt, i)); |
306b0c92 | 3041 | vect_get_slp_defs (vargs, slp_node, &vec_defs); |
37b5ec8f | 3042 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
3043 | |
3044 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3045 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
3046 | { |
3047 | size_t k; | |
9771b263 | 3048 | vargs.truncate (0); |
190c2236 JJ |
3049 | for (k = 0; k < nargs; k++) |
3050 | { | |
37b5ec8f | 3051 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
3052 | vargs.quick_push (vec_oprndsk[i]); |
3053 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 | 3054 | } |
a844293d | 3055 | gcall *call; |
70439f0d | 3056 | if (ifn != IFN_LAST) |
a844293d | 3057 | call = gimple_build_call_internal_vec (ifn, vargs); |
70439f0d | 3058 | else |
a844293d RS |
3059 | call = gimple_build_call_vec (fndecl, vargs); |
3060 | new_temp = make_ssa_name (vec_dest, call); | |
3061 | gimple_call_set_lhs (call, new_temp); | |
3062 | gimple_call_set_nothrow (call, true); | |
3063 | new_stmt = call; | |
190c2236 | 3064 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 3065 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
3066 | } |
3067 | ||
3068 | for (i = 0; i < nargs; i++) | |
3069 | { | |
37b5ec8f | 3070 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 3071 | vec_oprndsi.release (); |
190c2236 | 3072 | } |
190c2236 JJ |
3073 | continue; |
3074 | } | |
3075 | ||
ebfd146a IR |
3076 | for (i = 0; i < nargs; i++) |
3077 | { | |
3078 | op = gimple_call_arg (stmt, i); | |
3079 | if (j == 0) | |
3080 | { | |
3081 | vec_oprnd0 | |
81c40241 | 3082 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 3083 | vec_oprnd1 |
63827fb8 | 3084 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
3085 | } |
3086 | else | |
3087 | { | |
336ecb65 | 3088 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 3089 | vec_oprnd0 |
63827fb8 | 3090 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 3091 | vec_oprnd1 |
63827fb8 | 3092 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
3093 | } |
3094 | ||
9771b263 DN |
3095 | vargs.quick_push (vec_oprnd0); |
3096 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
3097 | } |
3098 | ||
b1b6836e | 3099 | new_stmt = gimple_build_call_vec (fndecl, vargs); |
ebfd146a IR |
3100 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3101 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
3102 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3103 | ||
3104 | if (j == 0) | |
3105 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3106 | else | |
3107 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3108 | ||
3109 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3110 | } | |
3111 | ||
3112 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a | 3113 | } |
b1b6836e RS |
3114 | else |
3115 | /* No current target implements this case. */ | |
3116 | return false; | |
ebfd146a | 3117 | |
9771b263 | 3118 | vargs.release (); |
ebfd146a | 3119 | |
ebfd146a IR |
3120 | /* The call in STMT might prevent it from being removed in dce. |
3121 | We however cannot remove it here, due to the way the ssa name | |
3122 | it defines is mapped to the new definition. So just replace | |
3123 | rhs of the statement with something harmless. */ | |
3124 | ||
dd34c087 JJ |
3125 | if (slp_node) |
3126 | return true; | |
3127 | ||
ebfd146a | 3128 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
3129 | if (is_pattern_stmt_p (stmt_info)) |
3130 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
3131 | else | |
3132 | lhs = gimple_call_lhs (stmt); | |
3cc2fa2a | 3133 | |
9d5e7640 | 3134 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); |
ebfd146a | 3135 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 3136 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
3137 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
3138 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
3139 | |
3140 | return true; | |
3141 | } | |
3142 | ||
3143 | ||
0136f8f0 AH |
3144 | struct simd_call_arg_info |
3145 | { | |
3146 | tree vectype; | |
3147 | tree op; | |
0136f8f0 | 3148 | HOST_WIDE_INT linear_step; |
34e82342 | 3149 | enum vect_def_type dt; |
0136f8f0 | 3150 | unsigned int align; |
17b658af | 3151 | bool simd_lane_linear; |
0136f8f0 AH |
3152 | }; |
3153 | ||
17b658af JJ |
3154 | /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME, |
3155 | is linear within simd lane (but not within whole loop), note it in | |
3156 | *ARGINFO. */ | |
3157 | ||
3158 | static void | |
3159 | vect_simd_lane_linear (tree op, struct loop *loop, | |
3160 | struct simd_call_arg_info *arginfo) | |
3161 | { | |
355fe088 | 3162 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
17b658af JJ |
3163 | |
3164 | if (!is_gimple_assign (def_stmt) | |
3165 | || gimple_assign_rhs_code (def_stmt) != POINTER_PLUS_EXPR | |
3166 | || !is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))) | |
3167 | return; | |
3168 | ||
3169 | tree base = gimple_assign_rhs1 (def_stmt); | |
3170 | HOST_WIDE_INT linear_step = 0; | |
3171 | tree v = gimple_assign_rhs2 (def_stmt); | |
3172 | while (TREE_CODE (v) == SSA_NAME) | |
3173 | { | |
3174 | tree t; | |
3175 | def_stmt = SSA_NAME_DEF_STMT (v); | |
3176 | if (is_gimple_assign (def_stmt)) | |
3177 | switch (gimple_assign_rhs_code (def_stmt)) | |
3178 | { | |
3179 | case PLUS_EXPR: | |
3180 | t = gimple_assign_rhs2 (def_stmt); | |
3181 | if (linear_step || TREE_CODE (t) != INTEGER_CST) | |
3182 | return; | |
3183 | base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t); | |
3184 | v = gimple_assign_rhs1 (def_stmt); | |
3185 | continue; | |
3186 | case MULT_EXPR: | |
3187 | t = gimple_assign_rhs2 (def_stmt); | |
3188 | if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t)) | |
3189 | return; | |
3190 | linear_step = tree_to_shwi (t); | |
3191 | v = gimple_assign_rhs1 (def_stmt); | |
3192 | continue; | |
3193 | CASE_CONVERT: | |
3194 | t = gimple_assign_rhs1 (def_stmt); | |
3195 | if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE | |
3196 | || (TYPE_PRECISION (TREE_TYPE (v)) | |
3197 | < TYPE_PRECISION (TREE_TYPE (t)))) | |
3198 | return; | |
3199 | if (!linear_step) | |
3200 | linear_step = 1; | |
3201 | v = t; | |
3202 | continue; | |
3203 | default: | |
3204 | return; | |
3205 | } | |
8e4284d0 | 3206 | else if (gimple_call_internal_p (def_stmt, IFN_GOMP_SIMD_LANE) |
17b658af JJ |
3207 | && loop->simduid |
3208 | && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME | |
3209 | && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0)) | |
3210 | == loop->simduid)) | |
3211 | { | |
3212 | if (!linear_step) | |
3213 | linear_step = 1; | |
3214 | arginfo->linear_step = linear_step; | |
3215 | arginfo->op = base; | |
3216 | arginfo->simd_lane_linear = true; | |
3217 | return; | |
3218 | } | |
3219 | } | |
3220 | } | |
3221 | ||
0136f8f0 AH |
3222 | /* Function vectorizable_simd_clone_call. |
3223 | ||
3224 | Check if STMT performs a function call that can be vectorized | |
3225 | by calling a simd clone of the function. | |
3226 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3227 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3228 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3229 | ||
3230 | static bool | |
355fe088 TS |
3231 | vectorizable_simd_clone_call (gimple *stmt, gimple_stmt_iterator *gsi, |
3232 | gimple **vec_stmt, slp_tree slp_node) | |
0136f8f0 AH |
3233 | { |
3234 | tree vec_dest; | |
3235 | tree scalar_dest; | |
3236 | tree op, type; | |
3237 | tree vec_oprnd0 = NULL_TREE; | |
3238 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
3239 | tree vectype; | |
3240 | unsigned int nunits; | |
3241 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3242 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
310213d4 | 3243 | vec_info *vinfo = stmt_info->vinfo; |
0136f8f0 | 3244 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
81c40241 | 3245 | tree fndecl, new_temp; |
355fe088 TS |
3246 | gimple *def_stmt; |
3247 | gimple *new_stmt = NULL; | |
0136f8f0 | 3248 | int ncopies, j; |
00426f9a | 3249 | auto_vec<simd_call_arg_info> arginfo; |
0136f8f0 AH |
3250 | vec<tree> vargs = vNULL; |
3251 | size_t i, nargs; | |
3252 | tree lhs, rtype, ratype; | |
e7a74006 | 3253 | vec<constructor_elt, va_gc> *ret_ctor_elts = NULL; |
0136f8f0 AH |
3254 | |
3255 | /* Is STMT a vectorizable call? */ | |
3256 | if (!is_gimple_call (stmt)) | |
3257 | return false; | |
3258 | ||
3259 | fndecl = gimple_call_fndecl (stmt); | |
3260 | if (fndecl == NULL_TREE) | |
3261 | return false; | |
3262 | ||
d52f5295 | 3263 | struct cgraph_node *node = cgraph_node::get (fndecl); |
0136f8f0 AH |
3264 | if (node == NULL || node->simd_clones == NULL) |
3265 | return false; | |
3266 | ||
3267 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
3268 | return false; | |
3269 | ||
66c16fd9 RB |
3270 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3271 | && ! vec_stmt) | |
0136f8f0 AH |
3272 | return false; |
3273 | ||
3274 | if (gimple_call_lhs (stmt) | |
3275 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
3276 | return false; | |
3277 | ||
3278 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); | |
3279 | ||
3280 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3281 | ||
3282 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt)) | |
3283 | return false; | |
3284 | ||
3285 | /* FORNOW */ | |
fce57248 | 3286 | if (slp_node) |
0136f8f0 AH |
3287 | return false; |
3288 | ||
3289 | /* Process function arguments. */ | |
3290 | nargs = gimple_call_num_args (stmt); | |
3291 | ||
3292 | /* Bail out if the function has zero arguments. */ | |
3293 | if (nargs == 0) | |
3294 | return false; | |
3295 | ||
00426f9a | 3296 | arginfo.reserve (nargs, true); |
0136f8f0 AH |
3297 | |
3298 | for (i = 0; i < nargs; i++) | |
3299 | { | |
3300 | simd_call_arg_info thisarginfo; | |
3301 | affine_iv iv; | |
3302 | ||
3303 | thisarginfo.linear_step = 0; | |
3304 | thisarginfo.align = 0; | |
3305 | thisarginfo.op = NULL_TREE; | |
17b658af | 3306 | thisarginfo.simd_lane_linear = false; |
0136f8f0 AH |
3307 | |
3308 | op = gimple_call_arg (stmt, i); | |
81c40241 RB |
3309 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &thisarginfo.dt, |
3310 | &thisarginfo.vectype) | |
0136f8f0 AH |
3311 | || thisarginfo.dt == vect_uninitialized_def) |
3312 | { | |
3313 | if (dump_enabled_p ()) | |
3314 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3315 | "use not simple.\n"); | |
0136f8f0 AH |
3316 | return false; |
3317 | } | |
3318 | ||
3319 | if (thisarginfo.dt == vect_constant_def | |
3320 | || thisarginfo.dt == vect_external_def) | |
3321 | gcc_assert (thisarginfo.vectype == NULL_TREE); | |
3322 | else | |
3323 | gcc_assert (thisarginfo.vectype != NULL_TREE); | |
3324 | ||
6c9e85fb JJ |
3325 | /* For linear arguments, the analyze phase should have saved |
3326 | the base and step in STMT_VINFO_SIMD_CLONE_INFO. */ | |
17b658af JJ |
3327 | if (i * 3 + 4 <= STMT_VINFO_SIMD_CLONE_INFO (stmt_info).length () |
3328 | && STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]) | |
6c9e85fb JJ |
3329 | { |
3330 | gcc_assert (vec_stmt); | |
3331 | thisarginfo.linear_step | |
17b658af | 3332 | = tree_to_shwi (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]); |
6c9e85fb | 3333 | thisarginfo.op |
17b658af JJ |
3334 | = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 1]; |
3335 | thisarginfo.simd_lane_linear | |
3336 | = (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 3] | |
3337 | == boolean_true_node); | |
6c9e85fb JJ |
3338 | /* If loop has been peeled for alignment, we need to adjust it. */ |
3339 | tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo); | |
3340 | tree n2 = LOOP_VINFO_NITERS (loop_vinfo); | |
17b658af | 3341 | if (n1 != n2 && !thisarginfo.simd_lane_linear) |
6c9e85fb JJ |
3342 | { |
3343 | tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2); | |
17b658af | 3344 | tree step = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]; |
6c9e85fb JJ |
3345 | tree opt = TREE_TYPE (thisarginfo.op); |
3346 | bias = fold_convert (TREE_TYPE (step), bias); | |
3347 | bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step); | |
3348 | thisarginfo.op | |
3349 | = fold_build2 (POINTER_TYPE_P (opt) | |
3350 | ? POINTER_PLUS_EXPR : PLUS_EXPR, opt, | |
3351 | thisarginfo.op, bias); | |
3352 | } | |
3353 | } | |
3354 | else if (!vec_stmt | |
3355 | && thisarginfo.dt != vect_constant_def | |
3356 | && thisarginfo.dt != vect_external_def | |
3357 | && loop_vinfo | |
3358 | && TREE_CODE (op) == SSA_NAME | |
3359 | && simple_iv (loop, loop_containing_stmt (stmt), op, | |
3360 | &iv, false) | |
3361 | && tree_fits_shwi_p (iv.step)) | |
0136f8f0 AH |
3362 | { |
3363 | thisarginfo.linear_step = tree_to_shwi (iv.step); | |
3364 | thisarginfo.op = iv.base; | |
3365 | } | |
3366 | else if ((thisarginfo.dt == vect_constant_def | |
3367 | || thisarginfo.dt == vect_external_def) | |
3368 | && POINTER_TYPE_P (TREE_TYPE (op))) | |
3369 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; | |
17b658af JJ |
3370 | /* Addresses of array elements indexed by GOMP_SIMD_LANE are |
3371 | linear too. */ | |
3372 | if (POINTER_TYPE_P (TREE_TYPE (op)) | |
3373 | && !thisarginfo.linear_step | |
3374 | && !vec_stmt | |
3375 | && thisarginfo.dt != vect_constant_def | |
3376 | && thisarginfo.dt != vect_external_def | |
3377 | && loop_vinfo | |
3378 | && !slp_node | |
3379 | && TREE_CODE (op) == SSA_NAME) | |
3380 | vect_simd_lane_linear (op, loop, &thisarginfo); | |
0136f8f0 AH |
3381 | |
3382 | arginfo.quick_push (thisarginfo); | |
3383 | } | |
3384 | ||
d9f21f6a RS |
3385 | unsigned HOST_WIDE_INT vf; |
3386 | if (!LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (&vf)) | |
3387 | { | |
3388 | if (dump_enabled_p ()) | |
3389 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3390 | "not considering SIMD clones; not yet supported" | |
3391 | " for variable-width vectors.\n"); | |
3392 | return NULL; | |
3393 | } | |
3394 | ||
0136f8f0 AH |
3395 | unsigned int badness = 0; |
3396 | struct cgraph_node *bestn = NULL; | |
6c9e85fb JJ |
3397 | if (STMT_VINFO_SIMD_CLONE_INFO (stmt_info).exists ()) |
3398 | bestn = cgraph_node::get (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[0]); | |
0136f8f0 AH |
3399 | else |
3400 | for (struct cgraph_node *n = node->simd_clones; n != NULL; | |
3401 | n = n->simdclone->next_clone) | |
3402 | { | |
3403 | unsigned int this_badness = 0; | |
d9f21f6a | 3404 | if (n->simdclone->simdlen > vf |
0136f8f0 AH |
3405 | || n->simdclone->nargs != nargs) |
3406 | continue; | |
d9f21f6a RS |
3407 | if (n->simdclone->simdlen < vf) |
3408 | this_badness += (exact_log2 (vf) | |
0136f8f0 AH |
3409 | - exact_log2 (n->simdclone->simdlen)) * 1024; |
3410 | if (n->simdclone->inbranch) | |
3411 | this_badness += 2048; | |
3412 | int target_badness = targetm.simd_clone.usable (n); | |
3413 | if (target_badness < 0) | |
3414 | continue; | |
3415 | this_badness += target_badness * 512; | |
3416 | /* FORNOW: Have to add code to add the mask argument. */ | |
3417 | if (n->simdclone->inbranch) | |
3418 | continue; | |
3419 | for (i = 0; i < nargs; i++) | |
3420 | { | |
3421 | switch (n->simdclone->args[i].arg_type) | |
3422 | { | |
3423 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3424 | if (!useless_type_conversion_p | |
3425 | (n->simdclone->args[i].orig_type, | |
3426 | TREE_TYPE (gimple_call_arg (stmt, i)))) | |
3427 | i = -1; | |
3428 | else if (arginfo[i].dt == vect_constant_def | |
3429 | || arginfo[i].dt == vect_external_def | |
3430 | || arginfo[i].linear_step) | |
3431 | this_badness += 64; | |
3432 | break; | |
3433 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3434 | if (arginfo[i].dt != vect_constant_def | |
3435 | && arginfo[i].dt != vect_external_def) | |
3436 | i = -1; | |
3437 | break; | |
3438 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
d9a6bd32 | 3439 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
3440 | if (arginfo[i].dt == vect_constant_def |
3441 | || arginfo[i].dt == vect_external_def | |
3442 | || (arginfo[i].linear_step | |
3443 | != n->simdclone->args[i].linear_step)) | |
3444 | i = -1; | |
3445 | break; | |
3446 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
d9a6bd32 JJ |
3447 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
3448 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
e01d41e5 JJ |
3449 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3450 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3451 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3452 | /* FORNOW */ |
3453 | i = -1; | |
3454 | break; | |
3455 | case SIMD_CLONE_ARG_TYPE_MASK: | |
3456 | gcc_unreachable (); | |
3457 | } | |
3458 | if (i == (size_t) -1) | |
3459 | break; | |
3460 | if (n->simdclone->args[i].alignment > arginfo[i].align) | |
3461 | { | |
3462 | i = -1; | |
3463 | break; | |
3464 | } | |
3465 | if (arginfo[i].align) | |
3466 | this_badness += (exact_log2 (arginfo[i].align) | |
3467 | - exact_log2 (n->simdclone->args[i].alignment)); | |
3468 | } | |
3469 | if (i == (size_t) -1) | |
3470 | continue; | |
3471 | if (bestn == NULL || this_badness < badness) | |
3472 | { | |
3473 | bestn = n; | |
3474 | badness = this_badness; | |
3475 | } | |
3476 | } | |
3477 | ||
3478 | if (bestn == NULL) | |
00426f9a | 3479 | return false; |
0136f8f0 AH |
3480 | |
3481 | for (i = 0; i < nargs; i++) | |
3482 | if ((arginfo[i].dt == vect_constant_def | |
3483 | || arginfo[i].dt == vect_external_def) | |
3484 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
3485 | { | |
3486 | arginfo[i].vectype | |
3487 | = get_vectype_for_scalar_type (TREE_TYPE (gimple_call_arg (stmt, | |
3488 | i))); | |
3489 | if (arginfo[i].vectype == NULL | |
3490 | || (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
3491 | > bestn->simdclone->simdlen)) | |
00426f9a | 3492 | return false; |
0136f8f0 AH |
3493 | } |
3494 | ||
3495 | fndecl = bestn->decl; | |
3496 | nunits = bestn->simdclone->simdlen; | |
d9f21f6a | 3497 | ncopies = vf / nunits; |
0136f8f0 AH |
3498 | |
3499 | /* If the function isn't const, only allow it in simd loops where user | |
3500 | has asserted that at least nunits consecutive iterations can be | |
3501 | performed using SIMD instructions. */ | |
3502 | if ((loop == NULL || (unsigned) loop->safelen < nunits) | |
3503 | && gimple_vuse (stmt)) | |
00426f9a | 3504 | return false; |
0136f8f0 AH |
3505 | |
3506 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
3507 | needs to be generated. */ | |
3508 | gcc_assert (ncopies >= 1); | |
3509 | ||
3510 | if (!vec_stmt) /* transformation not required. */ | |
3511 | { | |
6c9e85fb JJ |
3512 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (bestn->decl); |
3513 | for (i = 0; i < nargs; i++) | |
7adb26f2 JJ |
3514 | if ((bestn->simdclone->args[i].arg_type |
3515 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
3516 | || (bestn->simdclone->args[i].arg_type | |
3517 | == SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP)) | |
6c9e85fb | 3518 | { |
17b658af | 3519 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_grow_cleared (i * 3 |
6c9e85fb JJ |
3520 | + 1); |
3521 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (arginfo[i].op); | |
3522 | tree lst = POINTER_TYPE_P (TREE_TYPE (arginfo[i].op)) | |
3523 | ? size_type_node : TREE_TYPE (arginfo[i].op); | |
3524 | tree ls = build_int_cst (lst, arginfo[i].linear_step); | |
3525 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (ls); | |
17b658af JJ |
3526 | tree sll = arginfo[i].simd_lane_linear |
3527 | ? boolean_true_node : boolean_false_node; | |
3528 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (sll); | |
6c9e85fb | 3529 | } |
0136f8f0 AH |
3530 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; |
3531 | if (dump_enabled_p ()) | |
3532 | dump_printf_loc (MSG_NOTE, vect_location, | |
3533 | "=== vectorizable_simd_clone_call ===\n"); | |
3534 | /* vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); */ | |
0136f8f0 AH |
3535 | return true; |
3536 | } | |
3537 | ||
67b8dbac | 3538 | /* Transform. */ |
0136f8f0 AH |
3539 | |
3540 | if (dump_enabled_p ()) | |
3541 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); | |
3542 | ||
3543 | /* Handle def. */ | |
3544 | scalar_dest = gimple_call_lhs (stmt); | |
3545 | vec_dest = NULL_TREE; | |
3546 | rtype = NULL_TREE; | |
3547 | ratype = NULL_TREE; | |
3548 | if (scalar_dest) | |
3549 | { | |
3550 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3551 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); | |
3552 | if (TREE_CODE (rtype) == ARRAY_TYPE) | |
3553 | { | |
3554 | ratype = rtype; | |
3555 | rtype = TREE_TYPE (ratype); | |
3556 | } | |
3557 | } | |
3558 | ||
3559 | prev_stmt_info = NULL; | |
3560 | for (j = 0; j < ncopies; ++j) | |
3561 | { | |
3562 | /* Build argument list for the vectorized call. */ | |
3563 | if (j == 0) | |
3564 | vargs.create (nargs); | |
3565 | else | |
3566 | vargs.truncate (0); | |
3567 | ||
3568 | for (i = 0; i < nargs; i++) | |
3569 | { | |
3570 | unsigned int k, l, m, o; | |
3571 | tree atype; | |
3572 | op = gimple_call_arg (stmt, i); | |
3573 | switch (bestn->simdclone->args[i].arg_type) | |
3574 | { | |
3575 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3576 | atype = bestn->simdclone->args[i].vector_type; | |
3577 | o = nunits / TYPE_VECTOR_SUBPARTS (atype); | |
3578 | for (m = j * o; m < (j + 1) * o; m++) | |
3579 | { | |
3580 | if (TYPE_VECTOR_SUBPARTS (atype) | |
3581 | < TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)) | |
3582 | { | |
3583 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); | |
3584 | k = (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
3585 | / TYPE_VECTOR_SUBPARTS (atype)); | |
3586 | gcc_assert ((k & (k - 1)) == 0); | |
3587 | if (m == 0) | |
3588 | vec_oprnd0 | |
81c40241 | 3589 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3590 | else |
3591 | { | |
3592 | vec_oprnd0 = arginfo[i].op; | |
3593 | if ((m & (k - 1)) == 0) | |
3594 | vec_oprnd0 | |
3595 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3596 | vec_oprnd0); | |
3597 | } | |
3598 | arginfo[i].op = vec_oprnd0; | |
3599 | vec_oprnd0 | |
3600 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, | |
92e29a5e | 3601 | bitsize_int (prec), |
0136f8f0 AH |
3602 | bitsize_int ((m & (k - 1)) * prec)); |
3603 | new_stmt | |
b731b390 | 3604 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3605 | vec_oprnd0); |
3606 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3607 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3608 | } | |
3609 | else | |
3610 | { | |
3611 | k = (TYPE_VECTOR_SUBPARTS (atype) | |
3612 | / TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)); | |
3613 | gcc_assert ((k & (k - 1)) == 0); | |
3614 | vec<constructor_elt, va_gc> *ctor_elts; | |
3615 | if (k != 1) | |
3616 | vec_alloc (ctor_elts, k); | |
3617 | else | |
3618 | ctor_elts = NULL; | |
3619 | for (l = 0; l < k; l++) | |
3620 | { | |
3621 | if (m == 0 && l == 0) | |
3622 | vec_oprnd0 | |
81c40241 | 3623 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3624 | else |
3625 | vec_oprnd0 | |
3626 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3627 | arginfo[i].op); | |
3628 | arginfo[i].op = vec_oprnd0; | |
3629 | if (k == 1) | |
3630 | break; | |
3631 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, | |
3632 | vec_oprnd0); | |
3633 | } | |
3634 | if (k == 1) | |
3635 | vargs.safe_push (vec_oprnd0); | |
3636 | else | |
3637 | { | |
3638 | vec_oprnd0 = build_constructor (atype, ctor_elts); | |
3639 | new_stmt | |
b731b390 | 3640 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3641 | vec_oprnd0); |
3642 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3643 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3644 | } | |
3645 | } | |
3646 | } | |
3647 | break; | |
3648 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3649 | vargs.safe_push (op); | |
3650 | break; | |
3651 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
7adb26f2 | 3652 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
3653 | if (j == 0) |
3654 | { | |
3655 | gimple_seq stmts; | |
3656 | arginfo[i].op | |
3657 | = force_gimple_operand (arginfo[i].op, &stmts, true, | |
3658 | NULL_TREE); | |
3659 | if (stmts != NULL) | |
3660 | { | |
3661 | basic_block new_bb; | |
3662 | edge pe = loop_preheader_edge (loop); | |
3663 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
3664 | gcc_assert (!new_bb); | |
3665 | } | |
17b658af JJ |
3666 | if (arginfo[i].simd_lane_linear) |
3667 | { | |
3668 | vargs.safe_push (arginfo[i].op); | |
3669 | break; | |
3670 | } | |
b731b390 | 3671 | tree phi_res = copy_ssa_name (op); |
538dd0b7 | 3672 | gphi *new_phi = create_phi_node (phi_res, loop->header); |
0136f8f0 | 3673 | set_vinfo_for_stmt (new_phi, |
310213d4 | 3674 | new_stmt_vec_info (new_phi, loop_vinfo)); |
0136f8f0 AH |
3675 | add_phi_arg (new_phi, arginfo[i].op, |
3676 | loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
3677 | enum tree_code code | |
3678 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3679 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3680 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3681 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3682 | widest_int cst |
3683 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3684 | ncopies * nunits); | |
3685 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3686 | tree phi_arg = copy_ssa_name (op); |
0d0e4a03 JJ |
3687 | new_stmt |
3688 | = gimple_build_assign (phi_arg, code, phi_res, tcst); | |
0136f8f0 AH |
3689 | gimple_stmt_iterator si = gsi_after_labels (loop->header); |
3690 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); | |
3691 | set_vinfo_for_stmt (new_stmt, | |
310213d4 | 3692 | new_stmt_vec_info (new_stmt, loop_vinfo)); |
0136f8f0 AH |
3693 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), |
3694 | UNKNOWN_LOCATION); | |
3695 | arginfo[i].op = phi_res; | |
3696 | vargs.safe_push (phi_res); | |
3697 | } | |
3698 | else | |
3699 | { | |
3700 | enum tree_code code | |
3701 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3702 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3703 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3704 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3705 | widest_int cst |
3706 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3707 | j * nunits); | |
3708 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3709 | new_temp = make_ssa_name (TREE_TYPE (op)); |
0d0e4a03 JJ |
3710 | new_stmt = gimple_build_assign (new_temp, code, |
3711 | arginfo[i].op, tcst); | |
0136f8f0 AH |
3712 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3713 | vargs.safe_push (new_temp); | |
3714 | } | |
3715 | break; | |
7adb26f2 JJ |
3716 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
3717 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
0136f8f0 | 3718 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: |
e01d41e5 JJ |
3719 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3720 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3721 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3722 | default: |
3723 | gcc_unreachable (); | |
3724 | } | |
3725 | } | |
3726 | ||
3727 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
3728 | if (vec_dest) | |
3729 | { | |
3730 | gcc_assert (ratype || TYPE_VECTOR_SUBPARTS (rtype) == nunits); | |
3731 | if (ratype) | |
b731b390 | 3732 | new_temp = create_tmp_var (ratype); |
0136f8f0 AH |
3733 | else if (TYPE_VECTOR_SUBPARTS (vectype) |
3734 | == TYPE_VECTOR_SUBPARTS (rtype)) | |
3735 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3736 | else | |
3737 | new_temp = make_ssa_name (rtype, new_stmt); | |
3738 | gimple_call_set_lhs (new_stmt, new_temp); | |
3739 | } | |
3740 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3741 | ||
3742 | if (vec_dest) | |
3743 | { | |
3744 | if (TYPE_VECTOR_SUBPARTS (vectype) < nunits) | |
3745 | { | |
3746 | unsigned int k, l; | |
3747 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); | |
3748 | k = nunits / TYPE_VECTOR_SUBPARTS (vectype); | |
3749 | gcc_assert ((k & (k - 1)) == 0); | |
3750 | for (l = 0; l < k; l++) | |
3751 | { | |
3752 | tree t; | |
3753 | if (ratype) | |
3754 | { | |
3755 | t = build_fold_addr_expr (new_temp); | |
3756 | t = build2 (MEM_REF, vectype, t, | |
3757 | build_int_cst (TREE_TYPE (t), | |
3758 | l * prec / BITS_PER_UNIT)); | |
3759 | } | |
3760 | else | |
3761 | t = build3 (BIT_FIELD_REF, vectype, new_temp, | |
92e29a5e | 3762 | bitsize_int (prec), bitsize_int (l * prec)); |
0136f8f0 | 3763 | new_stmt |
b731b390 | 3764 | = gimple_build_assign (make_ssa_name (vectype), t); |
0136f8f0 AH |
3765 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3766 | if (j == 0 && l == 0) | |
3767 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3768 | else | |
3769 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3770 | ||
3771 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3772 | } | |
3773 | ||
3774 | if (ratype) | |
3775 | { | |
3776 | tree clobber = build_constructor (ratype, NULL); | |
3777 | TREE_THIS_VOLATILE (clobber) = 1; | |
3778 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3779 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3780 | } | |
3781 | continue; | |
3782 | } | |
3783 | else if (TYPE_VECTOR_SUBPARTS (vectype) > nunits) | |
3784 | { | |
3785 | unsigned int k = (TYPE_VECTOR_SUBPARTS (vectype) | |
3786 | / TYPE_VECTOR_SUBPARTS (rtype)); | |
3787 | gcc_assert ((k & (k - 1)) == 0); | |
3788 | if ((j & (k - 1)) == 0) | |
3789 | vec_alloc (ret_ctor_elts, k); | |
3790 | if (ratype) | |
3791 | { | |
3792 | unsigned int m, o = nunits / TYPE_VECTOR_SUBPARTS (rtype); | |
3793 | for (m = 0; m < o; m++) | |
3794 | { | |
3795 | tree tem = build4 (ARRAY_REF, rtype, new_temp, | |
3796 | size_int (m), NULL_TREE, NULL_TREE); | |
3797 | new_stmt | |
b731b390 | 3798 | = gimple_build_assign (make_ssa_name (rtype), tem); |
0136f8f0 AH |
3799 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3800 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, | |
3801 | gimple_assign_lhs (new_stmt)); | |
3802 | } | |
3803 | tree clobber = build_constructor (ratype, NULL); | |
3804 | TREE_THIS_VOLATILE (clobber) = 1; | |
3805 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3806 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3807 | } | |
3808 | else | |
3809 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); | |
3810 | if ((j & (k - 1)) != k - 1) | |
3811 | continue; | |
3812 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); | |
3813 | new_stmt | |
b731b390 | 3814 | = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0); |
0136f8f0 AH |
3815 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3816 | ||
3817 | if ((unsigned) j == k - 1) | |
3818 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3819 | else | |
3820 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3821 | ||
3822 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3823 | continue; | |
3824 | } | |
3825 | else if (ratype) | |
3826 | { | |
3827 | tree t = build_fold_addr_expr (new_temp); | |
3828 | t = build2 (MEM_REF, vectype, t, | |
3829 | build_int_cst (TREE_TYPE (t), 0)); | |
3830 | new_stmt | |
b731b390 | 3831 | = gimple_build_assign (make_ssa_name (vec_dest), t); |
0136f8f0 AH |
3832 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3833 | tree clobber = build_constructor (ratype, NULL); | |
3834 | TREE_THIS_VOLATILE (clobber) = 1; | |
3835 | vect_finish_stmt_generation (stmt, | |
3836 | gimple_build_assign (new_temp, | |
3837 | clobber), gsi); | |
3838 | } | |
3839 | } | |
3840 | ||
3841 | if (j == 0) | |
3842 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3843 | else | |
3844 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3845 | ||
3846 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3847 | } | |
3848 | ||
3849 | vargs.release (); | |
3850 | ||
3851 | /* The call in STMT might prevent it from being removed in dce. | |
3852 | We however cannot remove it here, due to the way the ssa name | |
3853 | it defines is mapped to the new definition. So just replace | |
3854 | rhs of the statement with something harmless. */ | |
3855 | ||
3856 | if (slp_node) | |
3857 | return true; | |
3858 | ||
3859 | if (scalar_dest) | |
3860 | { | |
3861 | type = TREE_TYPE (scalar_dest); | |
3862 | if (is_pattern_stmt_p (stmt_info)) | |
3863 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
3864 | else | |
3865 | lhs = gimple_call_lhs (stmt); | |
3866 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
3867 | } | |
3868 | else | |
3869 | new_stmt = gimple_build_nop (); | |
3870 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
3871 | set_vinfo_for_stmt (stmt, NULL); | |
3872 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2865f32a | 3873 | gsi_replace (gsi, new_stmt, true); |
0136f8f0 AH |
3874 | unlink_stmt_vdef (stmt); |
3875 | ||
3876 | return true; | |
3877 | } | |
3878 | ||
3879 | ||
ebfd146a IR |
3880 | /* Function vect_gen_widened_results_half |
3881 | ||
3882 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 3883 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 3884 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
3885 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
3886 | needs to be created (DECL is a function-decl of a target-builtin). | |
3887 | STMT is the original scalar stmt that we are vectorizing. */ | |
3888 | ||
355fe088 | 3889 | static gimple * |
ebfd146a IR |
3890 | vect_gen_widened_results_half (enum tree_code code, |
3891 | tree decl, | |
3892 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
3893 | tree vec_dest, gimple_stmt_iterator *gsi, | |
355fe088 | 3894 | gimple *stmt) |
b8698a0f | 3895 | { |
355fe088 | 3896 | gimple *new_stmt; |
b8698a0f L |
3897 | tree new_temp; |
3898 | ||
3899 | /* Generate half of the widened result: */ | |
3900 | if (code == CALL_EXPR) | |
3901 | { | |
3902 | /* Target specific support */ | |
ebfd146a IR |
3903 | if (op_type == binary_op) |
3904 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
3905 | else | |
3906 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
3907 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3908 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
3909 | } |
3910 | else | |
ebfd146a | 3911 | { |
b8698a0f L |
3912 | /* Generic support */ |
3913 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
3914 | if (op_type != binary_op) |
3915 | vec_oprnd1 = NULL; | |
0d0e4a03 | 3916 | new_stmt = gimple_build_assign (vec_dest, code, vec_oprnd0, vec_oprnd1); |
ebfd146a IR |
3917 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3918 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 3919 | } |
ebfd146a IR |
3920 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3921 | ||
ebfd146a IR |
3922 | return new_stmt; |
3923 | } | |
3924 | ||
4a00c761 JJ |
3925 | |
3926 | /* Get vectorized definitions for loop-based vectorization. For the first | |
3927 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
3928 | scalar operand), and for the rest we get a copy with | |
3929 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
3930 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
3931 | The vectors are collected into VEC_OPRNDS. */ | |
3932 | ||
3933 | static void | |
355fe088 | 3934 | vect_get_loop_based_defs (tree *oprnd, gimple *stmt, enum vect_def_type dt, |
9771b263 | 3935 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
3936 | { |
3937 | tree vec_oprnd; | |
3938 | ||
3939 | /* Get first vector operand. */ | |
3940 | /* All the vector operands except the very first one (that is scalar oprnd) | |
3941 | are stmt copies. */ | |
3942 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
81c40241 | 3943 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt); |
4a00c761 JJ |
3944 | else |
3945 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
3946 | ||
9771b263 | 3947 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3948 | |
3949 | /* Get second vector operand. */ | |
3950 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 3951 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3952 | |
3953 | *oprnd = vec_oprnd; | |
3954 | ||
3955 | /* For conversion in multiple steps, continue to get operands | |
3956 | recursively. */ | |
3957 | if (multi_step_cvt) | |
3958 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
3959 | } | |
3960 | ||
3961 | ||
3962 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
3963 | For multi-step conversions store the resulting vectors and call the function | |
3964 | recursively. */ | |
3965 | ||
3966 | static void | |
9771b263 | 3967 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
355fe088 | 3968 | int multi_step_cvt, gimple *stmt, |
9771b263 | 3969 | vec<tree> vec_dsts, |
4a00c761 JJ |
3970 | gimple_stmt_iterator *gsi, |
3971 | slp_tree slp_node, enum tree_code code, | |
3972 | stmt_vec_info *prev_stmt_info) | |
3973 | { | |
3974 | unsigned int i; | |
3975 | tree vop0, vop1, new_tmp, vec_dest; | |
355fe088 | 3976 | gimple *new_stmt; |
4a00c761 JJ |
3977 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
3978 | ||
9771b263 | 3979 | vec_dest = vec_dsts.pop (); |
4a00c761 | 3980 | |
9771b263 | 3981 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
3982 | { |
3983 | /* Create demotion operation. */ | |
9771b263 DN |
3984 | vop0 = (*vec_oprnds)[i]; |
3985 | vop1 = (*vec_oprnds)[i + 1]; | |
0d0e4a03 | 3986 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
4a00c761 JJ |
3987 | new_tmp = make_ssa_name (vec_dest, new_stmt); |
3988 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
3989 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3990 | ||
3991 | if (multi_step_cvt) | |
3992 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 3993 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
3994 | else |
3995 | { | |
3996 | /* This is the last step of the conversion sequence. Store the | |
3997 | vectors in SLP_NODE or in vector info of the scalar statement | |
3998 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
3999 | if (slp_node) | |
9771b263 | 4000 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 4001 | else |
c689ce1e RB |
4002 | { |
4003 | if (!*prev_stmt_info) | |
4004 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
4005 | else | |
4006 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
4a00c761 | 4007 | |
c689ce1e RB |
4008 | *prev_stmt_info = vinfo_for_stmt (new_stmt); |
4009 | } | |
4a00c761 JJ |
4010 | } |
4011 | } | |
4012 | ||
4013 | /* For multi-step demotion operations we first generate demotion operations | |
4014 | from the source type to the intermediate types, and then combine the | |
4015 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
4016 | type. */ | |
4017 | if (multi_step_cvt) | |
4018 | { | |
4019 | /* At each level of recursion we have half of the operands we had at the | |
4020 | previous level. */ | |
9771b263 | 4021 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
4022 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
4023 | stmt, vec_dsts, gsi, slp_node, | |
4024 | VEC_PACK_TRUNC_EXPR, | |
4025 | prev_stmt_info); | |
4026 | } | |
4027 | ||
9771b263 | 4028 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4029 | } |
4030 | ||
4031 | ||
4032 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
4033 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
4034 | the resulting vectors and call the function recursively. */ | |
4035 | ||
4036 | static void | |
9771b263 DN |
4037 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
4038 | vec<tree> *vec_oprnds1, | |
355fe088 | 4039 | gimple *stmt, tree vec_dest, |
4a00c761 JJ |
4040 | gimple_stmt_iterator *gsi, |
4041 | enum tree_code code1, | |
4042 | enum tree_code code2, tree decl1, | |
4043 | tree decl2, int op_type) | |
4044 | { | |
4045 | int i; | |
4046 | tree vop0, vop1, new_tmp1, new_tmp2; | |
355fe088 | 4047 | gimple *new_stmt1, *new_stmt2; |
6e1aa848 | 4048 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 4049 | |
9771b263 DN |
4050 | vec_tmp.create (vec_oprnds0->length () * 2); |
4051 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
4052 | { |
4053 | if (op_type == binary_op) | |
9771b263 | 4054 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
4055 | else |
4056 | vop1 = NULL_TREE; | |
4057 | ||
4058 | /* Generate the two halves of promotion operation. */ | |
4059 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
4060 | op_type, vec_dest, gsi, stmt); | |
4061 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
4062 | op_type, vec_dest, gsi, stmt); | |
4063 | if (is_gimple_call (new_stmt1)) | |
4064 | { | |
4065 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
4066 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
4067 | } | |
4068 | else | |
4069 | { | |
4070 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
4071 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
4072 | } | |
4073 | ||
4074 | /* Store the results for the next step. */ | |
9771b263 DN |
4075 | vec_tmp.quick_push (new_tmp1); |
4076 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
4077 | } |
4078 | ||
689eaba3 | 4079 | vec_oprnds0->release (); |
4a00c761 JJ |
4080 | *vec_oprnds0 = vec_tmp; |
4081 | } | |
4082 | ||
4083 | ||
b8698a0f L |
4084 | /* Check if STMT performs a conversion operation, that can be vectorized. |
4085 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 4086 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
4087 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
4088 | ||
4089 | static bool | |
355fe088 TS |
4090 | vectorizable_conversion (gimple *stmt, gimple_stmt_iterator *gsi, |
4091 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
4092 | { |
4093 | tree vec_dest; | |
4094 | tree scalar_dest; | |
4a00c761 | 4095 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
4096 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
4097 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4098 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4099 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 4100 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
4101 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
4102 | tree new_temp; | |
355fe088 | 4103 | gimple *def_stmt; |
ebfd146a | 4104 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 4105 | int ndts = 2; |
355fe088 | 4106 | gimple *new_stmt = NULL; |
ebfd146a IR |
4107 | stmt_vec_info prev_stmt_info; |
4108 | int nunits_in; | |
4109 | int nunits_out; | |
4110 | tree vectype_out, vectype_in; | |
4a00c761 JJ |
4111 | int ncopies, i, j; |
4112 | tree lhs_type, rhs_type; | |
ebfd146a | 4113 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
4114 | vec<tree> vec_oprnds0 = vNULL; |
4115 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 4116 | tree vop0; |
4a00c761 | 4117 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4118 | vec_info *vinfo = stmt_info->vinfo; |
4a00c761 | 4119 | int multi_step_cvt = 0; |
6e1aa848 | 4120 | vec<tree> interm_types = vNULL; |
4a00c761 JJ |
4121 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
4122 | int op_type; | |
4a00c761 | 4123 | unsigned short fltsz; |
ebfd146a IR |
4124 | |
4125 | /* Is STMT a vectorizable conversion? */ | |
4126 | ||
4a00c761 | 4127 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4128 | return false; |
4129 | ||
66c16fd9 RB |
4130 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4131 | && ! vec_stmt) | |
ebfd146a IR |
4132 | return false; |
4133 | ||
4134 | if (!is_gimple_assign (stmt)) | |
4135 | return false; | |
4136 | ||
4137 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4138 | return false; | |
4139 | ||
4140 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
4141 | if (!CONVERT_EXPR_CODE_P (code) |
4142 | && code != FIX_TRUNC_EXPR | |
4143 | && code != FLOAT_EXPR | |
4144 | && code != WIDEN_MULT_EXPR | |
4145 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
4146 | return false; |
4147 | ||
4a00c761 JJ |
4148 | op_type = TREE_CODE_LENGTH (code); |
4149 | ||
ebfd146a | 4150 | /* Check types of lhs and rhs. */ |
b690cc0f | 4151 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 4152 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
4153 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
4154 | ||
ebfd146a IR |
4155 | op0 = gimple_assign_rhs1 (stmt); |
4156 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
4157 | |
4158 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4159 | && !((INTEGRAL_TYPE_P (lhs_type) | |
4160 | && INTEGRAL_TYPE_P (rhs_type)) | |
4161 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
4162 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
4163 | return false; | |
4164 | ||
e6f5c25d IE |
4165 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4166 | && ((INTEGRAL_TYPE_P (lhs_type) | |
2be65d9e | 4167 | && !type_has_mode_precision_p (lhs_type)) |
e6f5c25d | 4168 | || (INTEGRAL_TYPE_P (rhs_type) |
2be65d9e | 4169 | && !type_has_mode_precision_p (rhs_type)))) |
4a00c761 | 4170 | { |
73fbfcad | 4171 | if (dump_enabled_p ()) |
78c60e3d | 4172 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
4173 | "type conversion to/from bit-precision unsupported." |
4174 | "\n"); | |
4a00c761 JJ |
4175 | return false; |
4176 | } | |
4177 | ||
b690cc0f | 4178 | /* Check the operands of the operation. */ |
81c40241 | 4179 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype_in)) |
b690cc0f | 4180 | { |
73fbfcad | 4181 | if (dump_enabled_p ()) |
78c60e3d | 4182 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4183 | "use not simple.\n"); |
b690cc0f RG |
4184 | return false; |
4185 | } | |
4a00c761 JJ |
4186 | if (op_type == binary_op) |
4187 | { | |
4188 | bool ok; | |
4189 | ||
4190 | op1 = gimple_assign_rhs2 (stmt); | |
4191 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
4192 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
4193 | OP1. */ | |
4194 | if (CONSTANT_CLASS_P (op0)) | |
81c40241 | 4195 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &vectype_in); |
4a00c761 | 4196 | else |
81c40241 | 4197 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1]); |
4a00c761 JJ |
4198 | |
4199 | if (!ok) | |
4200 | { | |
73fbfcad | 4201 | if (dump_enabled_p ()) |
78c60e3d | 4202 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4203 | "use not simple.\n"); |
4a00c761 JJ |
4204 | return false; |
4205 | } | |
4206 | } | |
4207 | ||
b690cc0f RG |
4208 | /* If op0 is an external or constant defs use a vector type of |
4209 | the same size as the output vector type. */ | |
ebfd146a | 4210 | if (!vectype_in) |
b690cc0f | 4211 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
4212 | if (vec_stmt) |
4213 | gcc_assert (vectype_in); | |
4214 | if (!vectype_in) | |
4215 | { | |
73fbfcad | 4216 | if (dump_enabled_p ()) |
4a00c761 | 4217 | { |
78c60e3d SS |
4218 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4219 | "no vectype for scalar type "); | |
4220 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 4221 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 4222 | } |
7d8930a0 IR |
4223 | |
4224 | return false; | |
4225 | } | |
ebfd146a | 4226 | |
e6f5c25d IE |
4227 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4228 | && !VECTOR_BOOLEAN_TYPE_P (vectype_in)) | |
4229 | { | |
4230 | if (dump_enabled_p ()) | |
4231 | { | |
4232 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
4233 | "can't convert between boolean and non " | |
4234 | "boolean vectors"); | |
4235 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
4236 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
4237 | } | |
4238 | ||
4239 | return false; | |
4240 | } | |
4241 | ||
b690cc0f RG |
4242 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
4243 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4a00c761 | 4244 | if (nunits_in < nunits_out) |
ebfd146a IR |
4245 | modifier = NARROW; |
4246 | else if (nunits_out == nunits_in) | |
4247 | modifier = NONE; | |
ebfd146a | 4248 | else |
4a00c761 | 4249 | modifier = WIDEN; |
ebfd146a | 4250 | |
ff802fa1 IR |
4251 | /* Multiple types in SLP are handled by creating the appropriate number of |
4252 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4253 | case of SLP. */ | |
fce57248 | 4254 | if (slp_node) |
ebfd146a | 4255 | ncopies = 1; |
4a00c761 | 4256 | else if (modifier == NARROW) |
e8f142e2 | 4257 | ncopies = vect_get_num_copies (loop_vinfo, vectype_out); |
4a00c761 | 4258 | else |
e8f142e2 | 4259 | ncopies = vect_get_num_copies (loop_vinfo, vectype_in); |
b8698a0f | 4260 | |
ebfd146a IR |
4261 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
4262 | needs to be generated. */ | |
4263 | gcc_assert (ncopies >= 1); | |
4264 | ||
16d22000 RS |
4265 | bool found_mode = false; |
4266 | scalar_mode lhs_mode = SCALAR_TYPE_MODE (lhs_type); | |
4267 | scalar_mode rhs_mode = SCALAR_TYPE_MODE (rhs_type); | |
4268 | opt_scalar_mode rhs_mode_iter; | |
b397965c | 4269 | |
ebfd146a | 4270 | /* Supportable by target? */ |
4a00c761 | 4271 | switch (modifier) |
ebfd146a | 4272 | { |
4a00c761 JJ |
4273 | case NONE: |
4274 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4275 | return false; | |
4276 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
4277 | &decl1, &code1)) | |
4278 | break; | |
4279 | /* FALLTHRU */ | |
4280 | unsupported: | |
73fbfcad | 4281 | if (dump_enabled_p ()) |
78c60e3d | 4282 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4283 | "conversion not supported by target.\n"); |
ebfd146a | 4284 | return false; |
ebfd146a | 4285 | |
4a00c761 JJ |
4286 | case WIDEN: |
4287 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
4288 | &code1, &code2, &multi_step_cvt, |
4289 | &interm_types)) | |
4a00c761 JJ |
4290 | { |
4291 | /* Binary widening operation can only be supported directly by the | |
4292 | architecture. */ | |
4293 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
4294 | break; | |
4295 | } | |
4296 | ||
4297 | if (code != FLOAT_EXPR | |
b397965c | 4298 | || GET_MODE_SIZE (lhs_mode) <= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4299 | goto unsupported; |
4300 | ||
b397965c | 4301 | fltsz = GET_MODE_SIZE (lhs_mode); |
16d22000 | 4302 | FOR_EACH_2XWIDER_MODE (rhs_mode_iter, rhs_mode) |
4a00c761 | 4303 | { |
16d22000 | 4304 | rhs_mode = rhs_mode_iter.require (); |
c94843d2 RS |
4305 | if (GET_MODE_SIZE (rhs_mode) > fltsz) |
4306 | break; | |
4307 | ||
4a00c761 JJ |
4308 | cvt_type |
4309 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4310 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4311 | if (cvt_type == NULL_TREE) | |
4312 | goto unsupported; | |
4313 | ||
4314 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4315 | { | |
4316 | if (!supportable_convert_operation (code, vectype_out, | |
4317 | cvt_type, &decl1, &codecvt1)) | |
4318 | goto unsupported; | |
4319 | } | |
4320 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
4321 | cvt_type, &codecvt1, |
4322 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
4323 | &interm_types)) |
4324 | continue; | |
4325 | else | |
4326 | gcc_assert (multi_step_cvt == 0); | |
4327 | ||
4328 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
4329 | vectype_in, &code1, &code2, |
4330 | &multi_step_cvt, &interm_types)) | |
16d22000 RS |
4331 | { |
4332 | found_mode = true; | |
4333 | break; | |
4334 | } | |
4a00c761 JJ |
4335 | } |
4336 | ||
16d22000 | 4337 | if (!found_mode) |
4a00c761 JJ |
4338 | goto unsupported; |
4339 | ||
4340 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4341 | codecvt2 = ERROR_MARK; | |
4342 | else | |
4343 | { | |
4344 | multi_step_cvt++; | |
9771b263 | 4345 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
4346 | cvt_type = NULL_TREE; |
4347 | } | |
4348 | break; | |
4349 | ||
4350 | case NARROW: | |
4351 | gcc_assert (op_type == unary_op); | |
4352 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
4353 | &code1, &multi_step_cvt, | |
4354 | &interm_types)) | |
4355 | break; | |
4356 | ||
4357 | if (code != FIX_TRUNC_EXPR | |
b397965c | 4358 | || GET_MODE_SIZE (lhs_mode) >= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4359 | goto unsupported; |
4360 | ||
4a00c761 JJ |
4361 | cvt_type |
4362 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4363 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4364 | if (cvt_type == NULL_TREE) | |
4365 | goto unsupported; | |
4366 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
4367 | &decl1, &codecvt1)) | |
4368 | goto unsupported; | |
4369 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
4370 | &code1, &multi_step_cvt, | |
4371 | &interm_types)) | |
4372 | break; | |
4373 | goto unsupported; | |
4374 | ||
4375 | default: | |
4376 | gcc_unreachable (); | |
ebfd146a IR |
4377 | } |
4378 | ||
4379 | if (!vec_stmt) /* transformation not required. */ | |
4380 | { | |
73fbfcad | 4381 | if (dump_enabled_p ()) |
78c60e3d | 4382 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4383 | "=== vectorizable_conversion ===\n"); |
4a00c761 | 4384 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
4385 | { |
4386 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
4fc5ebf1 | 4387 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); |
8bd37302 | 4388 | } |
4a00c761 JJ |
4389 | else if (modifier == NARROW) |
4390 | { | |
4391 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
8bd37302 | 4392 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 JJ |
4393 | } |
4394 | else | |
4395 | { | |
4396 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
8bd37302 | 4397 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 | 4398 | } |
9771b263 | 4399 | interm_types.release (); |
ebfd146a IR |
4400 | return true; |
4401 | } | |
4402 | ||
67b8dbac | 4403 | /* Transform. */ |
73fbfcad | 4404 | if (dump_enabled_p ()) |
78c60e3d | 4405 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4406 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 4407 | |
4a00c761 JJ |
4408 | if (op_type == binary_op) |
4409 | { | |
4410 | if (CONSTANT_CLASS_P (op0)) | |
4411 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
4412 | else if (CONSTANT_CLASS_P (op1)) | |
4413 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
4414 | } | |
4415 | ||
4416 | /* In case of multi-step conversion, we first generate conversion operations | |
4417 | to the intermediate types, and then from that types to the final one. | |
4418 | We create vector destinations for the intermediate type (TYPES) received | |
4419 | from supportable_*_operation, and store them in the correct order | |
4420 | for future use in vect_create_vectorized_*_stmts (). */ | |
8c681247 | 4421 | auto_vec<tree> vec_dsts (multi_step_cvt + 1); |
82294ec1 JJ |
4422 | vec_dest = vect_create_destination_var (scalar_dest, |
4423 | (cvt_type && modifier == WIDEN) | |
4424 | ? cvt_type : vectype_out); | |
9771b263 | 4425 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4426 | |
4427 | if (multi_step_cvt) | |
4428 | { | |
9771b263 DN |
4429 | for (i = interm_types.length () - 1; |
4430 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
4431 | { |
4432 | vec_dest = vect_create_destination_var (scalar_dest, | |
4433 | intermediate_type); | |
9771b263 | 4434 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4435 | } |
4436 | } | |
ebfd146a | 4437 | |
4a00c761 | 4438 | if (cvt_type) |
82294ec1 JJ |
4439 | vec_dest = vect_create_destination_var (scalar_dest, |
4440 | modifier == WIDEN | |
4441 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
4442 | |
4443 | if (!slp_node) | |
4444 | { | |
30862efc | 4445 | if (modifier == WIDEN) |
4a00c761 | 4446 | { |
c3284718 | 4447 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 4448 | if (op_type == binary_op) |
9771b263 | 4449 | vec_oprnds1.create (1); |
4a00c761 | 4450 | } |
30862efc | 4451 | else if (modifier == NARROW) |
9771b263 DN |
4452 | vec_oprnds0.create ( |
4453 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
4454 | } |
4455 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 4456 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 4457 | |
4a00c761 | 4458 | last_oprnd = op0; |
ebfd146a IR |
4459 | prev_stmt_info = NULL; |
4460 | switch (modifier) | |
4461 | { | |
4462 | case NONE: | |
4463 | for (j = 0; j < ncopies; j++) | |
4464 | { | |
ebfd146a | 4465 | if (j == 0) |
306b0c92 | 4466 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node); |
ebfd146a IR |
4467 | else |
4468 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
4469 | ||
9771b263 | 4470 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4471 | { |
4472 | /* Arguments are ready, create the new vector stmt. */ | |
4473 | if (code1 == CALL_EXPR) | |
4474 | { | |
4475 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4476 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4477 | gimple_call_set_lhs (new_stmt, new_temp); | |
4478 | } | |
4479 | else | |
4480 | { | |
4481 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
0d0e4a03 | 4482 | new_stmt = gimple_build_assign (vec_dest, code1, vop0); |
4a00c761 JJ |
4483 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4484 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4485 | } | |
4486 | ||
4487 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4488 | if (slp_node) | |
9771b263 | 4489 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
225ce44b RB |
4490 | else |
4491 | { | |
4492 | if (!prev_stmt_info) | |
4493 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4494 | else | |
4495 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4496 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4497 | } | |
4a00c761 | 4498 | } |
ebfd146a IR |
4499 | } |
4500 | break; | |
4501 | ||
4502 | case WIDEN: | |
4503 | /* In case the vectorization factor (VF) is bigger than the number | |
4504 | of elements that we can fit in a vectype (nunits), we have to | |
4505 | generate more than one vector stmt - i.e - we need to "unroll" | |
4506 | the vector stmt by a factor VF/nunits. */ | |
4507 | for (j = 0; j < ncopies; j++) | |
4508 | { | |
4a00c761 | 4509 | /* Handle uses. */ |
ebfd146a | 4510 | if (j == 0) |
4a00c761 JJ |
4511 | { |
4512 | if (slp_node) | |
4513 | { | |
4514 | if (code == WIDEN_LSHIFT_EXPR) | |
4515 | { | |
4516 | unsigned int k; | |
ebfd146a | 4517 | |
4a00c761 JJ |
4518 | vec_oprnd1 = op1; |
4519 | /* Store vec_oprnd1 for every vector stmt to be created | |
4520 | for SLP_NODE. We check during the analysis that all | |
4521 | the shift arguments are the same. */ | |
4522 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4523 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4524 | |
4525 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 4526 | slp_node); |
4a00c761 JJ |
4527 | } |
4528 | else | |
4529 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
306b0c92 | 4530 | &vec_oprnds1, slp_node); |
4a00c761 JJ |
4531 | } |
4532 | else | |
4533 | { | |
81c40241 | 4534 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 4535 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
4536 | if (op_type == binary_op) |
4537 | { | |
4538 | if (code == WIDEN_LSHIFT_EXPR) | |
4539 | vec_oprnd1 = op1; | |
4540 | else | |
81c40241 | 4541 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 4542 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4543 | } |
4544 | } | |
4545 | } | |
ebfd146a | 4546 | else |
4a00c761 JJ |
4547 | { |
4548 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
4549 | vec_oprnds0.truncate (0); |
4550 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
4551 | if (op_type == binary_op) |
4552 | { | |
4553 | if (code == WIDEN_LSHIFT_EXPR) | |
4554 | vec_oprnd1 = op1; | |
4555 | else | |
4556 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
4557 | vec_oprnd1); | |
9771b263 DN |
4558 | vec_oprnds1.truncate (0); |
4559 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
4560 | } |
4561 | } | |
ebfd146a | 4562 | |
4a00c761 JJ |
4563 | /* Arguments are ready. Create the new vector stmts. */ |
4564 | for (i = multi_step_cvt; i >= 0; i--) | |
4565 | { | |
9771b263 | 4566 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
4567 | enum tree_code c1 = code1, c2 = code2; |
4568 | if (i == 0 && codecvt2 != ERROR_MARK) | |
4569 | { | |
4570 | c1 = codecvt1; | |
4571 | c2 = codecvt2; | |
4572 | } | |
4573 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
4574 | &vec_oprnds1, | |
4575 | stmt, this_dest, gsi, | |
4576 | c1, c2, decl1, decl2, | |
4577 | op_type); | |
4578 | } | |
4579 | ||
9771b263 | 4580 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4581 | { |
4582 | if (cvt_type) | |
4583 | { | |
4584 | if (codecvt1 == CALL_EXPR) | |
4585 | { | |
4586 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4587 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4588 | gimple_call_set_lhs (new_stmt, new_temp); | |
4589 | } | |
4590 | else | |
4591 | { | |
4592 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4593 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4594 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4595 | vop0); | |
4a00c761 JJ |
4596 | } |
4597 | ||
4598 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4599 | } | |
4600 | else | |
4601 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
4602 | ||
4603 | if (slp_node) | |
9771b263 | 4604 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 4605 | else |
c689ce1e RB |
4606 | { |
4607 | if (!prev_stmt_info) | |
4608 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
4609 | else | |
4610 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4611 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4612 | } | |
4a00c761 | 4613 | } |
ebfd146a | 4614 | } |
4a00c761 JJ |
4615 | |
4616 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
4617 | break; |
4618 | ||
4619 | case NARROW: | |
4620 | /* In case the vectorization factor (VF) is bigger than the number | |
4621 | of elements that we can fit in a vectype (nunits), we have to | |
4622 | generate more than one vector stmt - i.e - we need to "unroll" | |
4623 | the vector stmt by a factor VF/nunits. */ | |
4624 | for (j = 0; j < ncopies; j++) | |
4625 | { | |
4626 | /* Handle uses. */ | |
4a00c761 JJ |
4627 | if (slp_node) |
4628 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 4629 | slp_node); |
ebfd146a IR |
4630 | else |
4631 | { | |
9771b263 | 4632 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
4633 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
4634 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
4635 | } |
4636 | ||
4a00c761 JJ |
4637 | /* Arguments are ready. Create the new vector stmts. */ |
4638 | if (cvt_type) | |
9771b263 | 4639 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4640 | { |
4641 | if (codecvt1 == CALL_EXPR) | |
4642 | { | |
4643 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4644 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4645 | gimple_call_set_lhs (new_stmt, new_temp); | |
4646 | } | |
4647 | else | |
4648 | { | |
4649 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4650 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4651 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4652 | vop0); | |
4a00c761 | 4653 | } |
ebfd146a | 4654 | |
4a00c761 | 4655 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 4656 | vec_oprnds0[i] = new_temp; |
4a00c761 | 4657 | } |
ebfd146a | 4658 | |
4a00c761 JJ |
4659 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
4660 | stmt, vec_dsts, gsi, | |
4661 | slp_node, code1, | |
4662 | &prev_stmt_info); | |
ebfd146a IR |
4663 | } |
4664 | ||
4665 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 4666 | break; |
ebfd146a IR |
4667 | } |
4668 | ||
9771b263 DN |
4669 | vec_oprnds0.release (); |
4670 | vec_oprnds1.release (); | |
9771b263 | 4671 | interm_types.release (); |
ebfd146a IR |
4672 | |
4673 | return true; | |
4674 | } | |
ff802fa1 IR |
4675 | |
4676 | ||
ebfd146a IR |
4677 | /* Function vectorizable_assignment. |
4678 | ||
b8698a0f L |
4679 | Check if STMT performs an assignment (copy) that can be vectorized. |
4680 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4681 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4682 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4683 | ||
4684 | static bool | |
355fe088 TS |
4685 | vectorizable_assignment (gimple *stmt, gimple_stmt_iterator *gsi, |
4686 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
4687 | { |
4688 | tree vec_dest; | |
4689 | tree scalar_dest; | |
4690 | tree op; | |
4691 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a IR |
4692 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4693 | tree new_temp; | |
355fe088 | 4694 | gimple *def_stmt; |
4fc5ebf1 JG |
4695 | enum vect_def_type dt[1] = {vect_unknown_def_type}; |
4696 | int ndts = 1; | |
ebfd146a | 4697 | int ncopies; |
f18b55bd | 4698 | int i, j; |
6e1aa848 | 4699 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 4700 | tree vop; |
a70d6342 | 4701 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4702 | vec_info *vinfo = stmt_info->vinfo; |
355fe088 | 4703 | gimple *new_stmt = NULL; |
f18b55bd | 4704 | stmt_vec_info prev_stmt_info = NULL; |
fde9c428 RG |
4705 | enum tree_code code; |
4706 | tree vectype_in; | |
ebfd146a | 4707 | |
a70d6342 | 4708 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4709 | return false; |
4710 | ||
66c16fd9 RB |
4711 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4712 | && ! vec_stmt) | |
ebfd146a IR |
4713 | return false; |
4714 | ||
4715 | /* Is vectorizable assignment? */ | |
4716 | if (!is_gimple_assign (stmt)) | |
4717 | return false; | |
4718 | ||
4719 | scalar_dest = gimple_assign_lhs (stmt); | |
4720 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4721 | return false; | |
4722 | ||
fde9c428 | 4723 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 4724 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
4725 | || code == PAREN_EXPR |
4726 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
4727 | op = gimple_assign_rhs1 (stmt); |
4728 | else | |
4729 | return false; | |
4730 | ||
7b7ec6c5 RG |
4731 | if (code == VIEW_CONVERT_EXPR) |
4732 | op = TREE_OPERAND (op, 0); | |
4733 | ||
465c8c19 JJ |
4734 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
4735 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4736 | ||
4737 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4738 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4739 | case of SLP. */ | |
fce57248 | 4740 | if (slp_node) |
465c8c19 JJ |
4741 | ncopies = 1; |
4742 | else | |
e8f142e2 | 4743 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
4744 | |
4745 | gcc_assert (ncopies >= 1); | |
4746 | ||
81c40241 | 4747 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[0], &vectype_in)) |
ebfd146a | 4748 | { |
73fbfcad | 4749 | if (dump_enabled_p ()) |
78c60e3d | 4750 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4751 | "use not simple.\n"); |
ebfd146a IR |
4752 | return false; |
4753 | } | |
4754 | ||
fde9c428 RG |
4755 | /* We can handle NOP_EXPR conversions that do not change the number |
4756 | of elements or the vector size. */ | |
7b7ec6c5 RG |
4757 | if ((CONVERT_EXPR_CODE_P (code) |
4758 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 RG |
4759 | && (!vectype_in |
4760 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
4761 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
4762 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
4763 | return false; | |
4764 | ||
7b7b1813 RG |
4765 | /* We do not handle bit-precision changes. */ |
4766 | if ((CONVERT_EXPR_CODE_P (code) | |
4767 | || code == VIEW_CONVERT_EXPR) | |
4768 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
2be65d9e RS |
4769 | && (!type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
4770 | || !type_has_mode_precision_p (TREE_TYPE (op))) | |
7b7b1813 RG |
4771 | /* But a conversion that does not change the bit-pattern is ok. */ |
4772 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4773 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2dab46d5 IE |
4774 | && TYPE_UNSIGNED (TREE_TYPE (op))) |
4775 | /* Conversion between boolean types of different sizes is | |
4776 | a simple assignment in case their vectypes are same | |
4777 | boolean vectors. */ | |
4778 | && (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
4779 | || !VECTOR_BOOLEAN_TYPE_P (vectype_in))) | |
7b7b1813 | 4780 | { |
73fbfcad | 4781 | if (dump_enabled_p ()) |
78c60e3d SS |
4782 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4783 | "type conversion to/from bit-precision " | |
e645e942 | 4784 | "unsupported.\n"); |
7b7b1813 RG |
4785 | return false; |
4786 | } | |
4787 | ||
ebfd146a IR |
4788 | if (!vec_stmt) /* transformation not required. */ |
4789 | { | |
4790 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 4791 | if (dump_enabled_p ()) |
78c60e3d | 4792 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4793 | "=== vectorizable_assignment ===\n"); |
4fc5ebf1 | 4794 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); |
ebfd146a IR |
4795 | return true; |
4796 | } | |
4797 | ||
67b8dbac | 4798 | /* Transform. */ |
73fbfcad | 4799 | if (dump_enabled_p ()) |
e645e942 | 4800 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
4801 | |
4802 | /* Handle def. */ | |
4803 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4804 | ||
4805 | /* Handle use. */ | |
f18b55bd | 4806 | for (j = 0; j < ncopies; j++) |
ebfd146a | 4807 | { |
f18b55bd IR |
4808 | /* Handle uses. */ |
4809 | if (j == 0) | |
306b0c92 | 4810 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); |
f18b55bd IR |
4811 | else |
4812 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
4813 | ||
4814 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 4815 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 4816 | { |
7b7ec6c5 RG |
4817 | if (CONVERT_EXPR_CODE_P (code) |
4818 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 4819 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
4820 | new_stmt = gimple_build_assign (vec_dest, vop); |
4821 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4822 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4823 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4824 | if (slp_node) | |
9771b263 | 4825 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 4826 | } |
ebfd146a IR |
4827 | |
4828 | if (slp_node) | |
f18b55bd IR |
4829 | continue; |
4830 | ||
4831 | if (j == 0) | |
4832 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4833 | else | |
4834 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4835 | ||
4836 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4837 | } | |
b8698a0f | 4838 | |
9771b263 | 4839 | vec_oprnds.release (); |
ebfd146a IR |
4840 | return true; |
4841 | } | |
4842 | ||
9dc3f7de | 4843 | |
1107f3ae IR |
4844 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
4845 | either as shift by a scalar or by a vector. */ | |
4846 | ||
4847 | bool | |
4848 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
4849 | { | |
4850 | ||
ef4bddc2 | 4851 | machine_mode vec_mode; |
1107f3ae IR |
4852 | optab optab; |
4853 | int icode; | |
4854 | tree vectype; | |
4855 | ||
4856 | vectype = get_vectype_for_scalar_type (scalar_type); | |
4857 | if (!vectype) | |
4858 | return false; | |
4859 | ||
4860 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4861 | if (!optab | |
4862 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
4863 | { | |
4864 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4865 | if (!optab | |
4866 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
4867 | == CODE_FOR_nothing)) | |
4868 | return false; | |
4869 | } | |
4870 | ||
4871 | vec_mode = TYPE_MODE (vectype); | |
4872 | icode = (int) optab_handler (optab, vec_mode); | |
4873 | if (icode == CODE_FOR_nothing) | |
4874 | return false; | |
4875 | ||
4876 | return true; | |
4877 | } | |
4878 | ||
4879 | ||
9dc3f7de IR |
4880 | /* Function vectorizable_shift. |
4881 | ||
4882 | Check if STMT performs a shift operation that can be vectorized. | |
4883 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4884 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
4885 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4886 | ||
4887 | static bool | |
355fe088 TS |
4888 | vectorizable_shift (gimple *stmt, gimple_stmt_iterator *gsi, |
4889 | gimple **vec_stmt, slp_tree slp_node) | |
9dc3f7de IR |
4890 | { |
4891 | tree vec_dest; | |
4892 | tree scalar_dest; | |
4893 | tree op0, op1 = NULL; | |
4894 | tree vec_oprnd1 = NULL_TREE; | |
4895 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4896 | tree vectype; | |
4897 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4898 | enum tree_code code; | |
ef4bddc2 | 4899 | machine_mode vec_mode; |
9dc3f7de IR |
4900 | tree new_temp; |
4901 | optab optab; | |
4902 | int icode; | |
ef4bddc2 | 4903 | machine_mode optab_op2_mode; |
355fe088 | 4904 | gimple *def_stmt; |
9dc3f7de | 4905 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 4906 | int ndts = 2; |
355fe088 | 4907 | gimple *new_stmt = NULL; |
9dc3f7de IR |
4908 | stmt_vec_info prev_stmt_info; |
4909 | int nunits_in; | |
4910 | int nunits_out; | |
4911 | tree vectype_out; | |
cede2577 | 4912 | tree op1_vectype; |
9dc3f7de IR |
4913 | int ncopies; |
4914 | int j, i; | |
6e1aa848 DN |
4915 | vec<tree> vec_oprnds0 = vNULL; |
4916 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
4917 | tree vop0, vop1; |
4918 | unsigned int k; | |
49eab32e | 4919 | bool scalar_shift_arg = true; |
9dc3f7de | 4920 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4921 | vec_info *vinfo = stmt_info->vinfo; |
9dc3f7de IR |
4922 | |
4923 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
4924 | return false; | |
4925 | ||
66c16fd9 RB |
4926 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4927 | && ! vec_stmt) | |
9dc3f7de IR |
4928 | return false; |
4929 | ||
4930 | /* Is STMT a vectorizable binary/unary operation? */ | |
4931 | if (!is_gimple_assign (stmt)) | |
4932 | return false; | |
4933 | ||
4934 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4935 | return false; | |
4936 | ||
4937 | code = gimple_assign_rhs_code (stmt); | |
4938 | ||
4939 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
4940 | || code == RROTATE_EXPR)) | |
4941 | return false; | |
4942 | ||
4943 | scalar_dest = gimple_assign_lhs (stmt); | |
4944 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2be65d9e | 4945 | if (!type_has_mode_precision_p (TREE_TYPE (scalar_dest))) |
7b7b1813 | 4946 | { |
73fbfcad | 4947 | if (dump_enabled_p ()) |
78c60e3d | 4948 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4949 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
4950 | return false; |
4951 | } | |
9dc3f7de IR |
4952 | |
4953 | op0 = gimple_assign_rhs1 (stmt); | |
81c40241 | 4954 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
9dc3f7de | 4955 | { |
73fbfcad | 4956 | if (dump_enabled_p ()) |
78c60e3d | 4957 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4958 | "use not simple.\n"); |
9dc3f7de IR |
4959 | return false; |
4960 | } | |
4961 | /* If op0 is an external or constant def use a vector type with | |
4962 | the same size as the output vector type. */ | |
4963 | if (!vectype) | |
4964 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
4965 | if (vec_stmt) | |
4966 | gcc_assert (vectype); | |
4967 | if (!vectype) | |
4968 | { | |
73fbfcad | 4969 | if (dump_enabled_p ()) |
78c60e3d | 4970 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4971 | "no vectype for scalar type\n"); |
9dc3f7de IR |
4972 | return false; |
4973 | } | |
4974 | ||
4975 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4976 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4977 | if (nunits_out != nunits_in) | |
4978 | return false; | |
4979 | ||
4980 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 4981 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &op1_vectype)) |
9dc3f7de | 4982 | { |
73fbfcad | 4983 | if (dump_enabled_p ()) |
78c60e3d | 4984 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4985 | "use not simple.\n"); |
9dc3f7de IR |
4986 | return false; |
4987 | } | |
4988 | ||
9dc3f7de IR |
4989 | /* Multiple types in SLP are handled by creating the appropriate number of |
4990 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4991 | case of SLP. */ | |
fce57248 | 4992 | if (slp_node) |
9dc3f7de IR |
4993 | ncopies = 1; |
4994 | else | |
e8f142e2 | 4995 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
9dc3f7de IR |
4996 | |
4997 | gcc_assert (ncopies >= 1); | |
4998 | ||
4999 | /* Determine whether the shift amount is a vector, or scalar. If the | |
5000 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
5001 | ||
dbfa87aa YR |
5002 | if ((dt[1] == vect_internal_def |
5003 | || dt[1] == vect_induction_def) | |
5004 | && !slp_node) | |
49eab32e JJ |
5005 | scalar_shift_arg = false; |
5006 | else if (dt[1] == vect_constant_def | |
5007 | || dt[1] == vect_external_def | |
5008 | || dt[1] == vect_internal_def) | |
5009 | { | |
5010 | /* In SLP, need to check whether the shift count is the same, | |
5011 | in loops if it is a constant or invariant, it is always | |
5012 | a scalar shift. */ | |
5013 | if (slp_node) | |
5014 | { | |
355fe088 TS |
5015 | vec<gimple *> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
5016 | gimple *slpstmt; | |
49eab32e | 5017 | |
9771b263 | 5018 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
5019 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
5020 | scalar_shift_arg = false; | |
5021 | } | |
60d393e8 RB |
5022 | |
5023 | /* If the shift amount is computed by a pattern stmt we cannot | |
5024 | use the scalar amount directly thus give up and use a vector | |
5025 | shift. */ | |
5026 | if (dt[1] == vect_internal_def) | |
5027 | { | |
5028 | gimple *def = SSA_NAME_DEF_STMT (op1); | |
5029 | if (is_pattern_stmt_p (vinfo_for_stmt (def))) | |
5030 | scalar_shift_arg = false; | |
5031 | } | |
49eab32e JJ |
5032 | } |
5033 | else | |
5034 | { | |
73fbfcad | 5035 | if (dump_enabled_p ()) |
78c60e3d | 5036 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5037 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
5038 | return false; |
5039 | } | |
5040 | ||
9dc3f7de | 5041 | /* Vector shifted by vector. */ |
49eab32e | 5042 | if (!scalar_shift_arg) |
9dc3f7de IR |
5043 | { |
5044 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 5045 | if (dump_enabled_p ()) |
78c60e3d | 5046 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5047 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 5048 | |
aa948027 JJ |
5049 | if (!op1_vectype) |
5050 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
5051 | if (op1_vectype == NULL_TREE | |
5052 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 5053 | { |
73fbfcad | 5054 | if (dump_enabled_p ()) |
78c60e3d SS |
5055 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5056 | "unusable type for last operand in" | |
e645e942 | 5057 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
5058 | return false; |
5059 | } | |
9dc3f7de IR |
5060 | } |
5061 | /* See if the machine has a vector shifted by scalar insn and if not | |
5062 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 5063 | else |
9dc3f7de IR |
5064 | { |
5065 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
5066 | if (optab | |
5067 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
5068 | { | |
73fbfcad | 5069 | if (dump_enabled_p ()) |
78c60e3d | 5070 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5071 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
5072 | } |
5073 | else | |
5074 | { | |
5075 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
5076 | if (optab | |
5077 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
5078 | != CODE_FOR_nothing)) | |
5079 | { | |
49eab32e JJ |
5080 | scalar_shift_arg = false; |
5081 | ||
73fbfcad | 5082 | if (dump_enabled_p ()) |
78c60e3d | 5083 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5084 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
5085 | |
5086 | /* Unlike the other binary operators, shifts/rotates have | |
5087 | the rhs being int, instead of the same type as the lhs, | |
5088 | so make sure the scalar is the right type if we are | |
aa948027 | 5089 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
5090 | if (dt[1] == vect_constant_def) |
5091 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
5092 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
5093 | TREE_TYPE (op1))) | |
5094 | { | |
5095 | if (slp_node | |
5096 | && TYPE_MODE (TREE_TYPE (vectype)) | |
5097 | != TYPE_MODE (TREE_TYPE (op1))) | |
5098 | { | |
73fbfcad | 5099 | if (dump_enabled_p ()) |
78c60e3d SS |
5100 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5101 | "unusable type for last operand in" | |
e645e942 | 5102 | " vector/vector shift/rotate.\n"); |
21c0a521 | 5103 | return false; |
aa948027 JJ |
5104 | } |
5105 | if (vec_stmt && !slp_node) | |
5106 | { | |
5107 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
5108 | op1 = vect_init_vector (stmt, op1, | |
5109 | TREE_TYPE (vectype), NULL); | |
5110 | } | |
5111 | } | |
9dc3f7de IR |
5112 | } |
5113 | } | |
5114 | } | |
9dc3f7de IR |
5115 | |
5116 | /* Supportable by target? */ | |
5117 | if (!optab) | |
5118 | { | |
73fbfcad | 5119 | if (dump_enabled_p ()) |
78c60e3d | 5120 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5121 | "no optab.\n"); |
9dc3f7de IR |
5122 | return false; |
5123 | } | |
5124 | vec_mode = TYPE_MODE (vectype); | |
5125 | icode = (int) optab_handler (optab, vec_mode); | |
5126 | if (icode == CODE_FOR_nothing) | |
5127 | { | |
73fbfcad | 5128 | if (dump_enabled_p ()) |
78c60e3d | 5129 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5130 | "op not supported by target.\n"); |
9dc3f7de IR |
5131 | /* Check only during analysis. */ |
5132 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
ca09abcb RS |
5133 | || (!vec_stmt |
5134 | && !vect_worthwhile_without_simd_p (vinfo, code))) | |
9dc3f7de | 5135 | return false; |
73fbfcad | 5136 | if (dump_enabled_p ()) |
e645e942 TJ |
5137 | dump_printf_loc (MSG_NOTE, vect_location, |
5138 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
5139 | } |
5140 | ||
5141 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
ca09abcb RS |
5142 | if (!vec_stmt |
5143 | && !VECTOR_MODE_P (TYPE_MODE (vectype)) | |
5144 | && !vect_worthwhile_without_simd_p (vinfo, code)) | |
9dc3f7de | 5145 | { |
73fbfcad | 5146 | if (dump_enabled_p ()) |
78c60e3d | 5147 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5148 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
5149 | return false; |
5150 | } | |
5151 | ||
5152 | if (!vec_stmt) /* transformation not required. */ | |
5153 | { | |
5154 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 5155 | if (dump_enabled_p ()) |
e645e942 TJ |
5156 | dump_printf_loc (MSG_NOTE, vect_location, |
5157 | "=== vectorizable_shift ===\n"); | |
4fc5ebf1 | 5158 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); |
9dc3f7de IR |
5159 | return true; |
5160 | } | |
5161 | ||
67b8dbac | 5162 | /* Transform. */ |
9dc3f7de | 5163 | |
73fbfcad | 5164 | if (dump_enabled_p ()) |
78c60e3d | 5165 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5166 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
5167 | |
5168 | /* Handle def. */ | |
5169 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5170 | ||
9dc3f7de IR |
5171 | prev_stmt_info = NULL; |
5172 | for (j = 0; j < ncopies; j++) | |
5173 | { | |
5174 | /* Handle uses. */ | |
5175 | if (j == 0) | |
5176 | { | |
5177 | if (scalar_shift_arg) | |
5178 | { | |
5179 | /* Vector shl and shr insn patterns can be defined with scalar | |
5180 | operand 2 (shift operand). In this case, use constant or loop | |
5181 | invariant op1 directly, without extending it to vector mode | |
5182 | first. */ | |
5183 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
5184 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
5185 | { | |
73fbfcad | 5186 | if (dump_enabled_p ()) |
78c60e3d | 5187 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5188 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 5189 | vec_oprnd1 = op1; |
8930f723 | 5190 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 5191 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5192 | if (slp_node) |
5193 | { | |
5194 | /* Store vec_oprnd1 for every vector stmt to be created | |
5195 | for SLP_NODE. We check during the analysis that all | |
5196 | the shift arguments are the same. | |
5197 | TODO: Allow different constants for different vector | |
5198 | stmts generated for an SLP instance. */ | |
5199 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 5200 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5201 | } |
5202 | } | |
5203 | } | |
5204 | ||
5205 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
5206 | (a special case for certain kind of vector shifts); otherwise, | |
5207 | operand 1 should be of a vector type (the usual case). */ | |
5208 | if (vec_oprnd1) | |
5209 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5210 | slp_node); |
9dc3f7de IR |
5211 | else |
5212 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
306b0c92 | 5213 | slp_node); |
9dc3f7de IR |
5214 | } |
5215 | else | |
5216 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5217 | ||
5218 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 5219 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 5220 | { |
9771b263 | 5221 | vop1 = vec_oprnds1[i]; |
0d0e4a03 | 5222 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
9dc3f7de IR |
5223 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5224 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5225 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5226 | if (slp_node) | |
9771b263 | 5227 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
5228 | } |
5229 | ||
5230 | if (slp_node) | |
5231 | continue; | |
5232 | ||
5233 | if (j == 0) | |
5234 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5235 | else | |
5236 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5237 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5238 | } | |
5239 | ||
9771b263 DN |
5240 | vec_oprnds0.release (); |
5241 | vec_oprnds1.release (); | |
9dc3f7de IR |
5242 | |
5243 | return true; | |
5244 | } | |
5245 | ||
5246 | ||
ebfd146a IR |
5247 | /* Function vectorizable_operation. |
5248 | ||
16949072 RG |
5249 | Check if STMT performs a binary, unary or ternary operation that can |
5250 | be vectorized. | |
b8698a0f | 5251 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
5252 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5253 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5254 | ||
5255 | static bool | |
355fe088 TS |
5256 | vectorizable_operation (gimple *stmt, gimple_stmt_iterator *gsi, |
5257 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a | 5258 | { |
00f07b86 | 5259 | tree vec_dest; |
ebfd146a | 5260 | tree scalar_dest; |
16949072 | 5261 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 5262 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 5263 | tree vectype; |
ebfd146a | 5264 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
0eb952ea | 5265 | enum tree_code code, orig_code; |
ef4bddc2 | 5266 | machine_mode vec_mode; |
ebfd146a IR |
5267 | tree new_temp; |
5268 | int op_type; | |
00f07b86 | 5269 | optab optab; |
523ba738 | 5270 | bool target_support_p; |
355fe088 | 5271 | gimple *def_stmt; |
16949072 RG |
5272 | enum vect_def_type dt[3] |
5273 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 5274 | int ndts = 3; |
355fe088 | 5275 | gimple *new_stmt = NULL; |
ebfd146a | 5276 | stmt_vec_info prev_stmt_info; |
b690cc0f | 5277 | int nunits_in; |
ebfd146a IR |
5278 | int nunits_out; |
5279 | tree vectype_out; | |
5280 | int ncopies; | |
5281 | int j, i; | |
6e1aa848 DN |
5282 | vec<tree> vec_oprnds0 = vNULL; |
5283 | vec<tree> vec_oprnds1 = vNULL; | |
5284 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 5285 | tree vop0, vop1, vop2; |
a70d6342 | 5286 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5287 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 5288 | |
a70d6342 | 5289 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5290 | return false; |
5291 | ||
66c16fd9 RB |
5292 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5293 | && ! vec_stmt) | |
ebfd146a IR |
5294 | return false; |
5295 | ||
5296 | /* Is STMT a vectorizable binary/unary operation? */ | |
5297 | if (!is_gimple_assign (stmt)) | |
5298 | return false; | |
5299 | ||
5300 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
5301 | return false; | |
5302 | ||
0eb952ea | 5303 | orig_code = code = gimple_assign_rhs_code (stmt); |
ebfd146a | 5304 | |
1af4ebf5 MG |
5305 | /* For pointer addition and subtraction, we should use the normal |
5306 | plus and minus for the vector operation. */ | |
ebfd146a IR |
5307 | if (code == POINTER_PLUS_EXPR) |
5308 | code = PLUS_EXPR; | |
1af4ebf5 MG |
5309 | if (code == POINTER_DIFF_EXPR) |
5310 | code = MINUS_EXPR; | |
ebfd146a IR |
5311 | |
5312 | /* Support only unary or binary operations. */ | |
5313 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 5314 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 5315 | { |
73fbfcad | 5316 | if (dump_enabled_p ()) |
78c60e3d | 5317 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5318 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 5319 | op_type); |
ebfd146a IR |
5320 | return false; |
5321 | } | |
5322 | ||
b690cc0f RG |
5323 | scalar_dest = gimple_assign_lhs (stmt); |
5324 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
5325 | ||
7b7b1813 RG |
5326 | /* Most operations cannot handle bit-precision types without extra |
5327 | truncations. */ | |
045c1278 | 5328 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
2be65d9e | 5329 | && !type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
7b7b1813 RG |
5330 | /* Exception are bitwise binary operations. */ |
5331 | && code != BIT_IOR_EXPR | |
5332 | && code != BIT_XOR_EXPR | |
5333 | && code != BIT_AND_EXPR) | |
5334 | { | |
73fbfcad | 5335 | if (dump_enabled_p ()) |
78c60e3d | 5336 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5337 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
5338 | return false; |
5339 | } | |
5340 | ||
ebfd146a | 5341 | op0 = gimple_assign_rhs1 (stmt); |
81c40241 | 5342 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
ebfd146a | 5343 | { |
73fbfcad | 5344 | if (dump_enabled_p ()) |
78c60e3d | 5345 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5346 | "use not simple.\n"); |
ebfd146a IR |
5347 | return false; |
5348 | } | |
b690cc0f RG |
5349 | /* If op0 is an external or constant def use a vector type with |
5350 | the same size as the output vector type. */ | |
5351 | if (!vectype) | |
b036c6c5 IE |
5352 | { |
5353 | /* For boolean type we cannot determine vectype by | |
5354 | invariant value (don't know whether it is a vector | |
5355 | of booleans or vector of integers). We use output | |
5356 | vectype because operations on boolean don't change | |
5357 | type. */ | |
2568d8a1 | 5358 | if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op0))) |
b036c6c5 | 5359 | { |
2568d8a1 | 5360 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (scalar_dest))) |
b036c6c5 IE |
5361 | { |
5362 | if (dump_enabled_p ()) | |
5363 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5364 | "not supported operation on bool value.\n"); | |
5365 | return false; | |
5366 | } | |
5367 | vectype = vectype_out; | |
5368 | } | |
5369 | else | |
5370 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
5371 | } | |
7d8930a0 IR |
5372 | if (vec_stmt) |
5373 | gcc_assert (vectype); | |
5374 | if (!vectype) | |
5375 | { | |
73fbfcad | 5376 | if (dump_enabled_p ()) |
7d8930a0 | 5377 | { |
78c60e3d SS |
5378 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5379 | "no vectype for scalar type "); | |
5380 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
5381 | TREE_TYPE (op0)); | |
e645e942 | 5382 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
5383 | } |
5384 | ||
5385 | return false; | |
5386 | } | |
b690cc0f RG |
5387 | |
5388 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
5389 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
5390 | if (nunits_out != nunits_in) | |
5391 | return false; | |
ebfd146a | 5392 | |
16949072 | 5393 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
5394 | { |
5395 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 5396 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1])) |
ebfd146a | 5397 | { |
73fbfcad | 5398 | if (dump_enabled_p ()) |
78c60e3d | 5399 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5400 | "use not simple.\n"); |
ebfd146a IR |
5401 | return false; |
5402 | } | |
5403 | } | |
16949072 RG |
5404 | if (op_type == ternary_op) |
5405 | { | |
5406 | op2 = gimple_assign_rhs3 (stmt); | |
81c40241 | 5407 | if (!vect_is_simple_use (op2, vinfo, &def_stmt, &dt[2])) |
16949072 | 5408 | { |
73fbfcad | 5409 | if (dump_enabled_p ()) |
78c60e3d | 5410 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5411 | "use not simple.\n"); |
16949072 RG |
5412 | return false; |
5413 | } | |
5414 | } | |
ebfd146a | 5415 | |
b690cc0f | 5416 | /* Multiple types in SLP are handled by creating the appropriate number of |
ff802fa1 | 5417 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 5418 | case of SLP. */ |
fce57248 | 5419 | if (slp_node) |
b690cc0f RG |
5420 | ncopies = 1; |
5421 | else | |
e8f142e2 | 5422 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
b690cc0f RG |
5423 | |
5424 | gcc_assert (ncopies >= 1); | |
5425 | ||
9dc3f7de | 5426 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
5427 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
5428 | || code == RROTATE_EXPR) | |
9dc3f7de | 5429 | return false; |
ebfd146a | 5430 | |
ebfd146a | 5431 | /* Supportable by target? */ |
00f07b86 RH |
5432 | |
5433 | vec_mode = TYPE_MODE (vectype); | |
5434 | if (code == MULT_HIGHPART_EXPR) | |
523ba738 | 5435 | target_support_p = can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype)); |
00f07b86 RH |
5436 | else |
5437 | { | |
5438 | optab = optab_for_tree_code (code, vectype, optab_default); | |
5439 | if (!optab) | |
5deb57cb | 5440 | { |
73fbfcad | 5441 | if (dump_enabled_p ()) |
78c60e3d | 5442 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5443 | "no optab.\n"); |
00f07b86 | 5444 | return false; |
5deb57cb | 5445 | } |
523ba738 RS |
5446 | target_support_p = (optab_handler (optab, vec_mode) |
5447 | != CODE_FOR_nothing); | |
5deb57cb JJ |
5448 | } |
5449 | ||
523ba738 | 5450 | if (!target_support_p) |
ebfd146a | 5451 | { |
73fbfcad | 5452 | if (dump_enabled_p ()) |
78c60e3d | 5453 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5454 | "op not supported by target.\n"); |
ebfd146a IR |
5455 | /* Check only during analysis. */ |
5456 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
ca09abcb | 5457 | || (!vec_stmt && !vect_worthwhile_without_simd_p (vinfo, code))) |
ebfd146a | 5458 | return false; |
73fbfcad | 5459 | if (dump_enabled_p ()) |
e645e942 TJ |
5460 | dump_printf_loc (MSG_NOTE, vect_location, |
5461 | "proceeding using word mode.\n"); | |
383d9c83 IR |
5462 | } |
5463 | ||
4a00c761 | 5464 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
5465 | if (!VECTOR_MODE_P (vec_mode) |
5466 | && !vec_stmt | |
ca09abcb | 5467 | && !vect_worthwhile_without_simd_p (vinfo, code)) |
7d8930a0 | 5468 | { |
73fbfcad | 5469 | if (dump_enabled_p ()) |
78c60e3d | 5470 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5471 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 5472 | return false; |
7d8930a0 | 5473 | } |
ebfd146a | 5474 | |
ebfd146a IR |
5475 | if (!vec_stmt) /* transformation not required. */ |
5476 | { | |
4a00c761 | 5477 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 5478 | if (dump_enabled_p ()) |
78c60e3d | 5479 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5480 | "=== vectorizable_operation ===\n"); |
4fc5ebf1 | 5481 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); |
ebfd146a IR |
5482 | return true; |
5483 | } | |
5484 | ||
67b8dbac | 5485 | /* Transform. */ |
ebfd146a | 5486 | |
73fbfcad | 5487 | if (dump_enabled_p ()) |
78c60e3d | 5488 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5489 | "transform binary/unary operation.\n"); |
383d9c83 | 5490 | |
ebfd146a | 5491 | /* Handle def. */ |
00f07b86 | 5492 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 5493 | |
0eb952ea JJ |
5494 | /* POINTER_DIFF_EXPR has pointer arguments which are vectorized as |
5495 | vectors with unsigned elements, but the result is signed. So, we | |
5496 | need to compute the MINUS_EXPR into vectype temporary and | |
5497 | VIEW_CONVERT_EXPR it into the final vectype_out result. */ | |
5498 | tree vec_cvt_dest = NULL_TREE; | |
5499 | if (orig_code == POINTER_DIFF_EXPR) | |
5500 | vec_cvt_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
5501 | ||
ebfd146a IR |
5502 | /* In case the vectorization factor (VF) is bigger than the number |
5503 | of elements that we can fit in a vectype (nunits), we have to generate | |
5504 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
5505 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
5506 | from one copy of the vector stmt to the next, in the field | |
5507 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
5508 | stages to find the correct vector defs to be used when vectorizing | |
5509 | stmts that use the defs of the current stmt. The example below | |
5510 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
5511 | we need to create 4 vectorized stmts): | |
5512 | ||
5513 | before vectorization: | |
5514 | RELATED_STMT VEC_STMT | |
5515 | S1: x = memref - - | |
5516 | S2: z = x + 1 - - | |
5517 | ||
5518 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
5519 | there): | |
5520 | RELATED_STMT VEC_STMT | |
5521 | VS1_0: vx0 = memref0 VS1_1 - | |
5522 | VS1_1: vx1 = memref1 VS1_2 - | |
5523 | VS1_2: vx2 = memref2 VS1_3 - | |
5524 | VS1_3: vx3 = memref3 - - | |
5525 | S1: x = load - VS1_0 | |
5526 | S2: z = x + 1 - - | |
5527 | ||
5528 | step2: vectorize stmt S2 (done here): | |
5529 | To vectorize stmt S2 we first need to find the relevant vector | |
5530 | def for the first operand 'x'. This is, as usual, obtained from | |
5531 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
5532 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
5533 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
5534 | the vector stmt VS2_0, and as usual, record it in the | |
5535 | STMT_VINFO_VEC_STMT of stmt S2. | |
5536 | When creating the second copy (VS2_1), we obtain the relevant vector | |
5537 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
5538 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
5539 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
5540 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
5541 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
5542 | chain of stmts and pointers: | |
5543 | RELATED_STMT VEC_STMT | |
5544 | VS1_0: vx0 = memref0 VS1_1 - | |
5545 | VS1_1: vx1 = memref1 VS1_2 - | |
5546 | VS1_2: vx2 = memref2 VS1_3 - | |
5547 | VS1_3: vx3 = memref3 - - | |
5548 | S1: x = load - VS1_0 | |
5549 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
5550 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
5551 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
5552 | VS2_3: vz3 = vx3 + v1 - - | |
5553 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
5554 | |
5555 | prev_stmt_info = NULL; | |
5556 | for (j = 0; j < ncopies; j++) | |
5557 | { | |
5558 | /* Handle uses. */ | |
5559 | if (j == 0) | |
4a00c761 JJ |
5560 | { |
5561 | if (op_type == binary_op || op_type == ternary_op) | |
5562 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
306b0c92 | 5563 | slp_node); |
4a00c761 JJ |
5564 | else |
5565 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5566 | slp_node); |
4a00c761 | 5567 | if (op_type == ternary_op) |
c392943c | 5568 | vect_get_vec_defs (op2, NULL_TREE, stmt, &vec_oprnds2, NULL, |
306b0c92 | 5569 | slp_node); |
4a00c761 | 5570 | } |
ebfd146a | 5571 | else |
4a00c761 JJ |
5572 | { |
5573 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5574 | if (op_type == ternary_op) | |
5575 | { | |
9771b263 DN |
5576 | tree vec_oprnd = vec_oprnds2.pop (); |
5577 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
5578 | vec_oprnd)); | |
4a00c761 JJ |
5579 | } |
5580 | } | |
5581 | ||
5582 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 5583 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 5584 | { |
4a00c761 | 5585 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 5586 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 5587 | vop2 = ((op_type == ternary_op) |
9771b263 | 5588 | ? vec_oprnds2[i] : NULL_TREE); |
0d0e4a03 | 5589 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1, vop2); |
4a00c761 JJ |
5590 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5591 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5592 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
0eb952ea JJ |
5593 | if (vec_cvt_dest) |
5594 | { | |
5595 | new_temp = build1 (VIEW_CONVERT_EXPR, vectype_out, new_temp); | |
5596 | new_stmt = gimple_build_assign (vec_cvt_dest, VIEW_CONVERT_EXPR, | |
5597 | new_temp); | |
5598 | new_temp = make_ssa_name (vec_cvt_dest, new_stmt); | |
5599 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5600 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5601 | } | |
4a00c761 | 5602 | if (slp_node) |
9771b263 | 5603 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
5604 | } |
5605 | ||
4a00c761 JJ |
5606 | if (slp_node) |
5607 | continue; | |
5608 | ||
5609 | if (j == 0) | |
5610 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5611 | else | |
5612 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5613 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
5614 | } |
5615 | ||
9771b263 DN |
5616 | vec_oprnds0.release (); |
5617 | vec_oprnds1.release (); | |
5618 | vec_oprnds2.release (); | |
ebfd146a | 5619 | |
ebfd146a IR |
5620 | return true; |
5621 | } | |
5622 | ||
f702e7d4 | 5623 | /* A helper function to ensure data reference DR's base alignment. */ |
c716e67f XDL |
5624 | |
5625 | static void | |
f702e7d4 | 5626 | ensure_base_align (struct data_reference *dr) |
c716e67f XDL |
5627 | { |
5628 | if (!dr->aux) | |
5629 | return; | |
5630 | ||
52639a61 | 5631 | if (DR_VECT_AUX (dr)->base_misaligned) |
c716e67f | 5632 | { |
52639a61 | 5633 | tree base_decl = DR_VECT_AUX (dr)->base_decl; |
c716e67f | 5634 | |
f702e7d4 RS |
5635 | unsigned int align_base_to = DR_TARGET_ALIGNMENT (dr) * BITS_PER_UNIT; |
5636 | ||
428f0c67 | 5637 | if (decl_in_symtab_p (base_decl)) |
f702e7d4 | 5638 | symtab_node::get (base_decl)->increase_alignment (align_base_to); |
428f0c67 JH |
5639 | else |
5640 | { | |
f702e7d4 | 5641 | SET_DECL_ALIGN (base_decl, align_base_to); |
428f0c67 JH |
5642 | DECL_USER_ALIGN (base_decl) = 1; |
5643 | } | |
52639a61 | 5644 | DR_VECT_AUX (dr)->base_misaligned = false; |
c716e67f XDL |
5645 | } |
5646 | } | |
5647 | ||
ebfd146a | 5648 | |
44fc7854 BE |
5649 | /* Function get_group_alias_ptr_type. |
5650 | ||
5651 | Return the alias type for the group starting at FIRST_STMT. */ | |
5652 | ||
5653 | static tree | |
5654 | get_group_alias_ptr_type (gimple *first_stmt) | |
5655 | { | |
5656 | struct data_reference *first_dr, *next_dr; | |
5657 | gimple *next_stmt; | |
5658 | ||
5659 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
5660 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (first_stmt)); | |
5661 | while (next_stmt) | |
5662 | { | |
5663 | next_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (next_stmt)); | |
5664 | if (get_alias_set (DR_REF (first_dr)) | |
5665 | != get_alias_set (DR_REF (next_dr))) | |
5666 | { | |
5667 | if (dump_enabled_p ()) | |
5668 | dump_printf_loc (MSG_NOTE, vect_location, | |
5669 | "conflicting alias set types.\n"); | |
5670 | return ptr_type_node; | |
5671 | } | |
5672 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); | |
5673 | } | |
5674 | return reference_alias_ptr_type (DR_REF (first_dr)); | |
5675 | } | |
5676 | ||
5677 | ||
ebfd146a IR |
5678 | /* Function vectorizable_store. |
5679 | ||
b8698a0f L |
5680 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
5681 | can be vectorized. | |
5682 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
5683 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5684 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5685 | ||
5686 | static bool | |
355fe088 | 5687 | vectorizable_store (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 5688 | slp_tree slp_node) |
ebfd146a IR |
5689 | { |
5690 | tree scalar_dest; | |
5691 | tree data_ref; | |
5692 | tree op; | |
5693 | tree vec_oprnd = NULL_TREE; | |
5694 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5695 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
272c6793 | 5696 | tree elem_type; |
ebfd146a | 5697 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 5698 | struct loop *loop = NULL; |
ef4bddc2 | 5699 | machine_mode vec_mode; |
ebfd146a IR |
5700 | tree dummy; |
5701 | enum dr_alignment_support alignment_support_scheme; | |
355fe088 | 5702 | gimple *def_stmt; |
ebfd146a IR |
5703 | enum vect_def_type dt; |
5704 | stmt_vec_info prev_stmt_info = NULL; | |
5705 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 5706 | tree dataref_offset = NULL_TREE; |
355fe088 | 5707 | gimple *ptr_incr = NULL; |
ebfd146a IR |
5708 | int ncopies; |
5709 | int j; | |
2de001ee RS |
5710 | gimple *next_stmt, *first_stmt; |
5711 | bool grouped_store; | |
ebfd146a | 5712 | unsigned int group_size, i; |
6e1aa848 DN |
5713 | vec<tree> oprnds = vNULL; |
5714 | vec<tree> result_chain = vNULL; | |
ebfd146a | 5715 | bool inv_p; |
09dfa495 | 5716 | tree offset = NULL_TREE; |
6e1aa848 | 5717 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 5718 | bool slp = (slp_node != NULL); |
ebfd146a | 5719 | unsigned int vec_num; |
a70d6342 | 5720 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5721 | vec_info *vinfo = stmt_info->vinfo; |
272c6793 | 5722 | tree aggr_type; |
134c85ca | 5723 | gather_scatter_info gs_info; |
3bab6342 | 5724 | enum vect_def_type scatter_src_dt = vect_unknown_def_type; |
355fe088 | 5725 | gimple *new_stmt; |
d9f21f6a | 5726 | poly_uint64 vf; |
2de001ee | 5727 | vec_load_store_type vls_type; |
44fc7854 | 5728 | tree ref_type; |
a70d6342 | 5729 | |
a70d6342 | 5730 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5731 | return false; |
5732 | ||
66c16fd9 RB |
5733 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5734 | && ! vec_stmt) | |
ebfd146a IR |
5735 | return false; |
5736 | ||
5737 | /* Is vectorizable store? */ | |
5738 | ||
5739 | if (!is_gimple_assign (stmt)) | |
5740 | return false; | |
5741 | ||
5742 | scalar_dest = gimple_assign_lhs (stmt); | |
ab0ef706 JJ |
5743 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
5744 | && is_pattern_stmt_p (stmt_info)) | |
5745 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
ebfd146a | 5746 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
38000232 | 5747 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF |
ebfd146a | 5748 | && TREE_CODE (scalar_dest) != INDIRECT_REF |
e9dbe7bb IR |
5749 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
5750 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
5751 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
5752 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
5753 | return false; |
5754 | ||
fce57248 RS |
5755 | /* Cannot have hybrid store SLP -- that would mean storing to the |
5756 | same location twice. */ | |
5757 | gcc_assert (slp == PURE_SLP_STMT (stmt_info)); | |
5758 | ||
ebfd146a | 5759 | gcc_assert (gimple_assign_single_p (stmt)); |
465c8c19 | 5760 | |
f4d09712 | 5761 | tree vectype = STMT_VINFO_VECTYPE (stmt_info), rhs_vectype = NULL_TREE; |
4d694b27 | 5762 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 JJ |
5763 | |
5764 | if (loop_vinfo) | |
b17dc4d4 RB |
5765 | { |
5766 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
5767 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
5768 | } | |
5769 | else | |
5770 | vf = 1; | |
465c8c19 JJ |
5771 | |
5772 | /* Multiple types in SLP are handled by creating the appropriate number of | |
5773 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5774 | case of SLP. */ | |
fce57248 | 5775 | if (slp) |
465c8c19 JJ |
5776 | ncopies = 1; |
5777 | else | |
e8f142e2 | 5778 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
5779 | |
5780 | gcc_assert (ncopies >= 1); | |
5781 | ||
5782 | /* FORNOW. This restriction should be relaxed. */ | |
5783 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) | |
5784 | { | |
5785 | if (dump_enabled_p ()) | |
5786 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5787 | "multiple types in nested loop.\n"); | |
5788 | return false; | |
5789 | } | |
5790 | ||
ebfd146a | 5791 | op = gimple_assign_rhs1 (stmt); |
f4d09712 | 5792 | |
2f391428 | 5793 | /* In the case this is a store from a constant make sure |
11a82e25 | 5794 | native_encode_expr can handle it. */ |
2f391428 | 5795 | if (CONSTANT_CLASS_P (op) && native_encode_expr (op, NULL, 64) == 0) |
11a82e25 RB |
5796 | return false; |
5797 | ||
f4d09712 | 5798 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt, &rhs_vectype)) |
ebfd146a | 5799 | { |
73fbfcad | 5800 | if (dump_enabled_p ()) |
78c60e3d | 5801 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5802 | "use not simple.\n"); |
ebfd146a IR |
5803 | return false; |
5804 | } | |
5805 | ||
2de001ee RS |
5806 | if (dt == vect_constant_def || dt == vect_external_def) |
5807 | vls_type = VLS_STORE_INVARIANT; | |
5808 | else | |
5809 | vls_type = VLS_STORE; | |
5810 | ||
f4d09712 KY |
5811 | if (rhs_vectype && !useless_type_conversion_p (vectype, rhs_vectype)) |
5812 | return false; | |
5813 | ||
272c6793 | 5814 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 5815 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 5816 | |
ebfd146a IR |
5817 | /* FORNOW. In some cases can vectorize even if data-type not supported |
5818 | (e.g. - array initialization with 0). */ | |
947131ba | 5819 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
5820 | return false; |
5821 | ||
5822 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
5823 | return false; | |
5824 | ||
2de001ee | 5825 | vect_memory_access_type memory_access_type; |
62da9e14 | 5826 | if (!get_load_store_type (stmt, vectype, slp, vls_type, ncopies, |
2de001ee RS |
5827 | &memory_access_type, &gs_info)) |
5828 | return false; | |
3bab6342 | 5829 | |
ebfd146a IR |
5830 | if (!vec_stmt) /* transformation not required. */ |
5831 | { | |
2de001ee | 5832 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; |
ebfd146a | 5833 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; |
2e8ab70c RB |
5834 | /* The SLP costs are calculated during SLP analysis. */ |
5835 | if (!PURE_SLP_STMT (stmt_info)) | |
2de001ee | 5836 | vect_model_store_cost (stmt_info, ncopies, memory_access_type, dt, |
2e8ab70c | 5837 | NULL, NULL, NULL); |
ebfd146a IR |
5838 | return true; |
5839 | } | |
2de001ee | 5840 | gcc_assert (memory_access_type == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); |
ebfd146a | 5841 | |
67b8dbac | 5842 | /* Transform. */ |
ebfd146a | 5843 | |
f702e7d4 | 5844 | ensure_base_align (dr); |
c716e67f | 5845 | |
2de001ee | 5846 | if (memory_access_type == VMAT_GATHER_SCATTER) |
3bab6342 AT |
5847 | { |
5848 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE, op, src; | |
134c85ca | 5849 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
5850 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; |
5851 | tree ptr, mask, var, scale, perm_mask = NULL_TREE; | |
5852 | edge pe = loop_preheader_edge (loop); | |
5853 | gimple_seq seq; | |
5854 | basic_block new_bb; | |
5855 | enum { NARROW, NONE, WIDEN } modifier; | |
4d694b27 RS |
5856 | poly_uint64 scatter_off_nunits |
5857 | = TYPE_VECTOR_SUBPARTS (gs_info.offset_vectype); | |
3bab6342 | 5858 | |
4d694b27 | 5859 | if (known_eq (nunits, scatter_off_nunits)) |
3bab6342 | 5860 | modifier = NONE; |
4d694b27 | 5861 | else if (known_eq (nunits * 2, scatter_off_nunits)) |
3bab6342 | 5862 | { |
3bab6342 AT |
5863 | modifier = WIDEN; |
5864 | ||
4d694b27 RS |
5865 | /* Currently gathers and scatters are only supported for |
5866 | fixed-length vectors. */ | |
5867 | unsigned int count = scatter_off_nunits.to_constant (); | |
5868 | vec_perm_builder sel (count, count, 1); | |
5869 | for (i = 0; i < (unsigned int) count; ++i) | |
5870 | sel.quick_push (i | (count / 2)); | |
3bab6342 | 5871 | |
4d694b27 | 5872 | vec_perm_indices indices (sel, 1, count); |
e3342de4 RS |
5873 | perm_mask = vect_gen_perm_mask_checked (gs_info.offset_vectype, |
5874 | indices); | |
3bab6342 AT |
5875 | gcc_assert (perm_mask != NULL_TREE); |
5876 | } | |
4d694b27 | 5877 | else if (known_eq (nunits, scatter_off_nunits * 2)) |
3bab6342 | 5878 | { |
3bab6342 AT |
5879 | modifier = NARROW; |
5880 | ||
4d694b27 RS |
5881 | /* Currently gathers and scatters are only supported for |
5882 | fixed-length vectors. */ | |
5883 | unsigned int count = nunits.to_constant (); | |
5884 | vec_perm_builder sel (count, count, 1); | |
5885 | for (i = 0; i < (unsigned int) count; ++i) | |
5886 | sel.quick_push (i | (count / 2)); | |
3bab6342 | 5887 | |
4d694b27 | 5888 | vec_perm_indices indices (sel, 2, count); |
e3342de4 | 5889 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); |
3bab6342 AT |
5890 | gcc_assert (perm_mask != NULL_TREE); |
5891 | ncopies *= 2; | |
5892 | } | |
5893 | else | |
5894 | gcc_unreachable (); | |
5895 | ||
134c85ca | 5896 | rettype = TREE_TYPE (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
5897 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
5898 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5899 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5900 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5901 | scaletype = TREE_VALUE (arglist); | |
5902 | ||
5903 | gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE | |
5904 | && TREE_CODE (rettype) == VOID_TYPE); | |
5905 | ||
134c85ca | 5906 | ptr = fold_convert (ptrtype, gs_info.base); |
3bab6342 AT |
5907 | if (!is_gimple_min_invariant (ptr)) |
5908 | { | |
5909 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
5910 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
5911 | gcc_assert (!new_bb); | |
5912 | } | |
5913 | ||
5914 | /* Currently we support only unconditional scatter stores, | |
5915 | so mask should be all ones. */ | |
5916 | mask = build_int_cst (masktype, -1); | |
5917 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
5918 | ||
134c85ca | 5919 | scale = build_int_cst (scaletype, gs_info.scale); |
3bab6342 AT |
5920 | |
5921 | prev_stmt_info = NULL; | |
5922 | for (j = 0; j < ncopies; ++j) | |
5923 | { | |
5924 | if (j == 0) | |
5925 | { | |
5926 | src = vec_oprnd1 | |
81c40241 | 5927 | = vect_get_vec_def_for_operand (gimple_assign_rhs1 (stmt), stmt); |
3bab6342 | 5928 | op = vec_oprnd0 |
134c85ca | 5929 | = vect_get_vec_def_for_operand (gs_info.offset, stmt); |
3bab6342 AT |
5930 | } |
5931 | else if (modifier != NONE && (j & 1)) | |
5932 | { | |
5933 | if (modifier == WIDEN) | |
5934 | { | |
5935 | src = vec_oprnd1 | |
5936 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5937 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, perm_mask, | |
5938 | stmt, gsi); | |
5939 | } | |
5940 | else if (modifier == NARROW) | |
5941 | { | |
5942 | src = permute_vec_elements (vec_oprnd1, vec_oprnd1, perm_mask, | |
5943 | stmt, gsi); | |
5944 | op = vec_oprnd0 | |
134c85ca RS |
5945 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
5946 | vec_oprnd0); | |
3bab6342 AT |
5947 | } |
5948 | else | |
5949 | gcc_unreachable (); | |
5950 | } | |
5951 | else | |
5952 | { | |
5953 | src = vec_oprnd1 | |
5954 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5955 | op = vec_oprnd0 | |
134c85ca RS |
5956 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
5957 | vec_oprnd0); | |
3bab6342 AT |
5958 | } |
5959 | ||
5960 | if (!useless_type_conversion_p (srctype, TREE_TYPE (src))) | |
5961 | { | |
5962 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)) | |
5963 | == TYPE_VECTOR_SUBPARTS (srctype)); | |
0e22bb5a | 5964 | var = vect_get_new_ssa_name (srctype, vect_simple_var); |
3bab6342 AT |
5965 | src = build1 (VIEW_CONVERT_EXPR, srctype, src); |
5966 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, src); | |
5967 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5968 | src = var; | |
5969 | } | |
5970 | ||
5971 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
5972 | { | |
5973 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
5974 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 5975 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
3bab6342 AT |
5976 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
5977 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
5978 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5979 | op = var; | |
5980 | } | |
5981 | ||
5982 | new_stmt | |
134c85ca | 5983 | = gimple_build_call (gs_info.decl, 5, ptr, mask, op, src, scale); |
3bab6342 AT |
5984 | |
5985 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5986 | ||
5987 | if (prev_stmt_info == NULL) | |
5988 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5989 | else | |
5990 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5991 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5992 | } | |
5993 | return true; | |
5994 | } | |
5995 | ||
2de001ee | 5996 | grouped_store = STMT_VINFO_GROUPED_ACCESS (stmt_info); |
0d0293ac | 5997 | if (grouped_store) |
ebfd146a | 5998 | { |
2de001ee | 5999 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a | 6000 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
e14c1050 | 6001 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 6002 | |
e14c1050 | 6003 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
6004 | |
6005 | /* FORNOW */ | |
a70d6342 | 6006 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
6007 | |
6008 | /* We vectorize all the stmts of the interleaving group when we | |
6009 | reach the last stmt in the group. */ | |
e14c1050 IR |
6010 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
6011 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
6012 | && !slp) |
6013 | { | |
6014 | *vec_stmt = NULL; | |
6015 | return true; | |
6016 | } | |
6017 | ||
6018 | if (slp) | |
4b5caab7 | 6019 | { |
0d0293ac | 6020 | grouped_store = false; |
4b5caab7 IR |
6021 | /* VEC_NUM is the number of vect stmts to be created for this |
6022 | group. */ | |
6023 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 6024 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
52eab378 | 6025 | gcc_assert (GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_stmt)) == first_stmt); |
4b5caab7 | 6026 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
d092494c | 6027 | op = gimple_assign_rhs1 (first_stmt); |
4b5caab7 | 6028 | } |
ebfd146a | 6029 | else |
4b5caab7 IR |
6030 | /* VEC_NUM is the number of vect stmts to be created for this |
6031 | group. */ | |
ebfd146a | 6032 | vec_num = group_size; |
44fc7854 BE |
6033 | |
6034 | ref_type = get_group_alias_ptr_type (first_stmt); | |
ebfd146a | 6035 | } |
b8698a0f | 6036 | else |
ebfd146a IR |
6037 | { |
6038 | first_stmt = stmt; | |
6039 | first_dr = dr; | |
6040 | group_size = vec_num = 1; | |
44fc7854 | 6041 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); |
ebfd146a | 6042 | } |
b8698a0f | 6043 | |
73fbfcad | 6044 | if (dump_enabled_p ()) |
78c60e3d | 6045 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6046 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 6047 | |
2de001ee RS |
6048 | if (memory_access_type == VMAT_ELEMENTWISE |
6049 | || memory_access_type == VMAT_STRIDED_SLP) | |
f2e2a985 MM |
6050 | { |
6051 | gimple_stmt_iterator incr_gsi; | |
6052 | bool insert_after; | |
355fe088 | 6053 | gimple *incr; |
f2e2a985 MM |
6054 | tree offvar; |
6055 | tree ivstep; | |
6056 | tree running_off; | |
6057 | gimple_seq stmts = NULL; | |
6058 | tree stride_base, stride_step, alias_off; | |
6059 | tree vec_oprnd; | |
f502d50e | 6060 | unsigned int g; |
4d694b27 RS |
6061 | /* Checked by get_load_store_type. */ |
6062 | unsigned int const_nunits = nunits.to_constant (); | |
f2e2a985 MM |
6063 | |
6064 | gcc_assert (!nested_in_vect_loop_p (loop, stmt)); | |
6065 | ||
6066 | stride_base | |
6067 | = fold_build_pointer_plus | |
f502d50e | 6068 | (unshare_expr (DR_BASE_ADDRESS (first_dr)), |
f2e2a985 | 6069 | size_binop (PLUS_EXPR, |
f502d50e | 6070 | convert_to_ptrofftype (unshare_expr (DR_OFFSET (first_dr))), |
44fc7854 | 6071 | convert_to_ptrofftype (DR_INIT (first_dr)))); |
f502d50e | 6072 | stride_step = fold_convert (sizetype, unshare_expr (DR_STEP (first_dr))); |
f2e2a985 MM |
6073 | |
6074 | /* For a store with loop-invariant (but other than power-of-2) | |
6075 | stride (i.e. not a grouped access) like so: | |
6076 | ||
6077 | for (i = 0; i < n; i += stride) | |
6078 | array[i] = ...; | |
6079 | ||
6080 | we generate a new induction variable and new stores from | |
6081 | the components of the (vectorized) rhs: | |
6082 | ||
6083 | for (j = 0; ; j += VF*stride) | |
6084 | vectemp = ...; | |
6085 | tmp1 = vectemp[0]; | |
6086 | array[j] = tmp1; | |
6087 | tmp2 = vectemp[1]; | |
6088 | array[j + stride] = tmp2; | |
6089 | ... | |
6090 | */ | |
6091 | ||
4d694b27 | 6092 | unsigned nstores = const_nunits; |
b17dc4d4 | 6093 | unsigned lnel = 1; |
cee62fee | 6094 | tree ltype = elem_type; |
04199738 | 6095 | tree lvectype = vectype; |
cee62fee MM |
6096 | if (slp) |
6097 | { | |
4d694b27 RS |
6098 | if (group_size < const_nunits |
6099 | && const_nunits % group_size == 0) | |
b17dc4d4 | 6100 | { |
4d694b27 | 6101 | nstores = const_nunits / group_size; |
b17dc4d4 RB |
6102 | lnel = group_size; |
6103 | ltype = build_vector_type (elem_type, group_size); | |
04199738 RB |
6104 | lvectype = vectype; |
6105 | ||
6106 | /* First check if vec_extract optab doesn't support extraction | |
6107 | of vector elts directly. */ | |
b397965c | 6108 | scalar_mode elmode = SCALAR_TYPE_MODE (elem_type); |
9da15d40 RS |
6109 | machine_mode vmode; |
6110 | if (!mode_for_vector (elmode, group_size).exists (&vmode) | |
6111 | || !VECTOR_MODE_P (vmode) | |
04199738 RB |
6112 | || (convert_optab_handler (vec_extract_optab, |
6113 | TYPE_MODE (vectype), vmode) | |
6114 | == CODE_FOR_nothing)) | |
6115 | { | |
6116 | /* Try to avoid emitting an extract of vector elements | |
6117 | by performing the extracts using an integer type of the | |
6118 | same size, extracting from a vector of those and then | |
6119 | re-interpreting it as the original vector type if | |
6120 | supported. */ | |
6121 | unsigned lsize | |
6122 | = group_size * GET_MODE_BITSIZE (elmode); | |
fffbab82 | 6123 | elmode = int_mode_for_size (lsize, 0).require (); |
4d694b27 | 6124 | unsigned int lnunits = const_nunits / group_size; |
04199738 RB |
6125 | /* If we can't construct such a vector fall back to |
6126 | element extracts from the original vector type and | |
6127 | element size stores. */ | |
4d694b27 | 6128 | if (mode_for_vector (elmode, lnunits).exists (&vmode) |
9da15d40 | 6129 | && VECTOR_MODE_P (vmode) |
04199738 RB |
6130 | && (convert_optab_handler (vec_extract_optab, |
6131 | vmode, elmode) | |
6132 | != CODE_FOR_nothing)) | |
6133 | { | |
4d694b27 | 6134 | nstores = lnunits; |
04199738 RB |
6135 | lnel = group_size; |
6136 | ltype = build_nonstandard_integer_type (lsize, 1); | |
6137 | lvectype = build_vector_type (ltype, nstores); | |
6138 | } | |
6139 | /* Else fall back to vector extraction anyway. | |
6140 | Fewer stores are more important than avoiding spilling | |
6141 | of the vector we extract from. Compared to the | |
6142 | construction case in vectorizable_load no store-forwarding | |
6143 | issue exists here for reasonable archs. */ | |
6144 | } | |
b17dc4d4 | 6145 | } |
4d694b27 RS |
6146 | else if (group_size >= const_nunits |
6147 | && group_size % const_nunits == 0) | |
b17dc4d4 RB |
6148 | { |
6149 | nstores = 1; | |
4d694b27 | 6150 | lnel = const_nunits; |
b17dc4d4 | 6151 | ltype = vectype; |
04199738 | 6152 | lvectype = vectype; |
b17dc4d4 | 6153 | } |
cee62fee MM |
6154 | ltype = build_aligned_type (ltype, TYPE_ALIGN (elem_type)); |
6155 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
6156 | } | |
6157 | ||
f2e2a985 MM |
6158 | ivstep = stride_step; |
6159 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
b17dc4d4 | 6160 | build_int_cst (TREE_TYPE (ivstep), vf)); |
f2e2a985 MM |
6161 | |
6162 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
6163 | ||
6164 | create_iv (stride_base, ivstep, NULL, | |
6165 | loop, &incr_gsi, insert_after, | |
6166 | &offvar, NULL); | |
6167 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 6168 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
f2e2a985 MM |
6169 | |
6170 | stride_step = force_gimple_operand (stride_step, &stmts, true, NULL_TREE); | |
6171 | if (stmts) | |
6172 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
6173 | ||
6174 | prev_stmt_info = NULL; | |
44fc7854 | 6175 | alias_off = build_int_cst (ref_type, 0); |
f502d50e MM |
6176 | next_stmt = first_stmt; |
6177 | for (g = 0; g < group_size; g++) | |
f2e2a985 | 6178 | { |
f502d50e MM |
6179 | running_off = offvar; |
6180 | if (g) | |
f2e2a985 | 6181 | { |
f502d50e MM |
6182 | tree size = TYPE_SIZE_UNIT (ltype); |
6183 | tree pos = fold_build2 (MULT_EXPR, sizetype, size_int (g), | |
f2e2a985 | 6184 | size); |
f502d50e | 6185 | tree newoff = copy_ssa_name (running_off, NULL); |
f2e2a985 | 6186 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
f502d50e | 6187 | running_off, pos); |
f2e2a985 | 6188 | vect_finish_stmt_generation (stmt, incr, gsi); |
f2e2a985 | 6189 | running_off = newoff; |
f502d50e | 6190 | } |
b17dc4d4 RB |
6191 | unsigned int group_el = 0; |
6192 | unsigned HOST_WIDE_INT | |
6193 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
f502d50e MM |
6194 | for (j = 0; j < ncopies; j++) |
6195 | { | |
6196 | /* We've set op and dt above, from gimple_assign_rhs1(stmt), | |
6197 | and first_stmt == stmt. */ | |
6198 | if (j == 0) | |
6199 | { | |
6200 | if (slp) | |
6201 | { | |
6202 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, NULL, | |
306b0c92 | 6203 | slp_node); |
f502d50e MM |
6204 | vec_oprnd = vec_oprnds[0]; |
6205 | } | |
6206 | else | |
6207 | { | |
6208 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
6209 | op = gimple_assign_rhs1 (next_stmt); | |
81c40241 | 6210 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
f502d50e MM |
6211 | } |
6212 | } | |
f2e2a985 | 6213 | else |
f502d50e MM |
6214 | { |
6215 | if (slp) | |
6216 | vec_oprnd = vec_oprnds[j]; | |
6217 | else | |
c079cbac | 6218 | { |
81c40241 | 6219 | vect_is_simple_use (vec_oprnd, vinfo, &def_stmt, &dt); |
c079cbac RB |
6220 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); |
6221 | } | |
f502d50e | 6222 | } |
04199738 RB |
6223 | /* Pun the vector to extract from if necessary. */ |
6224 | if (lvectype != vectype) | |
6225 | { | |
6226 | tree tem = make_ssa_name (lvectype); | |
6227 | gimple *pun | |
6228 | = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
6229 | lvectype, vec_oprnd)); | |
6230 | vect_finish_stmt_generation (stmt, pun, gsi); | |
6231 | vec_oprnd = tem; | |
6232 | } | |
f502d50e MM |
6233 | for (i = 0; i < nstores; i++) |
6234 | { | |
6235 | tree newref, newoff; | |
355fe088 | 6236 | gimple *incr, *assign; |
f502d50e MM |
6237 | tree size = TYPE_SIZE (ltype); |
6238 | /* Extract the i'th component. */ | |
6239 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, | |
6240 | bitsize_int (i), size); | |
6241 | tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd, | |
6242 | size, pos); | |
6243 | ||
6244 | elem = force_gimple_operand_gsi (gsi, elem, true, | |
6245 | NULL_TREE, true, | |
6246 | GSI_SAME_STMT); | |
6247 | ||
b17dc4d4 RB |
6248 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
6249 | group_el * elsz); | |
f502d50e | 6250 | newref = build2 (MEM_REF, ltype, |
b17dc4d4 | 6251 | running_off, this_off); |
f502d50e MM |
6252 | |
6253 | /* And store it to *running_off. */ | |
6254 | assign = gimple_build_assign (newref, elem); | |
6255 | vect_finish_stmt_generation (stmt, assign, gsi); | |
6256 | ||
b17dc4d4 RB |
6257 | group_el += lnel; |
6258 | if (! slp | |
6259 | || group_el == group_size) | |
6260 | { | |
6261 | newoff = copy_ssa_name (running_off, NULL); | |
6262 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
6263 | running_off, stride_step); | |
6264 | vect_finish_stmt_generation (stmt, incr, gsi); | |
f502d50e | 6265 | |
b17dc4d4 RB |
6266 | running_off = newoff; |
6267 | group_el = 0; | |
6268 | } | |
225ce44b RB |
6269 | if (g == group_size - 1 |
6270 | && !slp) | |
f502d50e MM |
6271 | { |
6272 | if (j == 0 && i == 0) | |
225ce44b RB |
6273 | STMT_VINFO_VEC_STMT (stmt_info) |
6274 | = *vec_stmt = assign; | |
f502d50e MM |
6275 | else |
6276 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = assign; | |
6277 | prev_stmt_info = vinfo_for_stmt (assign); | |
6278 | } | |
6279 | } | |
f2e2a985 | 6280 | } |
f502d50e | 6281 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
b17dc4d4 RB |
6282 | if (slp) |
6283 | break; | |
f2e2a985 | 6284 | } |
778dd3b6 RB |
6285 | |
6286 | vec_oprnds.release (); | |
f2e2a985 MM |
6287 | return true; |
6288 | } | |
6289 | ||
8c681247 | 6290 | auto_vec<tree> dr_chain (group_size); |
9771b263 | 6291 | oprnds.create (group_size); |
ebfd146a | 6292 | |
720f5239 | 6293 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 6294 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
6295 | /* Targets with store-lane instructions must not require explicit |
6296 | realignment. */ | |
2de001ee | 6297 | gcc_assert (memory_access_type != VMAT_LOAD_STORE_LANES |
272c6793 RS |
6298 | || alignment_support_scheme == dr_aligned |
6299 | || alignment_support_scheme == dr_unaligned_supported); | |
6300 | ||
62da9e14 RS |
6301 | if (memory_access_type == VMAT_CONTIGUOUS_DOWN |
6302 | || memory_access_type == VMAT_CONTIGUOUS_REVERSE) | |
09dfa495 BM |
6303 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
6304 | ||
2de001ee | 6305 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
272c6793 RS |
6306 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); |
6307 | else | |
6308 | aggr_type = vectype; | |
ebfd146a IR |
6309 | |
6310 | /* In case the vectorization factor (VF) is bigger than the number | |
6311 | of elements that we can fit in a vectype (nunits), we have to generate | |
6312 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 6313 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
6314 | vect_get_vec_def_for_copy_stmt. */ |
6315 | ||
0d0293ac | 6316 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
6317 | |
6318 | S1: &base + 2 = x2 | |
6319 | S2: &base = x0 | |
6320 | S3: &base + 1 = x1 | |
6321 | S4: &base + 3 = x3 | |
6322 | ||
6323 | We create vectorized stores starting from base address (the access of the | |
6324 | first stmt in the chain (S2 in the above example), when the last store stmt | |
6325 | of the chain (S4) is reached: | |
6326 | ||
6327 | VS1: &base = vx2 | |
6328 | VS2: &base + vec_size*1 = vx0 | |
6329 | VS3: &base + vec_size*2 = vx1 | |
6330 | VS4: &base + vec_size*3 = vx3 | |
6331 | ||
6332 | Then permutation statements are generated: | |
6333 | ||
3fcc1b55 JJ |
6334 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
6335 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 6336 | ... |
b8698a0f | 6337 | |
ebfd146a IR |
6338 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
6339 | (the order of the data-refs in the output of vect_permute_store_chain | |
6340 | corresponds to the order of scalar stmts in the interleaving chain - see | |
6341 | the documentation of vect_permute_store_chain()). | |
6342 | ||
6343 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 6344 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 6345 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 6346 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
6347 | */ |
6348 | ||
6349 | prev_stmt_info = NULL; | |
6350 | for (j = 0; j < ncopies; j++) | |
6351 | { | |
ebfd146a IR |
6352 | |
6353 | if (j == 0) | |
6354 | { | |
6355 | if (slp) | |
6356 | { | |
6357 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c | 6358 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
306b0c92 | 6359 | NULL, slp_node); |
ebfd146a | 6360 | |
9771b263 | 6361 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
6362 | } |
6363 | else | |
6364 | { | |
b8698a0f L |
6365 | /* For interleaved stores we collect vectorized defs for all the |
6366 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
6367 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
6368 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
6369 | ||
0d0293ac | 6370 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 6371 | OPRNDS are of size 1. */ |
b8698a0f | 6372 | next_stmt = first_stmt; |
ebfd146a IR |
6373 | for (i = 0; i < group_size; i++) |
6374 | { | |
b8698a0f L |
6375 | /* Since gaps are not supported for interleaved stores, |
6376 | GROUP_SIZE is the exact number of stmts in the chain. | |
6377 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
6378 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
6379 | iteration of the loop will be executed. */ |
6380 | gcc_assert (next_stmt | |
6381 | && gimple_assign_single_p (next_stmt)); | |
6382 | op = gimple_assign_rhs1 (next_stmt); | |
6383 | ||
81c40241 | 6384 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
9771b263 DN |
6385 | dr_chain.quick_push (vec_oprnd); |
6386 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 6387 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
6388 | } |
6389 | } | |
6390 | ||
6391 | /* We should have catched mismatched types earlier. */ | |
6392 | gcc_assert (useless_type_conversion_p (vectype, | |
6393 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
6394 | bool simd_lane_access_p |
6395 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
6396 | if (simd_lane_access_p | |
6397 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
6398 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
6399 | && integer_zerop (DR_OFFSET (first_dr)) | |
6400 | && integer_zerop (DR_INIT (first_dr)) | |
6401 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
44fc7854 | 6402 | get_alias_set (TREE_TYPE (ref_type)))) |
74bf76ed JJ |
6403 | { |
6404 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
44fc7854 | 6405 | dataref_offset = build_int_cst (ref_type, 0); |
8928eff3 | 6406 | inv_p = false; |
74bf76ed JJ |
6407 | } |
6408 | else | |
6409 | dataref_ptr | |
6410 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
6411 | simd_lane_access_p ? loop : NULL, | |
09dfa495 | 6412 | offset, &dummy, gsi, &ptr_incr, |
74bf76ed | 6413 | simd_lane_access_p, &inv_p); |
a70d6342 | 6414 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 6415 | } |
b8698a0f | 6416 | else |
ebfd146a | 6417 | { |
b8698a0f L |
6418 | /* For interleaved stores we created vectorized defs for all the |
6419 | defs stored in OPRNDS in the previous iteration (previous copy). | |
6420 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
6421 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
6422 | next copy. | |
0d0293ac | 6423 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
6424 | OPRNDS are of size 1. */ |
6425 | for (i = 0; i < group_size; i++) | |
6426 | { | |
9771b263 | 6427 | op = oprnds[i]; |
81c40241 | 6428 | vect_is_simple_use (op, vinfo, &def_stmt, &dt); |
b8698a0f | 6429 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
9771b263 DN |
6430 | dr_chain[i] = vec_oprnd; |
6431 | oprnds[i] = vec_oprnd; | |
ebfd146a | 6432 | } |
74bf76ed JJ |
6433 | if (dataref_offset) |
6434 | dataref_offset | |
6435 | = int_const_binop (PLUS_EXPR, dataref_offset, | |
6436 | TYPE_SIZE_UNIT (aggr_type)); | |
6437 | else | |
6438 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
6439 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
6440 | } |
6441 | ||
2de001ee | 6442 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
ebfd146a | 6443 | { |
272c6793 | 6444 | tree vec_array; |
267d3070 | 6445 | |
272c6793 RS |
6446 | /* Combine all the vectors into an array. */ |
6447 | vec_array = create_vector_array (vectype, vec_num); | |
6448 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 6449 | { |
9771b263 | 6450 | vec_oprnd = dr_chain[i]; |
272c6793 | 6451 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 6452 | } |
b8698a0f | 6453 | |
272c6793 RS |
6454 | /* Emit: |
6455 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
44fc7854 | 6456 | data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type); |
a844293d RS |
6457 | gcall *call = gimple_build_call_internal (IFN_STORE_LANES, 1, |
6458 | vec_array); | |
6459 | gimple_call_set_lhs (call, data_ref); | |
6460 | gimple_call_set_nothrow (call, true); | |
6461 | new_stmt = call; | |
267d3070 | 6462 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
6463 | } |
6464 | else | |
6465 | { | |
6466 | new_stmt = NULL; | |
0d0293ac | 6467 | if (grouped_store) |
272c6793 | 6468 | { |
b6b9227d JJ |
6469 | if (j == 0) |
6470 | result_chain.create (group_size); | |
272c6793 RS |
6471 | /* Permute. */ |
6472 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
6473 | &result_chain); | |
6474 | } | |
c2d7ab2a | 6475 | |
272c6793 RS |
6476 | next_stmt = first_stmt; |
6477 | for (i = 0; i < vec_num; i++) | |
6478 | { | |
644ffefd | 6479 | unsigned align, misalign; |
272c6793 RS |
6480 | |
6481 | if (i > 0) | |
6482 | /* Bump the vector pointer. */ | |
6483 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
6484 | stmt, NULL_TREE); | |
6485 | ||
6486 | if (slp) | |
9771b263 | 6487 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
6488 | else if (grouped_store) |
6489 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 6490 | vect_permute_store_chain(). */ |
9771b263 | 6491 | vec_oprnd = result_chain[i]; |
272c6793 | 6492 | |
69a2e8a1 | 6493 | data_ref = fold_build2 (MEM_REF, vectype, |
aed93b23 RB |
6494 | dataref_ptr, |
6495 | dataref_offset | |
6496 | ? dataref_offset | |
44fc7854 | 6497 | : build_int_cst (ref_type, 0)); |
f702e7d4 | 6498 | align = DR_TARGET_ALIGNMENT (first_dr); |
272c6793 | 6499 | if (aligned_access_p (first_dr)) |
644ffefd | 6500 | misalign = 0; |
272c6793 RS |
6501 | else if (DR_MISALIGNMENT (first_dr) == -1) |
6502 | { | |
25f68d90 | 6503 | align = dr_alignment (vect_dr_behavior (first_dr)); |
52639a61 | 6504 | misalign = 0; |
272c6793 RS |
6505 | TREE_TYPE (data_ref) |
6506 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 6507 | align * BITS_PER_UNIT); |
272c6793 RS |
6508 | } |
6509 | else | |
6510 | { | |
6511 | TREE_TYPE (data_ref) | |
6512 | = build_aligned_type (TREE_TYPE (data_ref), | |
6513 | TYPE_ALIGN (elem_type)); | |
644ffefd | 6514 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 6515 | } |
aed93b23 RB |
6516 | if (dataref_offset == NULL_TREE |
6517 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
6518 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
6519 | misalign); | |
c2d7ab2a | 6520 | |
62da9e14 | 6521 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
09dfa495 BM |
6522 | { |
6523 | tree perm_mask = perm_mask_for_reverse (vectype); | |
6524 | tree perm_dest | |
6525 | = vect_create_destination_var (gimple_assign_rhs1 (stmt), | |
6526 | vectype); | |
b731b390 | 6527 | tree new_temp = make_ssa_name (perm_dest); |
09dfa495 BM |
6528 | |
6529 | /* Generate the permute statement. */ | |
355fe088 | 6530 | gimple *perm_stmt |
0d0e4a03 JJ |
6531 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
6532 | vec_oprnd, perm_mask); | |
09dfa495 BM |
6533 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
6534 | ||
6535 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); | |
6536 | vec_oprnd = new_temp; | |
6537 | } | |
6538 | ||
272c6793 RS |
6539 | /* Arguments are ready. Create the new vector stmt. */ |
6540 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
6541 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
6542 | |
6543 | if (slp) | |
6544 | continue; | |
6545 | ||
e14c1050 | 6546 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
6547 | if (!next_stmt) |
6548 | break; | |
6549 | } | |
ebfd146a | 6550 | } |
1da0876c RS |
6551 | if (!slp) |
6552 | { | |
6553 | if (j == 0) | |
6554 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6555 | else | |
6556 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6557 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6558 | } | |
ebfd146a IR |
6559 | } |
6560 | ||
9771b263 DN |
6561 | oprnds.release (); |
6562 | result_chain.release (); | |
6563 | vec_oprnds.release (); | |
ebfd146a IR |
6564 | |
6565 | return true; | |
6566 | } | |
6567 | ||
557be5a8 AL |
6568 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
6569 | VECTOR_CST mask. No checks are made that the target platform supports the | |
7ac7e286 | 6570 | mask, so callers may wish to test can_vec_perm_const_p separately, or use |
557be5a8 | 6571 | vect_gen_perm_mask_checked. */ |
a1e53f3f | 6572 | |
3fcc1b55 | 6573 | tree |
4aae3cb3 | 6574 | vect_gen_perm_mask_any (tree vectype, const vec_perm_indices &sel) |
a1e53f3f | 6575 | { |
b00cb3bf | 6576 | tree mask_type; |
a1e53f3f | 6577 | |
b00cb3bf RS |
6578 | unsigned int nunits = sel.length (); |
6579 | gcc_assert (nunits == TYPE_VECTOR_SUBPARTS (vectype)); | |
6580 | ||
6581 | mask_type = build_vector_type (ssizetype, nunits); | |
736d0f28 | 6582 | return vec_perm_indices_to_tree (mask_type, sel); |
a1e53f3f L |
6583 | } |
6584 | ||
7ac7e286 | 6585 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_const_p, |
cf7aa6a3 | 6586 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ |
557be5a8 AL |
6587 | |
6588 | tree | |
4aae3cb3 | 6589 | vect_gen_perm_mask_checked (tree vectype, const vec_perm_indices &sel) |
557be5a8 | 6590 | { |
7ac7e286 | 6591 | gcc_assert (can_vec_perm_const_p (TYPE_MODE (vectype), sel)); |
557be5a8 AL |
6592 | return vect_gen_perm_mask_any (vectype, sel); |
6593 | } | |
6594 | ||
aec7ae7d JJ |
6595 | /* Given a vector variable X and Y, that was generated for the scalar |
6596 | STMT, generate instructions to permute the vector elements of X and Y | |
6597 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
6598 | permuted vector variable. */ | |
a1e53f3f L |
6599 | |
6600 | static tree | |
355fe088 | 6601 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple *stmt, |
aec7ae7d | 6602 | gimple_stmt_iterator *gsi) |
a1e53f3f L |
6603 | { |
6604 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 6605 | tree perm_dest, data_ref; |
355fe088 | 6606 | gimple *perm_stmt; |
a1e53f3f | 6607 | |
acdcd61b | 6608 | perm_dest = vect_create_destination_var (gimple_get_lhs (stmt), vectype); |
b731b390 | 6609 | data_ref = make_ssa_name (perm_dest); |
a1e53f3f L |
6610 | |
6611 | /* Generate the permute statement. */ | |
0d0e4a03 | 6612 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
a1e53f3f L |
6613 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
6614 | ||
6615 | return data_ref; | |
6616 | } | |
6617 | ||
6b916b36 RB |
6618 | /* Hoist the definitions of all SSA uses on STMT out of the loop LOOP, |
6619 | inserting them on the loops preheader edge. Returns true if we | |
6620 | were successful in doing so (and thus STMT can be moved then), | |
6621 | otherwise returns false. */ | |
6622 | ||
6623 | static bool | |
355fe088 | 6624 | hoist_defs_of_uses (gimple *stmt, struct loop *loop) |
6b916b36 RB |
6625 | { |
6626 | ssa_op_iter i; | |
6627 | tree op; | |
6628 | bool any = false; | |
6629 | ||
6630 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6631 | { | |
355fe088 | 6632 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6633 | if (!gimple_nop_p (def_stmt) |
6634 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6635 | { | |
6636 | /* Make sure we don't need to recurse. While we could do | |
6637 | so in simple cases when there are more complex use webs | |
6638 | we don't have an easy way to preserve stmt order to fulfil | |
6639 | dependencies within them. */ | |
6640 | tree op2; | |
6641 | ssa_op_iter i2; | |
d1417442 JJ |
6642 | if (gimple_code (def_stmt) == GIMPLE_PHI) |
6643 | return false; | |
6b916b36 RB |
6644 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
6645 | { | |
355fe088 | 6646 | gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2); |
6b916b36 RB |
6647 | if (!gimple_nop_p (def_stmt2) |
6648 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2))) | |
6649 | return false; | |
6650 | } | |
6651 | any = true; | |
6652 | } | |
6653 | } | |
6654 | ||
6655 | if (!any) | |
6656 | return true; | |
6657 | ||
6658 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6659 | { | |
355fe088 | 6660 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6661 | if (!gimple_nop_p (def_stmt) |
6662 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6663 | { | |
6664 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); | |
6665 | gsi_remove (&gsi, false); | |
6666 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); | |
6667 | } | |
6668 | } | |
6669 | ||
6670 | return true; | |
6671 | } | |
6672 | ||
ebfd146a IR |
6673 | /* vectorizable_load. |
6674 | ||
b8698a0f L |
6675 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
6676 | can be vectorized. | |
6677 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
6678 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
6679 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
6680 | ||
6681 | static bool | |
355fe088 | 6682 | vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 6683 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a IR |
6684 | { |
6685 | tree scalar_dest; | |
6686 | tree vec_dest = NULL; | |
6687 | tree data_ref = NULL; | |
6688 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 6689 | stmt_vec_info prev_stmt_info; |
ebfd146a | 6690 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 6691 | struct loop *loop = NULL; |
ebfd146a | 6692 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 6693 | bool nested_in_vect_loop = false; |
c716e67f | 6694 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
272c6793 | 6695 | tree elem_type; |
ebfd146a | 6696 | tree new_temp; |
ef4bddc2 | 6697 | machine_mode mode; |
355fe088 | 6698 | gimple *new_stmt = NULL; |
ebfd146a IR |
6699 | tree dummy; |
6700 | enum dr_alignment_support alignment_support_scheme; | |
6701 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 6702 | tree dataref_offset = NULL_TREE; |
355fe088 | 6703 | gimple *ptr_incr = NULL; |
ebfd146a | 6704 | int ncopies; |
4d694b27 RS |
6705 | int i, j; |
6706 | unsigned int group_size; | |
6707 | poly_uint64 group_gap_adj; | |
ebfd146a IR |
6708 | tree msq = NULL_TREE, lsq; |
6709 | tree offset = NULL_TREE; | |
356bbc4c | 6710 | tree byte_offset = NULL_TREE; |
ebfd146a | 6711 | tree realignment_token = NULL_TREE; |
538dd0b7 | 6712 | gphi *phi = NULL; |
6e1aa848 | 6713 | vec<tree> dr_chain = vNULL; |
0d0293ac | 6714 | bool grouped_load = false; |
355fe088 | 6715 | gimple *first_stmt; |
4f0a0218 | 6716 | gimple *first_stmt_for_drptr = NULL; |
ebfd146a IR |
6717 | bool inv_p; |
6718 | bool compute_in_loop = false; | |
6719 | struct loop *at_loop; | |
6720 | int vec_num; | |
6721 | bool slp = (slp_node != NULL); | |
6722 | bool slp_perm = false; | |
6723 | enum tree_code code; | |
a70d6342 | 6724 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
d9f21f6a | 6725 | poly_uint64 vf; |
272c6793 | 6726 | tree aggr_type; |
134c85ca | 6727 | gather_scatter_info gs_info; |
310213d4 | 6728 | vec_info *vinfo = stmt_info->vinfo; |
44fc7854 | 6729 | tree ref_type; |
a70d6342 | 6730 | |
465c8c19 JJ |
6731 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
6732 | return false; | |
6733 | ||
66c16fd9 RB |
6734 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
6735 | && ! vec_stmt) | |
465c8c19 JJ |
6736 | return false; |
6737 | ||
6738 | /* Is vectorizable load? */ | |
6739 | if (!is_gimple_assign (stmt)) | |
6740 | return false; | |
6741 | ||
6742 | scalar_dest = gimple_assign_lhs (stmt); | |
6743 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
6744 | return false; | |
6745 | ||
6746 | code = gimple_assign_rhs_code (stmt); | |
6747 | if (code != ARRAY_REF | |
6748 | && code != BIT_FIELD_REF | |
6749 | && code != INDIRECT_REF | |
6750 | && code != COMPONENT_REF | |
6751 | && code != IMAGPART_EXPR | |
6752 | && code != REALPART_EXPR | |
6753 | && code != MEM_REF | |
6754 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
6755 | return false; | |
6756 | ||
6757 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
6758 | return false; | |
6759 | ||
6760 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
4d694b27 | 6761 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 | 6762 | |
a70d6342 IR |
6763 | if (loop_vinfo) |
6764 | { | |
6765 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
6766 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
6767 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
6768 | } | |
6769 | else | |
3533e503 | 6770 | vf = 1; |
ebfd146a IR |
6771 | |
6772 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 6773 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 6774 | case of SLP. */ |
fce57248 | 6775 | if (slp) |
ebfd146a IR |
6776 | ncopies = 1; |
6777 | else | |
e8f142e2 | 6778 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
ebfd146a IR |
6779 | |
6780 | gcc_assert (ncopies >= 1); | |
6781 | ||
6782 | /* FORNOW. This restriction should be relaxed. */ | |
6783 | if (nested_in_vect_loop && ncopies > 1) | |
6784 | { | |
73fbfcad | 6785 | if (dump_enabled_p ()) |
78c60e3d | 6786 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6787 | "multiple types in nested loop.\n"); |
ebfd146a IR |
6788 | return false; |
6789 | } | |
6790 | ||
f2556b68 RB |
6791 | /* Invalidate assumptions made by dependence analysis when vectorization |
6792 | on the unrolled body effectively re-orders stmts. */ | |
6793 | if (ncopies > 1 | |
6794 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
d9f21f6a RS |
6795 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
6796 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
f2556b68 RB |
6797 | { |
6798 | if (dump_enabled_p ()) | |
6799 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6800 | "cannot perform implicit CSE when unrolling " | |
6801 | "with negative dependence distance\n"); | |
6802 | return false; | |
6803 | } | |
6804 | ||
7b7b1813 | 6805 | elem_type = TREE_TYPE (vectype); |
947131ba | 6806 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
6807 | |
6808 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
6809 | (e.g. - data copies). */ | |
947131ba | 6810 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 6811 | { |
73fbfcad | 6812 | if (dump_enabled_p ()) |
78c60e3d | 6813 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6814 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
6815 | return false; |
6816 | } | |
6817 | ||
ebfd146a | 6818 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 6819 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 6820 | { |
0d0293ac | 6821 | grouped_load = true; |
ebfd146a | 6822 | /* FORNOW */ |
2de001ee RS |
6823 | gcc_assert (!nested_in_vect_loop); |
6824 | gcc_assert (!STMT_VINFO_GATHER_SCATTER_P (stmt_info)); | |
ebfd146a | 6825 | |
e14c1050 | 6826 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
d3465d72 | 6827 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
d5f035ea | 6828 | |
b1af7da6 RB |
6829 | if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
6830 | slp_perm = true; | |
6831 | ||
f2556b68 RB |
6832 | /* Invalidate assumptions made by dependence analysis when vectorization |
6833 | on the unrolled body effectively re-orders stmts. */ | |
6834 | if (!PURE_SLP_STMT (stmt_info) | |
6835 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
d9f21f6a RS |
6836 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
6837 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
f2556b68 RB |
6838 | { |
6839 | if (dump_enabled_p ()) | |
6840 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6841 | "cannot perform implicit CSE when performing " | |
6842 | "group loads with negative dependence distance\n"); | |
6843 | return false; | |
6844 | } | |
96bb56b2 RB |
6845 | |
6846 | /* Similarly when the stmt is a load that is both part of a SLP | |
6847 | instance and a loop vectorized stmt via the same-dr mechanism | |
6848 | we have to give up. */ | |
6849 | if (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info) | |
6850 | && (STMT_SLP_TYPE (stmt_info) | |
6851 | != STMT_SLP_TYPE (vinfo_for_stmt | |
6852 | (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info))))) | |
6853 | { | |
6854 | if (dump_enabled_p ()) | |
6855 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6856 | "conflicting SLP types for CSEd load\n"); | |
6857 | return false; | |
6858 | } | |
ebfd146a IR |
6859 | } |
6860 | ||
2de001ee | 6861 | vect_memory_access_type memory_access_type; |
62da9e14 | 6862 | if (!get_load_store_type (stmt, vectype, slp, VLS_LOAD, ncopies, |
2de001ee RS |
6863 | &memory_access_type, &gs_info)) |
6864 | return false; | |
a1e53f3f | 6865 | |
ebfd146a IR |
6866 | if (!vec_stmt) /* transformation not required. */ |
6867 | { | |
2de001ee RS |
6868 | if (!slp) |
6869 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; | |
ebfd146a | 6870 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; |
2e8ab70c RB |
6871 | /* The SLP costs are calculated during SLP analysis. */ |
6872 | if (!PURE_SLP_STMT (stmt_info)) | |
2de001ee | 6873 | vect_model_load_cost (stmt_info, ncopies, memory_access_type, |
2e8ab70c | 6874 | NULL, NULL, NULL); |
ebfd146a IR |
6875 | return true; |
6876 | } | |
6877 | ||
2de001ee RS |
6878 | if (!slp) |
6879 | gcc_assert (memory_access_type | |
6880 | == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); | |
6881 | ||
73fbfcad | 6882 | if (dump_enabled_p ()) |
78c60e3d | 6883 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6884 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a | 6885 | |
67b8dbac | 6886 | /* Transform. */ |
ebfd146a | 6887 | |
f702e7d4 | 6888 | ensure_base_align (dr); |
c716e67f | 6889 | |
2de001ee | 6890 | if (memory_access_type == VMAT_GATHER_SCATTER) |
aec7ae7d JJ |
6891 | { |
6892 | tree vec_oprnd0 = NULL_TREE, op; | |
134c85ca | 6893 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); |
aec7ae7d | 6894 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; |
d3c2fee0 | 6895 | tree ptr, mask, var, scale, merge, perm_mask = NULL_TREE, prev_res = NULL_TREE; |
aec7ae7d JJ |
6896 | edge pe = loop_preheader_edge (loop); |
6897 | gimple_seq seq; | |
6898 | basic_block new_bb; | |
6899 | enum { NARROW, NONE, WIDEN } modifier; | |
4d694b27 RS |
6900 | poly_uint64 gather_off_nunits |
6901 | = TYPE_VECTOR_SUBPARTS (gs_info.offset_vectype); | |
aec7ae7d | 6902 | |
4d694b27 | 6903 | if (known_eq (nunits, gather_off_nunits)) |
aec7ae7d | 6904 | modifier = NONE; |
4d694b27 | 6905 | else if (known_eq (nunits * 2, gather_off_nunits)) |
aec7ae7d | 6906 | { |
aec7ae7d JJ |
6907 | modifier = WIDEN; |
6908 | ||
4d694b27 RS |
6909 | /* Currently widening gathers are only supported for |
6910 | fixed-length vectors. */ | |
6911 | int count = gather_off_nunits.to_constant (); | |
6912 | vec_perm_builder sel (count, count, 1); | |
6913 | for (i = 0; i < count; ++i) | |
6914 | sel.quick_push (i | (count / 2)); | |
aec7ae7d | 6915 | |
4d694b27 | 6916 | vec_perm_indices indices (sel, 1, count); |
e3342de4 RS |
6917 | perm_mask = vect_gen_perm_mask_checked (gs_info.offset_vectype, |
6918 | indices); | |
aec7ae7d | 6919 | } |
4d694b27 | 6920 | else if (known_eq (nunits, gather_off_nunits * 2)) |
aec7ae7d | 6921 | { |
aec7ae7d JJ |
6922 | modifier = NARROW; |
6923 | ||
4d694b27 RS |
6924 | /* Currently narrowing gathers are only supported for |
6925 | fixed-length vectors. */ | |
6926 | int count = nunits.to_constant (); | |
6927 | vec_perm_builder sel (count, count, 1); | |
6928 | for (i = 0; i < count; ++i) | |
6929 | sel.quick_push (i < count / 2 ? i : i + count / 2); | |
aec7ae7d | 6930 | |
4d694b27 | 6931 | vec_perm_indices indices (sel, 2, count); |
e3342de4 | 6932 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); |
aec7ae7d JJ |
6933 | ncopies *= 2; |
6934 | } | |
6935 | else | |
6936 | gcc_unreachable (); | |
6937 | ||
134c85ca | 6938 | rettype = TREE_TYPE (TREE_TYPE (gs_info.decl)); |
aec7ae7d JJ |
6939 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
6940 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6941 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6942 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6943 | scaletype = TREE_VALUE (arglist); | |
d3c2fee0 | 6944 | gcc_checking_assert (types_compatible_p (srctype, rettype)); |
aec7ae7d JJ |
6945 | |
6946 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
6947 | ||
134c85ca | 6948 | ptr = fold_convert (ptrtype, gs_info.base); |
aec7ae7d JJ |
6949 | if (!is_gimple_min_invariant (ptr)) |
6950 | { | |
6951 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
6952 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
6953 | gcc_assert (!new_bb); | |
6954 | } | |
6955 | ||
6956 | /* Currently we support only unconditional gather loads, | |
6957 | so mask should be all ones. */ | |
d3c2fee0 AI |
6958 | if (TREE_CODE (masktype) == INTEGER_TYPE) |
6959 | mask = build_int_cst (masktype, -1); | |
6960 | else if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
6961 | { | |
6962 | mask = build_int_cst (TREE_TYPE (masktype), -1); | |
6963 | mask = build_vector_from_val (masktype, mask); | |
03b9e8e4 | 6964 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
d3c2fee0 | 6965 | } |
aec7ae7d JJ |
6966 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) |
6967 | { | |
6968 | REAL_VALUE_TYPE r; | |
6969 | long tmp[6]; | |
6970 | for (j = 0; j < 6; ++j) | |
6971 | tmp[j] = -1; | |
6972 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
6973 | mask = build_real (TREE_TYPE (masktype), r); | |
d3c2fee0 | 6974 | mask = build_vector_from_val (masktype, mask); |
03b9e8e4 | 6975 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
aec7ae7d JJ |
6976 | } |
6977 | else | |
6978 | gcc_unreachable (); | |
aec7ae7d | 6979 | |
134c85ca | 6980 | scale = build_int_cst (scaletype, gs_info.scale); |
aec7ae7d | 6981 | |
d3c2fee0 AI |
6982 | if (TREE_CODE (TREE_TYPE (rettype)) == INTEGER_TYPE) |
6983 | merge = build_int_cst (TREE_TYPE (rettype), 0); | |
6984 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (rettype))) | |
6985 | { | |
6986 | REAL_VALUE_TYPE r; | |
6987 | long tmp[6]; | |
6988 | for (j = 0; j < 6; ++j) | |
6989 | tmp[j] = 0; | |
6990 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (rettype))); | |
6991 | merge = build_real (TREE_TYPE (rettype), r); | |
6992 | } | |
6993 | else | |
6994 | gcc_unreachable (); | |
6995 | merge = build_vector_from_val (rettype, merge); | |
6996 | merge = vect_init_vector (stmt, merge, rettype, NULL); | |
6997 | ||
aec7ae7d JJ |
6998 | prev_stmt_info = NULL; |
6999 | for (j = 0; j < ncopies; ++j) | |
7000 | { | |
7001 | if (modifier == WIDEN && (j & 1)) | |
7002 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
7003 | perm_mask, stmt, gsi); | |
7004 | else if (j == 0) | |
7005 | op = vec_oprnd0 | |
134c85ca | 7006 | = vect_get_vec_def_for_operand (gs_info.offset, stmt); |
aec7ae7d JJ |
7007 | else |
7008 | op = vec_oprnd0 | |
134c85ca | 7009 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, vec_oprnd0); |
aec7ae7d JJ |
7010 | |
7011 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
7012 | { | |
7013 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
7014 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 7015 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
aec7ae7d JJ |
7016 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
7017 | new_stmt | |
0d0e4a03 | 7018 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
7019 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7020 | op = var; | |
7021 | } | |
7022 | ||
7023 | new_stmt | |
134c85ca | 7024 | = gimple_build_call (gs_info.decl, 5, merge, ptr, op, mask, scale); |
aec7ae7d JJ |
7025 | |
7026 | if (!useless_type_conversion_p (vectype, rettype)) | |
7027 | { | |
7028 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
7029 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 7030 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
aec7ae7d JJ |
7031 | gimple_call_set_lhs (new_stmt, op); |
7032 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 7033 | var = make_ssa_name (vec_dest); |
aec7ae7d JJ |
7034 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
7035 | new_stmt | |
0d0e4a03 | 7036 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
7037 | } |
7038 | else | |
7039 | { | |
7040 | var = make_ssa_name (vec_dest, new_stmt); | |
7041 | gimple_call_set_lhs (new_stmt, var); | |
7042 | } | |
7043 | ||
7044 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7045 | ||
7046 | if (modifier == NARROW) | |
7047 | { | |
7048 | if ((j & 1) == 0) | |
7049 | { | |
7050 | prev_res = var; | |
7051 | continue; | |
7052 | } | |
7053 | var = permute_vec_elements (prev_res, var, | |
7054 | perm_mask, stmt, gsi); | |
7055 | new_stmt = SSA_NAME_DEF_STMT (var); | |
7056 | } | |
7057 | ||
7058 | if (prev_stmt_info == NULL) | |
7059 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7060 | else | |
7061 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7062 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7063 | } | |
7064 | return true; | |
7065 | } | |
2de001ee RS |
7066 | |
7067 | if (memory_access_type == VMAT_ELEMENTWISE | |
7068 | || memory_access_type == VMAT_STRIDED_SLP) | |
7d75abc8 MM |
7069 | { |
7070 | gimple_stmt_iterator incr_gsi; | |
7071 | bool insert_after; | |
355fe088 | 7072 | gimple *incr; |
7d75abc8 | 7073 | tree offvar; |
7d75abc8 MM |
7074 | tree ivstep; |
7075 | tree running_off; | |
9771b263 | 7076 | vec<constructor_elt, va_gc> *v = NULL; |
7d75abc8 | 7077 | gimple_seq stmts = NULL; |
14ac6aa2 | 7078 | tree stride_base, stride_step, alias_off; |
4d694b27 RS |
7079 | /* Checked by get_load_store_type. */ |
7080 | unsigned int const_nunits = nunits.to_constant (); | |
14ac6aa2 RB |
7081 | |
7082 | gcc_assert (!nested_in_vect_loop); | |
7d75abc8 | 7083 | |
f502d50e | 7084 | if (slp && grouped_load) |
44fc7854 BE |
7085 | { |
7086 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); | |
7087 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
7088 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
7089 | ref_type = get_group_alias_ptr_type (first_stmt); | |
7090 | } | |
ab313a8c | 7091 | else |
44fc7854 BE |
7092 | { |
7093 | first_stmt = stmt; | |
7094 | first_dr = dr; | |
7095 | group_size = 1; | |
7096 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); | |
7097 | } | |
ab313a8c | 7098 | |
14ac6aa2 RB |
7099 | stride_base |
7100 | = fold_build_pointer_plus | |
ab313a8c | 7101 | (DR_BASE_ADDRESS (first_dr), |
14ac6aa2 | 7102 | size_binop (PLUS_EXPR, |
ab313a8c RB |
7103 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
7104 | convert_to_ptrofftype (DR_INIT (first_dr)))); | |
7105 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); | |
7d75abc8 MM |
7106 | |
7107 | /* For a load with loop-invariant (but other than power-of-2) | |
7108 | stride (i.e. not a grouped access) like so: | |
7109 | ||
7110 | for (i = 0; i < n; i += stride) | |
7111 | ... = array[i]; | |
7112 | ||
7113 | we generate a new induction variable and new accesses to | |
7114 | form a new vector (or vectors, depending on ncopies): | |
7115 | ||
7116 | for (j = 0; ; j += VF*stride) | |
7117 | tmp1 = array[j]; | |
7118 | tmp2 = array[j + stride]; | |
7119 | ... | |
7120 | vectemp = {tmp1, tmp2, ...} | |
7121 | */ | |
7122 | ||
ab313a8c RB |
7123 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step, |
7124 | build_int_cst (TREE_TYPE (stride_step), vf)); | |
7d75abc8 MM |
7125 | |
7126 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
7127 | ||
ab313a8c | 7128 | create_iv (unshare_expr (stride_base), unshare_expr (ivstep), NULL, |
7d75abc8 MM |
7129 | loop, &incr_gsi, insert_after, |
7130 | &offvar, NULL); | |
7131 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 7132 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
7d75abc8 | 7133 | |
ab313a8c RB |
7134 | stride_step = force_gimple_operand (unshare_expr (stride_step), |
7135 | &stmts, true, NULL_TREE); | |
7d75abc8 MM |
7136 | if (stmts) |
7137 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
7138 | ||
7139 | prev_stmt_info = NULL; | |
7140 | running_off = offvar; | |
44fc7854 | 7141 | alias_off = build_int_cst (ref_type, 0); |
4d694b27 | 7142 | int nloads = const_nunits; |
e09b4c37 | 7143 | int lnel = 1; |
7b5fc413 | 7144 | tree ltype = TREE_TYPE (vectype); |
ea60dd34 | 7145 | tree lvectype = vectype; |
b266b968 | 7146 | auto_vec<tree> dr_chain; |
2de001ee | 7147 | if (memory_access_type == VMAT_STRIDED_SLP) |
7b5fc413 | 7148 | { |
4d694b27 | 7149 | if (group_size < const_nunits) |
e09b4c37 | 7150 | { |
ff03930a JJ |
7151 | /* First check if vec_init optab supports construction from |
7152 | vector elts directly. */ | |
b397965c | 7153 | scalar_mode elmode = SCALAR_TYPE_MODE (TREE_TYPE (vectype)); |
9da15d40 RS |
7154 | machine_mode vmode; |
7155 | if (mode_for_vector (elmode, group_size).exists (&vmode) | |
7156 | && VECTOR_MODE_P (vmode) | |
ff03930a JJ |
7157 | && (convert_optab_handler (vec_init_optab, |
7158 | TYPE_MODE (vectype), vmode) | |
7159 | != CODE_FOR_nothing)) | |
ea60dd34 | 7160 | { |
4d694b27 | 7161 | nloads = const_nunits / group_size; |
ea60dd34 | 7162 | lnel = group_size; |
ff03930a JJ |
7163 | ltype = build_vector_type (TREE_TYPE (vectype), group_size); |
7164 | } | |
7165 | else | |
7166 | { | |
7167 | /* Otherwise avoid emitting a constructor of vector elements | |
7168 | by performing the loads using an integer type of the same | |
7169 | size, constructing a vector of those and then | |
7170 | re-interpreting it as the original vector type. | |
7171 | This avoids a huge runtime penalty due to the general | |
7172 | inability to perform store forwarding from smaller stores | |
7173 | to a larger load. */ | |
7174 | unsigned lsize | |
7175 | = group_size * TYPE_PRECISION (TREE_TYPE (vectype)); | |
fffbab82 | 7176 | elmode = int_mode_for_size (lsize, 0).require (); |
4d694b27 | 7177 | unsigned int lnunits = const_nunits / group_size; |
ff03930a JJ |
7178 | /* If we can't construct such a vector fall back to |
7179 | element loads of the original vector type. */ | |
4d694b27 | 7180 | if (mode_for_vector (elmode, lnunits).exists (&vmode) |
9da15d40 | 7181 | && VECTOR_MODE_P (vmode) |
ff03930a JJ |
7182 | && (convert_optab_handler (vec_init_optab, vmode, elmode) |
7183 | != CODE_FOR_nothing)) | |
7184 | { | |
4d694b27 | 7185 | nloads = lnunits; |
ff03930a JJ |
7186 | lnel = group_size; |
7187 | ltype = build_nonstandard_integer_type (lsize, 1); | |
7188 | lvectype = build_vector_type (ltype, nloads); | |
7189 | } | |
ea60dd34 | 7190 | } |
e09b4c37 | 7191 | } |
2de001ee | 7192 | else |
e09b4c37 | 7193 | { |
ea60dd34 | 7194 | nloads = 1; |
4d694b27 | 7195 | lnel = const_nunits; |
e09b4c37 | 7196 | ltype = vectype; |
e09b4c37 | 7197 | } |
2de001ee RS |
7198 | ltype = build_aligned_type (ltype, TYPE_ALIGN (TREE_TYPE (vectype))); |
7199 | } | |
7200 | if (slp) | |
7201 | { | |
66c16fd9 RB |
7202 | /* For SLP permutation support we need to load the whole group, |
7203 | not only the number of vector stmts the permutation result | |
7204 | fits in. */ | |
b266b968 | 7205 | if (slp_perm) |
66c16fd9 | 7206 | { |
d9f21f6a RS |
7207 | /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for |
7208 | variable VF. */ | |
7209 | unsigned int const_vf = vf.to_constant (); | |
4d694b27 | 7210 | ncopies = CEIL (group_size * const_vf, const_nunits); |
66c16fd9 RB |
7211 | dr_chain.create (ncopies); |
7212 | } | |
7213 | else | |
7214 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
7b5fc413 | 7215 | } |
4d694b27 | 7216 | unsigned int group_el = 0; |
e09b4c37 RB |
7217 | unsigned HOST_WIDE_INT |
7218 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
7d75abc8 MM |
7219 | for (j = 0; j < ncopies; j++) |
7220 | { | |
7b5fc413 | 7221 | if (nloads > 1) |
e09b4c37 RB |
7222 | vec_alloc (v, nloads); |
7223 | for (i = 0; i < nloads; i++) | |
7b5fc413 | 7224 | { |
e09b4c37 RB |
7225 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
7226 | group_el * elsz); | |
7227 | new_stmt = gimple_build_assign (make_ssa_name (ltype), | |
7228 | build2 (MEM_REF, ltype, | |
7229 | running_off, this_off)); | |
7230 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7231 | if (nloads > 1) | |
7232 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, | |
7233 | gimple_assign_lhs (new_stmt)); | |
7234 | ||
7235 | group_el += lnel; | |
7236 | if (! slp | |
7237 | || group_el == group_size) | |
7b5fc413 | 7238 | { |
e09b4c37 RB |
7239 | tree newoff = copy_ssa_name (running_off); |
7240 | gimple *incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
7241 | running_off, stride_step); | |
7b5fc413 RB |
7242 | vect_finish_stmt_generation (stmt, incr, gsi); |
7243 | ||
7244 | running_off = newoff; | |
e09b4c37 | 7245 | group_el = 0; |
7b5fc413 | 7246 | } |
7b5fc413 | 7247 | } |
e09b4c37 | 7248 | if (nloads > 1) |
7d75abc8 | 7249 | { |
ea60dd34 RB |
7250 | tree vec_inv = build_constructor (lvectype, v); |
7251 | new_temp = vect_init_vector (stmt, vec_inv, lvectype, gsi); | |
e09b4c37 | 7252 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
ea60dd34 RB |
7253 | if (lvectype != vectype) |
7254 | { | |
7255 | new_stmt = gimple_build_assign (make_ssa_name (vectype), | |
7256 | VIEW_CONVERT_EXPR, | |
7257 | build1 (VIEW_CONVERT_EXPR, | |
7258 | vectype, new_temp)); | |
7259 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7260 | } | |
7d75abc8 MM |
7261 | } |
7262 | ||
7b5fc413 | 7263 | if (slp) |
b266b968 | 7264 | { |
b266b968 RB |
7265 | if (slp_perm) |
7266 | dr_chain.quick_push (gimple_assign_lhs (new_stmt)); | |
66c16fd9 RB |
7267 | else |
7268 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
b266b968 | 7269 | } |
7d75abc8 | 7270 | else |
225ce44b RB |
7271 | { |
7272 | if (j == 0) | |
7273 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7274 | else | |
7275 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7276 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7277 | } | |
7d75abc8 | 7278 | } |
b266b968 | 7279 | if (slp_perm) |
29afecdf RB |
7280 | { |
7281 | unsigned n_perms; | |
7282 | vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, | |
7283 | slp_node_instance, false, &n_perms); | |
7284 | } | |
7d75abc8 MM |
7285 | return true; |
7286 | } | |
aec7ae7d | 7287 | |
0d0293ac | 7288 | if (grouped_load) |
ebfd146a | 7289 | { |
e14c1050 | 7290 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
44fc7854 | 7291 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
4f0a0218 | 7292 | /* For SLP vectorization we directly vectorize a subchain |
52eab378 RB |
7293 | without permutation. */ |
7294 | if (slp && ! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
4f0a0218 RB |
7295 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
7296 | /* For BB vectorization always use the first stmt to base | |
7297 | the data ref pointer on. */ | |
7298 | if (bb_vinfo) | |
7299 | first_stmt_for_drptr = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 7300 | |
ebfd146a | 7301 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
7302 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
7303 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
7304 | ??? But we can only do so if there is exactly one | |
7305 | as we have no way to get at the rest. Leave the CSE | |
7306 | opportunity alone. | |
7307 | ??? With the group load eventually participating | |
7308 | in multiple different permutations (having multiple | |
7309 | slp nodes which refer to the same group) the CSE | |
7310 | is even wrong code. See PR56270. */ | |
7311 | && !slp) | |
ebfd146a IR |
7312 | { |
7313 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
7314 | return true; | |
7315 | } | |
7316 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
9b999e8c | 7317 | group_gap_adj = 0; |
ebfd146a IR |
7318 | |
7319 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
7320 | if (slp) | |
7321 | { | |
0d0293ac | 7322 | grouped_load = false; |
91ff1504 RB |
7323 | /* For SLP permutation support we need to load the whole group, |
7324 | not only the number of vector stmts the permutation result | |
7325 | fits in. */ | |
7326 | if (slp_perm) | |
b267968e | 7327 | { |
d9f21f6a RS |
7328 | /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for |
7329 | variable VF. */ | |
7330 | unsigned int const_vf = vf.to_constant (); | |
4d694b27 RS |
7331 | unsigned int const_nunits = nunits.to_constant (); |
7332 | vec_num = CEIL (group_size * const_vf, const_nunits); | |
b267968e RB |
7333 | group_gap_adj = vf * group_size - nunits * vec_num; |
7334 | } | |
91ff1504 | 7335 | else |
b267968e RB |
7336 | { |
7337 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
796bd467 RB |
7338 | group_gap_adj |
7339 | = group_size - SLP_INSTANCE_GROUP_SIZE (slp_node_instance); | |
b267968e | 7340 | } |
a70d6342 | 7341 | } |
ebfd146a | 7342 | else |
9b999e8c | 7343 | vec_num = group_size; |
44fc7854 BE |
7344 | |
7345 | ref_type = get_group_alias_ptr_type (first_stmt); | |
ebfd146a IR |
7346 | } |
7347 | else | |
7348 | { | |
7349 | first_stmt = stmt; | |
7350 | first_dr = dr; | |
7351 | group_size = vec_num = 1; | |
9b999e8c | 7352 | group_gap_adj = 0; |
44fc7854 | 7353 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); |
ebfd146a IR |
7354 | } |
7355 | ||
720f5239 | 7356 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 7357 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
7358 | /* Targets with load-lane instructions must not require explicit |
7359 | realignment. */ | |
2de001ee | 7360 | gcc_assert (memory_access_type != VMAT_LOAD_STORE_LANES |
272c6793 RS |
7361 | || alignment_support_scheme == dr_aligned |
7362 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
7363 | |
7364 | /* In case the vectorization factor (VF) is bigger than the number | |
7365 | of elements that we can fit in a vectype (nunits), we have to generate | |
7366 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 7367 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 7368 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 7369 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 7370 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
7371 | stmts that use the defs of the current stmt. The example below |
7372 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
7373 | need to create 4 vectorized stmts): | |
ebfd146a IR |
7374 | |
7375 | before vectorization: | |
7376 | RELATED_STMT VEC_STMT | |
7377 | S1: x = memref - - | |
7378 | S2: z = x + 1 - - | |
7379 | ||
7380 | step 1: vectorize stmt S1: | |
7381 | We first create the vector stmt VS1_0, and, as usual, record a | |
7382 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
7383 | Next, we create the vector stmt VS1_1, and record a pointer to | |
7384 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 7385 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
7386 | stmts and pointers: |
7387 | RELATED_STMT VEC_STMT | |
7388 | VS1_0: vx0 = memref0 VS1_1 - | |
7389 | VS1_1: vx1 = memref1 VS1_2 - | |
7390 | VS1_2: vx2 = memref2 VS1_3 - | |
7391 | VS1_3: vx3 = memref3 - - | |
7392 | S1: x = load - VS1_0 | |
7393 | S2: z = x + 1 - - | |
7394 | ||
b8698a0f L |
7395 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
7396 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
7397 | stmt S2. */ |
7398 | ||
0d0293ac | 7399 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
7400 | |
7401 | S1: x2 = &base + 2 | |
7402 | S2: x0 = &base | |
7403 | S3: x1 = &base + 1 | |
7404 | S4: x3 = &base + 3 | |
7405 | ||
b8698a0f | 7406 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
7407 | starting from the access of the first stmt of the chain: |
7408 | ||
7409 | VS1: vx0 = &base | |
7410 | VS2: vx1 = &base + vec_size*1 | |
7411 | VS3: vx3 = &base + vec_size*2 | |
7412 | VS4: vx4 = &base + vec_size*3 | |
7413 | ||
7414 | Then permutation statements are generated: | |
7415 | ||
e2c83630 RH |
7416 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
7417 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
7418 | ... |
7419 | ||
7420 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
7421 | (the order of the data-refs in the output of vect_permute_load_chain | |
7422 | corresponds to the order of scalar stmts in the interleaving chain - see | |
7423 | the documentation of vect_permute_load_chain()). | |
7424 | The generation of permutation stmts and recording them in | |
0d0293ac | 7425 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 7426 | |
b8698a0f | 7427 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
7428 | permutation stmts above are created for every copy. The result vector |
7429 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
7430 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
7431 | |
7432 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
7433 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
7434 | we generate the following code: | |
7435 | p = initial_addr; | |
7436 | indx = 0; | |
7437 | loop { | |
7438 | p = p + indx * vectype_size; | |
7439 | vec_dest = *(p); | |
7440 | indx = indx + 1; | |
7441 | } | |
7442 | ||
7443 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 7444 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
7445 | then generate the following code, in which the data in each iteration is |
7446 | obtained by two vector loads, one from the previous iteration, and one | |
7447 | from the current iteration: | |
7448 | p1 = initial_addr; | |
7449 | msq_init = *(floor(p1)) | |
7450 | p2 = initial_addr + VS - 1; | |
7451 | realignment_token = call target_builtin; | |
7452 | indx = 0; | |
7453 | loop { | |
7454 | p2 = p2 + indx * vectype_size | |
7455 | lsq = *(floor(p2)) | |
7456 | vec_dest = realign_load (msq, lsq, realignment_token) | |
7457 | indx = indx + 1; | |
7458 | msq = lsq; | |
7459 | } */ | |
7460 | ||
7461 | /* If the misalignment remains the same throughout the execution of the | |
7462 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 7463 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
7464 | This can only occur when vectorizing memory accesses in the inner-loop |
7465 | nested within an outer-loop that is being vectorized. */ | |
7466 | ||
d1e4b493 | 7467 | if (nested_in_vect_loop |
832b4117 | 7468 | && (DR_STEP_ALIGNMENT (dr) % GET_MODE_SIZE (TYPE_MODE (vectype))) != 0) |
ebfd146a IR |
7469 | { |
7470 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
7471 | compute_in_loop = true; | |
7472 | } | |
7473 | ||
7474 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
7475 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 7476 | && !compute_in_loop) |
ebfd146a IR |
7477 | { |
7478 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
7479 | alignment_support_scheme, NULL_TREE, | |
7480 | &at_loop); | |
7481 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7482 | { | |
538dd0b7 | 7483 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (msq)); |
356bbc4c JJ |
7484 | byte_offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
7485 | size_one_node); | |
ebfd146a IR |
7486 | } |
7487 | } | |
7488 | else | |
7489 | at_loop = loop; | |
7490 | ||
62da9e14 | 7491 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
a1e53f3f L |
7492 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
7493 | ||
2de001ee | 7494 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
272c6793 RS |
7495 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); |
7496 | else | |
7497 | aggr_type = vectype; | |
7498 | ||
ebfd146a | 7499 | prev_stmt_info = NULL; |
4d694b27 | 7500 | poly_uint64 group_elt = 0; |
ebfd146a | 7501 | for (j = 0; j < ncopies; j++) |
b8698a0f | 7502 | { |
272c6793 | 7503 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 7504 | if (j == 0) |
74bf76ed JJ |
7505 | { |
7506 | bool simd_lane_access_p | |
7507 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
7508 | if (simd_lane_access_p | |
7509 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
7510 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
7511 | && integer_zerop (DR_OFFSET (first_dr)) | |
7512 | && integer_zerop (DR_INIT (first_dr)) | |
7513 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
44fc7854 | 7514 | get_alias_set (TREE_TYPE (ref_type))) |
74bf76ed JJ |
7515 | && (alignment_support_scheme == dr_aligned |
7516 | || alignment_support_scheme == dr_unaligned_supported)) | |
7517 | { | |
7518 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
44fc7854 | 7519 | dataref_offset = build_int_cst (ref_type, 0); |
8928eff3 | 7520 | inv_p = false; |
74bf76ed | 7521 | } |
4f0a0218 RB |
7522 | else if (first_stmt_for_drptr |
7523 | && first_stmt != first_stmt_for_drptr) | |
7524 | { | |
7525 | dataref_ptr | |
7526 | = vect_create_data_ref_ptr (first_stmt_for_drptr, aggr_type, | |
7527 | at_loop, offset, &dummy, gsi, | |
7528 | &ptr_incr, simd_lane_access_p, | |
7529 | &inv_p, byte_offset); | |
7530 | /* Adjust the pointer by the difference to first_stmt. */ | |
7531 | data_reference_p ptrdr | |
7532 | = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt_for_drptr)); | |
7533 | tree diff = fold_convert (sizetype, | |
7534 | size_binop (MINUS_EXPR, | |
7535 | DR_INIT (first_dr), | |
7536 | DR_INIT (ptrdr))); | |
7537 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
7538 | stmt, diff); | |
7539 | } | |
74bf76ed JJ |
7540 | else |
7541 | dataref_ptr | |
7542 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
7543 | offset, &dummy, gsi, &ptr_incr, | |
356bbc4c JJ |
7544 | simd_lane_access_p, &inv_p, |
7545 | byte_offset); | |
74bf76ed JJ |
7546 | } |
7547 | else if (dataref_offset) | |
7548 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
7549 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 7550 | else |
272c6793 RS |
7551 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
7552 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 7553 | |
0d0293ac | 7554 | if (grouped_load || slp_perm) |
9771b263 | 7555 | dr_chain.create (vec_num); |
5ce1ee7f | 7556 | |
2de001ee | 7557 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
ebfd146a | 7558 | { |
272c6793 RS |
7559 | tree vec_array; |
7560 | ||
7561 | vec_array = create_vector_array (vectype, vec_num); | |
7562 | ||
7563 | /* Emit: | |
7564 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
44fc7854 | 7565 | data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type); |
a844293d RS |
7566 | gcall *call = gimple_build_call_internal (IFN_LOAD_LANES, 1, |
7567 | data_ref); | |
7568 | gimple_call_set_lhs (call, vec_array); | |
7569 | gimple_call_set_nothrow (call, true); | |
7570 | new_stmt = call; | |
272c6793 | 7571 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a | 7572 | |
272c6793 RS |
7573 | /* Extract each vector into an SSA_NAME. */ |
7574 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 7575 | { |
272c6793 RS |
7576 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
7577 | vec_array, i); | |
9771b263 | 7578 | dr_chain.quick_push (new_temp); |
272c6793 RS |
7579 | } |
7580 | ||
7581 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 7582 | vect_record_grouped_load_vectors (stmt, dr_chain); |
272c6793 RS |
7583 | } |
7584 | else | |
7585 | { | |
7586 | for (i = 0; i < vec_num; i++) | |
7587 | { | |
7588 | if (i > 0) | |
7589 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
7590 | stmt, NULL_TREE); | |
7591 | ||
7592 | /* 2. Create the vector-load in the loop. */ | |
7593 | switch (alignment_support_scheme) | |
7594 | { | |
7595 | case dr_aligned: | |
7596 | case dr_unaligned_supported: | |
be1ac4ec | 7597 | { |
644ffefd MJ |
7598 | unsigned int align, misalign; |
7599 | ||
272c6793 | 7600 | data_ref |
aed93b23 RB |
7601 | = fold_build2 (MEM_REF, vectype, dataref_ptr, |
7602 | dataref_offset | |
7603 | ? dataref_offset | |
44fc7854 | 7604 | : build_int_cst (ref_type, 0)); |
f702e7d4 | 7605 | align = DR_TARGET_ALIGNMENT (dr); |
272c6793 RS |
7606 | if (alignment_support_scheme == dr_aligned) |
7607 | { | |
7608 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 7609 | misalign = 0; |
272c6793 RS |
7610 | } |
7611 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
7612 | { | |
25f68d90 | 7613 | align = dr_alignment (vect_dr_behavior (first_dr)); |
52639a61 | 7614 | misalign = 0; |
272c6793 RS |
7615 | TREE_TYPE (data_ref) |
7616 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 7617 | align * BITS_PER_UNIT); |
272c6793 RS |
7618 | } |
7619 | else | |
7620 | { | |
7621 | TREE_TYPE (data_ref) | |
7622 | = build_aligned_type (TREE_TYPE (data_ref), | |
7623 | TYPE_ALIGN (elem_type)); | |
644ffefd | 7624 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 7625 | } |
aed93b23 RB |
7626 | if (dataref_offset == NULL_TREE |
7627 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
7628 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), |
7629 | align, misalign); | |
272c6793 | 7630 | break; |
be1ac4ec | 7631 | } |
272c6793 | 7632 | case dr_explicit_realign: |
267d3070 | 7633 | { |
272c6793 | 7634 | tree ptr, bump; |
272c6793 | 7635 | |
d88981fc | 7636 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
272c6793 RS |
7637 | |
7638 | if (compute_in_loop) | |
7639 | msq = vect_setup_realignment (first_stmt, gsi, | |
7640 | &realignment_token, | |
7641 | dr_explicit_realign, | |
7642 | dataref_ptr, NULL); | |
7643 | ||
aed93b23 RB |
7644 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
7645 | ptr = copy_ssa_name (dataref_ptr); | |
7646 | else | |
7647 | ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
f702e7d4 | 7648 | unsigned int align = DR_TARGET_ALIGNMENT (first_dr); |
0d0e4a03 JJ |
7649 | new_stmt = gimple_build_assign |
7650 | (ptr, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
7651 | build_int_cst |
7652 | (TREE_TYPE (dataref_ptr), | |
f702e7d4 | 7653 | -(HOST_WIDE_INT) align)); |
272c6793 RS |
7654 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7655 | data_ref | |
7656 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 7657 | build_int_cst (ref_type, 0)); |
272c6793 RS |
7658 | vec_dest = vect_create_destination_var (scalar_dest, |
7659 | vectype); | |
7660 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
7661 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
7662 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7663 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
7664 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
7665 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7666 | msq = new_temp; | |
7667 | ||
d88981fc | 7668 | bump = size_binop (MULT_EXPR, vs, |
7b7b1813 | 7669 | TYPE_SIZE_UNIT (elem_type)); |
d88981fc | 7670 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
272c6793 | 7671 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
0d0e4a03 JJ |
7672 | new_stmt = gimple_build_assign |
7673 | (NULL_TREE, BIT_AND_EXPR, ptr, | |
272c6793 | 7674 | build_int_cst |
f702e7d4 | 7675 | (TREE_TYPE (ptr), -(HOST_WIDE_INT) align)); |
aed93b23 | 7676 | ptr = copy_ssa_name (ptr, new_stmt); |
272c6793 RS |
7677 | gimple_assign_set_lhs (new_stmt, ptr); |
7678 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7679 | data_ref | |
7680 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 7681 | build_int_cst (ref_type, 0)); |
272c6793 | 7682 | break; |
267d3070 | 7683 | } |
272c6793 | 7684 | case dr_explicit_realign_optimized: |
f702e7d4 RS |
7685 | { |
7686 | if (TREE_CODE (dataref_ptr) == SSA_NAME) | |
7687 | new_temp = copy_ssa_name (dataref_ptr); | |
7688 | else | |
7689 | new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
7690 | unsigned int align = DR_TARGET_ALIGNMENT (first_dr); | |
7691 | new_stmt = gimple_build_assign | |
7692 | (new_temp, BIT_AND_EXPR, dataref_ptr, | |
7693 | build_int_cst (TREE_TYPE (dataref_ptr), | |
7694 | -(HOST_WIDE_INT) align)); | |
7695 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7696 | data_ref | |
7697 | = build2 (MEM_REF, vectype, new_temp, | |
7698 | build_int_cst (ref_type, 0)); | |
7699 | break; | |
7700 | } | |
272c6793 RS |
7701 | default: |
7702 | gcc_unreachable (); | |
7703 | } | |
ebfd146a | 7704 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 7705 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
7706 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7707 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7708 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7709 | ||
272c6793 RS |
7710 | /* 3. Handle explicit realignment if necessary/supported. |
7711 | Create in loop: | |
7712 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
7713 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
7714 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 7715 | { |
272c6793 RS |
7716 | lsq = gimple_assign_lhs (new_stmt); |
7717 | if (!realignment_token) | |
7718 | realignment_token = dataref_ptr; | |
7719 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
0d0e4a03 JJ |
7720 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, |
7721 | msq, lsq, realignment_token); | |
272c6793 RS |
7722 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7723 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7724 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7725 | ||
7726 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7727 | { | |
7728 | gcc_assert (phi); | |
7729 | if (i == vec_num - 1 && j == ncopies - 1) | |
7730 | add_phi_arg (phi, lsq, | |
7731 | loop_latch_edge (containing_loop), | |
9e227d60 | 7732 | UNKNOWN_LOCATION); |
272c6793 RS |
7733 | msq = lsq; |
7734 | } | |
ebfd146a | 7735 | } |
ebfd146a | 7736 | |
59fd17e3 RB |
7737 | /* 4. Handle invariant-load. */ |
7738 | if (inv_p && !bb_vinfo) | |
7739 | { | |
59fd17e3 | 7740 | gcc_assert (!grouped_load); |
d1417442 JJ |
7741 | /* If we have versioned for aliasing or the loop doesn't |
7742 | have any data dependencies that would preclude this, | |
7743 | then we are sure this is a loop invariant load and | |
7744 | thus we can insert it on the preheader edge. */ | |
7745 | if (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) | |
7746 | && !nested_in_vect_loop | |
6b916b36 | 7747 | && hoist_defs_of_uses (stmt, loop)) |
a0e35eb0 RB |
7748 | { |
7749 | if (dump_enabled_p ()) | |
7750 | { | |
7751 | dump_printf_loc (MSG_NOTE, vect_location, | |
7752 | "hoisting out of the vectorized " | |
7753 | "loop: "); | |
7754 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
a0e35eb0 | 7755 | } |
b731b390 | 7756 | tree tem = copy_ssa_name (scalar_dest); |
a0e35eb0 RB |
7757 | gsi_insert_on_edge_immediate |
7758 | (loop_preheader_edge (loop), | |
7759 | gimple_build_assign (tem, | |
7760 | unshare_expr | |
7761 | (gimple_assign_rhs1 (stmt)))); | |
7762 | new_temp = vect_init_vector (stmt, tem, vectype, NULL); | |
34cd48e5 RB |
7763 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
7764 | set_vinfo_for_stmt (new_stmt, | |
7765 | new_stmt_vec_info (new_stmt, vinfo)); | |
a0e35eb0 RB |
7766 | } |
7767 | else | |
7768 | { | |
7769 | gimple_stmt_iterator gsi2 = *gsi; | |
7770 | gsi_next (&gsi2); | |
7771 | new_temp = vect_init_vector (stmt, scalar_dest, | |
7772 | vectype, &gsi2); | |
34cd48e5 | 7773 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
a0e35eb0 | 7774 | } |
59fd17e3 RB |
7775 | } |
7776 | ||
62da9e14 | 7777 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
272c6793 | 7778 | { |
aec7ae7d JJ |
7779 | tree perm_mask = perm_mask_for_reverse (vectype); |
7780 | new_temp = permute_vec_elements (new_temp, new_temp, | |
7781 | perm_mask, stmt, gsi); | |
ebfd146a IR |
7782 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
7783 | } | |
267d3070 | 7784 | |
272c6793 | 7785 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
7786 | vect_transform_grouped_load (). */ |
7787 | if (grouped_load || slp_perm) | |
9771b263 | 7788 | dr_chain.quick_push (new_temp); |
267d3070 | 7789 | |
272c6793 RS |
7790 | /* Store vector loads in the corresponding SLP_NODE. */ |
7791 | if (slp && !slp_perm) | |
9771b263 | 7792 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
b267968e RB |
7793 | |
7794 | /* With SLP permutation we load the gaps as well, without | |
7795 | we need to skip the gaps after we manage to fully load | |
7796 | all elements. group_gap_adj is GROUP_SIZE here. */ | |
7797 | group_elt += nunits; | |
d9f21f6a RS |
7798 | if (maybe_ne (group_gap_adj, 0U) |
7799 | && !slp_perm | |
7800 | && known_eq (group_elt, group_size - group_gap_adj)) | |
b267968e | 7801 | { |
d9f21f6a RS |
7802 | poly_wide_int bump_val |
7803 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) | |
7804 | * group_gap_adj); | |
8e6cdc90 | 7805 | tree bump = wide_int_to_tree (sizetype, bump_val); |
b267968e RB |
7806 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
7807 | stmt, bump); | |
7808 | group_elt = 0; | |
7809 | } | |
272c6793 | 7810 | } |
9b999e8c RB |
7811 | /* Bump the vector pointer to account for a gap or for excess |
7812 | elements loaded for a permuted SLP load. */ | |
d9f21f6a | 7813 | if (maybe_ne (group_gap_adj, 0U) && slp_perm) |
a64b9c26 | 7814 | { |
d9f21f6a RS |
7815 | poly_wide_int bump_val |
7816 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) | |
7817 | * group_gap_adj); | |
8e6cdc90 | 7818 | tree bump = wide_int_to_tree (sizetype, bump_val); |
a64b9c26 RB |
7819 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
7820 | stmt, bump); | |
7821 | } | |
ebfd146a IR |
7822 | } |
7823 | ||
7824 | if (slp && !slp_perm) | |
7825 | continue; | |
7826 | ||
7827 | if (slp_perm) | |
7828 | { | |
29afecdf | 7829 | unsigned n_perms; |
01d8bf07 | 7830 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
29afecdf RB |
7831 | slp_node_instance, false, |
7832 | &n_perms)) | |
ebfd146a | 7833 | { |
9771b263 | 7834 | dr_chain.release (); |
ebfd146a IR |
7835 | return false; |
7836 | } | |
7837 | } | |
7838 | else | |
7839 | { | |
0d0293ac | 7840 | if (grouped_load) |
ebfd146a | 7841 | { |
2de001ee | 7842 | if (memory_access_type != VMAT_LOAD_STORE_LANES) |
0d0293ac | 7843 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 7844 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
7845 | } |
7846 | else | |
7847 | { | |
7848 | if (j == 0) | |
7849 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7850 | else | |
7851 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7852 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7853 | } | |
7854 | } | |
9771b263 | 7855 | dr_chain.release (); |
ebfd146a IR |
7856 | } |
7857 | ||
ebfd146a IR |
7858 | return true; |
7859 | } | |
7860 | ||
7861 | /* Function vect_is_simple_cond. | |
b8698a0f | 7862 | |
ebfd146a IR |
7863 | Input: |
7864 | LOOP - the loop that is being vectorized. | |
7865 | COND - Condition that is checked for simple use. | |
7866 | ||
e9e1d143 RG |
7867 | Output: |
7868 | *COMP_VECTYPE - the vector type for the comparison. | |
4fc5ebf1 | 7869 | *DTS - The def types for the arguments of the comparison |
e9e1d143 | 7870 | |
ebfd146a IR |
7871 | Returns whether a COND can be vectorized. Checks whether |
7872 | condition operands are supportable using vec_is_simple_use. */ | |
7873 | ||
87aab9b2 | 7874 | static bool |
4fc5ebf1 | 7875 | vect_is_simple_cond (tree cond, vec_info *vinfo, |
8da4c8d8 RB |
7876 | tree *comp_vectype, enum vect_def_type *dts, |
7877 | tree vectype) | |
ebfd146a IR |
7878 | { |
7879 | tree lhs, rhs; | |
e9e1d143 | 7880 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a | 7881 | |
a414c77f IE |
7882 | /* Mask case. */ |
7883 | if (TREE_CODE (cond) == SSA_NAME | |
2568d8a1 | 7884 | && VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (cond))) |
a414c77f IE |
7885 | { |
7886 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (cond); | |
7887 | if (!vect_is_simple_use (cond, vinfo, &lhs_def_stmt, | |
4fc5ebf1 | 7888 | &dts[0], comp_vectype) |
a414c77f IE |
7889 | || !*comp_vectype |
7890 | || !VECTOR_BOOLEAN_TYPE_P (*comp_vectype)) | |
7891 | return false; | |
7892 | return true; | |
7893 | } | |
7894 | ||
ebfd146a IR |
7895 | if (!COMPARISON_CLASS_P (cond)) |
7896 | return false; | |
7897 | ||
7898 | lhs = TREE_OPERAND (cond, 0); | |
7899 | rhs = TREE_OPERAND (cond, 1); | |
7900 | ||
7901 | if (TREE_CODE (lhs) == SSA_NAME) | |
7902 | { | |
355fe088 | 7903 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); |
4fc5ebf1 | 7904 | if (!vect_is_simple_use (lhs, vinfo, &lhs_def_stmt, &dts[0], &vectype1)) |
ebfd146a IR |
7905 | return false; |
7906 | } | |
4fc5ebf1 JG |
7907 | else if (TREE_CODE (lhs) == INTEGER_CST || TREE_CODE (lhs) == REAL_CST |
7908 | || TREE_CODE (lhs) == FIXED_CST) | |
7909 | dts[0] = vect_constant_def; | |
7910 | else | |
ebfd146a IR |
7911 | return false; |
7912 | ||
7913 | if (TREE_CODE (rhs) == SSA_NAME) | |
7914 | { | |
355fe088 | 7915 | gimple *rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); |
4fc5ebf1 | 7916 | if (!vect_is_simple_use (rhs, vinfo, &rhs_def_stmt, &dts[1], &vectype2)) |
ebfd146a IR |
7917 | return false; |
7918 | } | |
4fc5ebf1 JG |
7919 | else if (TREE_CODE (rhs) == INTEGER_CST || TREE_CODE (rhs) == REAL_CST |
7920 | || TREE_CODE (rhs) == FIXED_CST) | |
7921 | dts[1] = vect_constant_def; | |
7922 | else | |
ebfd146a IR |
7923 | return false; |
7924 | ||
28b33016 IE |
7925 | if (vectype1 && vectype2 |
7926 | && TYPE_VECTOR_SUBPARTS (vectype1) != TYPE_VECTOR_SUBPARTS (vectype2)) | |
7927 | return false; | |
7928 | ||
e9e1d143 | 7929 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
8da4c8d8 RB |
7930 | /* Invariant comparison. */ |
7931 | if (! *comp_vectype) | |
7932 | { | |
7933 | tree scalar_type = TREE_TYPE (lhs); | |
7934 | /* If we can widen the comparison to match vectype do so. */ | |
7935 | if (INTEGRAL_TYPE_P (scalar_type) | |
7936 | && tree_int_cst_lt (TYPE_SIZE (scalar_type), | |
7937 | TYPE_SIZE (TREE_TYPE (vectype)))) | |
7938 | scalar_type = build_nonstandard_integer_type | |
7939 | (tree_to_uhwi (TYPE_SIZE (TREE_TYPE (vectype))), | |
7940 | TYPE_UNSIGNED (scalar_type)); | |
7941 | *comp_vectype = get_vectype_for_scalar_type (scalar_type); | |
7942 | } | |
7943 | ||
ebfd146a IR |
7944 | return true; |
7945 | } | |
7946 | ||
7947 | /* vectorizable_condition. | |
7948 | ||
b8698a0f L |
7949 | Check if STMT is conditional modify expression that can be vectorized. |
7950 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
7951 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
7952 | at GSI. |
7953 | ||
7954 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
7955 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
0ad23163 | 7956 | else clause if it is 2). |
ebfd146a IR |
7957 | |
7958 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7959 | ||
4bbe8262 | 7960 | bool |
355fe088 TS |
7961 | vectorizable_condition (gimple *stmt, gimple_stmt_iterator *gsi, |
7962 | gimple **vec_stmt, tree reduc_def, int reduc_index, | |
f7e531cf | 7963 | slp_tree slp_node) |
ebfd146a IR |
7964 | { |
7965 | tree scalar_dest = NULL_TREE; | |
7966 | tree vec_dest = NULL_TREE; | |
01216d27 JJ |
7967 | tree cond_expr, cond_expr0 = NULL_TREE, cond_expr1 = NULL_TREE; |
7968 | tree then_clause, else_clause; | |
ebfd146a | 7969 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
df11cc78 | 7970 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
7971 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
7972 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
5958f9e2 | 7973 | tree vec_compare; |
ebfd146a IR |
7974 | tree new_temp; |
7975 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4fc5ebf1 JG |
7976 | enum vect_def_type dts[4] |
7977 | = {vect_unknown_def_type, vect_unknown_def_type, | |
7978 | vect_unknown_def_type, vect_unknown_def_type}; | |
7979 | int ndts = 4; | |
f7e531cf | 7980 | int ncopies; |
01216d27 | 7981 | enum tree_code code, cond_code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
a855b1b1 | 7982 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
7983 | int i, j; |
7984 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
7985 | vec<tree> vec_oprnds0 = vNULL; |
7986 | vec<tree> vec_oprnds1 = vNULL; | |
7987 | vec<tree> vec_oprnds2 = vNULL; | |
7988 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 7989 | tree vec_cmp_type; |
a414c77f | 7990 | bool masked = false; |
b8698a0f | 7991 | |
f7e531cf IR |
7992 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
7993 | return false; | |
7994 | ||
af29617a AH |
7995 | if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) == TREE_CODE_REDUCTION) |
7996 | { | |
7997 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
7998 | return false; | |
ebfd146a | 7999 | |
af29617a AH |
8000 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
8001 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
8002 | && reduc_def)) | |
8003 | return false; | |
ebfd146a | 8004 | |
af29617a AH |
8005 | /* FORNOW: not yet supported. */ |
8006 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
8007 | { | |
8008 | if (dump_enabled_p ()) | |
8009 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
8010 | "value used after loop.\n"); | |
8011 | return false; | |
8012 | } | |
ebfd146a IR |
8013 | } |
8014 | ||
8015 | /* Is vectorizable conditional operation? */ | |
8016 | if (!is_gimple_assign (stmt)) | |
8017 | return false; | |
8018 | ||
8019 | code = gimple_assign_rhs_code (stmt); | |
8020 | ||
8021 | if (code != COND_EXPR) | |
8022 | return false; | |
8023 | ||
465c8c19 | 8024 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
2947d3b2 | 8025 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
465c8c19 | 8026 | |
fce57248 | 8027 | if (slp_node) |
465c8c19 JJ |
8028 | ncopies = 1; |
8029 | else | |
e8f142e2 | 8030 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
8031 | |
8032 | gcc_assert (ncopies >= 1); | |
8033 | if (reduc_index && ncopies > 1) | |
8034 | return false; /* FORNOW */ | |
8035 | ||
4e71066d RG |
8036 | cond_expr = gimple_assign_rhs1 (stmt); |
8037 | then_clause = gimple_assign_rhs2 (stmt); | |
8038 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 8039 | |
4fc5ebf1 | 8040 | if (!vect_is_simple_cond (cond_expr, stmt_info->vinfo, |
8da4c8d8 | 8041 | &comp_vectype, &dts[0], vectype) |
e9e1d143 | 8042 | || !comp_vectype) |
ebfd146a IR |
8043 | return false; |
8044 | ||
81c40241 | 8045 | gimple *def_stmt; |
4fc5ebf1 | 8046 | if (!vect_is_simple_use (then_clause, stmt_info->vinfo, &def_stmt, &dts[2], |
2947d3b2 IE |
8047 | &vectype1)) |
8048 | return false; | |
4fc5ebf1 | 8049 | if (!vect_is_simple_use (else_clause, stmt_info->vinfo, &def_stmt, &dts[3], |
2947d3b2 | 8050 | &vectype2)) |
ebfd146a | 8051 | return false; |
2947d3b2 IE |
8052 | |
8053 | if (vectype1 && !useless_type_conversion_p (vectype, vectype1)) | |
8054 | return false; | |
8055 | ||
8056 | if (vectype2 && !useless_type_conversion_p (vectype, vectype2)) | |
ebfd146a IR |
8057 | return false; |
8058 | ||
28b33016 IE |
8059 | masked = !COMPARISON_CLASS_P (cond_expr); |
8060 | vec_cmp_type = build_same_sized_truth_vector_type (comp_vectype); | |
8061 | ||
74946978 MP |
8062 | if (vec_cmp_type == NULL_TREE) |
8063 | return false; | |
784fb9b3 | 8064 | |
01216d27 JJ |
8065 | cond_code = TREE_CODE (cond_expr); |
8066 | if (!masked) | |
8067 | { | |
8068 | cond_expr0 = TREE_OPERAND (cond_expr, 0); | |
8069 | cond_expr1 = TREE_OPERAND (cond_expr, 1); | |
8070 | } | |
8071 | ||
8072 | if (!masked && VECTOR_BOOLEAN_TYPE_P (comp_vectype)) | |
8073 | { | |
8074 | /* Boolean values may have another representation in vectors | |
8075 | and therefore we prefer bit operations over comparison for | |
8076 | them (which also works for scalar masks). We store opcodes | |
8077 | to use in bitop1 and bitop2. Statement is vectorized as | |
8078 | BITOP2 (rhs1 BITOP1 rhs2) or rhs1 BITOP2 (BITOP1 rhs2) | |
8079 | depending on bitop1 and bitop2 arity. */ | |
8080 | switch (cond_code) | |
8081 | { | |
8082 | case GT_EXPR: | |
8083 | bitop1 = BIT_NOT_EXPR; | |
8084 | bitop2 = BIT_AND_EXPR; | |
8085 | break; | |
8086 | case GE_EXPR: | |
8087 | bitop1 = BIT_NOT_EXPR; | |
8088 | bitop2 = BIT_IOR_EXPR; | |
8089 | break; | |
8090 | case LT_EXPR: | |
8091 | bitop1 = BIT_NOT_EXPR; | |
8092 | bitop2 = BIT_AND_EXPR; | |
8093 | std::swap (cond_expr0, cond_expr1); | |
8094 | break; | |
8095 | case LE_EXPR: | |
8096 | bitop1 = BIT_NOT_EXPR; | |
8097 | bitop2 = BIT_IOR_EXPR; | |
8098 | std::swap (cond_expr0, cond_expr1); | |
8099 | break; | |
8100 | case NE_EXPR: | |
8101 | bitop1 = BIT_XOR_EXPR; | |
8102 | break; | |
8103 | case EQ_EXPR: | |
8104 | bitop1 = BIT_XOR_EXPR; | |
8105 | bitop2 = BIT_NOT_EXPR; | |
8106 | break; | |
8107 | default: | |
8108 | return false; | |
8109 | } | |
8110 | cond_code = SSA_NAME; | |
8111 | } | |
8112 | ||
b8698a0f | 8113 | if (!vec_stmt) |
ebfd146a IR |
8114 | { |
8115 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
01216d27 JJ |
8116 | if (bitop1 != NOP_EXPR) |
8117 | { | |
8118 | machine_mode mode = TYPE_MODE (comp_vectype); | |
8119 | optab optab; | |
8120 | ||
8121 | optab = optab_for_tree_code (bitop1, comp_vectype, optab_default); | |
8122 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8123 | return false; | |
8124 | ||
8125 | if (bitop2 != NOP_EXPR) | |
8126 | { | |
8127 | optab = optab_for_tree_code (bitop2, comp_vectype, | |
8128 | optab_default); | |
8129 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8130 | return false; | |
8131 | } | |
8132 | } | |
4fc5ebf1 JG |
8133 | if (expand_vec_cond_expr_p (vectype, comp_vectype, |
8134 | cond_code)) | |
8135 | { | |
8136 | vect_model_simple_cost (stmt_info, ncopies, dts, ndts, NULL, NULL); | |
8137 | return true; | |
8138 | } | |
8139 | return false; | |
ebfd146a IR |
8140 | } |
8141 | ||
f7e531cf IR |
8142 | /* Transform. */ |
8143 | ||
8144 | if (!slp_node) | |
8145 | { | |
9771b263 DN |
8146 | vec_oprnds0.create (1); |
8147 | vec_oprnds1.create (1); | |
8148 | vec_oprnds2.create (1); | |
8149 | vec_oprnds3.create (1); | |
f7e531cf | 8150 | } |
ebfd146a IR |
8151 | |
8152 | /* Handle def. */ | |
8153 | scalar_dest = gimple_assign_lhs (stmt); | |
8154 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
8155 | ||
8156 | /* Handle cond expr. */ | |
a855b1b1 MM |
8157 | for (j = 0; j < ncopies; j++) |
8158 | { | |
538dd0b7 | 8159 | gassign *new_stmt = NULL; |
a855b1b1 MM |
8160 | if (j == 0) |
8161 | { | |
f7e531cf IR |
8162 | if (slp_node) |
8163 | { | |
00f96dc9 TS |
8164 | auto_vec<tree, 4> ops; |
8165 | auto_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 8166 | |
a414c77f | 8167 | if (masked) |
01216d27 | 8168 | ops.safe_push (cond_expr); |
a414c77f IE |
8169 | else |
8170 | { | |
01216d27 JJ |
8171 | ops.safe_push (cond_expr0); |
8172 | ops.safe_push (cond_expr1); | |
a414c77f | 8173 | } |
9771b263 DN |
8174 | ops.safe_push (then_clause); |
8175 | ops.safe_push (else_clause); | |
306b0c92 | 8176 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
37b5ec8f JJ |
8177 | vec_oprnds3 = vec_defs.pop (); |
8178 | vec_oprnds2 = vec_defs.pop (); | |
a414c77f IE |
8179 | if (!masked) |
8180 | vec_oprnds1 = vec_defs.pop (); | |
37b5ec8f | 8181 | vec_oprnds0 = vec_defs.pop (); |
f7e531cf IR |
8182 | } |
8183 | else | |
8184 | { | |
355fe088 | 8185 | gimple *gtemp; |
a414c77f IE |
8186 | if (masked) |
8187 | { | |
8188 | vec_cond_lhs | |
8189 | = vect_get_vec_def_for_operand (cond_expr, stmt, | |
8190 | comp_vectype); | |
8191 | vect_is_simple_use (cond_expr, stmt_info->vinfo, | |
8192 | >emp, &dts[0]); | |
8193 | } | |
8194 | else | |
8195 | { | |
01216d27 JJ |
8196 | vec_cond_lhs |
8197 | = vect_get_vec_def_for_operand (cond_expr0, | |
8198 | stmt, comp_vectype); | |
8199 | vect_is_simple_use (cond_expr0, loop_vinfo, >emp, &dts[0]); | |
8200 | ||
8201 | vec_cond_rhs | |
8202 | = vect_get_vec_def_for_operand (cond_expr1, | |
8203 | stmt, comp_vectype); | |
8204 | vect_is_simple_use (cond_expr1, loop_vinfo, >emp, &dts[1]); | |
a414c77f | 8205 | } |
f7e531cf IR |
8206 | if (reduc_index == 1) |
8207 | vec_then_clause = reduc_def; | |
8208 | else | |
8209 | { | |
8210 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
81c40241 RB |
8211 | stmt); |
8212 | vect_is_simple_use (then_clause, loop_vinfo, | |
8213 | >emp, &dts[2]); | |
f7e531cf IR |
8214 | } |
8215 | if (reduc_index == 2) | |
8216 | vec_else_clause = reduc_def; | |
8217 | else | |
8218 | { | |
8219 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
81c40241 RB |
8220 | stmt); |
8221 | vect_is_simple_use (else_clause, loop_vinfo, >emp, &dts[3]); | |
f7e531cf | 8222 | } |
a855b1b1 MM |
8223 | } |
8224 | } | |
8225 | else | |
8226 | { | |
a414c77f IE |
8227 | vec_cond_lhs |
8228 | = vect_get_vec_def_for_stmt_copy (dts[0], | |
8229 | vec_oprnds0.pop ()); | |
8230 | if (!masked) | |
8231 | vec_cond_rhs | |
8232 | = vect_get_vec_def_for_stmt_copy (dts[1], | |
8233 | vec_oprnds1.pop ()); | |
8234 | ||
a855b1b1 | 8235 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 8236 | vec_oprnds2.pop ()); |
a855b1b1 | 8237 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 8238 | vec_oprnds3.pop ()); |
f7e531cf IR |
8239 | } |
8240 | ||
8241 | if (!slp_node) | |
8242 | { | |
9771b263 | 8243 | vec_oprnds0.quick_push (vec_cond_lhs); |
a414c77f IE |
8244 | if (!masked) |
8245 | vec_oprnds1.quick_push (vec_cond_rhs); | |
9771b263 DN |
8246 | vec_oprnds2.quick_push (vec_then_clause); |
8247 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
8248 | } |
8249 | ||
9dc3f7de | 8250 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 8251 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 8252 | { |
9771b263 DN |
8253 | vec_then_clause = vec_oprnds2[i]; |
8254 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 8255 | |
a414c77f IE |
8256 | if (masked) |
8257 | vec_compare = vec_cond_lhs; | |
8258 | else | |
8259 | { | |
8260 | vec_cond_rhs = vec_oprnds1[i]; | |
01216d27 JJ |
8261 | if (bitop1 == NOP_EXPR) |
8262 | vec_compare = build2 (cond_code, vec_cmp_type, | |
8263 | vec_cond_lhs, vec_cond_rhs); | |
8264 | else | |
8265 | { | |
8266 | new_temp = make_ssa_name (vec_cmp_type); | |
8267 | if (bitop1 == BIT_NOT_EXPR) | |
8268 | new_stmt = gimple_build_assign (new_temp, bitop1, | |
8269 | vec_cond_rhs); | |
8270 | else | |
8271 | new_stmt | |
8272 | = gimple_build_assign (new_temp, bitop1, vec_cond_lhs, | |
8273 | vec_cond_rhs); | |
8274 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8275 | if (bitop2 == NOP_EXPR) | |
8276 | vec_compare = new_temp; | |
8277 | else if (bitop2 == BIT_NOT_EXPR) | |
8278 | { | |
8279 | /* Instead of doing ~x ? y : z do x ? z : y. */ | |
8280 | vec_compare = new_temp; | |
8281 | std::swap (vec_then_clause, vec_else_clause); | |
8282 | } | |
8283 | else | |
8284 | { | |
8285 | vec_compare = make_ssa_name (vec_cmp_type); | |
8286 | new_stmt | |
8287 | = gimple_build_assign (vec_compare, bitop2, | |
8288 | vec_cond_lhs, new_temp); | |
8289 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8290 | } | |
8291 | } | |
a414c77f | 8292 | } |
5958f9e2 JJ |
8293 | new_temp = make_ssa_name (vec_dest); |
8294 | new_stmt = gimple_build_assign (new_temp, VEC_COND_EXPR, | |
8295 | vec_compare, vec_then_clause, | |
8296 | vec_else_clause); | |
f7e531cf IR |
8297 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
8298 | if (slp_node) | |
9771b263 | 8299 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
8300 | } |
8301 | ||
8302 | if (slp_node) | |
8303 | continue; | |
8304 | ||
8305 | if (j == 0) | |
8306 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
8307 | else | |
8308 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
8309 | ||
8310 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 8311 | } |
b8698a0f | 8312 | |
9771b263 DN |
8313 | vec_oprnds0.release (); |
8314 | vec_oprnds1.release (); | |
8315 | vec_oprnds2.release (); | |
8316 | vec_oprnds3.release (); | |
f7e531cf | 8317 | |
ebfd146a IR |
8318 | return true; |
8319 | } | |
8320 | ||
42fd8198 IE |
8321 | /* vectorizable_comparison. |
8322 | ||
8323 | Check if STMT is comparison expression that can be vectorized. | |
8324 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
8325 | comparison, put it in VEC_STMT, and insert it at GSI. | |
8326 | ||
8327 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
8328 | ||
fce57248 | 8329 | static bool |
42fd8198 IE |
8330 | vectorizable_comparison (gimple *stmt, gimple_stmt_iterator *gsi, |
8331 | gimple **vec_stmt, tree reduc_def, | |
8332 | slp_tree slp_node) | |
8333 | { | |
8334 | tree lhs, rhs1, rhs2; | |
8335 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8336 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; | |
8337 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
8338 | tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE; | |
8339 | tree new_temp; | |
8340 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
8341 | enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 8342 | int ndts = 2; |
42fd8198 IE |
8343 | unsigned nunits; |
8344 | int ncopies; | |
49e76ff1 | 8345 | enum tree_code code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
42fd8198 IE |
8346 | stmt_vec_info prev_stmt_info = NULL; |
8347 | int i, j; | |
8348 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
8349 | vec<tree> vec_oprnds0 = vNULL; | |
8350 | vec<tree> vec_oprnds1 = vNULL; | |
8351 | gimple *def_stmt; | |
8352 | tree mask_type; | |
8353 | tree mask; | |
8354 | ||
c245362b IE |
8355 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
8356 | return false; | |
8357 | ||
30480bcd | 8358 | if (!vectype || !VECTOR_BOOLEAN_TYPE_P (vectype)) |
42fd8198 IE |
8359 | return false; |
8360 | ||
8361 | mask_type = vectype; | |
8362 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
8363 | ||
fce57248 | 8364 | if (slp_node) |
42fd8198 IE |
8365 | ncopies = 1; |
8366 | else | |
e8f142e2 | 8367 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
42fd8198 IE |
8368 | |
8369 | gcc_assert (ncopies >= 1); | |
42fd8198 IE |
8370 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
8371 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
8372 | && reduc_def)) | |
8373 | return false; | |
8374 | ||
8375 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
8376 | { | |
8377 | if (dump_enabled_p ()) | |
8378 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
8379 | "value used after loop.\n"); | |
8380 | return false; | |
8381 | } | |
8382 | ||
8383 | if (!is_gimple_assign (stmt)) | |
8384 | return false; | |
8385 | ||
8386 | code = gimple_assign_rhs_code (stmt); | |
8387 | ||
8388 | if (TREE_CODE_CLASS (code) != tcc_comparison) | |
8389 | return false; | |
8390 | ||
8391 | rhs1 = gimple_assign_rhs1 (stmt); | |
8392 | rhs2 = gimple_assign_rhs2 (stmt); | |
8393 | ||
8394 | if (!vect_is_simple_use (rhs1, stmt_info->vinfo, &def_stmt, | |
8395 | &dts[0], &vectype1)) | |
8396 | return false; | |
8397 | ||
8398 | if (!vect_is_simple_use (rhs2, stmt_info->vinfo, &def_stmt, | |
8399 | &dts[1], &vectype2)) | |
8400 | return false; | |
8401 | ||
8402 | if (vectype1 && vectype2 | |
8403 | && TYPE_VECTOR_SUBPARTS (vectype1) != TYPE_VECTOR_SUBPARTS (vectype2)) | |
8404 | return false; | |
8405 | ||
8406 | vectype = vectype1 ? vectype1 : vectype2; | |
8407 | ||
8408 | /* Invariant comparison. */ | |
8409 | if (!vectype) | |
8410 | { | |
69a9a66f RB |
8411 | vectype = get_vectype_for_scalar_type (TREE_TYPE (rhs1)); |
8412 | if (TYPE_VECTOR_SUBPARTS (vectype) != nunits) | |
42fd8198 IE |
8413 | return false; |
8414 | } | |
8415 | else if (nunits != TYPE_VECTOR_SUBPARTS (vectype)) | |
8416 | return false; | |
8417 | ||
49e76ff1 IE |
8418 | /* Can't compare mask and non-mask types. */ |
8419 | if (vectype1 && vectype2 | |
8420 | && (VECTOR_BOOLEAN_TYPE_P (vectype1) ^ VECTOR_BOOLEAN_TYPE_P (vectype2))) | |
8421 | return false; | |
8422 | ||
8423 | /* Boolean values may have another representation in vectors | |
8424 | and therefore we prefer bit operations over comparison for | |
8425 | them (which also works for scalar masks). We store opcodes | |
8426 | to use in bitop1 and bitop2. Statement is vectorized as | |
8427 | BITOP2 (rhs1 BITOP1 rhs2) or | |
8428 | rhs1 BITOP2 (BITOP1 rhs2) | |
8429 | depending on bitop1 and bitop2 arity. */ | |
8430 | if (VECTOR_BOOLEAN_TYPE_P (vectype)) | |
8431 | { | |
8432 | if (code == GT_EXPR) | |
8433 | { | |
8434 | bitop1 = BIT_NOT_EXPR; | |
8435 | bitop2 = BIT_AND_EXPR; | |
8436 | } | |
8437 | else if (code == GE_EXPR) | |
8438 | { | |
8439 | bitop1 = BIT_NOT_EXPR; | |
8440 | bitop2 = BIT_IOR_EXPR; | |
8441 | } | |
8442 | else if (code == LT_EXPR) | |
8443 | { | |
8444 | bitop1 = BIT_NOT_EXPR; | |
8445 | bitop2 = BIT_AND_EXPR; | |
8446 | std::swap (rhs1, rhs2); | |
264d951a | 8447 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
8448 | } |
8449 | else if (code == LE_EXPR) | |
8450 | { | |
8451 | bitop1 = BIT_NOT_EXPR; | |
8452 | bitop2 = BIT_IOR_EXPR; | |
8453 | std::swap (rhs1, rhs2); | |
264d951a | 8454 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
8455 | } |
8456 | else | |
8457 | { | |
8458 | bitop1 = BIT_XOR_EXPR; | |
8459 | if (code == EQ_EXPR) | |
8460 | bitop2 = BIT_NOT_EXPR; | |
8461 | } | |
8462 | } | |
8463 | ||
42fd8198 IE |
8464 | if (!vec_stmt) |
8465 | { | |
8466 | STMT_VINFO_TYPE (stmt_info) = comparison_vec_info_type; | |
49e76ff1 | 8467 | vect_model_simple_cost (stmt_info, ncopies * (1 + (bitop2 != NOP_EXPR)), |
4fc5ebf1 | 8468 | dts, ndts, NULL, NULL); |
49e76ff1 | 8469 | if (bitop1 == NOP_EXPR) |
96592eed | 8470 | return expand_vec_cmp_expr_p (vectype, mask_type, code); |
49e76ff1 IE |
8471 | else |
8472 | { | |
8473 | machine_mode mode = TYPE_MODE (vectype); | |
8474 | optab optab; | |
8475 | ||
8476 | optab = optab_for_tree_code (bitop1, vectype, optab_default); | |
8477 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8478 | return false; | |
8479 | ||
8480 | if (bitop2 != NOP_EXPR) | |
8481 | { | |
8482 | optab = optab_for_tree_code (bitop2, vectype, optab_default); | |
8483 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8484 | return false; | |
8485 | } | |
8486 | return true; | |
8487 | } | |
42fd8198 IE |
8488 | } |
8489 | ||
8490 | /* Transform. */ | |
8491 | if (!slp_node) | |
8492 | { | |
8493 | vec_oprnds0.create (1); | |
8494 | vec_oprnds1.create (1); | |
8495 | } | |
8496 | ||
8497 | /* Handle def. */ | |
8498 | lhs = gimple_assign_lhs (stmt); | |
8499 | mask = vect_create_destination_var (lhs, mask_type); | |
8500 | ||
8501 | /* Handle cmp expr. */ | |
8502 | for (j = 0; j < ncopies; j++) | |
8503 | { | |
8504 | gassign *new_stmt = NULL; | |
8505 | if (j == 0) | |
8506 | { | |
8507 | if (slp_node) | |
8508 | { | |
8509 | auto_vec<tree, 2> ops; | |
8510 | auto_vec<vec<tree>, 2> vec_defs; | |
8511 | ||
8512 | ops.safe_push (rhs1); | |
8513 | ops.safe_push (rhs2); | |
306b0c92 | 8514 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
42fd8198 IE |
8515 | vec_oprnds1 = vec_defs.pop (); |
8516 | vec_oprnds0 = vec_defs.pop (); | |
8517 | } | |
8518 | else | |
8519 | { | |
e4af0bc4 IE |
8520 | vec_rhs1 = vect_get_vec_def_for_operand (rhs1, stmt, vectype); |
8521 | vec_rhs2 = vect_get_vec_def_for_operand (rhs2, stmt, vectype); | |
42fd8198 IE |
8522 | } |
8523 | } | |
8524 | else | |
8525 | { | |
8526 | vec_rhs1 = vect_get_vec_def_for_stmt_copy (dts[0], | |
8527 | vec_oprnds0.pop ()); | |
8528 | vec_rhs2 = vect_get_vec_def_for_stmt_copy (dts[1], | |
8529 | vec_oprnds1.pop ()); | |
8530 | } | |
8531 | ||
8532 | if (!slp_node) | |
8533 | { | |
8534 | vec_oprnds0.quick_push (vec_rhs1); | |
8535 | vec_oprnds1.quick_push (vec_rhs2); | |
8536 | } | |
8537 | ||
8538 | /* Arguments are ready. Create the new vector stmt. */ | |
8539 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1) | |
8540 | { | |
8541 | vec_rhs2 = vec_oprnds1[i]; | |
8542 | ||
8543 | new_temp = make_ssa_name (mask); | |
49e76ff1 IE |
8544 | if (bitop1 == NOP_EXPR) |
8545 | { | |
8546 | new_stmt = gimple_build_assign (new_temp, code, | |
8547 | vec_rhs1, vec_rhs2); | |
8548 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8549 | } | |
8550 | else | |
8551 | { | |
8552 | if (bitop1 == BIT_NOT_EXPR) | |
8553 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs2); | |
8554 | else | |
8555 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs1, | |
8556 | vec_rhs2); | |
8557 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8558 | if (bitop2 != NOP_EXPR) | |
8559 | { | |
8560 | tree res = make_ssa_name (mask); | |
8561 | if (bitop2 == BIT_NOT_EXPR) | |
8562 | new_stmt = gimple_build_assign (res, bitop2, new_temp); | |
8563 | else | |
8564 | new_stmt = gimple_build_assign (res, bitop2, vec_rhs1, | |
8565 | new_temp); | |
8566 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8567 | } | |
8568 | } | |
42fd8198 IE |
8569 | if (slp_node) |
8570 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
8571 | } | |
8572 | ||
8573 | if (slp_node) | |
8574 | continue; | |
8575 | ||
8576 | if (j == 0) | |
8577 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
8578 | else | |
8579 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
8580 | ||
8581 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
8582 | } | |
8583 | ||
8584 | vec_oprnds0.release (); | |
8585 | vec_oprnds1.release (); | |
8586 | ||
8587 | return true; | |
8588 | } | |
ebfd146a | 8589 | |
68a0f2ff RS |
8590 | /* If SLP_NODE is nonnull, return true if vectorizable_live_operation |
8591 | can handle all live statements in the node. Otherwise return true | |
8592 | if STMT is not live or if vectorizable_live_operation can handle it. | |
8593 | GSI and VEC_STMT are as for vectorizable_live_operation. */ | |
8594 | ||
8595 | static bool | |
8596 | can_vectorize_live_stmts (gimple *stmt, gimple_stmt_iterator *gsi, | |
8597 | slp_tree slp_node, gimple **vec_stmt) | |
8598 | { | |
8599 | if (slp_node) | |
8600 | { | |
8601 | gimple *slp_stmt; | |
8602 | unsigned int i; | |
8603 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (slp_node), i, slp_stmt) | |
8604 | { | |
8605 | stmt_vec_info slp_stmt_info = vinfo_for_stmt (slp_stmt); | |
8606 | if (STMT_VINFO_LIVE_P (slp_stmt_info) | |
8607 | && !vectorizable_live_operation (slp_stmt, gsi, slp_node, i, | |
8608 | vec_stmt)) | |
8609 | return false; | |
8610 | } | |
8611 | } | |
8612 | else if (STMT_VINFO_LIVE_P (vinfo_for_stmt (stmt)) | |
8613 | && !vectorizable_live_operation (stmt, gsi, slp_node, -1, vec_stmt)) | |
8614 | return false; | |
8615 | ||
8616 | return true; | |
8617 | } | |
8618 | ||
8644a673 | 8619 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
8620 | |
8621 | bool | |
891ad31c RB |
8622 | vect_analyze_stmt (gimple *stmt, bool *need_to_vectorize, slp_tree node, |
8623 | slp_instance node_instance) | |
ebfd146a | 8624 | { |
8644a673 | 8625 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 8626 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 8627 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 8628 | bool ok; |
355fe088 | 8629 | gimple *pattern_stmt; |
363477c0 | 8630 | gimple_seq pattern_def_seq; |
ebfd146a | 8631 | |
73fbfcad | 8632 | if (dump_enabled_p ()) |
ebfd146a | 8633 | { |
78c60e3d SS |
8634 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
8635 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
8644a673 | 8636 | } |
ebfd146a | 8637 | |
1825a1f3 | 8638 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 8639 | { |
73fbfcad | 8640 | if (dump_enabled_p ()) |
78c60e3d | 8641 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8642 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
8643 | |
8644 | return false; | |
8645 | } | |
b8698a0f L |
8646 | |
8647 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
8648 | to include: |
8649 | - the COND_EXPR which is the loop exit condition | |
8650 | - any LABEL_EXPRs in the loop | |
b8698a0f | 8651 | - computations that are used only for array indexing or loop control. |
8644a673 | 8652 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 8653 | instance, therefore, all the statements are relevant. |
ebfd146a | 8654 | |
d092494c | 8655 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 8656 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
8657 | statements. In basic blocks we are called from some SLP instance |
8658 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
8659 | already will be part of SLP instance. */ | |
83197f37 IR |
8660 | |
8661 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 8662 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 8663 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 8664 | { |
9d5e7640 | 8665 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 8666 | && pattern_stmt |
9d5e7640 IR |
8667 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
8668 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
8669 | { | |
83197f37 | 8670 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
8671 | stmt = pattern_stmt; |
8672 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 8673 | if (dump_enabled_p ()) |
9d5e7640 | 8674 | { |
78c60e3d SS |
8675 | dump_printf_loc (MSG_NOTE, vect_location, |
8676 | "==> examining pattern statement: "); | |
8677 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
9d5e7640 IR |
8678 | } |
8679 | } | |
8680 | else | |
8681 | { | |
73fbfcad | 8682 | if (dump_enabled_p ()) |
e645e942 | 8683 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 8684 | |
9d5e7640 IR |
8685 | return true; |
8686 | } | |
8644a673 | 8687 | } |
83197f37 | 8688 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 8689 | && node == NULL |
83197f37 IR |
8690 | && pattern_stmt |
8691 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
8692 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
8693 | { | |
8694 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 8695 | if (dump_enabled_p ()) |
83197f37 | 8696 | { |
78c60e3d SS |
8697 | dump_printf_loc (MSG_NOTE, vect_location, |
8698 | "==> examining pattern statement: "); | |
8699 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
83197f37 IR |
8700 | } |
8701 | ||
891ad31c RB |
8702 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node, |
8703 | node_instance)) | |
83197f37 IR |
8704 | return false; |
8705 | } | |
ebfd146a | 8706 | |
1107f3ae | 8707 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 8708 | && node == NULL |
363477c0 | 8709 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 8710 | { |
363477c0 | 8711 | gimple_stmt_iterator si; |
1107f3ae | 8712 | |
363477c0 JJ |
8713 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
8714 | { | |
355fe088 | 8715 | gimple *pattern_def_stmt = gsi_stmt (si); |
363477c0 JJ |
8716 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) |
8717 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
8718 | { | |
8719 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 8720 | if (dump_enabled_p ()) |
363477c0 | 8721 | { |
78c60e3d SS |
8722 | dump_printf_loc (MSG_NOTE, vect_location, |
8723 | "==> examining pattern def statement: "); | |
8724 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
363477c0 | 8725 | } |
1107f3ae | 8726 | |
363477c0 | 8727 | if (!vect_analyze_stmt (pattern_def_stmt, |
891ad31c | 8728 | need_to_vectorize, node, node_instance)) |
363477c0 JJ |
8729 | return false; |
8730 | } | |
8731 | } | |
8732 | } | |
1107f3ae | 8733 | |
8644a673 IR |
8734 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
8735 | { | |
8736 | case vect_internal_def: | |
8737 | break; | |
ebfd146a | 8738 | |
8644a673 | 8739 | case vect_reduction_def: |
7c5222ff | 8740 | case vect_nested_cycle: |
14a61437 RB |
8741 | gcc_assert (!bb_vinfo |
8742 | && (relevance == vect_used_in_outer | |
8743 | || relevance == vect_used_in_outer_by_reduction | |
8744 | || relevance == vect_used_by_reduction | |
b28ead45 AH |
8745 | || relevance == vect_unused_in_scope |
8746 | || relevance == vect_used_only_live)); | |
8644a673 IR |
8747 | break; |
8748 | ||
8749 | case vect_induction_def: | |
e7baeb39 RB |
8750 | gcc_assert (!bb_vinfo); |
8751 | break; | |
8752 | ||
8644a673 IR |
8753 | case vect_constant_def: |
8754 | case vect_external_def: | |
8755 | case vect_unknown_def_type: | |
8756 | default: | |
8757 | gcc_unreachable (); | |
8758 | } | |
ebfd146a | 8759 | |
8644a673 | 8760 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 8761 | { |
8644a673 | 8762 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
0136f8f0 AH |
8763 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
8764 | || (is_gimple_call (stmt) | |
8765 | && gimple_call_lhs (stmt) == NULL_TREE)); | |
8644a673 | 8766 | *need_to_vectorize = true; |
ebfd146a IR |
8767 | } |
8768 | ||
b1af7da6 RB |
8769 | if (PURE_SLP_STMT (stmt_info) && !node) |
8770 | { | |
8771 | dump_printf_loc (MSG_NOTE, vect_location, | |
8772 | "handled only by SLP analysis\n"); | |
8773 | return true; | |
8774 | } | |
8775 | ||
8776 | ok = true; | |
8777 | if (!bb_vinfo | |
8778 | && (STMT_VINFO_RELEVANT_P (stmt_info) | |
8779 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
8780 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
8781 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
8782 | || vectorizable_shift (stmt, NULL, NULL, node) | |
8783 | || vectorizable_operation (stmt, NULL, NULL, node) | |
8784 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
8785 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
8786 | || vectorizable_call (stmt, NULL, NULL, node) | |
8787 | || vectorizable_store (stmt, NULL, NULL, node) | |
891ad31c | 8788 | || vectorizable_reduction (stmt, NULL, NULL, node, node_instance) |
e7baeb39 | 8789 | || vectorizable_induction (stmt, NULL, NULL, node) |
42fd8198 IE |
8790 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
8791 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 RB |
8792 | else |
8793 | { | |
8794 | if (bb_vinfo) | |
8795 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
8796 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
8797 | || vectorizable_shift (stmt, NULL, NULL, node) | |
8798 | || vectorizable_operation (stmt, NULL, NULL, node) | |
8799 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
8800 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
8801 | || vectorizable_call (stmt, NULL, NULL, node) | |
8802 | || vectorizable_store (stmt, NULL, NULL, node) | |
42fd8198 IE |
8803 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
8804 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 | 8805 | } |
8644a673 IR |
8806 | |
8807 | if (!ok) | |
ebfd146a | 8808 | { |
73fbfcad | 8809 | if (dump_enabled_p ()) |
8644a673 | 8810 | { |
78c60e3d SS |
8811 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8812 | "not vectorized: relevant stmt not "); | |
8813 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
8814 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 8815 | } |
b8698a0f | 8816 | |
ebfd146a IR |
8817 | return false; |
8818 | } | |
8819 | ||
a70d6342 IR |
8820 | if (bb_vinfo) |
8821 | return true; | |
8822 | ||
8644a673 IR |
8823 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
8824 | need extra handling, except for vectorizable reductions. */ | |
68a0f2ff RS |
8825 | if (STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
8826 | && !can_vectorize_live_stmts (stmt, NULL, node, NULL)) | |
ebfd146a | 8827 | { |
73fbfcad | 8828 | if (dump_enabled_p ()) |
8644a673 | 8829 | { |
78c60e3d | 8830 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
68a0f2ff | 8831 | "not vectorized: live stmt not supported: "); |
78c60e3d | 8832 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); |
8644a673 | 8833 | } |
b8698a0f | 8834 | |
8644a673 | 8835 | return false; |
ebfd146a IR |
8836 | } |
8837 | ||
ebfd146a IR |
8838 | return true; |
8839 | } | |
8840 | ||
8841 | ||
8842 | /* Function vect_transform_stmt. | |
8843 | ||
8844 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
8845 | ||
8846 | bool | |
355fe088 | 8847 | vect_transform_stmt (gimple *stmt, gimple_stmt_iterator *gsi, |
0d0293ac | 8848 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
8849 | slp_instance slp_node_instance) |
8850 | { | |
8851 | bool is_store = false; | |
355fe088 | 8852 | gimple *vec_stmt = NULL; |
ebfd146a | 8853 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
ebfd146a | 8854 | bool done; |
ebfd146a | 8855 | |
fce57248 | 8856 | gcc_assert (slp_node || !PURE_SLP_STMT (stmt_info)); |
355fe088 | 8857 | gimple *old_vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
225ce44b | 8858 | |
ebfd146a IR |
8859 | switch (STMT_VINFO_TYPE (stmt_info)) |
8860 | { | |
8861 | case type_demotion_vec_info_type: | |
ebfd146a | 8862 | case type_promotion_vec_info_type: |
ebfd146a IR |
8863 | case type_conversion_vec_info_type: |
8864 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
8865 | gcc_assert (done); | |
8866 | break; | |
8867 | ||
8868 | case induc_vec_info_type: | |
e7baeb39 | 8869 | done = vectorizable_induction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
8870 | gcc_assert (done); |
8871 | break; | |
8872 | ||
9dc3f7de IR |
8873 | case shift_vec_info_type: |
8874 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
8875 | gcc_assert (done); | |
8876 | break; | |
8877 | ||
ebfd146a IR |
8878 | case op_vec_info_type: |
8879 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
8880 | gcc_assert (done); | |
8881 | break; | |
8882 | ||
8883 | case assignment_vec_info_type: | |
8884 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
8885 | gcc_assert (done); | |
8886 | break; | |
8887 | ||
8888 | case load_vec_info_type: | |
b8698a0f | 8889 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
8890 | slp_node_instance); |
8891 | gcc_assert (done); | |
8892 | break; | |
8893 | ||
8894 | case store_vec_info_type: | |
8895 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
8896 | gcc_assert (done); | |
0d0293ac | 8897 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
8898 | { |
8899 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 8900 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
8901 | one are skipped, and there vec_stmt_info shouldn't be freed |
8902 | meanwhile. */ | |
0d0293ac | 8903 | *grouped_store = true; |
ebfd146a IR |
8904 | if (STMT_VINFO_VEC_STMT (stmt_info)) |
8905 | is_store = true; | |
8906 | } | |
8907 | else | |
8908 | is_store = true; | |
8909 | break; | |
8910 | ||
8911 | case condition_vec_info_type: | |
f7e531cf | 8912 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
8913 | gcc_assert (done); |
8914 | break; | |
8915 | ||
42fd8198 IE |
8916 | case comparison_vec_info_type: |
8917 | done = vectorizable_comparison (stmt, gsi, &vec_stmt, NULL, slp_node); | |
8918 | gcc_assert (done); | |
8919 | break; | |
8920 | ||
ebfd146a | 8921 | case call_vec_info_type: |
190c2236 | 8922 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 8923 | stmt = gsi_stmt (*gsi); |
8e4284d0 | 8924 | if (gimple_call_internal_p (stmt, IFN_MASK_STORE)) |
5ce9450f | 8925 | is_store = true; |
ebfd146a IR |
8926 | break; |
8927 | ||
0136f8f0 AH |
8928 | case call_simd_clone_vec_info_type: |
8929 | done = vectorizable_simd_clone_call (stmt, gsi, &vec_stmt, slp_node); | |
8930 | stmt = gsi_stmt (*gsi); | |
8931 | break; | |
8932 | ||
ebfd146a | 8933 | case reduc_vec_info_type: |
891ad31c RB |
8934 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node, |
8935 | slp_node_instance); | |
ebfd146a IR |
8936 | gcc_assert (done); |
8937 | break; | |
8938 | ||
8939 | default: | |
8940 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
8941 | { | |
73fbfcad | 8942 | if (dump_enabled_p ()) |
78c60e3d | 8943 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8944 | "stmt not supported.\n"); |
ebfd146a IR |
8945 | gcc_unreachable (); |
8946 | } | |
8947 | } | |
8948 | ||
225ce44b RB |
8949 | /* Verify SLP vectorization doesn't mess with STMT_VINFO_VEC_STMT. |
8950 | This would break hybrid SLP vectorization. */ | |
8951 | if (slp_node) | |
d90f8440 RB |
8952 | gcc_assert (!vec_stmt |
8953 | && STMT_VINFO_VEC_STMT (stmt_info) == old_vec_stmt); | |
225ce44b | 8954 | |
ebfd146a IR |
8955 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that |
8956 | is being vectorized, but outside the immediately enclosing loop. */ | |
8957 | if (vec_stmt | |
a70d6342 IR |
8958 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
8959 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
8960 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
8961 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
8962 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 8963 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 8964 | vect_used_in_outer_by_reduction)) |
ebfd146a | 8965 | { |
a70d6342 IR |
8966 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
8967 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
8968 | imm_use_iterator imm_iter; |
8969 | use_operand_p use_p; | |
8970 | tree scalar_dest; | |
355fe088 | 8971 | gimple *exit_phi; |
ebfd146a | 8972 | |
73fbfcad | 8973 | if (dump_enabled_p ()) |
78c60e3d | 8974 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 8975 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
8976 | |
8977 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
8978 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
8979 | STMT). */ | |
8980 | if (gimple_code (stmt) == GIMPLE_PHI) | |
8981 | scalar_dest = PHI_RESULT (stmt); | |
8982 | else | |
8983 | scalar_dest = gimple_assign_lhs (stmt); | |
8984 | ||
8985 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
8986 | { | |
8987 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
8988 | { | |
8989 | exit_phi = USE_STMT (use_p); | |
8990 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
8991 | } | |
8992 | } | |
8993 | } | |
8994 | ||
8995 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
8996 | being vectorized. */ | |
68a0f2ff | 8997 | if (STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) |
ebfd146a | 8998 | { |
68a0f2ff | 8999 | done = can_vectorize_live_stmts (stmt, gsi, slp_node, &vec_stmt); |
ebfd146a IR |
9000 | gcc_assert (done); |
9001 | } | |
9002 | ||
9003 | if (vec_stmt) | |
83197f37 | 9004 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 9005 | |
b8698a0f | 9006 | return is_store; |
ebfd146a IR |
9007 | } |
9008 | ||
9009 | ||
b8698a0f | 9010 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
9011 | stmt_vec_info. */ |
9012 | ||
9013 | void | |
355fe088 | 9014 | vect_remove_stores (gimple *first_stmt) |
ebfd146a | 9015 | { |
355fe088 TS |
9016 | gimple *next = first_stmt; |
9017 | gimple *tmp; | |
ebfd146a IR |
9018 | gimple_stmt_iterator next_si; |
9019 | ||
9020 | while (next) | |
9021 | { | |
78048b1c JJ |
9022 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
9023 | ||
9024 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
9025 | if (is_pattern_stmt_p (stmt_info)) | |
9026 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
9027 | /* Free the attached stmt_vec_info and remove the stmt. */ |
9028 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 9029 | unlink_stmt_vdef (next); |
ebfd146a | 9030 | gsi_remove (&next_si, true); |
3d3f2249 | 9031 | release_defs (next); |
ebfd146a IR |
9032 | free_stmt_vec_info (next); |
9033 | next = tmp; | |
9034 | } | |
9035 | } | |
9036 | ||
9037 | ||
9038 | /* Function new_stmt_vec_info. | |
9039 | ||
9040 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
9041 | ||
9042 | stmt_vec_info | |
310213d4 | 9043 | new_stmt_vec_info (gimple *stmt, vec_info *vinfo) |
ebfd146a IR |
9044 | { |
9045 | stmt_vec_info res; | |
9046 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
9047 | ||
9048 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
9049 | STMT_VINFO_STMT (res) = stmt; | |
310213d4 | 9050 | res->vinfo = vinfo; |
8644a673 | 9051 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
9052 | STMT_VINFO_LIVE_P (res) = false; |
9053 | STMT_VINFO_VECTYPE (res) = NULL; | |
9054 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 9055 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
9056 | STMT_VINFO_IN_PATTERN_P (res) = false; |
9057 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 9058 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a | 9059 | STMT_VINFO_DATA_REF (res) = NULL; |
af29617a | 9060 | STMT_VINFO_VEC_REDUCTION_TYPE (res) = TREE_CODE_REDUCTION; |
7e16ce79 | 9061 | STMT_VINFO_VEC_CONST_COND_REDUC_CODE (res) = ERROR_MARK; |
ebfd146a | 9062 | |
ebfd146a IR |
9063 | if (gimple_code (stmt) == GIMPLE_PHI |
9064 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
9065 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
9066 | else | |
8644a673 IR |
9067 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
9068 | ||
9771b263 | 9069 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 9070 | STMT_SLP_TYPE (res) = loop_vect; |
78810bd3 RB |
9071 | STMT_VINFO_NUM_SLP_USES (res) = 0; |
9072 | ||
e14c1050 IR |
9073 | GROUP_FIRST_ELEMENT (res) = NULL; |
9074 | GROUP_NEXT_ELEMENT (res) = NULL; | |
9075 | GROUP_SIZE (res) = 0; | |
9076 | GROUP_STORE_COUNT (res) = 0; | |
9077 | GROUP_GAP (res) = 0; | |
9078 | GROUP_SAME_DR_STMT (res) = NULL; | |
ebfd146a IR |
9079 | |
9080 | return res; | |
9081 | } | |
9082 | ||
9083 | ||
9084 | /* Create a hash table for stmt_vec_info. */ | |
9085 | ||
9086 | void | |
9087 | init_stmt_vec_info_vec (void) | |
9088 | { | |
9771b263 DN |
9089 | gcc_assert (!stmt_vec_info_vec.exists ()); |
9090 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
9091 | } |
9092 | ||
9093 | ||
9094 | /* Free hash table for stmt_vec_info. */ | |
9095 | ||
9096 | void | |
9097 | free_stmt_vec_info_vec (void) | |
9098 | { | |
93675444 | 9099 | unsigned int i; |
3161455c | 9100 | stmt_vec_info info; |
93675444 JJ |
9101 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) |
9102 | if (info != NULL) | |
3161455c | 9103 | free_stmt_vec_info (STMT_VINFO_STMT (info)); |
9771b263 DN |
9104 | gcc_assert (stmt_vec_info_vec.exists ()); |
9105 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
9106 | } |
9107 | ||
9108 | ||
9109 | /* Free stmt vectorization related info. */ | |
9110 | ||
9111 | void | |
355fe088 | 9112 | free_stmt_vec_info (gimple *stmt) |
ebfd146a IR |
9113 | { |
9114 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
9115 | ||
9116 | if (!stmt_info) | |
9117 | return; | |
9118 | ||
78048b1c JJ |
9119 | /* Check if this statement has a related "pattern stmt" |
9120 | (introduced by the vectorizer during the pattern recognition | |
9121 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
9122 | too. */ | |
9123 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
9124 | { | |
9125 | stmt_vec_info patt_info | |
9126 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
9127 | if (patt_info) | |
9128 | { | |
363477c0 | 9129 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
355fe088 | 9130 | gimple *patt_stmt = STMT_VINFO_STMT (patt_info); |
f0281fde RB |
9131 | gimple_set_bb (patt_stmt, NULL); |
9132 | tree lhs = gimple_get_lhs (patt_stmt); | |
e6f5c25d | 9133 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde | 9134 | release_ssa_name (lhs); |
363477c0 JJ |
9135 | if (seq) |
9136 | { | |
9137 | gimple_stmt_iterator si; | |
9138 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
f0281fde | 9139 | { |
355fe088 | 9140 | gimple *seq_stmt = gsi_stmt (si); |
f0281fde | 9141 | gimple_set_bb (seq_stmt, NULL); |
7532abf2 | 9142 | lhs = gimple_get_lhs (seq_stmt); |
e6f5c25d | 9143 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde RB |
9144 | release_ssa_name (lhs); |
9145 | free_stmt_vec_info (seq_stmt); | |
9146 | } | |
363477c0 | 9147 | } |
f0281fde | 9148 | free_stmt_vec_info (patt_stmt); |
78048b1c JJ |
9149 | } |
9150 | } | |
9151 | ||
9771b263 | 9152 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
6c9e85fb | 9153 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).release (); |
ebfd146a IR |
9154 | set_vinfo_for_stmt (stmt, NULL); |
9155 | free (stmt_info); | |
9156 | } | |
9157 | ||
9158 | ||
bb67d9c7 | 9159 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 9160 | |
bb67d9c7 | 9161 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
9162 | by the target. */ |
9163 | ||
bb67d9c7 | 9164 | static tree |
86e36728 | 9165 | get_vectype_for_scalar_type_and_size (tree scalar_type, poly_uint64 size) |
ebfd146a | 9166 | { |
c7d97b28 | 9167 | tree orig_scalar_type = scalar_type; |
3bd8f481 | 9168 | scalar_mode inner_mode; |
ef4bddc2 | 9169 | machine_mode simd_mode; |
86e36728 | 9170 | poly_uint64 nunits; |
ebfd146a IR |
9171 | tree vectype; |
9172 | ||
3bd8f481 RS |
9173 | if (!is_int_mode (TYPE_MODE (scalar_type), &inner_mode) |
9174 | && !is_float_mode (TYPE_MODE (scalar_type), &inner_mode)) | |
ebfd146a IR |
9175 | return NULL_TREE; |
9176 | ||
3bd8f481 | 9177 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
48f2e373 | 9178 | |
7b7b1813 RG |
9179 | /* For vector types of elements whose mode precision doesn't |
9180 | match their types precision we use a element type of mode | |
9181 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
9182 | they support the proper result truncation/extension. |
9183 | We also make sure to build vector types with INTEGER_TYPE | |
9184 | component type only. */ | |
6d7971b8 | 9185 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
9186 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
9187 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
9188 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
9189 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 9190 | |
ccbf5bb4 RG |
9191 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
9192 | When the component mode passes the above test simply use a type | |
9193 | corresponding to that mode. The theory is that any use that | |
9194 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 9195 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 9196 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
9197 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
9198 | ||
9199 | /* We can't build a vector type of elements with alignment bigger than | |
9200 | their size. */ | |
dfc2e2ac | 9201 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
9202 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
9203 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 9204 | |
dfc2e2ac RB |
9205 | /* If we felt back to using the mode fail if there was |
9206 | no scalar type for it. */ | |
9207 | if (scalar_type == NULL_TREE) | |
9208 | return NULL_TREE; | |
9209 | ||
bb67d9c7 RG |
9210 | /* If no size was supplied use the mode the target prefers. Otherwise |
9211 | lookup a vector mode of the specified size. */ | |
86e36728 | 9212 | if (known_eq (size, 0U)) |
bb67d9c7 | 9213 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); |
86e36728 RS |
9214 | else if (!multiple_p (size, nbytes, &nunits) |
9215 | || !mode_for_vector (inner_mode, nunits).exists (&simd_mode)) | |
9da15d40 | 9216 | return NULL_TREE; |
4c8fd8ac | 9217 | /* NOTE: nunits == 1 is allowed to support single element vector types. */ |
86e36728 | 9218 | if (!multiple_p (GET_MODE_SIZE (simd_mode), nbytes, &nunits)) |
cc4b5170 | 9219 | return NULL_TREE; |
ebfd146a IR |
9220 | |
9221 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
9222 | |
9223 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
9224 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 9225 | return NULL_TREE; |
ebfd146a | 9226 | |
c7d97b28 RB |
9227 | /* Re-attach the address-space qualifier if we canonicalized the scalar |
9228 | type. */ | |
9229 | if (TYPE_ADDR_SPACE (orig_scalar_type) != TYPE_ADDR_SPACE (vectype)) | |
9230 | return build_qualified_type | |
9231 | (vectype, KEEP_QUAL_ADDR_SPACE (TYPE_QUALS (orig_scalar_type))); | |
9232 | ||
ebfd146a IR |
9233 | return vectype; |
9234 | } | |
9235 | ||
86e36728 | 9236 | poly_uint64 current_vector_size; |
bb67d9c7 RG |
9237 | |
9238 | /* Function get_vectype_for_scalar_type. | |
9239 | ||
9240 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
9241 | by the target. */ | |
9242 | ||
9243 | tree | |
9244 | get_vectype_for_scalar_type (tree scalar_type) | |
9245 | { | |
9246 | tree vectype; | |
9247 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
9248 | current_vector_size); | |
9249 | if (vectype | |
86e36728 | 9250 | && known_eq (current_vector_size, 0U)) |
bb67d9c7 RG |
9251 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); |
9252 | return vectype; | |
9253 | } | |
9254 | ||
42fd8198 IE |
9255 | /* Function get_mask_type_for_scalar_type. |
9256 | ||
9257 | Returns the mask type corresponding to a result of comparison | |
9258 | of vectors of specified SCALAR_TYPE as supported by target. */ | |
9259 | ||
9260 | tree | |
9261 | get_mask_type_for_scalar_type (tree scalar_type) | |
9262 | { | |
9263 | tree vectype = get_vectype_for_scalar_type (scalar_type); | |
9264 | ||
9265 | if (!vectype) | |
9266 | return NULL; | |
9267 | ||
9268 | return build_truth_vector_type (TYPE_VECTOR_SUBPARTS (vectype), | |
9269 | current_vector_size); | |
9270 | } | |
9271 | ||
b690cc0f RG |
9272 | /* Function get_same_sized_vectype |
9273 | ||
9274 | Returns a vector type corresponding to SCALAR_TYPE of size | |
9275 | VECTOR_TYPE if supported by the target. */ | |
9276 | ||
9277 | tree | |
bb67d9c7 | 9278 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 9279 | { |
2568d8a1 | 9280 | if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type)) |
9f47c7e5 IE |
9281 | return build_same_sized_truth_vector_type (vector_type); |
9282 | ||
bb67d9c7 RG |
9283 | return get_vectype_for_scalar_type_and_size |
9284 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
9285 | } |
9286 | ||
ebfd146a IR |
9287 | /* Function vect_is_simple_use. |
9288 | ||
9289 | Input: | |
81c40241 RB |
9290 | VINFO - the vect info of the loop or basic block that is being vectorized. |
9291 | OPERAND - operand in the loop or bb. | |
9292 | Output: | |
9293 | DEF_STMT - the defining stmt in case OPERAND is an SSA_NAME. | |
9294 | DT - the type of definition | |
ebfd146a IR |
9295 | |
9296 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 9297 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 9298 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 9299 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
9300 | is the case in reduction/induction computations). |
9301 | For basic blocks, supportable operands are constants and bb invariants. | |
9302 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
9303 | |
9304 | bool | |
81c40241 RB |
9305 | vect_is_simple_use (tree operand, vec_info *vinfo, |
9306 | gimple **def_stmt, enum vect_def_type *dt) | |
b8698a0f | 9307 | { |
ebfd146a | 9308 | *def_stmt = NULL; |
3fc356dc | 9309 | *dt = vect_unknown_def_type; |
b8698a0f | 9310 | |
73fbfcad | 9311 | if (dump_enabled_p ()) |
ebfd146a | 9312 | { |
78c60e3d SS |
9313 | dump_printf_loc (MSG_NOTE, vect_location, |
9314 | "vect_is_simple_use: operand "); | |
9315 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
e645e942 | 9316 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 9317 | } |
b8698a0f | 9318 | |
b758f602 | 9319 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
9320 | { |
9321 | *dt = vect_constant_def; | |
9322 | return true; | |
9323 | } | |
b8698a0f | 9324 | |
ebfd146a IR |
9325 | if (is_gimple_min_invariant (operand)) |
9326 | { | |
8644a673 | 9327 | *dt = vect_external_def; |
ebfd146a IR |
9328 | return true; |
9329 | } | |
9330 | ||
ebfd146a IR |
9331 | if (TREE_CODE (operand) != SSA_NAME) |
9332 | { | |
73fbfcad | 9333 | if (dump_enabled_p ()) |
af29617a AH |
9334 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9335 | "not ssa-name.\n"); | |
ebfd146a IR |
9336 | return false; |
9337 | } | |
b8698a0f | 9338 | |
3fc356dc | 9339 | if (SSA_NAME_IS_DEFAULT_DEF (operand)) |
ebfd146a | 9340 | { |
3fc356dc RB |
9341 | *dt = vect_external_def; |
9342 | return true; | |
ebfd146a IR |
9343 | } |
9344 | ||
3fc356dc | 9345 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
73fbfcad | 9346 | if (dump_enabled_p ()) |
ebfd146a | 9347 | { |
78c60e3d SS |
9348 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
9349 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
ebfd146a IR |
9350 | } |
9351 | ||
61d371eb | 9352 | if (! vect_stmt_in_region_p (vinfo, *def_stmt)) |
8644a673 | 9353 | *dt = vect_external_def; |
ebfd146a IR |
9354 | else |
9355 | { | |
3fc356dc | 9356 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (*def_stmt); |
603cca93 | 9357 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
ebfd146a IR |
9358 | } |
9359 | ||
2e8ab70c RB |
9360 | if (dump_enabled_p ()) |
9361 | { | |
9362 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: "); | |
9363 | switch (*dt) | |
9364 | { | |
9365 | case vect_uninitialized_def: | |
9366 | dump_printf (MSG_NOTE, "uninitialized\n"); | |
9367 | break; | |
9368 | case vect_constant_def: | |
9369 | dump_printf (MSG_NOTE, "constant\n"); | |
9370 | break; | |
9371 | case vect_external_def: | |
9372 | dump_printf (MSG_NOTE, "external\n"); | |
9373 | break; | |
9374 | case vect_internal_def: | |
9375 | dump_printf (MSG_NOTE, "internal\n"); | |
9376 | break; | |
9377 | case vect_induction_def: | |
9378 | dump_printf (MSG_NOTE, "induction\n"); | |
9379 | break; | |
9380 | case vect_reduction_def: | |
9381 | dump_printf (MSG_NOTE, "reduction\n"); | |
9382 | break; | |
9383 | case vect_double_reduction_def: | |
9384 | dump_printf (MSG_NOTE, "double reduction\n"); | |
9385 | break; | |
9386 | case vect_nested_cycle: | |
9387 | dump_printf (MSG_NOTE, "nested cycle\n"); | |
9388 | break; | |
9389 | case vect_unknown_def_type: | |
9390 | dump_printf (MSG_NOTE, "unknown\n"); | |
9391 | break; | |
9392 | } | |
9393 | } | |
9394 | ||
81c40241 | 9395 | if (*dt == vect_unknown_def_type) |
ebfd146a | 9396 | { |
73fbfcad | 9397 | if (dump_enabled_p ()) |
78c60e3d | 9398 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9399 | "Unsupported pattern.\n"); |
ebfd146a IR |
9400 | return false; |
9401 | } | |
9402 | ||
ebfd146a IR |
9403 | switch (gimple_code (*def_stmt)) |
9404 | { | |
9405 | case GIMPLE_PHI: | |
ebfd146a | 9406 | case GIMPLE_ASSIGN: |
ebfd146a | 9407 | case GIMPLE_CALL: |
81c40241 | 9408 | break; |
ebfd146a | 9409 | default: |
73fbfcad | 9410 | if (dump_enabled_p ()) |
78c60e3d | 9411 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9412 | "unsupported defining stmt:\n"); |
ebfd146a IR |
9413 | return false; |
9414 | } | |
9415 | ||
9416 | return true; | |
9417 | } | |
9418 | ||
81c40241 | 9419 | /* Function vect_is_simple_use. |
b690cc0f | 9420 | |
81c40241 | 9421 | Same as vect_is_simple_use but also determines the vector operand |
b690cc0f RG |
9422 | type of OPERAND and stores it to *VECTYPE. If the definition of |
9423 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
9424 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
9425 | is responsible to compute the best suited vector type for the | |
9426 | scalar operand. */ | |
9427 | ||
9428 | bool | |
81c40241 RB |
9429 | vect_is_simple_use (tree operand, vec_info *vinfo, |
9430 | gimple **def_stmt, enum vect_def_type *dt, tree *vectype) | |
b690cc0f | 9431 | { |
81c40241 | 9432 | if (!vect_is_simple_use (operand, vinfo, def_stmt, dt)) |
b690cc0f RG |
9433 | return false; |
9434 | ||
9435 | /* Now get a vector type if the def is internal, otherwise supply | |
9436 | NULL_TREE and leave it up to the caller to figure out a proper | |
9437 | type for the use stmt. */ | |
9438 | if (*dt == vect_internal_def | |
9439 | || *dt == vect_induction_def | |
9440 | || *dt == vect_reduction_def | |
9441 | || *dt == vect_double_reduction_def | |
9442 | || *dt == vect_nested_cycle) | |
9443 | { | |
9444 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
9445 | |
9446 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
9447 | && !STMT_VINFO_RELEVANT (stmt_info) | |
9448 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 9449 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 9450 | |
b690cc0f RG |
9451 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
9452 | gcc_assert (*vectype != NULL_TREE); | |
9453 | } | |
9454 | else if (*dt == vect_uninitialized_def | |
9455 | || *dt == vect_constant_def | |
9456 | || *dt == vect_external_def) | |
9457 | *vectype = NULL_TREE; | |
9458 | else | |
9459 | gcc_unreachable (); | |
9460 | ||
9461 | return true; | |
9462 | } | |
9463 | ||
ebfd146a IR |
9464 | |
9465 | /* Function supportable_widening_operation | |
9466 | ||
b8698a0f L |
9467 | Check whether an operation represented by the code CODE is a |
9468 | widening operation that is supported by the target platform in | |
b690cc0f RG |
9469 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
9470 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 9471 | |
ebfd146a IR |
9472 | Widening operations we currently support are NOP (CONVERT), FLOAT |
9473 | and WIDEN_MULT. This function checks if these operations are supported | |
9474 | by the target platform either directly (via vector tree-codes), or via | |
9475 | target builtins. | |
9476 | ||
9477 | Output: | |
b8698a0f L |
9478 | - CODE1 and CODE2 are codes of vector operations to be used when |
9479 | vectorizing the operation, if available. | |
ebfd146a IR |
9480 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
9481 | case of multi-step conversion (like char->short->int - in that case | |
9482 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
9483 | - INTERM_TYPES contains the intermediate type required to perform the |
9484 | widening operation (short in the above example). */ | |
ebfd146a IR |
9485 | |
9486 | bool | |
355fe088 | 9487 | supportable_widening_operation (enum tree_code code, gimple *stmt, |
b690cc0f | 9488 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
9489 | enum tree_code *code1, enum tree_code *code2, |
9490 | int *multi_step_cvt, | |
9771b263 | 9491 | vec<tree> *interm_types) |
ebfd146a IR |
9492 | { |
9493 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
9494 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 9495 | struct loop *vect_loop = NULL; |
ef4bddc2 | 9496 | machine_mode vec_mode; |
81f40b79 | 9497 | enum insn_code icode1, icode2; |
ebfd146a | 9498 | optab optab1, optab2; |
b690cc0f RG |
9499 | tree vectype = vectype_in; |
9500 | tree wide_vectype = vectype_out; | |
ebfd146a | 9501 | enum tree_code c1, c2; |
4a00c761 JJ |
9502 | int i; |
9503 | tree prev_type, intermediate_type; | |
ef4bddc2 | 9504 | machine_mode intermediate_mode, prev_mode; |
4a00c761 | 9505 | optab optab3, optab4; |
ebfd146a | 9506 | |
4a00c761 | 9507 | *multi_step_cvt = 0; |
4ef69dfc IR |
9508 | if (loop_info) |
9509 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
9510 | ||
ebfd146a IR |
9511 | switch (code) |
9512 | { | |
9513 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
9514 | /* The result of a vectorized widening operation usually requires |
9515 | two vectors (because the widened results do not fit into one vector). | |
9516 | The generated vector results would normally be expected to be | |
9517 | generated in the same order as in the original scalar computation, | |
9518 | i.e. if 8 results are generated in each vector iteration, they are | |
9519 | to be organized as follows: | |
9520 | vect1: [res1,res2,res3,res4], | |
9521 | vect2: [res5,res6,res7,res8]. | |
9522 | ||
9523 | However, in the special case that the result of the widening | |
9524 | operation is used in a reduction computation only, the order doesn't | |
9525 | matter (because when vectorizing a reduction we change the order of | |
9526 | the computation). Some targets can take advantage of this and | |
9527 | generate more efficient code. For example, targets like Altivec, | |
9528 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
9529 | generate the following vectors: | |
9530 | vect1: [res1,res3,res5,res7], | |
9531 | vect2: [res2,res4,res6,res8]. | |
9532 | ||
9533 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
9534 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
9535 | iterations in parallel). We therefore don't allow to change the | |
9536 | order of the computation in the inner-loop during outer-loop | |
9537 | vectorization. */ | |
9538 | /* TODO: Another case in which order doesn't *really* matter is when we | |
9539 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
9540 | Normally, pack_trunc performs an even/odd permute, whereas the | |
9541 | repack from an even/odd expansion would be an interleave, which | |
9542 | would be significantly simpler for e.g. AVX2. */ | |
9543 | /* In any case, in order to avoid duplicating the code below, recurse | |
9544 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
9545 | are properly set up for the caller. If we fail, we'll continue with | |
9546 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
9547 | if (vect_loop | |
9548 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
9549 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
9550 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
9551 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
9552 | code1, code2, multi_step_cvt, |
9553 | interm_types)) | |
ebc047a2 CH |
9554 | { |
9555 | /* Elements in a vector with vect_used_by_reduction property cannot | |
9556 | be reordered if the use chain with this property does not have the | |
9557 | same operation. One such an example is s += a * b, where elements | |
9558 | in a and b cannot be reordered. Here we check if the vector defined | |
9559 | by STMT is only directly used in the reduction statement. */ | |
9560 | tree lhs = gimple_assign_lhs (stmt); | |
9561 | use_operand_p dummy; | |
355fe088 | 9562 | gimple *use_stmt; |
ebc047a2 CH |
9563 | stmt_vec_info use_stmt_info = NULL; |
9564 | if (single_imm_use (lhs, &dummy, &use_stmt) | |
9565 | && (use_stmt_info = vinfo_for_stmt (use_stmt)) | |
9566 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) | |
9567 | return true; | |
9568 | } | |
4a00c761 JJ |
9569 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
9570 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
9571 | break; |
9572 | ||
81c40241 RB |
9573 | case DOT_PROD_EXPR: |
9574 | c1 = DOT_PROD_EXPR; | |
9575 | c2 = DOT_PROD_EXPR; | |
9576 | break; | |
9577 | ||
9578 | case SAD_EXPR: | |
9579 | c1 = SAD_EXPR; | |
9580 | c2 = SAD_EXPR; | |
9581 | break; | |
9582 | ||
6ae6116f RH |
9583 | case VEC_WIDEN_MULT_EVEN_EXPR: |
9584 | /* Support the recursion induced just above. */ | |
9585 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
9586 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
9587 | break; | |
9588 | ||
36ba4aae | 9589 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
9590 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
9591 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
9592 | break; |
9593 | ||
ebfd146a | 9594 | CASE_CONVERT: |
4a00c761 JJ |
9595 | c1 = VEC_UNPACK_LO_EXPR; |
9596 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
9597 | break; |
9598 | ||
9599 | case FLOAT_EXPR: | |
4a00c761 JJ |
9600 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
9601 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
9602 | break; |
9603 | ||
9604 | case FIX_TRUNC_EXPR: | |
9605 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
9606 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
9607 | computing the operation. */ | |
9608 | return false; | |
9609 | ||
9610 | default: | |
9611 | gcc_unreachable (); | |
9612 | } | |
9613 | ||
6ae6116f | 9614 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
6b4db501 | 9615 | std::swap (c1, c2); |
4a00c761 | 9616 | |
ebfd146a IR |
9617 | if (code == FIX_TRUNC_EXPR) |
9618 | { | |
9619 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
9620 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
9621 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
9622 | } |
9623 | else | |
9624 | { | |
9625 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
9626 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
9627 | } | |
9628 | ||
9629 | if (!optab1 || !optab2) | |
9630 | return false; | |
9631 | ||
9632 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
9633 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
9634 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
9635 | return false; |
9636 | ||
4a00c761 JJ |
9637 | *code1 = c1; |
9638 | *code2 = c2; | |
9639 | ||
9640 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
9641 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff IE |
9642 | /* For scalar masks we may have different boolean |
9643 | vector types having the same QImode. Thus we | |
9644 | add additional check for elements number. */ | |
9645 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
9646 | || (TYPE_VECTOR_SUBPARTS (vectype) / 2 | |
9647 | == TYPE_VECTOR_SUBPARTS (wide_vectype))); | |
4a00c761 | 9648 | |
b8698a0f | 9649 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 9650 | types. */ |
ebfd146a | 9651 | |
4a00c761 JJ |
9652 | prev_type = vectype; |
9653 | prev_mode = vec_mode; | |
b8698a0f | 9654 | |
4a00c761 JJ |
9655 | if (!CONVERT_EXPR_CODE_P (code)) |
9656 | return false; | |
b8698a0f | 9657 | |
4a00c761 JJ |
9658 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
9659 | intermediate steps in promotion sequence. We try | |
9660 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
9661 | not. */ | |
9771b263 | 9662 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
9663 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
9664 | { | |
9665 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
9666 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
9667 | { | |
9668 | intermediate_type | |
9669 | = build_truth_vector_type (TYPE_VECTOR_SUBPARTS (prev_type) / 2, | |
9670 | current_vector_size); | |
9671 | if (intermediate_mode != TYPE_MODE (intermediate_type)) | |
9672 | return false; | |
9673 | } | |
9674 | else | |
9675 | intermediate_type | |
9676 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
9677 | TYPE_UNSIGNED (prev_type)); | |
9678 | ||
4a00c761 JJ |
9679 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); |
9680 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
9681 | ||
9682 | if (!optab3 || !optab4 | |
9683 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
9684 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
9685 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
9686 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
9687 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
9688 | == CODE_FOR_nothing) | |
9689 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
9690 | == CODE_FOR_nothing)) | |
9691 | break; | |
ebfd146a | 9692 | |
9771b263 | 9693 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
9694 | (*multi_step_cvt)++; |
9695 | ||
9696 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
9697 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff IE |
9698 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
9699 | || (TYPE_VECTOR_SUBPARTS (intermediate_type) / 2 | |
9700 | == TYPE_VECTOR_SUBPARTS (wide_vectype))); | |
4a00c761 JJ |
9701 | |
9702 | prev_type = intermediate_type; | |
9703 | prev_mode = intermediate_mode; | |
ebfd146a IR |
9704 | } |
9705 | ||
9771b263 | 9706 | interm_types->release (); |
4a00c761 | 9707 | return false; |
ebfd146a IR |
9708 | } |
9709 | ||
9710 | ||
9711 | /* Function supportable_narrowing_operation | |
9712 | ||
b8698a0f L |
9713 | Check whether an operation represented by the code CODE is a |
9714 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
9715 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
9716 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 9717 | |
ebfd146a | 9718 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 9719 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
9720 | the target platform directly via vector tree-codes. |
9721 | ||
9722 | Output: | |
b8698a0f L |
9723 | - CODE1 is the code of a vector operation to be used when |
9724 | vectorizing the operation, if available. | |
ebfd146a IR |
9725 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
9726 | case of multi-step conversion (like int->short->char - in that case | |
9727 | MULTI_STEP_CVT will be 1). | |
9728 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 9729 | narrowing operation (short in the above example). */ |
ebfd146a IR |
9730 | |
9731 | bool | |
9732 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 9733 | tree vectype_out, tree vectype_in, |
ebfd146a | 9734 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 9735 | vec<tree> *interm_types) |
ebfd146a | 9736 | { |
ef4bddc2 | 9737 | machine_mode vec_mode; |
ebfd146a IR |
9738 | enum insn_code icode1; |
9739 | optab optab1, interm_optab; | |
b690cc0f RG |
9740 | tree vectype = vectype_in; |
9741 | tree narrow_vectype = vectype_out; | |
ebfd146a | 9742 | enum tree_code c1; |
3ae0661a | 9743 | tree intermediate_type, prev_type; |
ef4bddc2 | 9744 | machine_mode intermediate_mode, prev_mode; |
ebfd146a | 9745 | int i; |
4a00c761 | 9746 | bool uns; |
ebfd146a | 9747 | |
4a00c761 | 9748 | *multi_step_cvt = 0; |
ebfd146a IR |
9749 | switch (code) |
9750 | { | |
9751 | CASE_CONVERT: | |
9752 | c1 = VEC_PACK_TRUNC_EXPR; | |
9753 | break; | |
9754 | ||
9755 | case FIX_TRUNC_EXPR: | |
9756 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
9757 | break; | |
9758 | ||
9759 | case FLOAT_EXPR: | |
9760 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
9761 | tree code and optabs used for computing the operation. */ | |
9762 | return false; | |
9763 | ||
9764 | default: | |
9765 | gcc_unreachable (); | |
9766 | } | |
9767 | ||
9768 | if (code == FIX_TRUNC_EXPR) | |
9769 | /* The signedness is determined from output operand. */ | |
b690cc0f | 9770 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
9771 | else |
9772 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
9773 | ||
9774 | if (!optab1) | |
9775 | return false; | |
9776 | ||
9777 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 9778 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
9779 | return false; |
9780 | ||
4a00c761 JJ |
9781 | *code1 = c1; |
9782 | ||
9783 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff IE |
9784 | /* For scalar masks we may have different boolean |
9785 | vector types having the same QImode. Thus we | |
9786 | add additional check for elements number. */ | |
9787 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
9788 | || (TYPE_VECTOR_SUBPARTS (vectype) * 2 | |
9789 | == TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 | 9790 | |
ebfd146a IR |
9791 | /* Check if it's a multi-step conversion that can be done using intermediate |
9792 | types. */ | |
4a00c761 | 9793 | prev_mode = vec_mode; |
3ae0661a | 9794 | prev_type = vectype; |
4a00c761 JJ |
9795 | if (code == FIX_TRUNC_EXPR) |
9796 | uns = TYPE_UNSIGNED (vectype_out); | |
9797 | else | |
9798 | uns = TYPE_UNSIGNED (vectype); | |
9799 | ||
9800 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
9801 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
9802 | costly than signed. */ | |
9803 | if (code == FIX_TRUNC_EXPR && uns) | |
9804 | { | |
9805 | enum insn_code icode2; | |
9806 | ||
9807 | intermediate_type | |
9808 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
9809 | interm_optab | |
9810 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 9811 | if (interm_optab != unknown_optab |
4a00c761 JJ |
9812 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
9813 | && insn_data[icode1].operand[0].mode | |
9814 | == insn_data[icode2].operand[0].mode) | |
9815 | { | |
9816 | uns = false; | |
9817 | optab1 = interm_optab; | |
9818 | icode1 = icode2; | |
9819 | } | |
9820 | } | |
ebfd146a | 9821 | |
4a00c761 JJ |
9822 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
9823 | intermediate steps in promotion sequence. We try | |
9824 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 9825 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
9826 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
9827 | { | |
9828 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
9829 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
9830 | { | |
9831 | intermediate_type | |
9832 | = build_truth_vector_type (TYPE_VECTOR_SUBPARTS (prev_type) * 2, | |
9833 | current_vector_size); | |
9834 | if (intermediate_mode != TYPE_MODE (intermediate_type)) | |
9835 | return false; | |
9836 | } | |
9837 | else | |
9838 | intermediate_type | |
9839 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
4a00c761 JJ |
9840 | interm_optab |
9841 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
9842 | optab_default); | |
9843 | if (!interm_optab | |
9844 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
9845 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
9846 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
9847 | == CODE_FOR_nothing)) | |
9848 | break; | |
9849 | ||
9771b263 | 9850 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
9851 | (*multi_step_cvt)++; |
9852 | ||
9853 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff IE |
9854 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
9855 | || (TYPE_VECTOR_SUBPARTS (intermediate_type) * 2 | |
9856 | == TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 JJ |
9857 | |
9858 | prev_mode = intermediate_mode; | |
3ae0661a | 9859 | prev_type = intermediate_type; |
4a00c761 | 9860 | optab1 = interm_optab; |
ebfd146a IR |
9861 | } |
9862 | ||
9771b263 | 9863 | interm_types->release (); |
4a00c761 | 9864 | return false; |
ebfd146a | 9865 | } |