]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
818ab71a | 2 | Copyright (C) 2003-2016 Free Software Foundation, Inc. |
b8698a0f | 3 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
4 | and Ira Rosen <irar@il.ibm.com> |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 3, or (at your option) any later | |
11 | version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
c7131fb2 | 25 | #include "backend.h" |
957060b5 AM |
26 | #include "target.h" |
27 | #include "rtl.h" | |
ebfd146a | 28 | #include "tree.h" |
c7131fb2 | 29 | #include "gimple.h" |
c7131fb2 | 30 | #include "ssa.h" |
957060b5 AM |
31 | #include "optabs-tree.h" |
32 | #include "insn-config.h" | |
33 | #include "recog.h" /* FIXME: for insn_data */ | |
34 | #include "cgraph.h" | |
957060b5 | 35 | #include "dumpfile.h" |
c7131fb2 | 36 | #include "alias.h" |
40e23961 | 37 | #include "fold-const.h" |
d8a2d370 | 38 | #include "stor-layout.h" |
2fb9a547 | 39 | #include "tree-eh.h" |
45b0be94 | 40 | #include "gimplify.h" |
5be5c238 | 41 | #include "gimple-iterator.h" |
18f429e2 | 42 | #include "gimplify-me.h" |
442b4905 | 43 | #include "tree-cfg.h" |
e28030cf | 44 | #include "tree-ssa-loop-manip.h" |
ebfd146a | 45 | #include "cfgloop.h" |
0136f8f0 AH |
46 | #include "tree-ssa-loop.h" |
47 | #include "tree-scalar-evolution.h" | |
ebfd146a | 48 | #include "tree-vectorizer.h" |
9b2b7279 | 49 | #include "builtins.h" |
70439f0d | 50 | #include "internal-fn.h" |
ebfd146a | 51 | |
7ee2468b SB |
52 | /* For lang_hooks.types.type_for_mode. */ |
53 | #include "langhooks.h" | |
ebfd146a | 54 | |
c3e7ee41 BS |
55 | /* Return the vectorized type for the given statement. */ |
56 | ||
57 | tree | |
58 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
59 | { | |
60 | return STMT_VINFO_VECTYPE (stmt_info); | |
61 | } | |
62 | ||
63 | /* Return TRUE iff the given statement is in an inner loop relative to | |
64 | the loop being vectorized. */ | |
65 | bool | |
66 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
67 | { | |
355fe088 | 68 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
c3e7ee41 BS |
69 | basic_block bb = gimple_bb (stmt); |
70 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
71 | struct loop* loop; | |
72 | ||
73 | if (!loop_vinfo) | |
74 | return false; | |
75 | ||
76 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
77 | ||
78 | return (bb->loop_father == loop->inner); | |
79 | } | |
80 | ||
81 | /* Record the cost of a statement, either by directly informing the | |
82 | target model or by saving it in a vector for later processing. | |
83 | Return a preliminary estimate of the statement's cost. */ | |
84 | ||
85 | unsigned | |
92345349 | 86 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 87 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 88 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 89 | { |
92345349 | 90 | if (body_cost_vec) |
c3e7ee41 | 91 | { |
92345349 | 92 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
ddf56386 RB |
93 | stmt_info_for_cost si = { count, kind, |
94 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
95 | misalign }; | |
96 | body_cost_vec->safe_push (si); | |
c3e7ee41 | 97 | return (unsigned) |
92345349 | 98 | (builtin_vectorization_cost (kind, vectype, misalign) * count); |
c3e7ee41 BS |
99 | } |
100 | else | |
310213d4 RB |
101 | return add_stmt_cost (stmt_info->vinfo->target_cost_data, |
102 | count, kind, stmt_info, misalign, where); | |
c3e7ee41 BS |
103 | } |
104 | ||
272c6793 RS |
105 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
106 | ||
107 | static tree | |
108 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
109 | { | |
110 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
111 | "vect_array"); | |
112 | } | |
113 | ||
114 | /* ARRAY is an array of vectors created by create_vector_array. | |
115 | Return an SSA_NAME for the vector in index N. The reference | |
116 | is part of the vectorization of STMT and the vector is associated | |
117 | with scalar destination SCALAR_DEST. */ | |
118 | ||
119 | static tree | |
355fe088 | 120 | read_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree scalar_dest, |
272c6793 RS |
121 | tree array, unsigned HOST_WIDE_INT n) |
122 | { | |
123 | tree vect_type, vect, vect_name, array_ref; | |
355fe088 | 124 | gimple *new_stmt; |
272c6793 RS |
125 | |
126 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
127 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
128 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
129 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
130 | build_int_cst (size_type_node, n), | |
131 | NULL_TREE, NULL_TREE); | |
132 | ||
133 | new_stmt = gimple_build_assign (vect, array_ref); | |
134 | vect_name = make_ssa_name (vect, new_stmt); | |
135 | gimple_assign_set_lhs (new_stmt, vect_name); | |
136 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
137 | |
138 | return vect_name; | |
139 | } | |
140 | ||
141 | /* ARRAY is an array of vectors created by create_vector_array. | |
142 | Emit code to store SSA_NAME VECT in index N of the array. | |
143 | The store is part of the vectorization of STMT. */ | |
144 | ||
145 | static void | |
355fe088 | 146 | write_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree vect, |
272c6793 RS |
147 | tree array, unsigned HOST_WIDE_INT n) |
148 | { | |
149 | tree array_ref; | |
355fe088 | 150 | gimple *new_stmt; |
272c6793 RS |
151 | |
152 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
153 | build_int_cst (size_type_node, n), | |
154 | NULL_TREE, NULL_TREE); | |
155 | ||
156 | new_stmt = gimple_build_assign (array_ref, vect); | |
157 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
158 | } |
159 | ||
160 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
161 | of *PTR. The memory reference replaces those in FIRST_DR | |
162 | (and its group). */ | |
163 | ||
164 | static tree | |
165 | create_array_ref (tree type, tree ptr, struct data_reference *first_dr) | |
166 | { | |
272c6793 RS |
167 | tree mem_ref, alias_ptr_type; |
168 | ||
169 | alias_ptr_type = reference_alias_ptr_type (DR_REF (first_dr)); | |
170 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); | |
171 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 172 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
173 | return mem_ref; |
174 | } | |
175 | ||
ebfd146a IR |
176 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
177 | ||
178 | /* Function vect_mark_relevant. | |
179 | ||
180 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
181 | ||
182 | static void | |
355fe088 | 183 | vect_mark_relevant (vec<gimple *> *worklist, gimple *stmt, |
97ecdb46 | 184 | enum vect_relevant relevant, bool live_p) |
ebfd146a IR |
185 | { |
186 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
187 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
188 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
355fe088 | 189 | gimple *pattern_stmt; |
ebfd146a | 190 | |
73fbfcad | 191 | if (dump_enabled_p ()) |
66c16fd9 RB |
192 | { |
193 | dump_printf_loc (MSG_NOTE, vect_location, | |
194 | "mark relevant %d, live %d: ", relevant, live_p); | |
195 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
196 | } | |
ebfd146a | 197 | |
83197f37 IR |
198 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
199 | related pattern stmt instead of the original stmt. However, such stmts | |
200 | may have their own uses that are not in any pattern, in such cases the | |
201 | stmt itself should be marked. */ | |
ebfd146a IR |
202 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
203 | { | |
97ecdb46 JJ |
204 | /* This is the last stmt in a sequence that was detected as a |
205 | pattern that can potentially be vectorized. Don't mark the stmt | |
206 | as relevant/live because it's not going to be vectorized. | |
207 | Instead mark the pattern-stmt that replaces it. */ | |
83197f37 | 208 | |
97ecdb46 JJ |
209 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); |
210 | ||
211 | if (dump_enabled_p ()) | |
212 | dump_printf_loc (MSG_NOTE, vect_location, | |
213 | "last stmt in pattern. don't mark" | |
214 | " relevant/live.\n"); | |
215 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
216 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
217 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
218 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
219 | stmt = pattern_stmt; | |
ebfd146a IR |
220 | } |
221 | ||
222 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
223 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
224 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
225 | ||
226 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
227 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
228 | { | |
73fbfcad | 229 | if (dump_enabled_p ()) |
78c60e3d | 230 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 231 | "already marked relevant/live.\n"); |
ebfd146a IR |
232 | return; |
233 | } | |
234 | ||
9771b263 | 235 | worklist->safe_push (stmt); |
ebfd146a IR |
236 | } |
237 | ||
238 | ||
b28ead45 AH |
239 | /* Function is_simple_and_all_uses_invariant |
240 | ||
241 | Return true if STMT is simple and all uses of it are invariant. */ | |
242 | ||
243 | bool | |
244 | is_simple_and_all_uses_invariant (gimple *stmt, loop_vec_info loop_vinfo) | |
245 | { | |
246 | tree op; | |
247 | gimple *def_stmt; | |
248 | ssa_op_iter iter; | |
249 | ||
250 | if (!is_gimple_assign (stmt)) | |
251 | return false; | |
252 | ||
253 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) | |
254 | { | |
255 | enum vect_def_type dt = vect_uninitialized_def; | |
256 | ||
257 | if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt)) | |
258 | { | |
259 | if (dump_enabled_p ()) | |
260 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
261 | "use not simple.\n"); | |
262 | return false; | |
263 | } | |
264 | ||
265 | if (dt != vect_external_def && dt != vect_constant_def) | |
266 | return false; | |
267 | } | |
268 | return true; | |
269 | } | |
270 | ||
ebfd146a IR |
271 | /* Function vect_stmt_relevant_p. |
272 | ||
273 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
274 | "relevant for vectorization". | |
275 | ||
276 | A stmt is considered "relevant for vectorization" if: | |
277 | - it has uses outside the loop. | |
278 | - it has vdefs (it alters memory). | |
279 | - control stmts in the loop (except for the exit condition). | |
280 | ||
281 | CHECKME: what other side effects would the vectorizer allow? */ | |
282 | ||
283 | static bool | |
355fe088 | 284 | vect_stmt_relevant_p (gimple *stmt, loop_vec_info loop_vinfo, |
ebfd146a IR |
285 | enum vect_relevant *relevant, bool *live_p) |
286 | { | |
287 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
288 | ssa_op_iter op_iter; | |
289 | imm_use_iterator imm_iter; | |
290 | use_operand_p use_p; | |
291 | def_operand_p def_p; | |
292 | ||
8644a673 | 293 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
294 | *live_p = false; |
295 | ||
296 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
297 | if (is_ctrl_stmt (stmt) |
298 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
299 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 300 | *relevant = vect_used_in_scope; |
ebfd146a IR |
301 | |
302 | /* changing memory. */ | |
303 | if (gimple_code (stmt) != GIMPLE_PHI) | |
ac6aeab4 RB |
304 | if (gimple_vdef (stmt) |
305 | && !gimple_clobber_p (stmt)) | |
ebfd146a | 306 | { |
73fbfcad | 307 | if (dump_enabled_p ()) |
78c60e3d | 308 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 309 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 310 | *relevant = vect_used_in_scope; |
ebfd146a IR |
311 | } |
312 | ||
313 | /* uses outside the loop. */ | |
314 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
315 | { | |
316 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
317 | { | |
318 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
319 | if (!flow_bb_inside_loop_p (loop, bb)) | |
320 | { | |
73fbfcad | 321 | if (dump_enabled_p ()) |
78c60e3d | 322 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 323 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 324 | |
3157b0c2 AO |
325 | if (is_gimple_debug (USE_STMT (use_p))) |
326 | continue; | |
327 | ||
ebfd146a IR |
328 | /* We expect all such uses to be in the loop exit phis |
329 | (because of loop closed form) */ | |
330 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
331 | gcc_assert (bb == single_exit (loop)->dest); | |
332 | ||
333 | *live_p = true; | |
334 | } | |
335 | } | |
336 | } | |
337 | ||
3a2edf4c AH |
338 | if (*live_p && *relevant == vect_unused_in_scope |
339 | && !is_simple_and_all_uses_invariant (stmt, loop_vinfo)) | |
b28ead45 AH |
340 | { |
341 | if (dump_enabled_p ()) | |
342 | dump_printf_loc (MSG_NOTE, vect_location, | |
343 | "vec_stmt_relevant_p: stmt live but not relevant.\n"); | |
344 | *relevant = vect_used_only_live; | |
345 | } | |
346 | ||
ebfd146a IR |
347 | return (*live_p || *relevant); |
348 | } | |
349 | ||
350 | ||
b8698a0f | 351 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 352 | |
ff802fa1 | 353 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
354 | used in STMT for anything other than indexing an array. */ |
355 | ||
356 | static bool | |
355fe088 | 357 | exist_non_indexing_operands_for_use_p (tree use, gimple *stmt) |
ebfd146a IR |
358 | { |
359 | tree operand; | |
360 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 361 | |
ff802fa1 | 362 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
363 | reference in STMT, then any operand that corresponds to USE |
364 | is not indexing an array. */ | |
365 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
366 | return true; | |
59a05b0c | 367 | |
ebfd146a IR |
368 | /* STMT has a data_ref. FORNOW this means that its of one of |
369 | the following forms: | |
370 | -1- ARRAY_REF = var | |
371 | -2- var = ARRAY_REF | |
372 | (This should have been verified in analyze_data_refs). | |
373 | ||
374 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 375 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
376 | for array indexing. |
377 | ||
378 | Therefore, all we need to check is if STMT falls into the | |
379 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
380 | |
381 | if (!gimple_assign_copy_p (stmt)) | |
5ce9450f JJ |
382 | { |
383 | if (is_gimple_call (stmt) | |
384 | && gimple_call_internal_p (stmt)) | |
385 | switch (gimple_call_internal_fn (stmt)) | |
386 | { | |
387 | case IFN_MASK_STORE: | |
388 | operand = gimple_call_arg (stmt, 3); | |
389 | if (operand == use) | |
390 | return true; | |
391 | /* FALLTHRU */ | |
392 | case IFN_MASK_LOAD: | |
393 | operand = gimple_call_arg (stmt, 2); | |
394 | if (operand == use) | |
395 | return true; | |
396 | break; | |
397 | default: | |
398 | break; | |
399 | } | |
400 | return false; | |
401 | } | |
402 | ||
59a05b0c EB |
403 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
404 | return false; | |
ebfd146a | 405 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
406 | if (TREE_CODE (operand) != SSA_NAME) |
407 | return false; | |
408 | ||
409 | if (operand == use) | |
410 | return true; | |
411 | ||
412 | return false; | |
413 | } | |
414 | ||
415 | ||
b8698a0f | 416 | /* |
ebfd146a IR |
417 | Function process_use. |
418 | ||
419 | Inputs: | |
420 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b28ead45 | 421 | - RELEVANT - enum value to be set in the STMT_VINFO of the stmt |
ff802fa1 | 422 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 423 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
424 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
425 | be performed. | |
ebfd146a IR |
426 | |
427 | Outputs: | |
428 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
429 | relevance info of the DEF_STMT of this USE: | |
430 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
431 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
432 | Exceptions: | |
433 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 434 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 435 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
436 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
437 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
438 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
439 | be modified accordingly. | |
440 | ||
441 | Return true if everything is as expected. Return false otherwise. */ | |
442 | ||
443 | static bool | |
b28ead45 | 444 | process_use (gimple *stmt, tree use, loop_vec_info loop_vinfo, |
355fe088 | 445 | enum vect_relevant relevant, vec<gimple *> *worklist, |
aec7ae7d | 446 | bool force) |
ebfd146a IR |
447 | { |
448 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
449 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
450 | stmt_vec_info dstmt_vinfo; | |
451 | basic_block bb, def_bb; | |
355fe088 | 452 | gimple *def_stmt; |
ebfd146a IR |
453 | enum vect_def_type dt; |
454 | ||
b8698a0f | 455 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 456 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 457 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
458 | return true; |
459 | ||
81c40241 | 460 | if (!vect_is_simple_use (use, loop_vinfo, &def_stmt, &dt)) |
b8698a0f | 461 | { |
73fbfcad | 462 | if (dump_enabled_p ()) |
78c60e3d | 463 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 464 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
465 | return false; |
466 | } | |
467 | ||
468 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
469 | return true; | |
470 | ||
471 | def_bb = gimple_bb (def_stmt); | |
472 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
473 | { | |
73fbfcad | 474 | if (dump_enabled_p ()) |
e645e942 | 475 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n"); |
ebfd146a IR |
476 | return true; |
477 | } | |
478 | ||
b8698a0f L |
479 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
480 | DEF_STMT must have already been processed, because this should be the | |
481 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
482 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
483 | check that everything is as expected, and we are done. */ |
484 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
485 | bb = gimple_bb (stmt); | |
486 | if (gimple_code (stmt) == GIMPLE_PHI | |
487 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
488 | && gimple_code (def_stmt) != GIMPLE_PHI | |
489 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
490 | && bb->loop_father == def_bb->loop_father) | |
491 | { | |
73fbfcad | 492 | if (dump_enabled_p ()) |
78c60e3d | 493 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 494 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a IR |
495 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
496 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
497 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 498 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 499 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
500 | return true; |
501 | } | |
502 | ||
503 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
504 | outer-loop-header-bb: | |
505 | d = def_stmt | |
506 | inner-loop: | |
507 | stmt # use (d) | |
508 | outer-loop-tail-bb: | |
509 | ... */ | |
510 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
511 | { | |
73fbfcad | 512 | if (dump_enabled_p ()) |
78c60e3d | 513 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 514 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 515 | |
ebfd146a IR |
516 | switch (relevant) |
517 | { | |
8644a673 | 518 | case vect_unused_in_scope: |
7c5222ff IR |
519 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
520 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 521 | break; |
7c5222ff | 522 | |
ebfd146a | 523 | case vect_used_in_outer_by_reduction: |
7c5222ff | 524 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
525 | relevant = vect_used_by_reduction; |
526 | break; | |
7c5222ff | 527 | |
ebfd146a | 528 | case vect_used_in_outer: |
7c5222ff | 529 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 530 | relevant = vect_used_in_scope; |
ebfd146a | 531 | break; |
7c5222ff | 532 | |
8644a673 | 533 | case vect_used_in_scope: |
ebfd146a IR |
534 | break; |
535 | ||
536 | default: | |
537 | gcc_unreachable (); | |
b8698a0f | 538 | } |
ebfd146a IR |
539 | } |
540 | ||
541 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
542 | outer-loop-header-bb: | |
543 | ... | |
544 | inner-loop: | |
545 | d = def_stmt | |
06066f92 | 546 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
547 | stmt # use (d) */ |
548 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
549 | { | |
73fbfcad | 550 | if (dump_enabled_p ()) |
78c60e3d | 551 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 552 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 553 | |
ebfd146a IR |
554 | switch (relevant) |
555 | { | |
8644a673 | 556 | case vect_unused_in_scope: |
b8698a0f | 557 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 558 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 559 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
560 | break; |
561 | ||
ebfd146a | 562 | case vect_used_by_reduction: |
b28ead45 | 563 | case vect_used_only_live: |
ebfd146a IR |
564 | relevant = vect_used_in_outer_by_reduction; |
565 | break; | |
566 | ||
8644a673 | 567 | case vect_used_in_scope: |
ebfd146a IR |
568 | relevant = vect_used_in_outer; |
569 | break; | |
570 | ||
571 | default: | |
572 | gcc_unreachable (); | |
573 | } | |
574 | } | |
575 | ||
b28ead45 | 576 | vect_mark_relevant (worklist, def_stmt, relevant, false); |
ebfd146a IR |
577 | return true; |
578 | } | |
579 | ||
580 | ||
581 | /* Function vect_mark_stmts_to_be_vectorized. | |
582 | ||
583 | Not all stmts in the loop need to be vectorized. For example: | |
584 | ||
585 | for i... | |
586 | for j... | |
587 | 1. T0 = i + j | |
588 | 2. T1 = a[T0] | |
589 | ||
590 | 3. j = j + 1 | |
591 | ||
592 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
593 | addressing of vectorized data-refs are handled differently. | |
594 | ||
595 | This pass detects such stmts. */ | |
596 | ||
597 | bool | |
598 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
599 | { | |
ebfd146a IR |
600 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
601 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
602 | unsigned int nbbs = loop->num_nodes; | |
603 | gimple_stmt_iterator si; | |
355fe088 | 604 | gimple *stmt; |
ebfd146a IR |
605 | unsigned int i; |
606 | stmt_vec_info stmt_vinfo; | |
607 | basic_block bb; | |
355fe088 | 608 | gimple *phi; |
ebfd146a | 609 | bool live_p; |
b28ead45 | 610 | enum vect_relevant relevant; |
ebfd146a | 611 | |
73fbfcad | 612 | if (dump_enabled_p ()) |
78c60e3d | 613 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 614 | "=== vect_mark_stmts_to_be_vectorized ===\n"); |
ebfd146a | 615 | |
355fe088 | 616 | auto_vec<gimple *, 64> worklist; |
ebfd146a IR |
617 | |
618 | /* 1. Init worklist. */ | |
619 | for (i = 0; i < nbbs; i++) | |
620 | { | |
621 | bb = bbs[i]; | |
622 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 623 | { |
ebfd146a | 624 | phi = gsi_stmt (si); |
73fbfcad | 625 | if (dump_enabled_p ()) |
ebfd146a | 626 | { |
78c60e3d SS |
627 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
628 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
629 | } |
630 | ||
631 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 632 | vect_mark_relevant (&worklist, phi, relevant, live_p); |
ebfd146a IR |
633 | } |
634 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
635 | { | |
636 | stmt = gsi_stmt (si); | |
73fbfcad | 637 | if (dump_enabled_p ()) |
ebfd146a | 638 | { |
78c60e3d SS |
639 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
640 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
b8698a0f | 641 | } |
ebfd146a IR |
642 | |
643 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 644 | vect_mark_relevant (&worklist, stmt, relevant, live_p); |
ebfd146a IR |
645 | } |
646 | } | |
647 | ||
648 | /* 2. Process_worklist */ | |
9771b263 | 649 | while (worklist.length () > 0) |
ebfd146a IR |
650 | { |
651 | use_operand_p use_p; | |
652 | ssa_op_iter iter; | |
653 | ||
9771b263 | 654 | stmt = worklist.pop (); |
73fbfcad | 655 | if (dump_enabled_p ()) |
ebfd146a | 656 | { |
78c60e3d SS |
657 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
658 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
ebfd146a IR |
659 | } |
660 | ||
b8698a0f | 661 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
b28ead45 AH |
662 | (DEF_STMT) as relevant/irrelevant according to the relevance property |
663 | of STMT. */ | |
ebfd146a IR |
664 | stmt_vinfo = vinfo_for_stmt (stmt); |
665 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
ebfd146a | 666 | |
b28ead45 AH |
667 | /* Generally, the relevance property of STMT (in STMT_VINFO_RELEVANT) is |
668 | propagated as is to the DEF_STMTs of its USEs. | |
ebfd146a IR |
669 | |
670 | One exception is when STMT has been identified as defining a reduction | |
b28ead45 | 671 | variable; in this case we set the relevance to vect_used_by_reduction. |
ebfd146a | 672 | This is because we distinguish between two kinds of relevant stmts - |
b8698a0f | 673 | those that are used by a reduction computation, and those that are |
ff802fa1 | 674 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 675 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 676 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 677 | |
b28ead45 | 678 | switch (STMT_VINFO_DEF_TYPE (stmt_vinfo)) |
ebfd146a | 679 | { |
06066f92 | 680 | case vect_reduction_def: |
b28ead45 AH |
681 | gcc_assert (relevant != vect_unused_in_scope); |
682 | if (relevant != vect_unused_in_scope | |
683 | && relevant != vect_used_in_scope | |
684 | && relevant != vect_used_by_reduction | |
685 | && relevant != vect_used_only_live) | |
06066f92 | 686 | { |
b28ead45 AH |
687 | if (dump_enabled_p ()) |
688 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
689 | "unsupported use of reduction.\n"); | |
690 | return false; | |
06066f92 | 691 | } |
06066f92 | 692 | break; |
b8698a0f | 693 | |
06066f92 | 694 | case vect_nested_cycle: |
b28ead45 AH |
695 | if (relevant != vect_unused_in_scope |
696 | && relevant != vect_used_in_outer_by_reduction | |
697 | && relevant != vect_used_in_outer) | |
06066f92 | 698 | { |
73fbfcad | 699 | if (dump_enabled_p ()) |
78c60e3d | 700 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 701 | "unsupported use of nested cycle.\n"); |
7c5222ff | 702 | |
06066f92 IR |
703 | return false; |
704 | } | |
b8698a0f L |
705 | break; |
706 | ||
06066f92 | 707 | case vect_double_reduction_def: |
b28ead45 AH |
708 | if (relevant != vect_unused_in_scope |
709 | && relevant != vect_used_by_reduction | |
710 | && relevant != vect_used_only_live) | |
06066f92 | 711 | { |
73fbfcad | 712 | if (dump_enabled_p ()) |
78c60e3d | 713 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 714 | "unsupported use of double reduction.\n"); |
7c5222ff | 715 | |
7c5222ff | 716 | return false; |
06066f92 | 717 | } |
b8698a0f | 718 | break; |
7c5222ff | 719 | |
06066f92 IR |
720 | default: |
721 | break; | |
7c5222ff | 722 | } |
b8698a0f | 723 | |
aec7ae7d | 724 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
725 | { |
726 | /* Pattern statements are not inserted into the code, so | |
727 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
728 | have to scan the RHS or function arguments instead. */ | |
729 | if (is_gimple_assign (stmt)) | |
730 | { | |
69d2aade JJ |
731 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
732 | tree op = gimple_assign_rhs1 (stmt); | |
733 | ||
734 | i = 1; | |
735 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
736 | { | |
737 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
b28ead45 | 738 | relevant, &worklist, false) |
69d2aade | 739 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
b28ead45 | 740 | relevant, &worklist, false)) |
566d377a | 741 | return false; |
69d2aade JJ |
742 | i = 2; |
743 | } | |
744 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 745 | { |
69d2aade | 746 | op = gimple_op (stmt, i); |
afbe6325 | 747 | if (TREE_CODE (op) == SSA_NAME |
b28ead45 | 748 | && !process_use (stmt, op, loop_vinfo, relevant, |
afbe6325 | 749 | &worklist, false)) |
07687835 | 750 | return false; |
9d5e7640 IR |
751 | } |
752 | } | |
753 | else if (is_gimple_call (stmt)) | |
754 | { | |
755 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
756 | { | |
757 | tree arg = gimple_call_arg (stmt, i); | |
b28ead45 | 758 | if (!process_use (stmt, arg, loop_vinfo, relevant, |
aec7ae7d | 759 | &worklist, false)) |
07687835 | 760 | return false; |
9d5e7640 IR |
761 | } |
762 | } | |
763 | } | |
764 | else | |
765 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
766 | { | |
767 | tree op = USE_FROM_PTR (use_p); | |
b28ead45 | 768 | if (!process_use (stmt, op, loop_vinfo, relevant, |
aec7ae7d | 769 | &worklist, false)) |
07687835 | 770 | return false; |
9d5e7640 | 771 | } |
aec7ae7d | 772 | |
3bab6342 | 773 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo)) |
aec7ae7d JJ |
774 | { |
775 | tree off; | |
3bab6342 | 776 | tree decl = vect_check_gather_scatter (stmt, loop_vinfo, NULL, &off, NULL); |
aec7ae7d | 777 | gcc_assert (decl); |
b28ead45 | 778 | if (!process_use (stmt, off, loop_vinfo, relevant, &worklist, true)) |
566d377a | 779 | return false; |
aec7ae7d | 780 | } |
ebfd146a IR |
781 | } /* while worklist */ |
782 | ||
ebfd146a IR |
783 | return true; |
784 | } | |
785 | ||
786 | ||
b8698a0f | 787 | /* Function vect_model_simple_cost. |
ebfd146a | 788 | |
b8698a0f | 789 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
790 | single op. Right now, this does not account for multiple insns that could |
791 | be generated for the single vector op. We will handle that shortly. */ | |
792 | ||
793 | void | |
b8698a0f | 794 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 BS |
795 | enum vect_def_type *dt, |
796 | stmt_vector_for_cost *prologue_cost_vec, | |
797 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
798 | { |
799 | int i; | |
92345349 | 800 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
801 | |
802 | /* The SLP costs were already calculated during SLP tree build. */ | |
803 | if (PURE_SLP_STMT (stmt_info)) | |
804 | return; | |
805 | ||
ebfd146a IR |
806 | /* FORNOW: Assuming maximum 2 args per stmts. */ |
807 | for (i = 0; i < 2; i++) | |
92345349 BS |
808 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
809 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, vector_stmt, | |
810 | stmt_info, 0, vect_prologue); | |
c3e7ee41 BS |
811 | |
812 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
92345349 BS |
813 | inside_cost = record_stmt_cost (body_cost_vec, ncopies, vector_stmt, |
814 | stmt_info, 0, vect_body); | |
c3e7ee41 | 815 | |
73fbfcad | 816 | if (dump_enabled_p ()) |
78c60e3d SS |
817 | dump_printf_loc (MSG_NOTE, vect_location, |
818 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 819 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
820 | } |
821 | ||
822 | ||
8bd37302 BS |
823 | /* Model cost for type demotion and promotion operations. PWR is normally |
824 | zero for single-step promotions and demotions. It will be one if | |
825 | two-step promotion/demotion is required, and so on. Each additional | |
826 | step doubles the number of instructions required. */ | |
827 | ||
828 | static void | |
829 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
830 | enum vect_def_type *dt, int pwr) | |
831 | { | |
832 | int i, tmp; | |
92345349 | 833 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 BS |
834 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
835 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
836 | void *target_cost_data; | |
8bd37302 BS |
837 | |
838 | /* The SLP costs were already calculated during SLP tree build. */ | |
839 | if (PURE_SLP_STMT (stmt_info)) | |
840 | return; | |
841 | ||
c3e7ee41 BS |
842 | if (loop_vinfo) |
843 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
844 | else | |
845 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
846 | ||
8bd37302 BS |
847 | for (i = 0; i < pwr + 1; i++) |
848 | { | |
849 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
850 | (i + 1) : i; | |
c3e7ee41 | 851 | inside_cost += add_stmt_cost (target_cost_data, vect_pow2 (tmp), |
92345349 BS |
852 | vec_promote_demote, stmt_info, 0, |
853 | vect_body); | |
8bd37302 BS |
854 | } |
855 | ||
856 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
857 | for (i = 0; i < 2; i++) | |
92345349 BS |
858 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
859 | prologue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt, | |
860 | stmt_info, 0, vect_prologue); | |
8bd37302 | 861 | |
73fbfcad | 862 | if (dump_enabled_p ()) |
78c60e3d SS |
863 | dump_printf_loc (MSG_NOTE, vect_location, |
864 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 865 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
866 | } |
867 | ||
ebfd146a IR |
868 | /* Function vect_model_store_cost |
869 | ||
0d0293ac MM |
870 | Models cost for stores. In the case of grouped accesses, one access |
871 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a IR |
872 | |
873 | void | |
b8698a0f | 874 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
272c6793 | 875 | bool store_lanes_p, enum vect_def_type dt, |
92345349 BS |
876 | slp_tree slp_node, |
877 | stmt_vector_for_cost *prologue_cost_vec, | |
878 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a | 879 | { |
92345349 | 880 | unsigned int inside_cost = 0, prologue_cost = 0; |
892a981f RS |
881 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
882 | gimple *first_stmt = STMT_VINFO_STMT (stmt_info); | |
883 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); | |
ebfd146a | 884 | |
8644a673 | 885 | if (dt == vect_constant_def || dt == vect_external_def) |
92345349 BS |
886 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
887 | stmt_info, 0, vect_prologue); | |
ebfd146a | 888 | |
892a981f RS |
889 | /* Grouped stores update all elements in the group at once, |
890 | so we want the DR for the first statement. */ | |
891 | if (!slp_node && grouped_access_p) | |
720f5239 | 892 | { |
892a981f RS |
893 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
894 | dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
720f5239 | 895 | } |
ebfd146a | 896 | |
892a981f RS |
897 | /* True if we should include any once-per-group costs as well as |
898 | the cost of the statement itself. For SLP we only get called | |
899 | once per group anyhow. */ | |
900 | bool first_stmt_p = (first_stmt == STMT_VINFO_STMT (stmt_info)); | |
901 | ||
272c6793 | 902 | /* We assume that the cost of a single store-lanes instruction is |
0d0293ac | 903 | equivalent to the cost of GROUP_SIZE separate stores. If a grouped |
272c6793 | 904 | access is instead being provided by a permute-and-store operation, |
892a981f RS |
905 | include the cost of the permutes. |
906 | ||
907 | For SLP, the caller has already counted the permutation, if any. */ | |
908 | if (grouped_access_p | |
909 | && first_stmt_p | |
910 | && !store_lanes_p | |
911 | && !STMT_VINFO_STRIDED_P (stmt_info) | |
912 | && !slp_node) | |
ebfd146a | 913 | { |
e1377713 ES |
914 | /* Uses a high and low interleave or shuffle operations for each |
915 | needed permute. */ | |
892a981f | 916 | int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
e1377713 | 917 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
92345349 BS |
918 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
919 | stmt_info, 0, vect_body); | |
ebfd146a | 920 | |
73fbfcad | 921 | if (dump_enabled_p ()) |
78c60e3d | 922 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 923 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 924 | group_size); |
ebfd146a IR |
925 | } |
926 | ||
cee62fee | 927 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a | 928 | /* Costs of the stores. */ |
071e8018 | 929 | if (STMT_VINFO_STRIDED_P (stmt_info) && !slp_node) |
f2e2a985 MM |
930 | { |
931 | /* N scalar stores plus extracting the elements. */ | |
f2e2a985 MM |
932 | inside_cost += record_stmt_cost (body_cost_vec, |
933 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), | |
934 | scalar_store, stmt_info, 0, vect_body); | |
f2e2a985 MM |
935 | } |
936 | else | |
892a981f | 937 | vect_get_store_cost (dr, ncopies, &inside_cost, body_cost_vec); |
ebfd146a | 938 | |
cee62fee MM |
939 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
940 | inside_cost += record_stmt_cost (body_cost_vec, | |
941 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), | |
942 | vec_to_scalar, stmt_info, 0, vect_body); | |
943 | ||
73fbfcad | 944 | if (dump_enabled_p ()) |
78c60e3d SS |
945 | dump_printf_loc (MSG_NOTE, vect_location, |
946 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 947 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
948 | } |
949 | ||
950 | ||
720f5239 IR |
951 | /* Calculate cost of DR's memory access. */ |
952 | void | |
953 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 954 | unsigned int *inside_cost, |
92345349 | 955 | stmt_vector_for_cost *body_cost_vec) |
720f5239 IR |
956 | { |
957 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 958 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 959 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
960 | |
961 | switch (alignment_support_scheme) | |
962 | { | |
963 | case dr_aligned: | |
964 | { | |
92345349 BS |
965 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
966 | vector_store, stmt_info, 0, | |
967 | vect_body); | |
720f5239 | 968 | |
73fbfcad | 969 | if (dump_enabled_p ()) |
78c60e3d | 970 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 971 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
972 | break; |
973 | } | |
974 | ||
975 | case dr_unaligned_supported: | |
976 | { | |
720f5239 | 977 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 978 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 979 | unaligned_store, stmt_info, |
92345349 | 980 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 981 | if (dump_enabled_p ()) |
78c60e3d SS |
982 | dump_printf_loc (MSG_NOTE, vect_location, |
983 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 984 | "hardware.\n"); |
720f5239 IR |
985 | break; |
986 | } | |
987 | ||
38eec4c6 UW |
988 | case dr_unaligned_unsupported: |
989 | { | |
990 | *inside_cost = VECT_MAX_COST; | |
991 | ||
73fbfcad | 992 | if (dump_enabled_p ()) |
78c60e3d | 993 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 994 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
995 | break; |
996 | } | |
997 | ||
720f5239 IR |
998 | default: |
999 | gcc_unreachable (); | |
1000 | } | |
1001 | } | |
1002 | ||
1003 | ||
ebfd146a IR |
1004 | /* Function vect_model_load_cost |
1005 | ||
892a981f RS |
1006 | Models cost for loads. In the case of grouped accesses, one access has |
1007 | the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1008 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1009 | access scheme chosen. */ |
1010 | ||
1011 | void | |
92345349 BS |
1012 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, |
1013 | bool load_lanes_p, slp_tree slp_node, | |
1014 | stmt_vector_for_cost *prologue_cost_vec, | |
1015 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a | 1016 | { |
892a981f RS |
1017 | gimple *first_stmt = STMT_VINFO_STMT (stmt_info); |
1018 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
92345349 | 1019 | unsigned int inside_cost = 0, prologue_cost = 0; |
892a981f | 1020 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); |
ebfd146a | 1021 | |
892a981f RS |
1022 | /* Grouped loads read all elements in the group at once, |
1023 | so we want the DR for the first statement. */ | |
1024 | if (!slp_node && grouped_access_p) | |
ebfd146a | 1025 | { |
892a981f RS |
1026 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
1027 | dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
ebfd146a IR |
1028 | } |
1029 | ||
892a981f RS |
1030 | /* True if we should include any once-per-group costs as well as |
1031 | the cost of the statement itself. For SLP we only get called | |
1032 | once per group anyhow. */ | |
1033 | bool first_stmt_p = (first_stmt == STMT_VINFO_STMT (stmt_info)); | |
1034 | ||
272c6793 | 1035 | /* We assume that the cost of a single load-lanes instruction is |
0d0293ac | 1036 | equivalent to the cost of GROUP_SIZE separate loads. If a grouped |
272c6793 | 1037 | access is instead being provided by a load-and-permute operation, |
892a981f RS |
1038 | include the cost of the permutes. |
1039 | ||
1040 | For SLP, the caller has already counted the permutation, if any. */ | |
1041 | if (grouped_access_p | |
1042 | && first_stmt_p | |
1043 | && !load_lanes_p | |
1044 | && !STMT_VINFO_STRIDED_P (stmt_info) | |
1045 | && !slp_node) | |
ebfd146a | 1046 | { |
2c23db6d ES |
1047 | /* Uses an even and odd extract operations or shuffle operations |
1048 | for each needed permute. */ | |
892a981f | 1049 | int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
2c23db6d ES |
1050 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
1051 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, | |
1052 | stmt_info, 0, vect_body); | |
ebfd146a | 1053 | |
73fbfcad | 1054 | if (dump_enabled_p ()) |
e645e942 TJ |
1055 | dump_printf_loc (MSG_NOTE, vect_location, |
1056 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1057 | group_size); |
ebfd146a IR |
1058 | } |
1059 | ||
1060 | /* The loads themselves. */ | |
071e8018 | 1061 | if (STMT_VINFO_STRIDED_P (stmt_info) && !slp_node) |
a82960aa | 1062 | { |
a21892ad BS |
1063 | /* N scalar loads plus gathering them into a vector. */ |
1064 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
92345349 | 1065 | inside_cost += record_stmt_cost (body_cost_vec, |
c3e7ee41 | 1066 | ncopies * TYPE_VECTOR_SUBPARTS (vectype), |
92345349 | 1067 | scalar_load, stmt_info, 0, vect_body); |
a82960aa RG |
1068 | } |
1069 | else | |
892a981f | 1070 | vect_get_load_cost (dr, ncopies, first_stmt_p, |
92345349 BS |
1071 | &inside_cost, &prologue_cost, |
1072 | prologue_cost_vec, body_cost_vec, true); | |
f2e2a985 | 1073 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
892a981f RS |
1074 | inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct, |
1075 | stmt_info, 0, vect_body); | |
720f5239 | 1076 | |
73fbfcad | 1077 | if (dump_enabled_p ()) |
78c60e3d SS |
1078 | dump_printf_loc (MSG_NOTE, vect_location, |
1079 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1080 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1081 | } |
1082 | ||
1083 | ||
1084 | /* Calculate cost of DR's memory access. */ | |
1085 | void | |
1086 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1087 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1088 | unsigned int *prologue_cost, |
1089 | stmt_vector_for_cost *prologue_cost_vec, | |
1090 | stmt_vector_for_cost *body_cost_vec, | |
1091 | bool record_prologue_costs) | |
720f5239 IR |
1092 | { |
1093 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 1094 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 1095 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
1096 | |
1097 | switch (alignment_support_scheme) | |
ebfd146a IR |
1098 | { |
1099 | case dr_aligned: | |
1100 | { | |
92345349 BS |
1101 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1102 | stmt_info, 0, vect_body); | |
ebfd146a | 1103 | |
73fbfcad | 1104 | if (dump_enabled_p ()) |
78c60e3d | 1105 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1106 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1107 | |
1108 | break; | |
1109 | } | |
1110 | case dr_unaligned_supported: | |
1111 | { | |
720f5239 | 1112 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1113 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1114 | unaligned_load, stmt_info, |
92345349 | 1115 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1116 | |
73fbfcad | 1117 | if (dump_enabled_p ()) |
78c60e3d SS |
1118 | dump_printf_loc (MSG_NOTE, vect_location, |
1119 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1120 | "hardware.\n"); |
ebfd146a IR |
1121 | |
1122 | break; | |
1123 | } | |
1124 | case dr_explicit_realign: | |
1125 | { | |
92345349 BS |
1126 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1127 | vector_load, stmt_info, 0, vect_body); | |
1128 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1129 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1130 | |
1131 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1132 | the containing loop, the following cost should be added to the | |
92345349 | 1133 | prologue costs. */ |
ebfd146a | 1134 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1135 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1136 | stmt_info, 0, vect_body); | |
ebfd146a | 1137 | |
73fbfcad | 1138 | if (dump_enabled_p ()) |
e645e942 TJ |
1139 | dump_printf_loc (MSG_NOTE, vect_location, |
1140 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1141 | |
ebfd146a IR |
1142 | break; |
1143 | } | |
1144 | case dr_explicit_realign_optimized: | |
1145 | { | |
73fbfcad | 1146 | if (dump_enabled_p ()) |
e645e942 | 1147 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1148 | "vect_model_load_cost: unaligned software " |
e645e942 | 1149 | "pipelined.\n"); |
ebfd146a IR |
1150 | |
1151 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1152 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1153 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1154 | access, then the above cost should only be considered for one |
ff802fa1 | 1155 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1156 | and a realignment op. */ |
1157 | ||
92345349 | 1158 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1159 | { |
92345349 BS |
1160 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1161 | vector_stmt, stmt_info, | |
1162 | 0, vect_prologue); | |
ebfd146a | 1163 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1164 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1165 | vector_stmt, stmt_info, | |
1166 | 0, vect_prologue); | |
ebfd146a IR |
1167 | } |
1168 | ||
92345349 BS |
1169 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1170 | stmt_info, 0, vect_body); | |
1171 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1172 | stmt_info, 0, vect_body); | |
8bd37302 | 1173 | |
73fbfcad | 1174 | if (dump_enabled_p ()) |
78c60e3d | 1175 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1176 | "vect_model_load_cost: explicit realign optimized" |
1177 | "\n"); | |
8bd37302 | 1178 | |
ebfd146a IR |
1179 | break; |
1180 | } | |
1181 | ||
38eec4c6 UW |
1182 | case dr_unaligned_unsupported: |
1183 | { | |
1184 | *inside_cost = VECT_MAX_COST; | |
1185 | ||
73fbfcad | 1186 | if (dump_enabled_p ()) |
78c60e3d | 1187 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1188 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1189 | break; |
1190 | } | |
1191 | ||
ebfd146a IR |
1192 | default: |
1193 | gcc_unreachable (); | |
1194 | } | |
ebfd146a IR |
1195 | } |
1196 | ||
418b7df3 RG |
1197 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1198 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1199 | |
418b7df3 | 1200 | static void |
355fe088 | 1201 | vect_init_vector_1 (gimple *stmt, gimple *new_stmt, gimple_stmt_iterator *gsi) |
ebfd146a | 1202 | { |
ebfd146a | 1203 | if (gsi) |
418b7df3 | 1204 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1205 | else |
1206 | { | |
418b7df3 | 1207 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1208 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1209 | |
a70d6342 IR |
1210 | if (loop_vinfo) |
1211 | { | |
1212 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1213 | basic_block new_bb; |
1214 | edge pe; | |
a70d6342 IR |
1215 | |
1216 | if (nested_in_vect_loop_p (loop, stmt)) | |
1217 | loop = loop->inner; | |
b8698a0f | 1218 | |
a70d6342 | 1219 | pe = loop_preheader_edge (loop); |
418b7df3 | 1220 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1221 | gcc_assert (!new_bb); |
1222 | } | |
1223 | else | |
1224 | { | |
1225 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1226 | basic_block bb; | |
1227 | gimple_stmt_iterator gsi_bb_start; | |
1228 | ||
1229 | gcc_assert (bb_vinfo); | |
1230 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1231 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1232 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1233 | } |
ebfd146a IR |
1234 | } |
1235 | ||
73fbfcad | 1236 | if (dump_enabled_p ()) |
ebfd146a | 1237 | { |
78c60e3d SS |
1238 | dump_printf_loc (MSG_NOTE, vect_location, |
1239 | "created new init_stmt: "); | |
1240 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
ebfd146a | 1241 | } |
418b7df3 RG |
1242 | } |
1243 | ||
1244 | /* Function vect_init_vector. | |
ebfd146a | 1245 | |
5467ee52 RG |
1246 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1247 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1248 | vector type a vector with all elements equal to VAL is created first. | |
1249 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1250 | initialization at the loop preheader. | |
418b7df3 RG |
1251 | Return the DEF of INIT_STMT. |
1252 | It will be used in the vectorization of STMT. */ | |
1253 | ||
1254 | tree | |
355fe088 | 1255 | vect_init_vector (gimple *stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 | 1256 | { |
355fe088 | 1257 | gimple *init_stmt; |
418b7df3 RG |
1258 | tree new_temp; |
1259 | ||
e412ece4 RB |
1260 | /* We abuse this function to push sth to a SSA name with initial 'val'. */ |
1261 | if (! useless_type_conversion_p (type, TREE_TYPE (val))) | |
418b7df3 | 1262 | { |
e412ece4 RB |
1263 | gcc_assert (TREE_CODE (type) == VECTOR_TYPE); |
1264 | if (! types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) | |
418b7df3 | 1265 | { |
5a308cf1 IE |
1266 | /* Scalar boolean value should be transformed into |
1267 | all zeros or all ones value before building a vector. */ | |
1268 | if (VECTOR_BOOLEAN_TYPE_P (type)) | |
1269 | { | |
b3d51f23 IE |
1270 | tree true_val = build_all_ones_cst (TREE_TYPE (type)); |
1271 | tree false_val = build_zero_cst (TREE_TYPE (type)); | |
5a308cf1 IE |
1272 | |
1273 | if (CONSTANT_CLASS_P (val)) | |
1274 | val = integer_zerop (val) ? false_val : true_val; | |
1275 | else | |
1276 | { | |
1277 | new_temp = make_ssa_name (TREE_TYPE (type)); | |
1278 | init_stmt = gimple_build_assign (new_temp, COND_EXPR, | |
1279 | val, true_val, false_val); | |
1280 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
1281 | val = new_temp; | |
1282 | } | |
1283 | } | |
1284 | else if (CONSTANT_CLASS_P (val)) | |
42fd8198 | 1285 | val = fold_convert (TREE_TYPE (type), val); |
418b7df3 RG |
1286 | else |
1287 | { | |
b731b390 | 1288 | new_temp = make_ssa_name (TREE_TYPE (type)); |
e412ece4 RB |
1289 | if (! INTEGRAL_TYPE_P (TREE_TYPE (val))) |
1290 | init_stmt = gimple_build_assign (new_temp, | |
1291 | fold_build1 (VIEW_CONVERT_EXPR, | |
1292 | TREE_TYPE (type), | |
1293 | val)); | |
1294 | else | |
1295 | init_stmt = gimple_build_assign (new_temp, NOP_EXPR, val); | |
418b7df3 | 1296 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1297 | val = new_temp; |
418b7df3 RG |
1298 | } |
1299 | } | |
5467ee52 | 1300 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1301 | } |
1302 | ||
0e22bb5a RB |
1303 | new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_"); |
1304 | init_stmt = gimple_build_assign (new_temp, val); | |
418b7df3 | 1305 | vect_init_vector_1 (stmt, init_stmt, gsi); |
0e22bb5a | 1306 | return new_temp; |
ebfd146a IR |
1307 | } |
1308 | ||
c83a894c | 1309 | /* Function vect_get_vec_def_for_operand_1. |
a70d6342 | 1310 | |
c83a894c AH |
1311 | For a defining stmt DEF_STMT of a scalar stmt, return a vector def with type |
1312 | DT that will be used in the vectorized stmt. */ | |
ebfd146a IR |
1313 | |
1314 | tree | |
c83a894c | 1315 | vect_get_vec_def_for_operand_1 (gimple *def_stmt, enum vect_def_type dt) |
ebfd146a IR |
1316 | { |
1317 | tree vec_oprnd; | |
355fe088 | 1318 | gimple *vec_stmt; |
ebfd146a | 1319 | stmt_vec_info def_stmt_info = NULL; |
ebfd146a IR |
1320 | |
1321 | switch (dt) | |
1322 | { | |
81c40241 | 1323 | /* operand is a constant or a loop invariant. */ |
ebfd146a | 1324 | case vect_constant_def: |
81c40241 | 1325 | case vect_external_def: |
c83a894c AH |
1326 | /* Code should use vect_get_vec_def_for_operand. */ |
1327 | gcc_unreachable (); | |
ebfd146a | 1328 | |
81c40241 | 1329 | /* operand is defined inside the loop. */ |
8644a673 | 1330 | case vect_internal_def: |
ebfd146a | 1331 | { |
ebfd146a IR |
1332 | /* Get the def from the vectorized stmt. */ |
1333 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1334 | |
ebfd146a | 1335 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1336 | /* Get vectorized pattern statement. */ |
1337 | if (!vec_stmt | |
1338 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1339 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1340 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1341 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1342 | gcc_assert (vec_stmt); |
1343 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1344 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1345 | else if (is_gimple_call (vec_stmt)) | |
1346 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1347 | else | |
1348 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1349 | return vec_oprnd; | |
1350 | } | |
1351 | ||
81c40241 | 1352 | /* operand is defined by a loop header phi - reduction */ |
ebfd146a | 1353 | case vect_reduction_def: |
06066f92 | 1354 | case vect_double_reduction_def: |
7c5222ff | 1355 | case vect_nested_cycle: |
81c40241 RB |
1356 | /* Code should use get_initial_def_for_reduction. */ |
1357 | gcc_unreachable (); | |
ebfd146a | 1358 | |
81c40241 | 1359 | /* operand is defined by loop-header phi - induction. */ |
ebfd146a IR |
1360 | case vect_induction_def: |
1361 | { | |
1362 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1363 | ||
1364 | /* Get the def from the vectorized stmt. */ | |
1365 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1366 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1367 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1368 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1369 | else | |
1370 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1371 | return vec_oprnd; |
1372 | } | |
1373 | ||
1374 | default: | |
1375 | gcc_unreachable (); | |
1376 | } | |
1377 | } | |
1378 | ||
1379 | ||
c83a894c AH |
1380 | /* Function vect_get_vec_def_for_operand. |
1381 | ||
1382 | OP is an operand in STMT. This function returns a (vector) def that will be | |
1383 | used in the vectorized stmt for STMT. | |
1384 | ||
1385 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1386 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1387 | ||
1388 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1389 | needs to be introduced. VECTYPE may be used to specify a required type for | |
1390 | vector invariant. */ | |
1391 | ||
1392 | tree | |
1393 | vect_get_vec_def_for_operand (tree op, gimple *stmt, tree vectype) | |
1394 | { | |
1395 | gimple *def_stmt; | |
1396 | enum vect_def_type dt; | |
1397 | bool is_simple_use; | |
1398 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
1399 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
1400 | ||
1401 | if (dump_enabled_p ()) | |
1402 | { | |
1403 | dump_printf_loc (MSG_NOTE, vect_location, | |
1404 | "vect_get_vec_def_for_operand: "); | |
1405 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
1406 | dump_printf (MSG_NOTE, "\n"); | |
1407 | } | |
1408 | ||
1409 | is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt); | |
1410 | gcc_assert (is_simple_use); | |
1411 | if (def_stmt && dump_enabled_p ()) | |
1412 | { | |
1413 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1414 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
1415 | } | |
1416 | ||
1417 | if (dt == vect_constant_def || dt == vect_external_def) | |
1418 | { | |
1419 | tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo); | |
1420 | tree vector_type; | |
1421 | ||
1422 | if (vectype) | |
1423 | vector_type = vectype; | |
1424 | else if (TREE_CODE (TREE_TYPE (op)) == BOOLEAN_TYPE | |
1425 | && VECTOR_BOOLEAN_TYPE_P (stmt_vectype)) | |
1426 | vector_type = build_same_sized_truth_vector_type (stmt_vectype); | |
1427 | else | |
1428 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
1429 | ||
1430 | gcc_assert (vector_type); | |
1431 | return vect_init_vector (stmt, op, vector_type, NULL); | |
1432 | } | |
1433 | else | |
1434 | return vect_get_vec_def_for_operand_1 (def_stmt, dt); | |
1435 | } | |
1436 | ||
1437 | ||
ebfd146a IR |
1438 | /* Function vect_get_vec_def_for_stmt_copy |
1439 | ||
ff802fa1 | 1440 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1441 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1442 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1443 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1444 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1445 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1446 | DT is the type of the vector def VEC_OPRND. |
1447 | ||
1448 | Context: | |
1449 | In case the vectorization factor (VF) is bigger than the number | |
1450 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1451 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1452 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1453 | smallest data-type determines the VF, and as a result, when vectorizing |
1454 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1455 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1456 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1457 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1458 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1459 | ||
1460 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1461 | |
ebfd146a IR |
1462 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1463 | VS1.1: vx.1 = memref1 VS1.2 | |
1464 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1465 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1466 | |
1467 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1468 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1469 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1470 | VSnew.3: vz3 = vx.3 + ... | |
1471 | ||
1472 | The vectorization of S1 is explained in vectorizable_load. | |
1473 | The vectorization of S2: | |
b8698a0f L |
1474 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1475 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1476 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1477 | returns the vector-def 'vx.0'. |
1478 | ||
b8698a0f L |
1479 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1480 | function is called to get the relevant vector-def for each operand. It is | |
1481 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1482 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1483 | ||
b8698a0f L |
1484 | For example, to obtain the vector-def 'vx.1' in order to create the |
1485 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1486 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1487 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1488 | and return its def ('vx.1'). | |
1489 | Overall, to create the above sequence this function will be called 3 times: | |
1490 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1491 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1492 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1493 | ||
1494 | tree | |
1495 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1496 | { | |
355fe088 | 1497 | gimple *vec_stmt_for_operand; |
ebfd146a IR |
1498 | stmt_vec_info def_stmt_info; |
1499 | ||
1500 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1501 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1502 | return vec_oprnd; |
1503 | ||
1504 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1505 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1506 | gcc_assert (def_stmt_info); | |
1507 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1508 | gcc_assert (vec_stmt_for_operand); | |
ebfd146a IR |
1509 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) |
1510 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1511 | else | |
1512 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1513 | return vec_oprnd; | |
1514 | } | |
1515 | ||
1516 | ||
1517 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1518 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a IR |
1519 | |
1520 | static void | |
b8698a0f | 1521 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1522 | vec<tree> *vec_oprnds0, |
1523 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1524 | { |
9771b263 | 1525 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1526 | |
1527 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1528 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1529 | |
9771b263 | 1530 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1531 | { |
9771b263 | 1532 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1533 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1534 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1535 | } |
1536 | } | |
1537 | ||
1538 | ||
d092494c IR |
1539 | /* Get vectorized definitions for OP0 and OP1. |
1540 | REDUC_INDEX is the index of reduction operand in case of reduction, | |
1541 | and -1 otherwise. */ | |
ebfd146a | 1542 | |
d092494c | 1543 | void |
355fe088 | 1544 | vect_get_vec_defs (tree op0, tree op1, gimple *stmt, |
9771b263 DN |
1545 | vec<tree> *vec_oprnds0, |
1546 | vec<tree> *vec_oprnds1, | |
d092494c | 1547 | slp_tree slp_node, int reduc_index) |
ebfd146a IR |
1548 | { |
1549 | if (slp_node) | |
d092494c IR |
1550 | { |
1551 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1552 | auto_vec<tree> ops (nops); |
1553 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1554 | |
9771b263 | 1555 | ops.quick_push (op0); |
d092494c | 1556 | if (op1) |
9771b263 | 1557 | ops.quick_push (op1); |
d092494c IR |
1558 | |
1559 | vect_get_slp_defs (ops, slp_node, &vec_defs, reduc_index); | |
1560 | ||
37b5ec8f | 1561 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1562 | if (op1) |
37b5ec8f | 1563 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1564 | } |
ebfd146a IR |
1565 | else |
1566 | { | |
1567 | tree vec_oprnd; | |
1568 | ||
9771b263 | 1569 | vec_oprnds0->create (1); |
81c40241 | 1570 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 1571 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1572 | |
1573 | if (op1) | |
1574 | { | |
9771b263 | 1575 | vec_oprnds1->create (1); |
81c40241 | 1576 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 1577 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1578 | } |
1579 | } | |
1580 | } | |
1581 | ||
1582 | ||
1583 | /* Function vect_finish_stmt_generation. | |
1584 | ||
1585 | Insert a new stmt. */ | |
1586 | ||
1587 | void | |
355fe088 | 1588 | vect_finish_stmt_generation (gimple *stmt, gimple *vec_stmt, |
ebfd146a IR |
1589 | gimple_stmt_iterator *gsi) |
1590 | { | |
1591 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
310213d4 | 1592 | vec_info *vinfo = stmt_info->vinfo; |
ebfd146a IR |
1593 | |
1594 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); | |
1595 | ||
54e8e2c3 RG |
1596 | if (!gsi_end_p (*gsi) |
1597 | && gimple_has_mem_ops (vec_stmt)) | |
1598 | { | |
355fe088 | 1599 | gimple *at_stmt = gsi_stmt (*gsi); |
54e8e2c3 RG |
1600 | tree vuse = gimple_vuse (at_stmt); |
1601 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1602 | { | |
1603 | tree vdef = gimple_vdef (at_stmt); | |
1604 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1605 | /* If we have an SSA vuse and insert a store, update virtual | |
1606 | SSA form to avoid triggering the renamer. Do so only | |
1607 | if we can easily see all uses - which is what almost always | |
1608 | happens with the way vectorized stmts are inserted. */ | |
1609 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1610 | && ((is_gimple_assign (vec_stmt) | |
1611 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1612 | || (is_gimple_call (vec_stmt) | |
1613 | && !(gimple_call_flags (vec_stmt) | |
1614 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1615 | { | |
1616 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1617 | gimple_set_vdef (vec_stmt, new_vdef); | |
1618 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1619 | } | |
1620 | } | |
1621 | } | |
ebfd146a IR |
1622 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
1623 | ||
310213d4 | 1624 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, vinfo)); |
ebfd146a | 1625 | |
73fbfcad | 1626 | if (dump_enabled_p ()) |
ebfd146a | 1627 | { |
78c60e3d SS |
1628 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); |
1629 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
ebfd146a IR |
1630 | } |
1631 | ||
ad885386 | 1632 | gimple_set_location (vec_stmt, gimple_location (stmt)); |
8e91d222 JJ |
1633 | |
1634 | /* While EH edges will generally prevent vectorization, stmt might | |
1635 | e.g. be in a must-not-throw region. Ensure newly created stmts | |
1636 | that could throw are part of the same region. */ | |
1637 | int lp_nr = lookup_stmt_eh_lp (stmt); | |
1638 | if (lp_nr != 0 && stmt_could_throw_p (vec_stmt)) | |
1639 | add_stmt_to_eh_lp (vec_stmt, lp_nr); | |
ebfd146a IR |
1640 | } |
1641 | ||
70439f0d RS |
1642 | /* We want to vectorize a call to combined function CFN with function |
1643 | decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN | |
1644 | as the types of all inputs. Check whether this is possible using | |
1645 | an internal function, returning its code if so or IFN_LAST if not. */ | |
ebfd146a | 1646 | |
70439f0d RS |
1647 | static internal_fn |
1648 | vectorizable_internal_function (combined_fn cfn, tree fndecl, | |
1649 | tree vectype_out, tree vectype_in) | |
ebfd146a | 1650 | { |
70439f0d RS |
1651 | internal_fn ifn; |
1652 | if (internal_fn_p (cfn)) | |
1653 | ifn = as_internal_fn (cfn); | |
1654 | else | |
1655 | ifn = associated_internal_fn (fndecl); | |
1656 | if (ifn != IFN_LAST && direct_internal_fn_p (ifn)) | |
1657 | { | |
1658 | const direct_internal_fn_info &info = direct_internal_fn (ifn); | |
1659 | if (info.vectorizable) | |
1660 | { | |
1661 | tree type0 = (info.type0 < 0 ? vectype_out : vectype_in); | |
1662 | tree type1 = (info.type1 < 0 ? vectype_out : vectype_in); | |
d95ab70a RS |
1663 | if (direct_internal_fn_supported_p (ifn, tree_pair (type0, type1), |
1664 | OPTIMIZE_FOR_SPEED)) | |
70439f0d RS |
1665 | return ifn; |
1666 | } | |
1667 | } | |
1668 | return IFN_LAST; | |
ebfd146a IR |
1669 | } |
1670 | ||
5ce9450f | 1671 | |
355fe088 | 1672 | static tree permute_vec_elements (tree, tree, tree, gimple *, |
5ce9450f JJ |
1673 | gimple_stmt_iterator *); |
1674 | ||
1675 | ||
1676 | /* Function vectorizable_mask_load_store. | |
1677 | ||
1678 | Check if STMT performs a conditional load or store that can be vectorized. | |
1679 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
1680 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. | |
1681 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
1682 | ||
1683 | static bool | |
355fe088 TS |
1684 | vectorizable_mask_load_store (gimple *stmt, gimple_stmt_iterator *gsi, |
1685 | gimple **vec_stmt, slp_tree slp_node) | |
5ce9450f JJ |
1686 | { |
1687 | tree vec_dest = NULL; | |
1688 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1689 | stmt_vec_info prev_stmt_info; | |
1690 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
1691 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1692 | bool nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
1693 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
1694 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
57e2f6ad | 1695 | tree rhs_vectype = NULL_TREE; |
045c1278 | 1696 | tree mask_vectype; |
5ce9450f | 1697 | tree elem_type; |
355fe088 | 1698 | gimple *new_stmt; |
5ce9450f JJ |
1699 | tree dummy; |
1700 | tree dataref_ptr = NULL_TREE; | |
355fe088 | 1701 | gimple *ptr_incr; |
5ce9450f JJ |
1702 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
1703 | int ncopies; | |
1704 | int i, j; | |
1705 | bool inv_p; | |
1706 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; | |
1707 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
1708 | int gather_scale = 1; | |
1709 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
1710 | bool is_store; | |
1711 | tree mask; | |
355fe088 | 1712 | gimple *def_stmt; |
5ce9450f JJ |
1713 | enum vect_def_type dt; |
1714 | ||
1715 | if (slp_node != NULL) | |
1716 | return false; | |
1717 | ||
1718 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
1719 | gcc_assert (ncopies >= 1); | |
1720 | ||
1721 | is_store = gimple_call_internal_fn (stmt) == IFN_MASK_STORE; | |
1722 | mask = gimple_call_arg (stmt, 2); | |
045c1278 IE |
1723 | |
1724 | if (TREE_CODE (TREE_TYPE (mask)) != BOOLEAN_TYPE) | |
5ce9450f JJ |
1725 | return false; |
1726 | ||
1727 | /* FORNOW. This restriction should be relaxed. */ | |
1728 | if (nested_in_vect_loop && ncopies > 1) | |
1729 | { | |
1730 | if (dump_enabled_p ()) | |
1731 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1732 | "multiple types in nested loop."); | |
1733 | return false; | |
1734 | } | |
1735 | ||
1736 | if (!STMT_VINFO_RELEVANT_P (stmt_info)) | |
1737 | return false; | |
1738 | ||
66c16fd9 RB |
1739 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
1740 | && ! vec_stmt) | |
5ce9450f JJ |
1741 | return false; |
1742 | ||
1743 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
1744 | return false; | |
1745 | ||
1746 | elem_type = TREE_TYPE (vectype); | |
1747 | ||
1748 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
1749 | return false; | |
1750 | ||
f2e2a985 | 1751 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
5ce9450f JJ |
1752 | return false; |
1753 | ||
045c1278 IE |
1754 | if (TREE_CODE (mask) != SSA_NAME) |
1755 | return false; | |
1756 | ||
1757 | if (!vect_is_simple_use (mask, loop_vinfo, &def_stmt, &dt, &mask_vectype)) | |
1758 | return false; | |
1759 | ||
1760 | if (!mask_vectype) | |
1761 | mask_vectype = get_mask_type_for_scalar_type (TREE_TYPE (vectype)); | |
1762 | ||
dc6a3147 IE |
1763 | if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype) |
1764 | || TYPE_VECTOR_SUBPARTS (mask_vectype) != TYPE_VECTOR_SUBPARTS (vectype)) | |
045c1278 IE |
1765 | return false; |
1766 | ||
57e2f6ad IE |
1767 | if (is_store) |
1768 | { | |
1769 | tree rhs = gimple_call_arg (stmt, 3); | |
1770 | if (!vect_is_simple_use (rhs, loop_vinfo, &def_stmt, &dt, &rhs_vectype)) | |
1771 | return false; | |
1772 | } | |
1773 | ||
3bab6342 | 1774 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5ce9450f | 1775 | { |
355fe088 | 1776 | gimple *def_stmt; |
3bab6342 | 1777 | gather_decl = vect_check_gather_scatter (stmt, loop_vinfo, &gather_base, |
5ce9450f JJ |
1778 | &gather_off, &gather_scale); |
1779 | gcc_assert (gather_decl); | |
81c40241 RB |
1780 | if (!vect_is_simple_use (gather_off, loop_vinfo, &def_stmt, &gather_dt, |
1781 | &gather_off_vectype)) | |
5ce9450f JJ |
1782 | { |
1783 | if (dump_enabled_p ()) | |
1784 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1785 | "gather index use not simple."); | |
1786 | return false; | |
1787 | } | |
03b9e8e4 JJ |
1788 | |
1789 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
1790 | tree masktype | |
1791 | = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (TREE_CHAIN (arglist)))); | |
1792 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
1793 | { | |
1794 | if (dump_enabled_p ()) | |
1795 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1796 | "masked gather with integer mask not supported."); | |
1797 | return false; | |
1798 | } | |
5ce9450f JJ |
1799 | } |
1800 | else if (tree_int_cst_compare (nested_in_vect_loop | |
1801 | ? STMT_VINFO_DR_STEP (stmt_info) | |
1802 | : DR_STEP (dr), size_zero_node) <= 0) | |
1803 | return false; | |
1804 | else if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
045c1278 IE |
1805 | || !can_vec_mask_load_store_p (TYPE_MODE (vectype), |
1806 | TYPE_MODE (mask_vectype), | |
57e2f6ad IE |
1807 | !is_store) |
1808 | || (rhs_vectype | |
1809 | && !useless_type_conversion_p (vectype, rhs_vectype))) | |
5ce9450f JJ |
1810 | return false; |
1811 | ||
5ce9450f JJ |
1812 | if (!vec_stmt) /* transformation not required. */ |
1813 | { | |
1814 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
1815 | if (is_store) | |
1816 | vect_model_store_cost (stmt_info, ncopies, false, dt, | |
1817 | NULL, NULL, NULL); | |
1818 | else | |
1819 | vect_model_load_cost (stmt_info, ncopies, false, NULL, NULL, NULL); | |
1820 | return true; | |
1821 | } | |
1822 | ||
1823 | /** Transform. **/ | |
1824 | ||
3bab6342 | 1825 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5ce9450f JJ |
1826 | { |
1827 | tree vec_oprnd0 = NULL_TREE, op; | |
1828 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
1829 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
acdcd61b | 1830 | tree ptr, vec_mask = NULL_TREE, mask_op = NULL_TREE, var, scale; |
5ce9450f | 1831 | tree perm_mask = NULL_TREE, prev_res = NULL_TREE; |
acdcd61b | 1832 | tree mask_perm_mask = NULL_TREE; |
5ce9450f JJ |
1833 | edge pe = loop_preheader_edge (loop); |
1834 | gimple_seq seq; | |
1835 | basic_block new_bb; | |
1836 | enum { NARROW, NONE, WIDEN } modifier; | |
1837 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
1838 | ||
acdcd61b JJ |
1839 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); |
1840 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1841 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1842 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1843 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
1844 | scaletype = TREE_VALUE (arglist); | |
1845 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
1846 | && types_compatible_p (srctype, masktype)); | |
1847 | ||
5ce9450f JJ |
1848 | if (nunits == gather_off_nunits) |
1849 | modifier = NONE; | |
1850 | else if (nunits == gather_off_nunits / 2) | |
1851 | { | |
1852 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
1853 | modifier = WIDEN; | |
1854 | ||
1855 | for (i = 0; i < gather_off_nunits; ++i) | |
1856 | sel[i] = i | nunits; | |
1857 | ||
557be5a8 | 1858 | perm_mask = vect_gen_perm_mask_checked (gather_off_vectype, sel); |
5ce9450f JJ |
1859 | } |
1860 | else if (nunits == gather_off_nunits * 2) | |
1861 | { | |
1862 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
1863 | modifier = NARROW; | |
1864 | ||
1865 | for (i = 0; i < nunits; ++i) | |
1866 | sel[i] = i < gather_off_nunits | |
1867 | ? i : i + nunits - gather_off_nunits; | |
1868 | ||
557be5a8 | 1869 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); |
5ce9450f | 1870 | ncopies *= 2; |
acdcd61b JJ |
1871 | for (i = 0; i < nunits; ++i) |
1872 | sel[i] = i | gather_off_nunits; | |
557be5a8 | 1873 | mask_perm_mask = vect_gen_perm_mask_checked (masktype, sel); |
5ce9450f JJ |
1874 | } |
1875 | else | |
1876 | gcc_unreachable (); | |
1877 | ||
5ce9450f JJ |
1878 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); |
1879 | ||
1880 | ptr = fold_convert (ptrtype, gather_base); | |
1881 | if (!is_gimple_min_invariant (ptr)) | |
1882 | { | |
1883 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
1884 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
1885 | gcc_assert (!new_bb); | |
1886 | } | |
1887 | ||
1888 | scale = build_int_cst (scaletype, gather_scale); | |
1889 | ||
1890 | prev_stmt_info = NULL; | |
1891 | for (j = 0; j < ncopies; ++j) | |
1892 | { | |
1893 | if (modifier == WIDEN && (j & 1)) | |
1894 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
1895 | perm_mask, stmt, gsi); | |
1896 | else if (j == 0) | |
1897 | op = vec_oprnd0 | |
81c40241 | 1898 | = vect_get_vec_def_for_operand (gather_off, stmt); |
5ce9450f JJ |
1899 | else |
1900 | op = vec_oprnd0 | |
1901 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
1902 | ||
1903 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
1904 | { | |
1905 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
1906 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 1907 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
5ce9450f JJ |
1908 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
1909 | new_stmt | |
0d0e4a03 | 1910 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
1911 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1912 | op = var; | |
1913 | } | |
1914 | ||
acdcd61b JJ |
1915 | if (mask_perm_mask && (j & 1)) |
1916 | mask_op = permute_vec_elements (mask_op, mask_op, | |
1917 | mask_perm_mask, stmt, gsi); | |
5ce9450f JJ |
1918 | else |
1919 | { | |
acdcd61b | 1920 | if (j == 0) |
81c40241 | 1921 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); |
acdcd61b JJ |
1922 | else |
1923 | { | |
81c40241 | 1924 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
acdcd61b JJ |
1925 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
1926 | } | |
5ce9450f | 1927 | |
acdcd61b JJ |
1928 | mask_op = vec_mask; |
1929 | if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask))) | |
1930 | { | |
1931 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)) | |
1932 | == TYPE_VECTOR_SUBPARTS (masktype)); | |
0e22bb5a | 1933 | var = vect_get_new_ssa_name (masktype, vect_simple_var); |
acdcd61b JJ |
1934 | mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op); |
1935 | new_stmt | |
0d0e4a03 | 1936 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, mask_op); |
acdcd61b JJ |
1937 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
1938 | mask_op = var; | |
1939 | } | |
5ce9450f JJ |
1940 | } |
1941 | ||
1942 | new_stmt | |
1943 | = gimple_build_call (gather_decl, 5, mask_op, ptr, op, mask_op, | |
1944 | scale); | |
1945 | ||
1946 | if (!useless_type_conversion_p (vectype, rettype)) | |
1947 | { | |
1948 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
1949 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 1950 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
5ce9450f JJ |
1951 | gimple_call_set_lhs (new_stmt, op); |
1952 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 1953 | var = make_ssa_name (vec_dest); |
5ce9450f | 1954 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
0d0e4a03 | 1955 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
5ce9450f JJ |
1956 | } |
1957 | else | |
1958 | { | |
1959 | var = make_ssa_name (vec_dest, new_stmt); | |
1960 | gimple_call_set_lhs (new_stmt, var); | |
1961 | } | |
1962 | ||
1963 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
1964 | ||
1965 | if (modifier == NARROW) | |
1966 | { | |
1967 | if ((j & 1) == 0) | |
1968 | { | |
1969 | prev_res = var; | |
1970 | continue; | |
1971 | } | |
1972 | var = permute_vec_elements (prev_res, var, | |
1973 | perm_mask, stmt, gsi); | |
1974 | new_stmt = SSA_NAME_DEF_STMT (var); | |
1975 | } | |
1976 | ||
1977 | if (prev_stmt_info == NULL) | |
1978 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
1979 | else | |
1980 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
1981 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
1982 | } | |
3efe2e2c JJ |
1983 | |
1984 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
1985 | from the IL. */ | |
e6f5c25d IE |
1986 | if (STMT_VINFO_RELATED_STMT (stmt_info)) |
1987 | { | |
1988 | stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
1989 | stmt_info = vinfo_for_stmt (stmt); | |
1990 | } | |
3efe2e2c JJ |
1991 | tree lhs = gimple_call_lhs (stmt); |
1992 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
1993 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
1994 | set_vinfo_for_stmt (stmt, NULL); | |
1995 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
1996 | gsi_replace (gsi, new_stmt, true); | |
5ce9450f JJ |
1997 | return true; |
1998 | } | |
1999 | else if (is_store) | |
2000 | { | |
2001 | tree vec_rhs = NULL_TREE, vec_mask = NULL_TREE; | |
2002 | prev_stmt_info = NULL; | |
2d4dc223 | 2003 | LOOP_VINFO_HAS_MASK_STORE (loop_vinfo) = true; |
5ce9450f JJ |
2004 | for (i = 0; i < ncopies; i++) |
2005 | { | |
2006 | unsigned align, misalign; | |
2007 | ||
2008 | if (i == 0) | |
2009 | { | |
2010 | tree rhs = gimple_call_arg (stmt, 3); | |
81c40241 RB |
2011 | vec_rhs = vect_get_vec_def_for_operand (rhs, stmt); |
2012 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); | |
5ce9450f JJ |
2013 | /* We should have catched mismatched types earlier. */ |
2014 | gcc_assert (useless_type_conversion_p (vectype, | |
2015 | TREE_TYPE (vec_rhs))); | |
2016 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, | |
2017 | NULL_TREE, &dummy, gsi, | |
2018 | &ptr_incr, false, &inv_p); | |
2019 | gcc_assert (!inv_p); | |
2020 | } | |
2021 | else | |
2022 | { | |
81c40241 | 2023 | vect_is_simple_use (vec_rhs, loop_vinfo, &def_stmt, &dt); |
5ce9450f | 2024 | vec_rhs = vect_get_vec_def_for_stmt_copy (dt, vec_rhs); |
81c40241 | 2025 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2026 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2027 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2028 | TYPE_SIZE_UNIT (vectype)); | |
2029 | } | |
2030 | ||
2031 | align = TYPE_ALIGN_UNIT (vectype); | |
2032 | if (aligned_access_p (dr)) | |
2033 | misalign = 0; | |
2034 | else if (DR_MISALIGNMENT (dr) == -1) | |
2035 | { | |
2036 | align = TYPE_ALIGN_UNIT (elem_type); | |
2037 | misalign = 0; | |
2038 | } | |
2039 | else | |
2040 | misalign = DR_MISALIGNMENT (dr); | |
2041 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2042 | misalign); | |
08554c26 JJ |
2043 | tree ptr = build_int_cst (TREE_TYPE (gimple_call_arg (stmt, 1)), |
2044 | misalign ? misalign & -misalign : align); | |
5ce9450f JJ |
2045 | new_stmt |
2046 | = gimple_build_call_internal (IFN_MASK_STORE, 4, dataref_ptr, | |
08554c26 | 2047 | ptr, vec_mask, vec_rhs); |
5ce9450f JJ |
2048 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2049 | if (i == 0) | |
2050 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2051 | else | |
2052 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2053 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2054 | } | |
2055 | } | |
2056 | else | |
2057 | { | |
2058 | tree vec_mask = NULL_TREE; | |
2059 | prev_stmt_info = NULL; | |
2060 | vec_dest = vect_create_destination_var (gimple_call_lhs (stmt), vectype); | |
2061 | for (i = 0; i < ncopies; i++) | |
2062 | { | |
2063 | unsigned align, misalign; | |
2064 | ||
2065 | if (i == 0) | |
2066 | { | |
81c40241 | 2067 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); |
5ce9450f JJ |
2068 | dataref_ptr = vect_create_data_ref_ptr (stmt, vectype, NULL, |
2069 | NULL_TREE, &dummy, gsi, | |
2070 | &ptr_incr, false, &inv_p); | |
2071 | gcc_assert (!inv_p); | |
2072 | } | |
2073 | else | |
2074 | { | |
81c40241 | 2075 | vect_is_simple_use (vec_mask, loop_vinfo, &def_stmt, &dt); |
5ce9450f JJ |
2076 | vec_mask = vect_get_vec_def_for_stmt_copy (dt, vec_mask); |
2077 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
2078 | TYPE_SIZE_UNIT (vectype)); | |
2079 | } | |
2080 | ||
2081 | align = TYPE_ALIGN_UNIT (vectype); | |
2082 | if (aligned_access_p (dr)) | |
2083 | misalign = 0; | |
2084 | else if (DR_MISALIGNMENT (dr) == -1) | |
2085 | { | |
2086 | align = TYPE_ALIGN_UNIT (elem_type); | |
2087 | misalign = 0; | |
2088 | } | |
2089 | else | |
2090 | misalign = DR_MISALIGNMENT (dr); | |
2091 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, | |
2092 | misalign); | |
08554c26 JJ |
2093 | tree ptr = build_int_cst (TREE_TYPE (gimple_call_arg (stmt, 1)), |
2094 | misalign ? misalign & -misalign : align); | |
5ce9450f JJ |
2095 | new_stmt |
2096 | = gimple_build_call_internal (IFN_MASK_LOAD, 3, dataref_ptr, | |
08554c26 | 2097 | ptr, vec_mask); |
b731b390 | 2098 | gimple_call_set_lhs (new_stmt, make_ssa_name (vec_dest)); |
5ce9450f JJ |
2099 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2100 | if (i == 0) | |
2101 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2102 | else | |
2103 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2104 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2105 | } | |
2106 | } | |
2107 | ||
3efe2e2c JJ |
2108 | if (!is_store) |
2109 | { | |
2110 | /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed | |
2111 | from the IL. */ | |
e6f5c25d IE |
2112 | if (STMT_VINFO_RELATED_STMT (stmt_info)) |
2113 | { | |
2114 | stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
2115 | stmt_info = vinfo_for_stmt (stmt); | |
2116 | } | |
3efe2e2c JJ |
2117 | tree lhs = gimple_call_lhs (stmt); |
2118 | new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs))); | |
2119 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
2120 | set_vinfo_for_stmt (stmt, NULL); | |
2121 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2122 | gsi_replace (gsi, new_stmt, true); | |
2123 | } | |
2124 | ||
5ce9450f JJ |
2125 | return true; |
2126 | } | |
2127 | ||
b1b6836e RS |
2128 | /* Return true if vector types VECTYPE_IN and VECTYPE_OUT have |
2129 | integer elements and if we can narrow VECTYPE_IN to VECTYPE_OUT | |
2130 | in a single step. On success, store the binary pack code in | |
2131 | *CONVERT_CODE. */ | |
2132 | ||
2133 | static bool | |
2134 | simple_integer_narrowing (tree vectype_out, tree vectype_in, | |
2135 | tree_code *convert_code) | |
2136 | { | |
2137 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_out)) | |
2138 | || !INTEGRAL_TYPE_P (TREE_TYPE (vectype_in))) | |
2139 | return false; | |
2140 | ||
2141 | tree_code code; | |
2142 | int multi_step_cvt = 0; | |
2143 | auto_vec <tree, 8> interm_types; | |
2144 | if (!supportable_narrowing_operation (NOP_EXPR, vectype_out, vectype_in, | |
2145 | &code, &multi_step_cvt, | |
2146 | &interm_types) | |
2147 | || multi_step_cvt) | |
2148 | return false; | |
2149 | ||
2150 | *convert_code = code; | |
2151 | return true; | |
2152 | } | |
5ce9450f | 2153 | |
ebfd146a IR |
2154 | /* Function vectorizable_call. |
2155 | ||
538dd0b7 | 2156 | Check if GS performs a function call that can be vectorized. |
b8698a0f | 2157 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
2158 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2159 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2160 | ||
2161 | static bool | |
355fe088 | 2162 | vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
190c2236 | 2163 | slp_tree slp_node) |
ebfd146a | 2164 | { |
538dd0b7 | 2165 | gcall *stmt; |
ebfd146a IR |
2166 | tree vec_dest; |
2167 | tree scalar_dest; | |
2168 | tree op, type; | |
2169 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
538dd0b7 | 2170 | stmt_vec_info stmt_info = vinfo_for_stmt (gs), prev_stmt_info; |
ebfd146a IR |
2171 | tree vectype_out, vectype_in; |
2172 | int nunits_in; | |
2173 | int nunits_out; | |
2174 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
190c2236 | 2175 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 2176 | vec_info *vinfo = stmt_info->vinfo; |
81c40241 | 2177 | tree fndecl, new_temp, rhs_type; |
355fe088 | 2178 | gimple *def_stmt; |
0502fb85 UB |
2179 | enum vect_def_type dt[3] |
2180 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
355fe088 | 2181 | gimple *new_stmt = NULL; |
ebfd146a | 2182 | int ncopies, j; |
6e1aa848 | 2183 | vec<tree> vargs = vNULL; |
ebfd146a IR |
2184 | enum { NARROW, NONE, WIDEN } modifier; |
2185 | size_t i, nargs; | |
9d5e7640 | 2186 | tree lhs; |
ebfd146a | 2187 | |
190c2236 | 2188 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
2189 | return false; |
2190 | ||
66c16fd9 RB |
2191 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
2192 | && ! vec_stmt) | |
ebfd146a IR |
2193 | return false; |
2194 | ||
538dd0b7 DM |
2195 | /* Is GS a vectorizable call? */ |
2196 | stmt = dyn_cast <gcall *> (gs); | |
2197 | if (!stmt) | |
ebfd146a IR |
2198 | return false; |
2199 | ||
5ce9450f JJ |
2200 | if (gimple_call_internal_p (stmt) |
2201 | && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD | |
2202 | || gimple_call_internal_fn (stmt) == IFN_MASK_STORE)) | |
2203 | return vectorizable_mask_load_store (stmt, gsi, vec_stmt, | |
2204 | slp_node); | |
2205 | ||
0136f8f0 AH |
2206 | if (gimple_call_lhs (stmt) == NULL_TREE |
2207 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
ebfd146a IR |
2208 | return false; |
2209 | ||
0136f8f0 | 2210 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); |
5a2c1986 | 2211 | |
b690cc0f RG |
2212 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
2213 | ||
ebfd146a IR |
2214 | /* Process function arguments. */ |
2215 | rhs_type = NULL_TREE; | |
b690cc0f | 2216 | vectype_in = NULL_TREE; |
ebfd146a IR |
2217 | nargs = gimple_call_num_args (stmt); |
2218 | ||
1b1562a5 MM |
2219 | /* Bail out if the function has more than three arguments, we do not have |
2220 | interesting builtin functions to vectorize with more than two arguments | |
2221 | except for fma. No arguments is also not good. */ | |
2222 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
2223 | return false; |
2224 | ||
74bf76ed JJ |
2225 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
2226 | if (gimple_call_internal_p (stmt) | |
2227 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2228 | { | |
2229 | nargs = 0; | |
2230 | rhs_type = unsigned_type_node; | |
2231 | } | |
2232 | ||
ebfd146a IR |
2233 | for (i = 0; i < nargs; i++) |
2234 | { | |
b690cc0f RG |
2235 | tree opvectype; |
2236 | ||
ebfd146a IR |
2237 | op = gimple_call_arg (stmt, i); |
2238 | ||
2239 | /* We can only handle calls with arguments of the same type. */ | |
2240 | if (rhs_type | |
8533c9d8 | 2241 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 2242 | { |
73fbfcad | 2243 | if (dump_enabled_p ()) |
78c60e3d | 2244 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2245 | "argument types differ.\n"); |
ebfd146a IR |
2246 | return false; |
2247 | } | |
b690cc0f RG |
2248 | if (!rhs_type) |
2249 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 2250 | |
81c40241 | 2251 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[i], &opvectype)) |
ebfd146a | 2252 | { |
73fbfcad | 2253 | if (dump_enabled_p ()) |
78c60e3d | 2254 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2255 | "use not simple.\n"); |
ebfd146a IR |
2256 | return false; |
2257 | } | |
ebfd146a | 2258 | |
b690cc0f RG |
2259 | if (!vectype_in) |
2260 | vectype_in = opvectype; | |
2261 | else if (opvectype | |
2262 | && opvectype != vectype_in) | |
2263 | { | |
73fbfcad | 2264 | if (dump_enabled_p ()) |
78c60e3d | 2265 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 2266 | "argument vector types differ.\n"); |
b690cc0f RG |
2267 | return false; |
2268 | } | |
2269 | } | |
2270 | /* If all arguments are external or constant defs use a vector type with | |
2271 | the same size as the output vector type. */ | |
ebfd146a | 2272 | if (!vectype_in) |
b690cc0f | 2273 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
2274 | if (vec_stmt) |
2275 | gcc_assert (vectype_in); | |
2276 | if (!vectype_in) | |
2277 | { | |
73fbfcad | 2278 | if (dump_enabled_p ()) |
7d8930a0 | 2279 | { |
78c60e3d SS |
2280 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
2281 | "no vectype for scalar type "); | |
2282 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 2283 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
2284 | } |
2285 | ||
2286 | return false; | |
2287 | } | |
ebfd146a IR |
2288 | |
2289 | /* FORNOW */ | |
b690cc0f RG |
2290 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
2291 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
ebfd146a IR |
2292 | if (nunits_in == nunits_out / 2) |
2293 | modifier = NARROW; | |
2294 | else if (nunits_out == nunits_in) | |
2295 | modifier = NONE; | |
2296 | else if (nunits_out == nunits_in / 2) | |
2297 | modifier = WIDEN; | |
2298 | else | |
2299 | return false; | |
2300 | ||
70439f0d RS |
2301 | /* We only handle functions that do not read or clobber memory. */ |
2302 | if (gimple_vuse (stmt)) | |
2303 | { | |
2304 | if (dump_enabled_p ()) | |
2305 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2306 | "function reads from or writes to memory.\n"); | |
2307 | return false; | |
2308 | } | |
2309 | ||
ebfd146a IR |
2310 | /* For now, we only vectorize functions if a target specific builtin |
2311 | is available. TODO -- in some cases, it might be profitable to | |
2312 | insert the calls for pieces of the vector, in order to be able | |
2313 | to vectorize other operations in the loop. */ | |
70439f0d RS |
2314 | fndecl = NULL_TREE; |
2315 | internal_fn ifn = IFN_LAST; | |
2316 | combined_fn cfn = gimple_call_combined_fn (stmt); | |
2317 | tree callee = gimple_call_fndecl (stmt); | |
2318 | ||
2319 | /* First try using an internal function. */ | |
b1b6836e RS |
2320 | tree_code convert_code = ERROR_MARK; |
2321 | if (cfn != CFN_LAST | |
2322 | && (modifier == NONE | |
2323 | || (modifier == NARROW | |
2324 | && simple_integer_narrowing (vectype_out, vectype_in, | |
2325 | &convert_code)))) | |
70439f0d RS |
2326 | ifn = vectorizable_internal_function (cfn, callee, vectype_out, |
2327 | vectype_in); | |
2328 | ||
2329 | /* If that fails, try asking for a target-specific built-in function. */ | |
2330 | if (ifn == IFN_LAST) | |
2331 | { | |
2332 | if (cfn != CFN_LAST) | |
2333 | fndecl = targetm.vectorize.builtin_vectorized_function | |
2334 | (cfn, vectype_out, vectype_in); | |
2335 | else | |
2336 | fndecl = targetm.vectorize.builtin_md_vectorized_function | |
2337 | (callee, vectype_out, vectype_in); | |
2338 | } | |
2339 | ||
2340 | if (ifn == IFN_LAST && !fndecl) | |
ebfd146a | 2341 | { |
70439f0d | 2342 | if (cfn == CFN_GOMP_SIMD_LANE |
74bf76ed JJ |
2343 | && !slp_node |
2344 | && loop_vinfo | |
2345 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2346 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
2347 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
2348 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
2349 | { | |
2350 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
2351 | { 0, 1, 2, ... vf - 1 } vector. */ | |
2352 | gcc_assert (nargs == 0); | |
2353 | } | |
2354 | else | |
2355 | { | |
2356 | if (dump_enabled_p ()) | |
2357 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 2358 | "function is not vectorizable.\n"); |
74bf76ed JJ |
2359 | return false; |
2360 | } | |
ebfd146a IR |
2361 | } |
2362 | ||
fce57248 | 2363 | if (slp_node) |
190c2236 | 2364 | ncopies = 1; |
b1b6836e | 2365 | else if (modifier == NARROW && ifn == IFN_LAST) |
ebfd146a IR |
2366 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; |
2367 | else | |
2368 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
2369 | ||
2370 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
2371 | needs to be generated. */ | |
2372 | gcc_assert (ncopies >= 1); | |
2373 | ||
2374 | if (!vec_stmt) /* transformation not required. */ | |
2375 | { | |
2376 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 2377 | if (dump_enabled_p ()) |
e645e942 TJ |
2378 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ===" |
2379 | "\n"); | |
c3e7ee41 | 2380 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
b1b6836e RS |
2381 | if (ifn != IFN_LAST && modifier == NARROW && !slp_node) |
2382 | add_stmt_cost (stmt_info->vinfo->target_cost_data, ncopies / 2, | |
2383 | vec_promote_demote, stmt_info, 0, vect_body); | |
2384 | ||
ebfd146a IR |
2385 | return true; |
2386 | } | |
2387 | ||
2388 | /** Transform. **/ | |
2389 | ||
73fbfcad | 2390 | if (dump_enabled_p ()) |
e645e942 | 2391 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
2392 | |
2393 | /* Handle def. */ | |
2394 | scalar_dest = gimple_call_lhs (stmt); | |
2395 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
2396 | ||
2397 | prev_stmt_info = NULL; | |
b1b6836e | 2398 | if (modifier == NONE || ifn != IFN_LAST) |
ebfd146a | 2399 | { |
b1b6836e | 2400 | tree prev_res = NULL_TREE; |
ebfd146a IR |
2401 | for (j = 0; j < ncopies; ++j) |
2402 | { | |
2403 | /* Build argument list for the vectorized call. */ | |
2404 | if (j == 0) | |
9771b263 | 2405 | vargs.create (nargs); |
ebfd146a | 2406 | else |
9771b263 | 2407 | vargs.truncate (0); |
ebfd146a | 2408 | |
190c2236 JJ |
2409 | if (slp_node) |
2410 | { | |
ef062b13 | 2411 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2412 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2413 | |
2414 | for (i = 0; i < nargs; i++) | |
9771b263 | 2415 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2416 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2417 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2418 | |
2419 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2420 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
2421 | { |
2422 | size_t k; | |
2423 | for (k = 0; k < nargs; k++) | |
2424 | { | |
37b5ec8f | 2425 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 2426 | vargs[k] = vec_oprndsk[i]; |
190c2236 | 2427 | } |
b1b6836e RS |
2428 | if (modifier == NARROW) |
2429 | { | |
2430 | tree half_res = make_ssa_name (vectype_in); | |
2431 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2432 | gimple_call_set_lhs (new_stmt, half_res); | |
2433 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2434 | if ((i & 1) == 0) | |
2435 | { | |
2436 | prev_res = half_res; | |
2437 | continue; | |
2438 | } | |
2439 | new_temp = make_ssa_name (vec_dest); | |
2440 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
2441 | prev_res, half_res); | |
2442 | } | |
70439f0d | 2443 | else |
b1b6836e RS |
2444 | { |
2445 | if (ifn != IFN_LAST) | |
2446 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2447 | else | |
2448 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
2449 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2450 | gimple_call_set_lhs (new_stmt, new_temp); | |
2451 | } | |
190c2236 | 2452 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 2453 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2454 | } |
2455 | ||
2456 | for (i = 0; i < nargs; i++) | |
2457 | { | |
37b5ec8f | 2458 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2459 | vec_oprndsi.release (); |
190c2236 | 2460 | } |
190c2236 JJ |
2461 | continue; |
2462 | } | |
2463 | ||
ebfd146a IR |
2464 | for (i = 0; i < nargs; i++) |
2465 | { | |
2466 | op = gimple_call_arg (stmt, i); | |
2467 | if (j == 0) | |
2468 | vec_oprnd0 | |
81c40241 | 2469 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 2470 | else |
63827fb8 IR |
2471 | { |
2472 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
2473 | vec_oprnd0 | |
2474 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
2475 | } | |
ebfd146a | 2476 | |
9771b263 | 2477 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
2478 | } |
2479 | ||
74bf76ed JJ |
2480 | if (gimple_call_internal_p (stmt) |
2481 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
2482 | { | |
2483 | tree *v = XALLOCAVEC (tree, nunits_out); | |
2484 | int k; | |
2485 | for (k = 0; k < nunits_out; ++k) | |
2486 | v[k] = build_int_cst (unsigned_type_node, j * nunits_out + k); | |
2487 | tree cst = build_vector (vectype_out, v); | |
2488 | tree new_var | |
0e22bb5a | 2489 | = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_"); |
355fe088 | 2490 | gimple *init_stmt = gimple_build_assign (new_var, cst); |
74bf76ed | 2491 | vect_init_vector_1 (stmt, init_stmt, NULL); |
b731b390 | 2492 | new_temp = make_ssa_name (vec_dest); |
0e22bb5a | 2493 | new_stmt = gimple_build_assign (new_temp, new_var); |
74bf76ed | 2494 | } |
b1b6836e RS |
2495 | else if (modifier == NARROW) |
2496 | { | |
2497 | tree half_res = make_ssa_name (vectype_in); | |
2498 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2499 | gimple_call_set_lhs (new_stmt, half_res); | |
2500 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2501 | if ((j & 1) == 0) | |
2502 | { | |
2503 | prev_res = half_res; | |
2504 | continue; | |
2505 | } | |
2506 | new_temp = make_ssa_name (vec_dest); | |
2507 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
2508 | prev_res, half_res); | |
2509 | } | |
74bf76ed JJ |
2510 | else |
2511 | { | |
70439f0d RS |
2512 | if (ifn != IFN_LAST) |
2513 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2514 | else | |
2515 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
74bf76ed JJ |
2516 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2517 | gimple_call_set_lhs (new_stmt, new_temp); | |
2518 | } | |
ebfd146a IR |
2519 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2520 | ||
b1b6836e | 2521 | if (j == (modifier == NARROW ? 1 : 0)) |
ebfd146a IR |
2522 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; |
2523 | else | |
2524 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2525 | ||
2526 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2527 | } | |
b1b6836e RS |
2528 | } |
2529 | else if (modifier == NARROW) | |
2530 | { | |
ebfd146a IR |
2531 | for (j = 0; j < ncopies; ++j) |
2532 | { | |
2533 | /* Build argument list for the vectorized call. */ | |
2534 | if (j == 0) | |
9771b263 | 2535 | vargs.create (nargs * 2); |
ebfd146a | 2536 | else |
9771b263 | 2537 | vargs.truncate (0); |
ebfd146a | 2538 | |
190c2236 JJ |
2539 | if (slp_node) |
2540 | { | |
ef062b13 | 2541 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 2542 | vec<tree> vec_oprnds0; |
190c2236 JJ |
2543 | |
2544 | for (i = 0; i < nargs; i++) | |
9771b263 | 2545 | vargs.quick_push (gimple_call_arg (stmt, i)); |
190c2236 | 2546 | vect_get_slp_defs (vargs, slp_node, &vec_defs, -1); |
37b5ec8f | 2547 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
2548 | |
2549 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 2550 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
2551 | { |
2552 | size_t k; | |
9771b263 | 2553 | vargs.truncate (0); |
190c2236 JJ |
2554 | for (k = 0; k < nargs; k++) |
2555 | { | |
37b5ec8f | 2556 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
2557 | vargs.quick_push (vec_oprndsk[i]); |
2558 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 | 2559 | } |
70439f0d RS |
2560 | if (ifn != IFN_LAST) |
2561 | new_stmt = gimple_build_call_internal_vec (ifn, vargs); | |
2562 | else | |
2563 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
190c2236 JJ |
2564 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2565 | gimple_call_set_lhs (new_stmt, new_temp); | |
2566 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9771b263 | 2567 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
2568 | } |
2569 | ||
2570 | for (i = 0; i < nargs; i++) | |
2571 | { | |
37b5ec8f | 2572 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 2573 | vec_oprndsi.release (); |
190c2236 | 2574 | } |
190c2236 JJ |
2575 | continue; |
2576 | } | |
2577 | ||
ebfd146a IR |
2578 | for (i = 0; i < nargs; i++) |
2579 | { | |
2580 | op = gimple_call_arg (stmt, i); | |
2581 | if (j == 0) | |
2582 | { | |
2583 | vec_oprnd0 | |
81c40241 | 2584 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 2585 | vec_oprnd1 |
63827fb8 | 2586 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2587 | } |
2588 | else | |
2589 | { | |
336ecb65 | 2590 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 2591 | vec_oprnd0 |
63827fb8 | 2592 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 2593 | vec_oprnd1 |
63827fb8 | 2594 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
2595 | } |
2596 | ||
9771b263 DN |
2597 | vargs.quick_push (vec_oprnd0); |
2598 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
2599 | } |
2600 | ||
b1b6836e | 2601 | new_stmt = gimple_build_call_vec (fndecl, vargs); |
ebfd146a IR |
2602 | new_temp = make_ssa_name (vec_dest, new_stmt); |
2603 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
2604 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
2605 | ||
2606 | if (j == 0) | |
2607 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
2608 | else | |
2609 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2610 | ||
2611 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2612 | } | |
2613 | ||
2614 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a | 2615 | } |
b1b6836e RS |
2616 | else |
2617 | /* No current target implements this case. */ | |
2618 | return false; | |
ebfd146a | 2619 | |
9771b263 | 2620 | vargs.release (); |
ebfd146a | 2621 | |
ebfd146a IR |
2622 | /* The call in STMT might prevent it from being removed in dce. |
2623 | We however cannot remove it here, due to the way the ssa name | |
2624 | it defines is mapped to the new definition. So just replace | |
2625 | rhs of the statement with something harmless. */ | |
2626 | ||
dd34c087 JJ |
2627 | if (slp_node) |
2628 | return true; | |
2629 | ||
ebfd146a | 2630 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
2631 | if (is_pattern_stmt_p (stmt_info)) |
2632 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
2633 | else | |
2634 | lhs = gimple_call_lhs (stmt); | |
3cc2fa2a | 2635 | |
9d5e7640 | 2636 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); |
ebfd146a | 2637 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 2638 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
2639 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
2640 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
2641 | |
2642 | return true; | |
2643 | } | |
2644 | ||
2645 | ||
0136f8f0 AH |
2646 | struct simd_call_arg_info |
2647 | { | |
2648 | tree vectype; | |
2649 | tree op; | |
2650 | enum vect_def_type dt; | |
2651 | HOST_WIDE_INT linear_step; | |
2652 | unsigned int align; | |
17b658af | 2653 | bool simd_lane_linear; |
0136f8f0 AH |
2654 | }; |
2655 | ||
17b658af JJ |
2656 | /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME, |
2657 | is linear within simd lane (but not within whole loop), note it in | |
2658 | *ARGINFO. */ | |
2659 | ||
2660 | static void | |
2661 | vect_simd_lane_linear (tree op, struct loop *loop, | |
2662 | struct simd_call_arg_info *arginfo) | |
2663 | { | |
355fe088 | 2664 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
17b658af JJ |
2665 | |
2666 | if (!is_gimple_assign (def_stmt) | |
2667 | || gimple_assign_rhs_code (def_stmt) != POINTER_PLUS_EXPR | |
2668 | || !is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))) | |
2669 | return; | |
2670 | ||
2671 | tree base = gimple_assign_rhs1 (def_stmt); | |
2672 | HOST_WIDE_INT linear_step = 0; | |
2673 | tree v = gimple_assign_rhs2 (def_stmt); | |
2674 | while (TREE_CODE (v) == SSA_NAME) | |
2675 | { | |
2676 | tree t; | |
2677 | def_stmt = SSA_NAME_DEF_STMT (v); | |
2678 | if (is_gimple_assign (def_stmt)) | |
2679 | switch (gimple_assign_rhs_code (def_stmt)) | |
2680 | { | |
2681 | case PLUS_EXPR: | |
2682 | t = gimple_assign_rhs2 (def_stmt); | |
2683 | if (linear_step || TREE_CODE (t) != INTEGER_CST) | |
2684 | return; | |
2685 | base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t); | |
2686 | v = gimple_assign_rhs1 (def_stmt); | |
2687 | continue; | |
2688 | case MULT_EXPR: | |
2689 | t = gimple_assign_rhs2 (def_stmt); | |
2690 | if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t)) | |
2691 | return; | |
2692 | linear_step = tree_to_shwi (t); | |
2693 | v = gimple_assign_rhs1 (def_stmt); | |
2694 | continue; | |
2695 | CASE_CONVERT: | |
2696 | t = gimple_assign_rhs1 (def_stmt); | |
2697 | if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE | |
2698 | || (TYPE_PRECISION (TREE_TYPE (v)) | |
2699 | < TYPE_PRECISION (TREE_TYPE (t)))) | |
2700 | return; | |
2701 | if (!linear_step) | |
2702 | linear_step = 1; | |
2703 | v = t; | |
2704 | continue; | |
2705 | default: | |
2706 | return; | |
2707 | } | |
2708 | else if (is_gimple_call (def_stmt) | |
2709 | && gimple_call_internal_p (def_stmt) | |
2710 | && gimple_call_internal_fn (def_stmt) == IFN_GOMP_SIMD_LANE | |
2711 | && loop->simduid | |
2712 | && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME | |
2713 | && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0)) | |
2714 | == loop->simduid)) | |
2715 | { | |
2716 | if (!linear_step) | |
2717 | linear_step = 1; | |
2718 | arginfo->linear_step = linear_step; | |
2719 | arginfo->op = base; | |
2720 | arginfo->simd_lane_linear = true; | |
2721 | return; | |
2722 | } | |
2723 | } | |
2724 | } | |
2725 | ||
0136f8f0 AH |
2726 | /* Function vectorizable_simd_clone_call. |
2727 | ||
2728 | Check if STMT performs a function call that can be vectorized | |
2729 | by calling a simd clone of the function. | |
2730 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
2731 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
2732 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2733 | ||
2734 | static bool | |
355fe088 TS |
2735 | vectorizable_simd_clone_call (gimple *stmt, gimple_stmt_iterator *gsi, |
2736 | gimple **vec_stmt, slp_tree slp_node) | |
0136f8f0 AH |
2737 | { |
2738 | tree vec_dest; | |
2739 | tree scalar_dest; | |
2740 | tree op, type; | |
2741 | tree vec_oprnd0 = NULL_TREE; | |
2742 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
2743 | tree vectype; | |
2744 | unsigned int nunits; | |
2745 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2746 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
310213d4 | 2747 | vec_info *vinfo = stmt_info->vinfo; |
0136f8f0 | 2748 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
81c40241 | 2749 | tree fndecl, new_temp; |
355fe088 TS |
2750 | gimple *def_stmt; |
2751 | gimple *new_stmt = NULL; | |
0136f8f0 | 2752 | int ncopies, j; |
00426f9a | 2753 | auto_vec<simd_call_arg_info> arginfo; |
0136f8f0 AH |
2754 | vec<tree> vargs = vNULL; |
2755 | size_t i, nargs; | |
2756 | tree lhs, rtype, ratype; | |
2757 | vec<constructor_elt, va_gc> *ret_ctor_elts; | |
2758 | ||
2759 | /* Is STMT a vectorizable call? */ | |
2760 | if (!is_gimple_call (stmt)) | |
2761 | return false; | |
2762 | ||
2763 | fndecl = gimple_call_fndecl (stmt); | |
2764 | if (fndecl == NULL_TREE) | |
2765 | return false; | |
2766 | ||
d52f5295 | 2767 | struct cgraph_node *node = cgraph_node::get (fndecl); |
0136f8f0 AH |
2768 | if (node == NULL || node->simd_clones == NULL) |
2769 | return false; | |
2770 | ||
2771 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
2772 | return false; | |
2773 | ||
66c16fd9 RB |
2774 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
2775 | && ! vec_stmt) | |
0136f8f0 AH |
2776 | return false; |
2777 | ||
2778 | if (gimple_call_lhs (stmt) | |
2779 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
2780 | return false; | |
2781 | ||
2782 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); | |
2783 | ||
2784 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2785 | ||
2786 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt)) | |
2787 | return false; | |
2788 | ||
2789 | /* FORNOW */ | |
fce57248 | 2790 | if (slp_node) |
0136f8f0 AH |
2791 | return false; |
2792 | ||
2793 | /* Process function arguments. */ | |
2794 | nargs = gimple_call_num_args (stmt); | |
2795 | ||
2796 | /* Bail out if the function has zero arguments. */ | |
2797 | if (nargs == 0) | |
2798 | return false; | |
2799 | ||
00426f9a | 2800 | arginfo.reserve (nargs, true); |
0136f8f0 AH |
2801 | |
2802 | for (i = 0; i < nargs; i++) | |
2803 | { | |
2804 | simd_call_arg_info thisarginfo; | |
2805 | affine_iv iv; | |
2806 | ||
2807 | thisarginfo.linear_step = 0; | |
2808 | thisarginfo.align = 0; | |
2809 | thisarginfo.op = NULL_TREE; | |
17b658af | 2810 | thisarginfo.simd_lane_linear = false; |
0136f8f0 AH |
2811 | |
2812 | op = gimple_call_arg (stmt, i); | |
81c40241 RB |
2813 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &thisarginfo.dt, |
2814 | &thisarginfo.vectype) | |
0136f8f0 AH |
2815 | || thisarginfo.dt == vect_uninitialized_def) |
2816 | { | |
2817 | if (dump_enabled_p ()) | |
2818 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2819 | "use not simple.\n"); | |
0136f8f0 AH |
2820 | return false; |
2821 | } | |
2822 | ||
2823 | if (thisarginfo.dt == vect_constant_def | |
2824 | || thisarginfo.dt == vect_external_def) | |
2825 | gcc_assert (thisarginfo.vectype == NULL_TREE); | |
2826 | else | |
2827 | gcc_assert (thisarginfo.vectype != NULL_TREE); | |
2828 | ||
6c9e85fb JJ |
2829 | /* For linear arguments, the analyze phase should have saved |
2830 | the base and step in STMT_VINFO_SIMD_CLONE_INFO. */ | |
17b658af JJ |
2831 | if (i * 3 + 4 <= STMT_VINFO_SIMD_CLONE_INFO (stmt_info).length () |
2832 | && STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]) | |
6c9e85fb JJ |
2833 | { |
2834 | gcc_assert (vec_stmt); | |
2835 | thisarginfo.linear_step | |
17b658af | 2836 | = tree_to_shwi (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]); |
6c9e85fb | 2837 | thisarginfo.op |
17b658af JJ |
2838 | = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 1]; |
2839 | thisarginfo.simd_lane_linear | |
2840 | = (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 3] | |
2841 | == boolean_true_node); | |
6c9e85fb JJ |
2842 | /* If loop has been peeled for alignment, we need to adjust it. */ |
2843 | tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo); | |
2844 | tree n2 = LOOP_VINFO_NITERS (loop_vinfo); | |
17b658af | 2845 | if (n1 != n2 && !thisarginfo.simd_lane_linear) |
6c9e85fb JJ |
2846 | { |
2847 | tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2); | |
17b658af | 2848 | tree step = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]; |
6c9e85fb JJ |
2849 | tree opt = TREE_TYPE (thisarginfo.op); |
2850 | bias = fold_convert (TREE_TYPE (step), bias); | |
2851 | bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step); | |
2852 | thisarginfo.op | |
2853 | = fold_build2 (POINTER_TYPE_P (opt) | |
2854 | ? POINTER_PLUS_EXPR : PLUS_EXPR, opt, | |
2855 | thisarginfo.op, bias); | |
2856 | } | |
2857 | } | |
2858 | else if (!vec_stmt | |
2859 | && thisarginfo.dt != vect_constant_def | |
2860 | && thisarginfo.dt != vect_external_def | |
2861 | && loop_vinfo | |
2862 | && TREE_CODE (op) == SSA_NAME | |
2863 | && simple_iv (loop, loop_containing_stmt (stmt), op, | |
2864 | &iv, false) | |
2865 | && tree_fits_shwi_p (iv.step)) | |
0136f8f0 AH |
2866 | { |
2867 | thisarginfo.linear_step = tree_to_shwi (iv.step); | |
2868 | thisarginfo.op = iv.base; | |
2869 | } | |
2870 | else if ((thisarginfo.dt == vect_constant_def | |
2871 | || thisarginfo.dt == vect_external_def) | |
2872 | && POINTER_TYPE_P (TREE_TYPE (op))) | |
2873 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; | |
17b658af JJ |
2874 | /* Addresses of array elements indexed by GOMP_SIMD_LANE are |
2875 | linear too. */ | |
2876 | if (POINTER_TYPE_P (TREE_TYPE (op)) | |
2877 | && !thisarginfo.linear_step | |
2878 | && !vec_stmt | |
2879 | && thisarginfo.dt != vect_constant_def | |
2880 | && thisarginfo.dt != vect_external_def | |
2881 | && loop_vinfo | |
2882 | && !slp_node | |
2883 | && TREE_CODE (op) == SSA_NAME) | |
2884 | vect_simd_lane_linear (op, loop, &thisarginfo); | |
0136f8f0 AH |
2885 | |
2886 | arginfo.quick_push (thisarginfo); | |
2887 | } | |
2888 | ||
2889 | unsigned int badness = 0; | |
2890 | struct cgraph_node *bestn = NULL; | |
6c9e85fb JJ |
2891 | if (STMT_VINFO_SIMD_CLONE_INFO (stmt_info).exists ()) |
2892 | bestn = cgraph_node::get (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[0]); | |
0136f8f0 AH |
2893 | else |
2894 | for (struct cgraph_node *n = node->simd_clones; n != NULL; | |
2895 | n = n->simdclone->next_clone) | |
2896 | { | |
2897 | unsigned int this_badness = 0; | |
2898 | if (n->simdclone->simdlen | |
2899 | > (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
2900 | || n->simdclone->nargs != nargs) | |
2901 | continue; | |
2902 | if (n->simdclone->simdlen | |
2903 | < (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
2904 | this_badness += (exact_log2 (LOOP_VINFO_VECT_FACTOR (loop_vinfo)) | |
2905 | - exact_log2 (n->simdclone->simdlen)) * 1024; | |
2906 | if (n->simdclone->inbranch) | |
2907 | this_badness += 2048; | |
2908 | int target_badness = targetm.simd_clone.usable (n); | |
2909 | if (target_badness < 0) | |
2910 | continue; | |
2911 | this_badness += target_badness * 512; | |
2912 | /* FORNOW: Have to add code to add the mask argument. */ | |
2913 | if (n->simdclone->inbranch) | |
2914 | continue; | |
2915 | for (i = 0; i < nargs; i++) | |
2916 | { | |
2917 | switch (n->simdclone->args[i].arg_type) | |
2918 | { | |
2919 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
2920 | if (!useless_type_conversion_p | |
2921 | (n->simdclone->args[i].orig_type, | |
2922 | TREE_TYPE (gimple_call_arg (stmt, i)))) | |
2923 | i = -1; | |
2924 | else if (arginfo[i].dt == vect_constant_def | |
2925 | || arginfo[i].dt == vect_external_def | |
2926 | || arginfo[i].linear_step) | |
2927 | this_badness += 64; | |
2928 | break; | |
2929 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
2930 | if (arginfo[i].dt != vect_constant_def | |
2931 | && arginfo[i].dt != vect_external_def) | |
2932 | i = -1; | |
2933 | break; | |
2934 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
d9a6bd32 | 2935 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
2936 | if (arginfo[i].dt == vect_constant_def |
2937 | || arginfo[i].dt == vect_external_def | |
2938 | || (arginfo[i].linear_step | |
2939 | != n->simdclone->args[i].linear_step)) | |
2940 | i = -1; | |
2941 | break; | |
2942 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
d9a6bd32 JJ |
2943 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
2944 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
e01d41e5 JJ |
2945 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
2946 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
2947 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
2948 | /* FORNOW */ |
2949 | i = -1; | |
2950 | break; | |
2951 | case SIMD_CLONE_ARG_TYPE_MASK: | |
2952 | gcc_unreachable (); | |
2953 | } | |
2954 | if (i == (size_t) -1) | |
2955 | break; | |
2956 | if (n->simdclone->args[i].alignment > arginfo[i].align) | |
2957 | { | |
2958 | i = -1; | |
2959 | break; | |
2960 | } | |
2961 | if (arginfo[i].align) | |
2962 | this_badness += (exact_log2 (arginfo[i].align) | |
2963 | - exact_log2 (n->simdclone->args[i].alignment)); | |
2964 | } | |
2965 | if (i == (size_t) -1) | |
2966 | continue; | |
2967 | if (bestn == NULL || this_badness < badness) | |
2968 | { | |
2969 | bestn = n; | |
2970 | badness = this_badness; | |
2971 | } | |
2972 | } | |
2973 | ||
2974 | if (bestn == NULL) | |
00426f9a | 2975 | return false; |
0136f8f0 AH |
2976 | |
2977 | for (i = 0; i < nargs; i++) | |
2978 | if ((arginfo[i].dt == vect_constant_def | |
2979 | || arginfo[i].dt == vect_external_def) | |
2980 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
2981 | { | |
2982 | arginfo[i].vectype | |
2983 | = get_vectype_for_scalar_type (TREE_TYPE (gimple_call_arg (stmt, | |
2984 | i))); | |
2985 | if (arginfo[i].vectype == NULL | |
2986 | || (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
2987 | > bestn->simdclone->simdlen)) | |
00426f9a | 2988 | return false; |
0136f8f0 AH |
2989 | } |
2990 | ||
2991 | fndecl = bestn->decl; | |
2992 | nunits = bestn->simdclone->simdlen; | |
2993 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
2994 | ||
2995 | /* If the function isn't const, only allow it in simd loops where user | |
2996 | has asserted that at least nunits consecutive iterations can be | |
2997 | performed using SIMD instructions. */ | |
2998 | if ((loop == NULL || (unsigned) loop->safelen < nunits) | |
2999 | && gimple_vuse (stmt)) | |
00426f9a | 3000 | return false; |
0136f8f0 AH |
3001 | |
3002 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
3003 | needs to be generated. */ | |
3004 | gcc_assert (ncopies >= 1); | |
3005 | ||
3006 | if (!vec_stmt) /* transformation not required. */ | |
3007 | { | |
6c9e85fb JJ |
3008 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (bestn->decl); |
3009 | for (i = 0; i < nargs; i++) | |
7adb26f2 JJ |
3010 | if ((bestn->simdclone->args[i].arg_type |
3011 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
3012 | || (bestn->simdclone->args[i].arg_type | |
3013 | == SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP)) | |
6c9e85fb | 3014 | { |
17b658af | 3015 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_grow_cleared (i * 3 |
6c9e85fb JJ |
3016 | + 1); |
3017 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (arginfo[i].op); | |
3018 | tree lst = POINTER_TYPE_P (TREE_TYPE (arginfo[i].op)) | |
3019 | ? size_type_node : TREE_TYPE (arginfo[i].op); | |
3020 | tree ls = build_int_cst (lst, arginfo[i].linear_step); | |
3021 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (ls); | |
17b658af JJ |
3022 | tree sll = arginfo[i].simd_lane_linear |
3023 | ? boolean_true_node : boolean_false_node; | |
3024 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (sll); | |
6c9e85fb | 3025 | } |
0136f8f0 AH |
3026 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; |
3027 | if (dump_enabled_p ()) | |
3028 | dump_printf_loc (MSG_NOTE, vect_location, | |
3029 | "=== vectorizable_simd_clone_call ===\n"); | |
3030 | /* vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); */ | |
0136f8f0 AH |
3031 | return true; |
3032 | } | |
3033 | ||
3034 | /** Transform. **/ | |
3035 | ||
3036 | if (dump_enabled_p ()) | |
3037 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); | |
3038 | ||
3039 | /* Handle def. */ | |
3040 | scalar_dest = gimple_call_lhs (stmt); | |
3041 | vec_dest = NULL_TREE; | |
3042 | rtype = NULL_TREE; | |
3043 | ratype = NULL_TREE; | |
3044 | if (scalar_dest) | |
3045 | { | |
3046 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3047 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); | |
3048 | if (TREE_CODE (rtype) == ARRAY_TYPE) | |
3049 | { | |
3050 | ratype = rtype; | |
3051 | rtype = TREE_TYPE (ratype); | |
3052 | } | |
3053 | } | |
3054 | ||
3055 | prev_stmt_info = NULL; | |
3056 | for (j = 0; j < ncopies; ++j) | |
3057 | { | |
3058 | /* Build argument list for the vectorized call. */ | |
3059 | if (j == 0) | |
3060 | vargs.create (nargs); | |
3061 | else | |
3062 | vargs.truncate (0); | |
3063 | ||
3064 | for (i = 0; i < nargs; i++) | |
3065 | { | |
3066 | unsigned int k, l, m, o; | |
3067 | tree atype; | |
3068 | op = gimple_call_arg (stmt, i); | |
3069 | switch (bestn->simdclone->args[i].arg_type) | |
3070 | { | |
3071 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3072 | atype = bestn->simdclone->args[i].vector_type; | |
3073 | o = nunits / TYPE_VECTOR_SUBPARTS (atype); | |
3074 | for (m = j * o; m < (j + 1) * o; m++) | |
3075 | { | |
3076 | if (TYPE_VECTOR_SUBPARTS (atype) | |
3077 | < TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)) | |
3078 | { | |
3079 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); | |
3080 | k = (TYPE_VECTOR_SUBPARTS (arginfo[i].vectype) | |
3081 | / TYPE_VECTOR_SUBPARTS (atype)); | |
3082 | gcc_assert ((k & (k - 1)) == 0); | |
3083 | if (m == 0) | |
3084 | vec_oprnd0 | |
81c40241 | 3085 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3086 | else |
3087 | { | |
3088 | vec_oprnd0 = arginfo[i].op; | |
3089 | if ((m & (k - 1)) == 0) | |
3090 | vec_oprnd0 | |
3091 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3092 | vec_oprnd0); | |
3093 | } | |
3094 | arginfo[i].op = vec_oprnd0; | |
3095 | vec_oprnd0 | |
3096 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, | |
3097 | size_int (prec), | |
3098 | bitsize_int ((m & (k - 1)) * prec)); | |
3099 | new_stmt | |
b731b390 | 3100 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3101 | vec_oprnd0); |
3102 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3103 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3104 | } | |
3105 | else | |
3106 | { | |
3107 | k = (TYPE_VECTOR_SUBPARTS (atype) | |
3108 | / TYPE_VECTOR_SUBPARTS (arginfo[i].vectype)); | |
3109 | gcc_assert ((k & (k - 1)) == 0); | |
3110 | vec<constructor_elt, va_gc> *ctor_elts; | |
3111 | if (k != 1) | |
3112 | vec_alloc (ctor_elts, k); | |
3113 | else | |
3114 | ctor_elts = NULL; | |
3115 | for (l = 0; l < k; l++) | |
3116 | { | |
3117 | if (m == 0 && l == 0) | |
3118 | vec_oprnd0 | |
81c40241 | 3119 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3120 | else |
3121 | vec_oprnd0 | |
3122 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3123 | arginfo[i].op); | |
3124 | arginfo[i].op = vec_oprnd0; | |
3125 | if (k == 1) | |
3126 | break; | |
3127 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, | |
3128 | vec_oprnd0); | |
3129 | } | |
3130 | if (k == 1) | |
3131 | vargs.safe_push (vec_oprnd0); | |
3132 | else | |
3133 | { | |
3134 | vec_oprnd0 = build_constructor (atype, ctor_elts); | |
3135 | new_stmt | |
b731b390 | 3136 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3137 | vec_oprnd0); |
3138 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3139 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3140 | } | |
3141 | } | |
3142 | } | |
3143 | break; | |
3144 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3145 | vargs.safe_push (op); | |
3146 | break; | |
3147 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
7adb26f2 | 3148 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
3149 | if (j == 0) |
3150 | { | |
3151 | gimple_seq stmts; | |
3152 | arginfo[i].op | |
3153 | = force_gimple_operand (arginfo[i].op, &stmts, true, | |
3154 | NULL_TREE); | |
3155 | if (stmts != NULL) | |
3156 | { | |
3157 | basic_block new_bb; | |
3158 | edge pe = loop_preheader_edge (loop); | |
3159 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
3160 | gcc_assert (!new_bb); | |
3161 | } | |
17b658af JJ |
3162 | if (arginfo[i].simd_lane_linear) |
3163 | { | |
3164 | vargs.safe_push (arginfo[i].op); | |
3165 | break; | |
3166 | } | |
b731b390 | 3167 | tree phi_res = copy_ssa_name (op); |
538dd0b7 | 3168 | gphi *new_phi = create_phi_node (phi_res, loop->header); |
0136f8f0 | 3169 | set_vinfo_for_stmt (new_phi, |
310213d4 | 3170 | new_stmt_vec_info (new_phi, loop_vinfo)); |
0136f8f0 AH |
3171 | add_phi_arg (new_phi, arginfo[i].op, |
3172 | loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
3173 | enum tree_code code | |
3174 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3175 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3176 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3177 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3178 | widest_int cst |
3179 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3180 | ncopies * nunits); | |
3181 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3182 | tree phi_arg = copy_ssa_name (op); |
0d0e4a03 JJ |
3183 | new_stmt |
3184 | = gimple_build_assign (phi_arg, code, phi_res, tcst); | |
0136f8f0 AH |
3185 | gimple_stmt_iterator si = gsi_after_labels (loop->header); |
3186 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); | |
3187 | set_vinfo_for_stmt (new_stmt, | |
310213d4 | 3188 | new_stmt_vec_info (new_stmt, loop_vinfo)); |
0136f8f0 AH |
3189 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), |
3190 | UNKNOWN_LOCATION); | |
3191 | arginfo[i].op = phi_res; | |
3192 | vargs.safe_push (phi_res); | |
3193 | } | |
3194 | else | |
3195 | { | |
3196 | enum tree_code code | |
3197 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
3198 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
3199 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
3200 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
3201 | widest_int cst |
3202 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
3203 | j * nunits); | |
3204 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 3205 | new_temp = make_ssa_name (TREE_TYPE (op)); |
0d0e4a03 JJ |
3206 | new_stmt = gimple_build_assign (new_temp, code, |
3207 | arginfo[i].op, tcst); | |
0136f8f0 AH |
3208 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3209 | vargs.safe_push (new_temp); | |
3210 | } | |
3211 | break; | |
7adb26f2 JJ |
3212 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
3213 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
0136f8f0 | 3214 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: |
e01d41e5 JJ |
3215 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3216 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3217 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3218 | default: |
3219 | gcc_unreachable (); | |
3220 | } | |
3221 | } | |
3222 | ||
3223 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
3224 | if (vec_dest) | |
3225 | { | |
3226 | gcc_assert (ratype || TYPE_VECTOR_SUBPARTS (rtype) == nunits); | |
3227 | if (ratype) | |
b731b390 | 3228 | new_temp = create_tmp_var (ratype); |
0136f8f0 AH |
3229 | else if (TYPE_VECTOR_SUBPARTS (vectype) |
3230 | == TYPE_VECTOR_SUBPARTS (rtype)) | |
3231 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3232 | else | |
3233 | new_temp = make_ssa_name (rtype, new_stmt); | |
3234 | gimple_call_set_lhs (new_stmt, new_temp); | |
3235 | } | |
3236 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3237 | ||
3238 | if (vec_dest) | |
3239 | { | |
3240 | if (TYPE_VECTOR_SUBPARTS (vectype) < nunits) | |
3241 | { | |
3242 | unsigned int k, l; | |
3243 | unsigned int prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); | |
3244 | k = nunits / TYPE_VECTOR_SUBPARTS (vectype); | |
3245 | gcc_assert ((k & (k - 1)) == 0); | |
3246 | for (l = 0; l < k; l++) | |
3247 | { | |
3248 | tree t; | |
3249 | if (ratype) | |
3250 | { | |
3251 | t = build_fold_addr_expr (new_temp); | |
3252 | t = build2 (MEM_REF, vectype, t, | |
3253 | build_int_cst (TREE_TYPE (t), | |
3254 | l * prec / BITS_PER_UNIT)); | |
3255 | } | |
3256 | else | |
3257 | t = build3 (BIT_FIELD_REF, vectype, new_temp, | |
3258 | size_int (prec), bitsize_int (l * prec)); | |
3259 | new_stmt | |
b731b390 | 3260 | = gimple_build_assign (make_ssa_name (vectype), t); |
0136f8f0 AH |
3261 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3262 | if (j == 0 && l == 0) | |
3263 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3264 | else | |
3265 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3266 | ||
3267 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3268 | } | |
3269 | ||
3270 | if (ratype) | |
3271 | { | |
3272 | tree clobber = build_constructor (ratype, NULL); | |
3273 | TREE_THIS_VOLATILE (clobber) = 1; | |
3274 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3275 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3276 | } | |
3277 | continue; | |
3278 | } | |
3279 | else if (TYPE_VECTOR_SUBPARTS (vectype) > nunits) | |
3280 | { | |
3281 | unsigned int k = (TYPE_VECTOR_SUBPARTS (vectype) | |
3282 | / TYPE_VECTOR_SUBPARTS (rtype)); | |
3283 | gcc_assert ((k & (k - 1)) == 0); | |
3284 | if ((j & (k - 1)) == 0) | |
3285 | vec_alloc (ret_ctor_elts, k); | |
3286 | if (ratype) | |
3287 | { | |
3288 | unsigned int m, o = nunits / TYPE_VECTOR_SUBPARTS (rtype); | |
3289 | for (m = 0; m < o; m++) | |
3290 | { | |
3291 | tree tem = build4 (ARRAY_REF, rtype, new_temp, | |
3292 | size_int (m), NULL_TREE, NULL_TREE); | |
3293 | new_stmt | |
b731b390 | 3294 | = gimple_build_assign (make_ssa_name (rtype), tem); |
0136f8f0 AH |
3295 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3296 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, | |
3297 | gimple_assign_lhs (new_stmt)); | |
3298 | } | |
3299 | tree clobber = build_constructor (ratype, NULL); | |
3300 | TREE_THIS_VOLATILE (clobber) = 1; | |
3301 | new_stmt = gimple_build_assign (new_temp, clobber); | |
3302 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3303 | } | |
3304 | else | |
3305 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); | |
3306 | if ((j & (k - 1)) != k - 1) | |
3307 | continue; | |
3308 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); | |
3309 | new_stmt | |
b731b390 | 3310 | = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0); |
0136f8f0 AH |
3311 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3312 | ||
3313 | if ((unsigned) j == k - 1) | |
3314 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3315 | else | |
3316 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3317 | ||
3318 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3319 | continue; | |
3320 | } | |
3321 | else if (ratype) | |
3322 | { | |
3323 | tree t = build_fold_addr_expr (new_temp); | |
3324 | t = build2 (MEM_REF, vectype, t, | |
3325 | build_int_cst (TREE_TYPE (t), 0)); | |
3326 | new_stmt | |
b731b390 | 3327 | = gimple_build_assign (make_ssa_name (vec_dest), t); |
0136f8f0 AH |
3328 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3329 | tree clobber = build_constructor (ratype, NULL); | |
3330 | TREE_THIS_VOLATILE (clobber) = 1; | |
3331 | vect_finish_stmt_generation (stmt, | |
3332 | gimple_build_assign (new_temp, | |
3333 | clobber), gsi); | |
3334 | } | |
3335 | } | |
3336 | ||
3337 | if (j == 0) | |
3338 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3339 | else | |
3340 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3341 | ||
3342 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3343 | } | |
3344 | ||
3345 | vargs.release (); | |
3346 | ||
3347 | /* The call in STMT might prevent it from being removed in dce. | |
3348 | We however cannot remove it here, due to the way the ssa name | |
3349 | it defines is mapped to the new definition. So just replace | |
3350 | rhs of the statement with something harmless. */ | |
3351 | ||
3352 | if (slp_node) | |
3353 | return true; | |
3354 | ||
3355 | if (scalar_dest) | |
3356 | { | |
3357 | type = TREE_TYPE (scalar_dest); | |
3358 | if (is_pattern_stmt_p (stmt_info)) | |
3359 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
3360 | else | |
3361 | lhs = gimple_call_lhs (stmt); | |
3362 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
3363 | } | |
3364 | else | |
3365 | new_stmt = gimple_build_nop (); | |
3366 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
3367 | set_vinfo_for_stmt (stmt, NULL); | |
3368 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2865f32a | 3369 | gsi_replace (gsi, new_stmt, true); |
0136f8f0 AH |
3370 | unlink_stmt_vdef (stmt); |
3371 | ||
3372 | return true; | |
3373 | } | |
3374 | ||
3375 | ||
ebfd146a IR |
3376 | /* Function vect_gen_widened_results_half |
3377 | ||
3378 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 3379 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 3380 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
3381 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
3382 | needs to be created (DECL is a function-decl of a target-builtin). | |
3383 | STMT is the original scalar stmt that we are vectorizing. */ | |
3384 | ||
355fe088 | 3385 | static gimple * |
ebfd146a IR |
3386 | vect_gen_widened_results_half (enum tree_code code, |
3387 | tree decl, | |
3388 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
3389 | tree vec_dest, gimple_stmt_iterator *gsi, | |
355fe088 | 3390 | gimple *stmt) |
b8698a0f | 3391 | { |
355fe088 | 3392 | gimple *new_stmt; |
b8698a0f L |
3393 | tree new_temp; |
3394 | ||
3395 | /* Generate half of the widened result: */ | |
3396 | if (code == CALL_EXPR) | |
3397 | { | |
3398 | /* Target specific support */ | |
ebfd146a IR |
3399 | if (op_type == binary_op) |
3400 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
3401 | else | |
3402 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
3403 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3404 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
3405 | } |
3406 | else | |
ebfd146a | 3407 | { |
b8698a0f L |
3408 | /* Generic support */ |
3409 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
3410 | if (op_type != binary_op) |
3411 | vec_oprnd1 = NULL; | |
0d0e4a03 | 3412 | new_stmt = gimple_build_assign (vec_dest, code, vec_oprnd0, vec_oprnd1); |
ebfd146a IR |
3413 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3414 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 3415 | } |
ebfd146a IR |
3416 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3417 | ||
ebfd146a IR |
3418 | return new_stmt; |
3419 | } | |
3420 | ||
4a00c761 JJ |
3421 | |
3422 | /* Get vectorized definitions for loop-based vectorization. For the first | |
3423 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
3424 | scalar operand), and for the rest we get a copy with | |
3425 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
3426 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
3427 | The vectors are collected into VEC_OPRNDS. */ | |
3428 | ||
3429 | static void | |
355fe088 | 3430 | vect_get_loop_based_defs (tree *oprnd, gimple *stmt, enum vect_def_type dt, |
9771b263 | 3431 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
3432 | { |
3433 | tree vec_oprnd; | |
3434 | ||
3435 | /* Get first vector operand. */ | |
3436 | /* All the vector operands except the very first one (that is scalar oprnd) | |
3437 | are stmt copies. */ | |
3438 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
81c40241 | 3439 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt); |
4a00c761 JJ |
3440 | else |
3441 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
3442 | ||
9771b263 | 3443 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3444 | |
3445 | /* Get second vector operand. */ | |
3446 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 3447 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
3448 | |
3449 | *oprnd = vec_oprnd; | |
3450 | ||
3451 | /* For conversion in multiple steps, continue to get operands | |
3452 | recursively. */ | |
3453 | if (multi_step_cvt) | |
3454 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
3455 | } | |
3456 | ||
3457 | ||
3458 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
3459 | For multi-step conversions store the resulting vectors and call the function | |
3460 | recursively. */ | |
3461 | ||
3462 | static void | |
9771b263 | 3463 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
355fe088 | 3464 | int multi_step_cvt, gimple *stmt, |
9771b263 | 3465 | vec<tree> vec_dsts, |
4a00c761 JJ |
3466 | gimple_stmt_iterator *gsi, |
3467 | slp_tree slp_node, enum tree_code code, | |
3468 | stmt_vec_info *prev_stmt_info) | |
3469 | { | |
3470 | unsigned int i; | |
3471 | tree vop0, vop1, new_tmp, vec_dest; | |
355fe088 | 3472 | gimple *new_stmt; |
4a00c761 JJ |
3473 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
3474 | ||
9771b263 | 3475 | vec_dest = vec_dsts.pop (); |
4a00c761 | 3476 | |
9771b263 | 3477 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
3478 | { |
3479 | /* Create demotion operation. */ | |
9771b263 DN |
3480 | vop0 = (*vec_oprnds)[i]; |
3481 | vop1 = (*vec_oprnds)[i + 1]; | |
0d0e4a03 | 3482 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
4a00c761 JJ |
3483 | new_tmp = make_ssa_name (vec_dest, new_stmt); |
3484 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
3485 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3486 | ||
3487 | if (multi_step_cvt) | |
3488 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 3489 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
3490 | else |
3491 | { | |
3492 | /* This is the last step of the conversion sequence. Store the | |
3493 | vectors in SLP_NODE or in vector info of the scalar statement | |
3494 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
3495 | if (slp_node) | |
9771b263 | 3496 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 3497 | else |
c689ce1e RB |
3498 | { |
3499 | if (!*prev_stmt_info) | |
3500 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3501 | else | |
3502 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
4a00c761 | 3503 | |
c689ce1e RB |
3504 | *prev_stmt_info = vinfo_for_stmt (new_stmt); |
3505 | } | |
4a00c761 JJ |
3506 | } |
3507 | } | |
3508 | ||
3509 | /* For multi-step demotion operations we first generate demotion operations | |
3510 | from the source type to the intermediate types, and then combine the | |
3511 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
3512 | type. */ | |
3513 | if (multi_step_cvt) | |
3514 | { | |
3515 | /* At each level of recursion we have half of the operands we had at the | |
3516 | previous level. */ | |
9771b263 | 3517 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
3518 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
3519 | stmt, vec_dsts, gsi, slp_node, | |
3520 | VEC_PACK_TRUNC_EXPR, | |
3521 | prev_stmt_info); | |
3522 | } | |
3523 | ||
9771b263 | 3524 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3525 | } |
3526 | ||
3527 | ||
3528 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
3529 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
3530 | the resulting vectors and call the function recursively. */ | |
3531 | ||
3532 | static void | |
9771b263 DN |
3533 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
3534 | vec<tree> *vec_oprnds1, | |
355fe088 | 3535 | gimple *stmt, tree vec_dest, |
4a00c761 JJ |
3536 | gimple_stmt_iterator *gsi, |
3537 | enum tree_code code1, | |
3538 | enum tree_code code2, tree decl1, | |
3539 | tree decl2, int op_type) | |
3540 | { | |
3541 | int i; | |
3542 | tree vop0, vop1, new_tmp1, new_tmp2; | |
355fe088 | 3543 | gimple *new_stmt1, *new_stmt2; |
6e1aa848 | 3544 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 3545 | |
9771b263 DN |
3546 | vec_tmp.create (vec_oprnds0->length () * 2); |
3547 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
3548 | { |
3549 | if (op_type == binary_op) | |
9771b263 | 3550 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
3551 | else |
3552 | vop1 = NULL_TREE; | |
3553 | ||
3554 | /* Generate the two halves of promotion operation. */ | |
3555 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
3556 | op_type, vec_dest, gsi, stmt); | |
3557 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
3558 | op_type, vec_dest, gsi, stmt); | |
3559 | if (is_gimple_call (new_stmt1)) | |
3560 | { | |
3561 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
3562 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
3563 | } | |
3564 | else | |
3565 | { | |
3566 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
3567 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
3568 | } | |
3569 | ||
3570 | /* Store the results for the next step. */ | |
9771b263 DN |
3571 | vec_tmp.quick_push (new_tmp1); |
3572 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
3573 | } |
3574 | ||
689eaba3 | 3575 | vec_oprnds0->release (); |
4a00c761 JJ |
3576 | *vec_oprnds0 = vec_tmp; |
3577 | } | |
3578 | ||
3579 | ||
b8698a0f L |
3580 | /* Check if STMT performs a conversion operation, that can be vectorized. |
3581 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 3582 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
3583 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
3584 | ||
3585 | static bool | |
355fe088 TS |
3586 | vectorizable_conversion (gimple *stmt, gimple_stmt_iterator *gsi, |
3587 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
3588 | { |
3589 | tree vec_dest; | |
3590 | tree scalar_dest; | |
4a00c761 | 3591 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
3592 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
3593 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3594 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3595 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 3596 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
3597 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
3598 | tree new_temp; | |
355fe088 | 3599 | gimple *def_stmt; |
ebfd146a | 3600 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
355fe088 | 3601 | gimple *new_stmt = NULL; |
ebfd146a IR |
3602 | stmt_vec_info prev_stmt_info; |
3603 | int nunits_in; | |
3604 | int nunits_out; | |
3605 | tree vectype_out, vectype_in; | |
4a00c761 JJ |
3606 | int ncopies, i, j; |
3607 | tree lhs_type, rhs_type; | |
ebfd146a | 3608 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
3609 | vec<tree> vec_oprnds0 = vNULL; |
3610 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 3611 | tree vop0; |
4a00c761 | 3612 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 3613 | vec_info *vinfo = stmt_info->vinfo; |
4a00c761 | 3614 | int multi_step_cvt = 0; |
6e1aa848 DN |
3615 | vec<tree> vec_dsts = vNULL; |
3616 | vec<tree> interm_types = vNULL; | |
4a00c761 JJ |
3617 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
3618 | int op_type; | |
ef4bddc2 | 3619 | machine_mode rhs_mode; |
4a00c761 | 3620 | unsigned short fltsz; |
ebfd146a IR |
3621 | |
3622 | /* Is STMT a vectorizable conversion? */ | |
3623 | ||
4a00c761 | 3624 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3625 | return false; |
3626 | ||
66c16fd9 RB |
3627 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3628 | && ! vec_stmt) | |
ebfd146a IR |
3629 | return false; |
3630 | ||
3631 | if (!is_gimple_assign (stmt)) | |
3632 | return false; | |
3633 | ||
3634 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
3635 | return false; | |
3636 | ||
3637 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
3638 | if (!CONVERT_EXPR_CODE_P (code) |
3639 | && code != FIX_TRUNC_EXPR | |
3640 | && code != FLOAT_EXPR | |
3641 | && code != WIDEN_MULT_EXPR | |
3642 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
3643 | return false; |
3644 | ||
4a00c761 JJ |
3645 | op_type = TREE_CODE_LENGTH (code); |
3646 | ||
ebfd146a | 3647 | /* Check types of lhs and rhs. */ |
b690cc0f | 3648 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 3649 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
3650 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
3651 | ||
ebfd146a IR |
3652 | op0 = gimple_assign_rhs1 (stmt); |
3653 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
3654 | |
3655 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
3656 | && !((INTEGRAL_TYPE_P (lhs_type) | |
3657 | && INTEGRAL_TYPE_P (rhs_type)) | |
3658 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
3659 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
3660 | return false; | |
3661 | ||
e6f5c25d IE |
3662 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
3663 | && ((INTEGRAL_TYPE_P (lhs_type) | |
3664 | && (TYPE_PRECISION (lhs_type) | |
3665 | != GET_MODE_PRECISION (TYPE_MODE (lhs_type)))) | |
3666 | || (INTEGRAL_TYPE_P (rhs_type) | |
3667 | && (TYPE_PRECISION (rhs_type) | |
3668 | != GET_MODE_PRECISION (TYPE_MODE (rhs_type)))))) | |
4a00c761 | 3669 | { |
73fbfcad | 3670 | if (dump_enabled_p ()) |
78c60e3d | 3671 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
3672 | "type conversion to/from bit-precision unsupported." |
3673 | "\n"); | |
4a00c761 JJ |
3674 | return false; |
3675 | } | |
3676 | ||
b690cc0f | 3677 | /* Check the operands of the operation. */ |
81c40241 | 3678 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype_in)) |
b690cc0f | 3679 | { |
73fbfcad | 3680 | if (dump_enabled_p ()) |
78c60e3d | 3681 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3682 | "use not simple.\n"); |
b690cc0f RG |
3683 | return false; |
3684 | } | |
4a00c761 JJ |
3685 | if (op_type == binary_op) |
3686 | { | |
3687 | bool ok; | |
3688 | ||
3689 | op1 = gimple_assign_rhs2 (stmt); | |
3690 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
3691 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
3692 | OP1. */ | |
3693 | if (CONSTANT_CLASS_P (op0)) | |
81c40241 | 3694 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &vectype_in); |
4a00c761 | 3695 | else |
81c40241 | 3696 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1]); |
4a00c761 JJ |
3697 | |
3698 | if (!ok) | |
3699 | { | |
73fbfcad | 3700 | if (dump_enabled_p ()) |
78c60e3d | 3701 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3702 | "use not simple.\n"); |
4a00c761 JJ |
3703 | return false; |
3704 | } | |
3705 | } | |
3706 | ||
b690cc0f RG |
3707 | /* If op0 is an external or constant defs use a vector type of |
3708 | the same size as the output vector type. */ | |
ebfd146a | 3709 | if (!vectype_in) |
b690cc0f | 3710 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
3711 | if (vec_stmt) |
3712 | gcc_assert (vectype_in); | |
3713 | if (!vectype_in) | |
3714 | { | |
73fbfcad | 3715 | if (dump_enabled_p ()) |
4a00c761 | 3716 | { |
78c60e3d SS |
3717 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3718 | "no vectype for scalar type "); | |
3719 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 3720 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 3721 | } |
7d8930a0 IR |
3722 | |
3723 | return false; | |
3724 | } | |
ebfd146a | 3725 | |
e6f5c25d IE |
3726 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
3727 | && !VECTOR_BOOLEAN_TYPE_P (vectype_in)) | |
3728 | { | |
3729 | if (dump_enabled_p ()) | |
3730 | { | |
3731 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3732 | "can't convert between boolean and non " | |
3733 | "boolean vectors"); | |
3734 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
3735 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
3736 | } | |
3737 | ||
3738 | return false; | |
3739 | } | |
3740 | ||
b690cc0f RG |
3741 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
3742 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4a00c761 | 3743 | if (nunits_in < nunits_out) |
ebfd146a IR |
3744 | modifier = NARROW; |
3745 | else if (nunits_out == nunits_in) | |
3746 | modifier = NONE; | |
ebfd146a | 3747 | else |
4a00c761 | 3748 | modifier = WIDEN; |
ebfd146a | 3749 | |
ff802fa1 IR |
3750 | /* Multiple types in SLP are handled by creating the appropriate number of |
3751 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3752 | case of SLP. */ | |
fce57248 | 3753 | if (slp_node) |
ebfd146a | 3754 | ncopies = 1; |
4a00c761 JJ |
3755 | else if (modifier == NARROW) |
3756 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; | |
3757 | else | |
3758 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
b8698a0f | 3759 | |
ebfd146a IR |
3760 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
3761 | needs to be generated. */ | |
3762 | gcc_assert (ncopies >= 1); | |
3763 | ||
ebfd146a | 3764 | /* Supportable by target? */ |
4a00c761 | 3765 | switch (modifier) |
ebfd146a | 3766 | { |
4a00c761 JJ |
3767 | case NONE: |
3768 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
3769 | return false; | |
3770 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
3771 | &decl1, &code1)) | |
3772 | break; | |
3773 | /* FALLTHRU */ | |
3774 | unsupported: | |
73fbfcad | 3775 | if (dump_enabled_p ()) |
78c60e3d | 3776 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3777 | "conversion not supported by target.\n"); |
ebfd146a | 3778 | return false; |
ebfd146a | 3779 | |
4a00c761 JJ |
3780 | case WIDEN: |
3781 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
3782 | &code1, &code2, &multi_step_cvt, |
3783 | &interm_types)) | |
4a00c761 JJ |
3784 | { |
3785 | /* Binary widening operation can only be supported directly by the | |
3786 | architecture. */ | |
3787 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
3788 | break; | |
3789 | } | |
3790 | ||
3791 | if (code != FLOAT_EXPR | |
3792 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
3793 | <= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
3794 | goto unsupported; | |
3795 | ||
3796 | rhs_mode = TYPE_MODE (rhs_type); | |
3797 | fltsz = GET_MODE_SIZE (TYPE_MODE (lhs_type)); | |
3798 | for (rhs_mode = GET_MODE_2XWIDER_MODE (TYPE_MODE (rhs_type)); | |
3799 | rhs_mode != VOIDmode && GET_MODE_SIZE (rhs_mode) <= fltsz; | |
3800 | rhs_mode = GET_MODE_2XWIDER_MODE (rhs_mode)) | |
3801 | { | |
3802 | cvt_type | |
3803 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
3804 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
3805 | if (cvt_type == NULL_TREE) | |
3806 | goto unsupported; | |
3807 | ||
3808 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
3809 | { | |
3810 | if (!supportable_convert_operation (code, vectype_out, | |
3811 | cvt_type, &decl1, &codecvt1)) | |
3812 | goto unsupported; | |
3813 | } | |
3814 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
3815 | cvt_type, &codecvt1, |
3816 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
3817 | &interm_types)) |
3818 | continue; | |
3819 | else | |
3820 | gcc_assert (multi_step_cvt == 0); | |
3821 | ||
3822 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
3823 | vectype_in, &code1, &code2, |
3824 | &multi_step_cvt, &interm_types)) | |
4a00c761 JJ |
3825 | break; |
3826 | } | |
3827 | ||
3828 | if (rhs_mode == VOIDmode || GET_MODE_SIZE (rhs_mode) > fltsz) | |
3829 | goto unsupported; | |
3830 | ||
3831 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
3832 | codecvt2 = ERROR_MARK; | |
3833 | else | |
3834 | { | |
3835 | multi_step_cvt++; | |
9771b263 | 3836 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
3837 | cvt_type = NULL_TREE; |
3838 | } | |
3839 | break; | |
3840 | ||
3841 | case NARROW: | |
3842 | gcc_assert (op_type == unary_op); | |
3843 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
3844 | &code1, &multi_step_cvt, | |
3845 | &interm_types)) | |
3846 | break; | |
3847 | ||
3848 | if (code != FIX_TRUNC_EXPR | |
3849 | || (GET_MODE_SIZE (TYPE_MODE (lhs_type)) | |
3850 | >= GET_MODE_SIZE (TYPE_MODE (rhs_type)))) | |
3851 | goto unsupported; | |
3852 | ||
3853 | rhs_mode = TYPE_MODE (rhs_type); | |
3854 | cvt_type | |
3855 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
3856 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
3857 | if (cvt_type == NULL_TREE) | |
3858 | goto unsupported; | |
3859 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
3860 | &decl1, &codecvt1)) | |
3861 | goto unsupported; | |
3862 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
3863 | &code1, &multi_step_cvt, | |
3864 | &interm_types)) | |
3865 | break; | |
3866 | goto unsupported; | |
3867 | ||
3868 | default: | |
3869 | gcc_unreachable (); | |
ebfd146a IR |
3870 | } |
3871 | ||
3872 | if (!vec_stmt) /* transformation not required. */ | |
3873 | { | |
73fbfcad | 3874 | if (dump_enabled_p ()) |
78c60e3d | 3875 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3876 | "=== vectorizable_conversion ===\n"); |
4a00c761 | 3877 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
3878 | { |
3879 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
c3e7ee41 | 3880 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
8bd37302 | 3881 | } |
4a00c761 JJ |
3882 | else if (modifier == NARROW) |
3883 | { | |
3884 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
8bd37302 | 3885 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 JJ |
3886 | } |
3887 | else | |
3888 | { | |
3889 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
8bd37302 | 3890 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); |
4a00c761 | 3891 | } |
9771b263 | 3892 | interm_types.release (); |
ebfd146a IR |
3893 | return true; |
3894 | } | |
3895 | ||
3896 | /** Transform. **/ | |
73fbfcad | 3897 | if (dump_enabled_p ()) |
78c60e3d | 3898 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 3899 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 3900 | |
4a00c761 JJ |
3901 | if (op_type == binary_op) |
3902 | { | |
3903 | if (CONSTANT_CLASS_P (op0)) | |
3904 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
3905 | else if (CONSTANT_CLASS_P (op1)) | |
3906 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
3907 | } | |
3908 | ||
3909 | /* In case of multi-step conversion, we first generate conversion operations | |
3910 | to the intermediate types, and then from that types to the final one. | |
3911 | We create vector destinations for the intermediate type (TYPES) received | |
3912 | from supportable_*_operation, and store them in the correct order | |
3913 | for future use in vect_create_vectorized_*_stmts (). */ | |
9771b263 | 3914 | vec_dsts.create (multi_step_cvt + 1); |
82294ec1 JJ |
3915 | vec_dest = vect_create_destination_var (scalar_dest, |
3916 | (cvt_type && modifier == WIDEN) | |
3917 | ? cvt_type : vectype_out); | |
9771b263 | 3918 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3919 | |
3920 | if (multi_step_cvt) | |
3921 | { | |
9771b263 DN |
3922 | for (i = interm_types.length () - 1; |
3923 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
3924 | { |
3925 | vec_dest = vect_create_destination_var (scalar_dest, | |
3926 | intermediate_type); | |
9771b263 | 3927 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
3928 | } |
3929 | } | |
ebfd146a | 3930 | |
4a00c761 | 3931 | if (cvt_type) |
82294ec1 JJ |
3932 | vec_dest = vect_create_destination_var (scalar_dest, |
3933 | modifier == WIDEN | |
3934 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
3935 | |
3936 | if (!slp_node) | |
3937 | { | |
30862efc | 3938 | if (modifier == WIDEN) |
4a00c761 | 3939 | { |
c3284718 | 3940 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 3941 | if (op_type == binary_op) |
9771b263 | 3942 | vec_oprnds1.create (1); |
4a00c761 | 3943 | } |
30862efc | 3944 | else if (modifier == NARROW) |
9771b263 DN |
3945 | vec_oprnds0.create ( |
3946 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
3947 | } |
3948 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 3949 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 3950 | |
4a00c761 | 3951 | last_oprnd = op0; |
ebfd146a IR |
3952 | prev_stmt_info = NULL; |
3953 | switch (modifier) | |
3954 | { | |
3955 | case NONE: | |
3956 | for (j = 0; j < ncopies; j++) | |
3957 | { | |
ebfd146a | 3958 | if (j == 0) |
d092494c IR |
3959 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node, |
3960 | -1); | |
ebfd146a IR |
3961 | else |
3962 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
3963 | ||
9771b263 | 3964 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
3965 | { |
3966 | /* Arguments are ready, create the new vector stmt. */ | |
3967 | if (code1 == CALL_EXPR) | |
3968 | { | |
3969 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
3970 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3971 | gimple_call_set_lhs (new_stmt, new_temp); | |
3972 | } | |
3973 | else | |
3974 | { | |
3975 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
0d0e4a03 | 3976 | new_stmt = gimple_build_assign (vec_dest, code1, vop0); |
4a00c761 JJ |
3977 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3978 | gimple_assign_set_lhs (new_stmt, new_temp); | |
3979 | } | |
3980 | ||
3981 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3982 | if (slp_node) | |
9771b263 | 3983 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
225ce44b RB |
3984 | else |
3985 | { | |
3986 | if (!prev_stmt_info) | |
3987 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3988 | else | |
3989 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3990 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3991 | } | |
4a00c761 | 3992 | } |
ebfd146a IR |
3993 | } |
3994 | break; | |
3995 | ||
3996 | case WIDEN: | |
3997 | /* In case the vectorization factor (VF) is bigger than the number | |
3998 | of elements that we can fit in a vectype (nunits), we have to | |
3999 | generate more than one vector stmt - i.e - we need to "unroll" | |
4000 | the vector stmt by a factor VF/nunits. */ | |
4001 | for (j = 0; j < ncopies; j++) | |
4002 | { | |
4a00c761 | 4003 | /* Handle uses. */ |
ebfd146a | 4004 | if (j == 0) |
4a00c761 JJ |
4005 | { |
4006 | if (slp_node) | |
4007 | { | |
4008 | if (code == WIDEN_LSHIFT_EXPR) | |
4009 | { | |
4010 | unsigned int k; | |
ebfd146a | 4011 | |
4a00c761 JJ |
4012 | vec_oprnd1 = op1; |
4013 | /* Store vec_oprnd1 for every vector stmt to be created | |
4014 | for SLP_NODE. We check during the analysis that all | |
4015 | the shift arguments are the same. */ | |
4016 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4017 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4018 | |
4019 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
4020 | slp_node, -1); | |
4021 | } | |
4022 | else | |
4023 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
4024 | &vec_oprnds1, slp_node, -1); | |
4025 | } | |
4026 | else | |
4027 | { | |
81c40241 | 4028 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 4029 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
4030 | if (op_type == binary_op) |
4031 | { | |
4032 | if (code == WIDEN_LSHIFT_EXPR) | |
4033 | vec_oprnd1 = op1; | |
4034 | else | |
81c40241 | 4035 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 4036 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4037 | } |
4038 | } | |
4039 | } | |
ebfd146a | 4040 | else |
4a00c761 JJ |
4041 | { |
4042 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
4043 | vec_oprnds0.truncate (0); |
4044 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
4045 | if (op_type == binary_op) |
4046 | { | |
4047 | if (code == WIDEN_LSHIFT_EXPR) | |
4048 | vec_oprnd1 = op1; | |
4049 | else | |
4050 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
4051 | vec_oprnd1); | |
9771b263 DN |
4052 | vec_oprnds1.truncate (0); |
4053 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
4054 | } |
4055 | } | |
ebfd146a | 4056 | |
4a00c761 JJ |
4057 | /* Arguments are ready. Create the new vector stmts. */ |
4058 | for (i = multi_step_cvt; i >= 0; i--) | |
4059 | { | |
9771b263 | 4060 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
4061 | enum tree_code c1 = code1, c2 = code2; |
4062 | if (i == 0 && codecvt2 != ERROR_MARK) | |
4063 | { | |
4064 | c1 = codecvt1; | |
4065 | c2 = codecvt2; | |
4066 | } | |
4067 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
4068 | &vec_oprnds1, | |
4069 | stmt, this_dest, gsi, | |
4070 | c1, c2, decl1, decl2, | |
4071 | op_type); | |
4072 | } | |
4073 | ||
9771b263 | 4074 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4075 | { |
4076 | if (cvt_type) | |
4077 | { | |
4078 | if (codecvt1 == CALL_EXPR) | |
4079 | { | |
4080 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4081 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4082 | gimple_call_set_lhs (new_stmt, new_temp); | |
4083 | } | |
4084 | else | |
4085 | { | |
4086 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4087 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4088 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4089 | vop0); | |
4a00c761 JJ |
4090 | } |
4091 | ||
4092 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4093 | } | |
4094 | else | |
4095 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
4096 | ||
4097 | if (slp_node) | |
9771b263 | 4098 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 4099 | else |
c689ce1e RB |
4100 | { |
4101 | if (!prev_stmt_info) | |
4102 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
4103 | else | |
4104 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4105 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4106 | } | |
4a00c761 | 4107 | } |
ebfd146a | 4108 | } |
4a00c761 JJ |
4109 | |
4110 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
4111 | break; |
4112 | ||
4113 | case NARROW: | |
4114 | /* In case the vectorization factor (VF) is bigger than the number | |
4115 | of elements that we can fit in a vectype (nunits), we have to | |
4116 | generate more than one vector stmt - i.e - we need to "unroll" | |
4117 | the vector stmt by a factor VF/nunits. */ | |
4118 | for (j = 0; j < ncopies; j++) | |
4119 | { | |
4120 | /* Handle uses. */ | |
4a00c761 JJ |
4121 | if (slp_node) |
4122 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
4123 | slp_node, -1); | |
ebfd146a IR |
4124 | else |
4125 | { | |
9771b263 | 4126 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
4127 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
4128 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
4129 | } |
4130 | ||
4a00c761 JJ |
4131 | /* Arguments are ready. Create the new vector stmts. */ |
4132 | if (cvt_type) | |
9771b263 | 4133 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4134 | { |
4135 | if (codecvt1 == CALL_EXPR) | |
4136 | { | |
4137 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4138 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4139 | gimple_call_set_lhs (new_stmt, new_temp); | |
4140 | } | |
4141 | else | |
4142 | { | |
4143 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4144 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4145 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4146 | vop0); | |
4a00c761 | 4147 | } |
ebfd146a | 4148 | |
4a00c761 | 4149 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 4150 | vec_oprnds0[i] = new_temp; |
4a00c761 | 4151 | } |
ebfd146a | 4152 | |
4a00c761 JJ |
4153 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
4154 | stmt, vec_dsts, gsi, | |
4155 | slp_node, code1, | |
4156 | &prev_stmt_info); | |
ebfd146a IR |
4157 | } |
4158 | ||
4159 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 4160 | break; |
ebfd146a IR |
4161 | } |
4162 | ||
9771b263 DN |
4163 | vec_oprnds0.release (); |
4164 | vec_oprnds1.release (); | |
4165 | vec_dsts.release (); | |
4166 | interm_types.release (); | |
ebfd146a IR |
4167 | |
4168 | return true; | |
4169 | } | |
ff802fa1 IR |
4170 | |
4171 | ||
ebfd146a IR |
4172 | /* Function vectorizable_assignment. |
4173 | ||
b8698a0f L |
4174 | Check if STMT performs an assignment (copy) that can be vectorized. |
4175 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
4176 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4177 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4178 | ||
4179 | static bool | |
355fe088 TS |
4180 | vectorizable_assignment (gimple *stmt, gimple_stmt_iterator *gsi, |
4181 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
4182 | { |
4183 | tree vec_dest; | |
4184 | tree scalar_dest; | |
4185 | tree op; | |
4186 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a IR |
4187 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4188 | tree new_temp; | |
355fe088 | 4189 | gimple *def_stmt; |
ebfd146a | 4190 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
ebfd146a | 4191 | int ncopies; |
f18b55bd | 4192 | int i, j; |
6e1aa848 | 4193 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 4194 | tree vop; |
a70d6342 | 4195 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4196 | vec_info *vinfo = stmt_info->vinfo; |
355fe088 | 4197 | gimple *new_stmt = NULL; |
f18b55bd | 4198 | stmt_vec_info prev_stmt_info = NULL; |
fde9c428 RG |
4199 | enum tree_code code; |
4200 | tree vectype_in; | |
ebfd146a | 4201 | |
a70d6342 | 4202 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4203 | return false; |
4204 | ||
66c16fd9 RB |
4205 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4206 | && ! vec_stmt) | |
ebfd146a IR |
4207 | return false; |
4208 | ||
4209 | /* Is vectorizable assignment? */ | |
4210 | if (!is_gimple_assign (stmt)) | |
4211 | return false; | |
4212 | ||
4213 | scalar_dest = gimple_assign_lhs (stmt); | |
4214 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
4215 | return false; | |
4216 | ||
fde9c428 | 4217 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 4218 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
4219 | || code == PAREN_EXPR |
4220 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
4221 | op = gimple_assign_rhs1 (stmt); |
4222 | else | |
4223 | return false; | |
4224 | ||
7b7ec6c5 RG |
4225 | if (code == VIEW_CONVERT_EXPR) |
4226 | op = TREE_OPERAND (op, 0); | |
4227 | ||
465c8c19 JJ |
4228 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
4229 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
4230 | ||
4231 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4232 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4233 | case of SLP. */ | |
fce57248 | 4234 | if (slp_node) |
465c8c19 JJ |
4235 | ncopies = 1; |
4236 | else | |
4237 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
4238 | ||
4239 | gcc_assert (ncopies >= 1); | |
4240 | ||
81c40241 | 4241 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[0], &vectype_in)) |
ebfd146a | 4242 | { |
73fbfcad | 4243 | if (dump_enabled_p ()) |
78c60e3d | 4244 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4245 | "use not simple.\n"); |
ebfd146a IR |
4246 | return false; |
4247 | } | |
4248 | ||
fde9c428 RG |
4249 | /* We can handle NOP_EXPR conversions that do not change the number |
4250 | of elements or the vector size. */ | |
7b7ec6c5 RG |
4251 | if ((CONVERT_EXPR_CODE_P (code) |
4252 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 RG |
4253 | && (!vectype_in |
4254 | || TYPE_VECTOR_SUBPARTS (vectype_in) != nunits | |
4255 | || (GET_MODE_SIZE (TYPE_MODE (vectype)) | |
4256 | != GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
4257 | return false; | |
4258 | ||
7b7b1813 RG |
4259 | /* We do not handle bit-precision changes. */ |
4260 | if ((CONVERT_EXPR_CODE_P (code) | |
4261 | || code == VIEW_CONVERT_EXPR) | |
4262 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
4263 | && ((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4264 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4265 | || ((TYPE_PRECISION (TREE_TYPE (op)) | |
4266 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op)))))) | |
4267 | /* But a conversion that does not change the bit-pattern is ok. */ | |
4268 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4269 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2dab46d5 IE |
4270 | && TYPE_UNSIGNED (TREE_TYPE (op))) |
4271 | /* Conversion between boolean types of different sizes is | |
4272 | a simple assignment in case their vectypes are same | |
4273 | boolean vectors. */ | |
4274 | && (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
4275 | || !VECTOR_BOOLEAN_TYPE_P (vectype_in))) | |
7b7b1813 | 4276 | { |
73fbfcad | 4277 | if (dump_enabled_p ()) |
78c60e3d SS |
4278 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4279 | "type conversion to/from bit-precision " | |
e645e942 | 4280 | "unsupported.\n"); |
7b7b1813 RG |
4281 | return false; |
4282 | } | |
4283 | ||
ebfd146a IR |
4284 | if (!vec_stmt) /* transformation not required. */ |
4285 | { | |
4286 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 4287 | if (dump_enabled_p ()) |
78c60e3d | 4288 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4289 | "=== vectorizable_assignment ===\n"); |
c3e7ee41 | 4290 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
4291 | return true; |
4292 | } | |
4293 | ||
4294 | /** Transform. **/ | |
73fbfcad | 4295 | if (dump_enabled_p ()) |
e645e942 | 4296 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
4297 | |
4298 | /* Handle def. */ | |
4299 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4300 | ||
4301 | /* Handle use. */ | |
f18b55bd | 4302 | for (j = 0; j < ncopies; j++) |
ebfd146a | 4303 | { |
f18b55bd IR |
4304 | /* Handle uses. */ |
4305 | if (j == 0) | |
d092494c | 4306 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node, -1); |
f18b55bd IR |
4307 | else |
4308 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
4309 | ||
4310 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 4311 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 4312 | { |
7b7ec6c5 RG |
4313 | if (CONVERT_EXPR_CODE_P (code) |
4314 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 4315 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
4316 | new_stmt = gimple_build_assign (vec_dest, vop); |
4317 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4318 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4319 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4320 | if (slp_node) | |
9771b263 | 4321 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 4322 | } |
ebfd146a IR |
4323 | |
4324 | if (slp_node) | |
f18b55bd IR |
4325 | continue; |
4326 | ||
4327 | if (j == 0) | |
4328 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4329 | else | |
4330 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4331 | ||
4332 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4333 | } | |
b8698a0f | 4334 | |
9771b263 | 4335 | vec_oprnds.release (); |
ebfd146a IR |
4336 | return true; |
4337 | } | |
4338 | ||
9dc3f7de | 4339 | |
1107f3ae IR |
4340 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
4341 | either as shift by a scalar or by a vector. */ | |
4342 | ||
4343 | bool | |
4344 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
4345 | { | |
4346 | ||
ef4bddc2 | 4347 | machine_mode vec_mode; |
1107f3ae IR |
4348 | optab optab; |
4349 | int icode; | |
4350 | tree vectype; | |
4351 | ||
4352 | vectype = get_vectype_for_scalar_type (scalar_type); | |
4353 | if (!vectype) | |
4354 | return false; | |
4355 | ||
4356 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4357 | if (!optab | |
4358 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
4359 | { | |
4360 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4361 | if (!optab | |
4362 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
4363 | == CODE_FOR_nothing)) | |
4364 | return false; | |
4365 | } | |
4366 | ||
4367 | vec_mode = TYPE_MODE (vectype); | |
4368 | icode = (int) optab_handler (optab, vec_mode); | |
4369 | if (icode == CODE_FOR_nothing) | |
4370 | return false; | |
4371 | ||
4372 | return true; | |
4373 | } | |
4374 | ||
4375 | ||
9dc3f7de IR |
4376 | /* Function vectorizable_shift. |
4377 | ||
4378 | Check if STMT performs a shift operation that can be vectorized. | |
4379 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4380 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
4381 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4382 | ||
4383 | static bool | |
355fe088 TS |
4384 | vectorizable_shift (gimple *stmt, gimple_stmt_iterator *gsi, |
4385 | gimple **vec_stmt, slp_tree slp_node) | |
9dc3f7de IR |
4386 | { |
4387 | tree vec_dest; | |
4388 | tree scalar_dest; | |
4389 | tree op0, op1 = NULL; | |
4390 | tree vec_oprnd1 = NULL_TREE; | |
4391 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4392 | tree vectype; | |
4393 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4394 | enum tree_code code; | |
ef4bddc2 | 4395 | machine_mode vec_mode; |
9dc3f7de IR |
4396 | tree new_temp; |
4397 | optab optab; | |
4398 | int icode; | |
ef4bddc2 | 4399 | machine_mode optab_op2_mode; |
355fe088 | 4400 | gimple *def_stmt; |
9dc3f7de | 4401 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
355fe088 | 4402 | gimple *new_stmt = NULL; |
9dc3f7de IR |
4403 | stmt_vec_info prev_stmt_info; |
4404 | int nunits_in; | |
4405 | int nunits_out; | |
4406 | tree vectype_out; | |
cede2577 | 4407 | tree op1_vectype; |
9dc3f7de IR |
4408 | int ncopies; |
4409 | int j, i; | |
6e1aa848 DN |
4410 | vec<tree> vec_oprnds0 = vNULL; |
4411 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
4412 | tree vop0, vop1; |
4413 | unsigned int k; | |
49eab32e | 4414 | bool scalar_shift_arg = true; |
9dc3f7de | 4415 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4416 | vec_info *vinfo = stmt_info->vinfo; |
9dc3f7de IR |
4417 | int vf; |
4418 | ||
4419 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
4420 | return false; | |
4421 | ||
66c16fd9 RB |
4422 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4423 | && ! vec_stmt) | |
9dc3f7de IR |
4424 | return false; |
4425 | ||
4426 | /* Is STMT a vectorizable binary/unary operation? */ | |
4427 | if (!is_gimple_assign (stmt)) | |
4428 | return false; | |
4429 | ||
4430 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4431 | return false; | |
4432 | ||
4433 | code = gimple_assign_rhs_code (stmt); | |
4434 | ||
4435 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
4436 | || code == RROTATE_EXPR)) | |
4437 | return false; | |
4438 | ||
4439 | scalar_dest = gimple_assign_lhs (stmt); | |
4440 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
7b7b1813 RG |
4441 | if (TYPE_PRECISION (TREE_TYPE (scalar_dest)) |
4442 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
4443 | { | |
73fbfcad | 4444 | if (dump_enabled_p ()) |
78c60e3d | 4445 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4446 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
4447 | return false; |
4448 | } | |
9dc3f7de IR |
4449 | |
4450 | op0 = gimple_assign_rhs1 (stmt); | |
81c40241 | 4451 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
9dc3f7de | 4452 | { |
73fbfcad | 4453 | if (dump_enabled_p ()) |
78c60e3d | 4454 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4455 | "use not simple.\n"); |
9dc3f7de IR |
4456 | return false; |
4457 | } | |
4458 | /* If op0 is an external or constant def use a vector type with | |
4459 | the same size as the output vector type. */ | |
4460 | if (!vectype) | |
4461 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
4462 | if (vec_stmt) | |
4463 | gcc_assert (vectype); | |
4464 | if (!vectype) | |
4465 | { | |
73fbfcad | 4466 | if (dump_enabled_p ()) |
78c60e3d | 4467 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4468 | "no vectype for scalar type\n"); |
9dc3f7de IR |
4469 | return false; |
4470 | } | |
4471 | ||
4472 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4473 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4474 | if (nunits_out != nunits_in) | |
4475 | return false; | |
4476 | ||
4477 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 4478 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &op1_vectype)) |
9dc3f7de | 4479 | { |
73fbfcad | 4480 | if (dump_enabled_p ()) |
78c60e3d | 4481 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4482 | "use not simple.\n"); |
9dc3f7de IR |
4483 | return false; |
4484 | } | |
4485 | ||
4486 | if (loop_vinfo) | |
4487 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4488 | else | |
4489 | vf = 1; | |
4490 | ||
4491 | /* Multiple types in SLP are handled by creating the appropriate number of | |
4492 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4493 | case of SLP. */ | |
fce57248 | 4494 | if (slp_node) |
9dc3f7de IR |
4495 | ncopies = 1; |
4496 | else | |
4497 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
4498 | ||
4499 | gcc_assert (ncopies >= 1); | |
4500 | ||
4501 | /* Determine whether the shift amount is a vector, or scalar. If the | |
4502 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
4503 | ||
dbfa87aa YR |
4504 | if ((dt[1] == vect_internal_def |
4505 | || dt[1] == vect_induction_def) | |
4506 | && !slp_node) | |
49eab32e JJ |
4507 | scalar_shift_arg = false; |
4508 | else if (dt[1] == vect_constant_def | |
4509 | || dt[1] == vect_external_def | |
4510 | || dt[1] == vect_internal_def) | |
4511 | { | |
4512 | /* In SLP, need to check whether the shift count is the same, | |
4513 | in loops if it is a constant or invariant, it is always | |
4514 | a scalar shift. */ | |
4515 | if (slp_node) | |
4516 | { | |
355fe088 TS |
4517 | vec<gimple *> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
4518 | gimple *slpstmt; | |
49eab32e | 4519 | |
9771b263 | 4520 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
4521 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
4522 | scalar_shift_arg = false; | |
4523 | } | |
60d393e8 RB |
4524 | |
4525 | /* If the shift amount is computed by a pattern stmt we cannot | |
4526 | use the scalar amount directly thus give up and use a vector | |
4527 | shift. */ | |
4528 | if (dt[1] == vect_internal_def) | |
4529 | { | |
4530 | gimple *def = SSA_NAME_DEF_STMT (op1); | |
4531 | if (is_pattern_stmt_p (vinfo_for_stmt (def))) | |
4532 | scalar_shift_arg = false; | |
4533 | } | |
49eab32e JJ |
4534 | } |
4535 | else | |
4536 | { | |
73fbfcad | 4537 | if (dump_enabled_p ()) |
78c60e3d | 4538 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4539 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
4540 | return false; |
4541 | } | |
4542 | ||
9dc3f7de | 4543 | /* Vector shifted by vector. */ |
49eab32e | 4544 | if (!scalar_shift_arg) |
9dc3f7de IR |
4545 | { |
4546 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 4547 | if (dump_enabled_p ()) |
78c60e3d | 4548 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4549 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 4550 | |
aa948027 JJ |
4551 | if (!op1_vectype) |
4552 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
4553 | if (op1_vectype == NULL_TREE | |
4554 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 4555 | { |
73fbfcad | 4556 | if (dump_enabled_p ()) |
78c60e3d SS |
4557 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4558 | "unusable type for last operand in" | |
e645e942 | 4559 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
4560 | return false; |
4561 | } | |
9dc3f7de IR |
4562 | } |
4563 | /* See if the machine has a vector shifted by scalar insn and if not | |
4564 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 4565 | else |
9dc3f7de IR |
4566 | { |
4567 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
4568 | if (optab | |
4569 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
4570 | { | |
73fbfcad | 4571 | if (dump_enabled_p ()) |
78c60e3d | 4572 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4573 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
4574 | } |
4575 | else | |
4576 | { | |
4577 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
4578 | if (optab | |
4579 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
4580 | != CODE_FOR_nothing)) | |
4581 | { | |
49eab32e JJ |
4582 | scalar_shift_arg = false; |
4583 | ||
73fbfcad | 4584 | if (dump_enabled_p ()) |
78c60e3d | 4585 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4586 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
4587 | |
4588 | /* Unlike the other binary operators, shifts/rotates have | |
4589 | the rhs being int, instead of the same type as the lhs, | |
4590 | so make sure the scalar is the right type if we are | |
aa948027 | 4591 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
4592 | if (dt[1] == vect_constant_def) |
4593 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
4594 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
4595 | TREE_TYPE (op1))) | |
4596 | { | |
4597 | if (slp_node | |
4598 | && TYPE_MODE (TREE_TYPE (vectype)) | |
4599 | != TYPE_MODE (TREE_TYPE (op1))) | |
4600 | { | |
73fbfcad | 4601 | if (dump_enabled_p ()) |
78c60e3d SS |
4602 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4603 | "unusable type for last operand in" | |
e645e942 | 4604 | " vector/vector shift/rotate.\n"); |
21c0a521 | 4605 | return false; |
aa948027 JJ |
4606 | } |
4607 | if (vec_stmt && !slp_node) | |
4608 | { | |
4609 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
4610 | op1 = vect_init_vector (stmt, op1, | |
4611 | TREE_TYPE (vectype), NULL); | |
4612 | } | |
4613 | } | |
9dc3f7de IR |
4614 | } |
4615 | } | |
4616 | } | |
9dc3f7de IR |
4617 | |
4618 | /* Supportable by target? */ | |
4619 | if (!optab) | |
4620 | { | |
73fbfcad | 4621 | if (dump_enabled_p ()) |
78c60e3d | 4622 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4623 | "no optab.\n"); |
9dc3f7de IR |
4624 | return false; |
4625 | } | |
4626 | vec_mode = TYPE_MODE (vectype); | |
4627 | icode = (int) optab_handler (optab, vec_mode); | |
4628 | if (icode == CODE_FOR_nothing) | |
4629 | { | |
73fbfcad | 4630 | if (dump_enabled_p ()) |
78c60e3d | 4631 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4632 | "op not supported by target.\n"); |
9dc3f7de IR |
4633 | /* Check only during analysis. */ |
4634 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
4635 | || (vf < vect_min_worthwhile_factor (code) | |
4636 | && !vec_stmt)) | |
4637 | return false; | |
73fbfcad | 4638 | if (dump_enabled_p ()) |
e645e942 TJ |
4639 | dump_printf_loc (MSG_NOTE, vect_location, |
4640 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
4641 | } |
4642 | ||
4643 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
4644 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
4645 | && vf < vect_min_worthwhile_factor (code) | |
4646 | && !vec_stmt) | |
4647 | { | |
73fbfcad | 4648 | if (dump_enabled_p ()) |
78c60e3d | 4649 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4650 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
4651 | return false; |
4652 | } | |
4653 | ||
4654 | if (!vec_stmt) /* transformation not required. */ | |
4655 | { | |
4656 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 4657 | if (dump_enabled_p ()) |
e645e942 TJ |
4658 | dump_printf_loc (MSG_NOTE, vect_location, |
4659 | "=== vectorizable_shift ===\n"); | |
c3e7ee41 | 4660 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
9dc3f7de IR |
4661 | return true; |
4662 | } | |
4663 | ||
4664 | /** Transform. **/ | |
4665 | ||
73fbfcad | 4666 | if (dump_enabled_p ()) |
78c60e3d | 4667 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4668 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
4669 | |
4670 | /* Handle def. */ | |
4671 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4672 | ||
9dc3f7de IR |
4673 | prev_stmt_info = NULL; |
4674 | for (j = 0; j < ncopies; j++) | |
4675 | { | |
4676 | /* Handle uses. */ | |
4677 | if (j == 0) | |
4678 | { | |
4679 | if (scalar_shift_arg) | |
4680 | { | |
4681 | /* Vector shl and shr insn patterns can be defined with scalar | |
4682 | operand 2 (shift operand). In this case, use constant or loop | |
4683 | invariant op1 directly, without extending it to vector mode | |
4684 | first. */ | |
4685 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
4686 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
4687 | { | |
73fbfcad | 4688 | if (dump_enabled_p ()) |
78c60e3d | 4689 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4690 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 4691 | vec_oprnd1 = op1; |
8930f723 | 4692 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 4693 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
4694 | if (slp_node) |
4695 | { | |
4696 | /* Store vec_oprnd1 for every vector stmt to be created | |
4697 | for SLP_NODE. We check during the analysis that all | |
4698 | the shift arguments are the same. | |
4699 | TODO: Allow different constants for different vector | |
4700 | stmts generated for an SLP instance. */ | |
4701 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4702 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
4703 | } |
4704 | } | |
4705 | } | |
4706 | ||
4707 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
4708 | (a special case for certain kind of vector shifts); otherwise, | |
4709 | operand 1 should be of a vector type (the usual case). */ | |
4710 | if (vec_oprnd1) | |
4711 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
d092494c | 4712 | slp_node, -1); |
9dc3f7de IR |
4713 | else |
4714 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
d092494c | 4715 | slp_node, -1); |
9dc3f7de IR |
4716 | } |
4717 | else | |
4718 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
4719 | ||
4720 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 4721 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 4722 | { |
9771b263 | 4723 | vop1 = vec_oprnds1[i]; |
0d0e4a03 | 4724 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
9dc3f7de IR |
4725 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4726 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4727 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4728 | if (slp_node) | |
9771b263 | 4729 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
4730 | } |
4731 | ||
4732 | if (slp_node) | |
4733 | continue; | |
4734 | ||
4735 | if (j == 0) | |
4736 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4737 | else | |
4738 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4739 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4740 | } | |
4741 | ||
9771b263 DN |
4742 | vec_oprnds0.release (); |
4743 | vec_oprnds1.release (); | |
9dc3f7de IR |
4744 | |
4745 | return true; | |
4746 | } | |
4747 | ||
4748 | ||
ebfd146a IR |
4749 | /* Function vectorizable_operation. |
4750 | ||
16949072 RG |
4751 | Check if STMT performs a binary, unary or ternary operation that can |
4752 | be vectorized. | |
b8698a0f | 4753 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
4754 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
4755 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
4756 | ||
4757 | static bool | |
355fe088 TS |
4758 | vectorizable_operation (gimple *stmt, gimple_stmt_iterator *gsi, |
4759 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a | 4760 | { |
00f07b86 | 4761 | tree vec_dest; |
ebfd146a | 4762 | tree scalar_dest; |
16949072 | 4763 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 4764 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 4765 | tree vectype; |
ebfd146a IR |
4766 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
4767 | enum tree_code code; | |
ef4bddc2 | 4768 | machine_mode vec_mode; |
ebfd146a IR |
4769 | tree new_temp; |
4770 | int op_type; | |
00f07b86 | 4771 | optab optab; |
523ba738 | 4772 | bool target_support_p; |
355fe088 | 4773 | gimple *def_stmt; |
16949072 RG |
4774 | enum vect_def_type dt[3] |
4775 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
355fe088 | 4776 | gimple *new_stmt = NULL; |
ebfd146a | 4777 | stmt_vec_info prev_stmt_info; |
b690cc0f | 4778 | int nunits_in; |
ebfd146a IR |
4779 | int nunits_out; |
4780 | tree vectype_out; | |
4781 | int ncopies; | |
4782 | int j, i; | |
6e1aa848 DN |
4783 | vec<tree> vec_oprnds0 = vNULL; |
4784 | vec<tree> vec_oprnds1 = vNULL; | |
4785 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 4786 | tree vop0, vop1, vop2; |
a70d6342 | 4787 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4788 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 IR |
4789 | int vf; |
4790 | ||
a70d6342 | 4791 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4792 | return false; |
4793 | ||
66c16fd9 RB |
4794 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4795 | && ! vec_stmt) | |
ebfd146a IR |
4796 | return false; |
4797 | ||
4798 | /* Is STMT a vectorizable binary/unary operation? */ | |
4799 | if (!is_gimple_assign (stmt)) | |
4800 | return false; | |
4801 | ||
4802 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4803 | return false; | |
4804 | ||
ebfd146a IR |
4805 | code = gimple_assign_rhs_code (stmt); |
4806 | ||
4807 | /* For pointer addition, we should use the normal plus for | |
4808 | the vector addition. */ | |
4809 | if (code == POINTER_PLUS_EXPR) | |
4810 | code = PLUS_EXPR; | |
4811 | ||
4812 | /* Support only unary or binary operations. */ | |
4813 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 4814 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 4815 | { |
73fbfcad | 4816 | if (dump_enabled_p ()) |
78c60e3d | 4817 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4818 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 4819 | op_type); |
ebfd146a IR |
4820 | return false; |
4821 | } | |
4822 | ||
b690cc0f RG |
4823 | scalar_dest = gimple_assign_lhs (stmt); |
4824 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
4825 | ||
7b7b1813 RG |
4826 | /* Most operations cannot handle bit-precision types without extra |
4827 | truncations. */ | |
045c1278 IE |
4828 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4829 | && (TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
4830 | != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (scalar_dest)))) | |
7b7b1813 RG |
4831 | /* Exception are bitwise binary operations. */ |
4832 | && code != BIT_IOR_EXPR | |
4833 | && code != BIT_XOR_EXPR | |
4834 | && code != BIT_AND_EXPR) | |
4835 | { | |
73fbfcad | 4836 | if (dump_enabled_p ()) |
78c60e3d | 4837 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4838 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
4839 | return false; |
4840 | } | |
4841 | ||
ebfd146a | 4842 | op0 = gimple_assign_rhs1 (stmt); |
81c40241 | 4843 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
ebfd146a | 4844 | { |
73fbfcad | 4845 | if (dump_enabled_p ()) |
78c60e3d | 4846 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4847 | "use not simple.\n"); |
ebfd146a IR |
4848 | return false; |
4849 | } | |
b690cc0f RG |
4850 | /* If op0 is an external or constant def use a vector type with |
4851 | the same size as the output vector type. */ | |
4852 | if (!vectype) | |
b036c6c5 IE |
4853 | { |
4854 | /* For boolean type we cannot determine vectype by | |
4855 | invariant value (don't know whether it is a vector | |
4856 | of booleans or vector of integers). We use output | |
4857 | vectype because operations on boolean don't change | |
4858 | type. */ | |
4859 | if (TREE_CODE (TREE_TYPE (op0)) == BOOLEAN_TYPE) | |
4860 | { | |
4861 | if (TREE_CODE (TREE_TYPE (scalar_dest)) != BOOLEAN_TYPE) | |
4862 | { | |
4863 | if (dump_enabled_p ()) | |
4864 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
4865 | "not supported operation on bool value.\n"); | |
4866 | return false; | |
4867 | } | |
4868 | vectype = vectype_out; | |
4869 | } | |
4870 | else | |
4871 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
4872 | } | |
7d8930a0 IR |
4873 | if (vec_stmt) |
4874 | gcc_assert (vectype); | |
4875 | if (!vectype) | |
4876 | { | |
73fbfcad | 4877 | if (dump_enabled_p ()) |
7d8930a0 | 4878 | { |
78c60e3d SS |
4879 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4880 | "no vectype for scalar type "); | |
4881 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
4882 | TREE_TYPE (op0)); | |
e645e942 | 4883 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
4884 | } |
4885 | ||
4886 | return false; | |
4887 | } | |
b690cc0f RG |
4888 | |
4889 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
4890 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
4891 | if (nunits_out != nunits_in) | |
4892 | return false; | |
ebfd146a | 4893 | |
16949072 | 4894 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
4895 | { |
4896 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 4897 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1])) |
ebfd146a | 4898 | { |
73fbfcad | 4899 | if (dump_enabled_p ()) |
78c60e3d | 4900 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4901 | "use not simple.\n"); |
ebfd146a IR |
4902 | return false; |
4903 | } | |
4904 | } | |
16949072 RG |
4905 | if (op_type == ternary_op) |
4906 | { | |
4907 | op2 = gimple_assign_rhs3 (stmt); | |
81c40241 | 4908 | if (!vect_is_simple_use (op2, vinfo, &def_stmt, &dt[2])) |
16949072 | 4909 | { |
73fbfcad | 4910 | if (dump_enabled_p ()) |
78c60e3d | 4911 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4912 | "use not simple.\n"); |
16949072 RG |
4913 | return false; |
4914 | } | |
4915 | } | |
ebfd146a | 4916 | |
b690cc0f RG |
4917 | if (loop_vinfo) |
4918 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
4919 | else | |
4920 | vf = 1; | |
4921 | ||
4922 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 4923 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 4924 | case of SLP. */ |
fce57248 | 4925 | if (slp_node) |
b690cc0f RG |
4926 | ncopies = 1; |
4927 | else | |
4928 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; | |
4929 | ||
4930 | gcc_assert (ncopies >= 1); | |
4931 | ||
9dc3f7de | 4932 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
4933 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
4934 | || code == RROTATE_EXPR) | |
9dc3f7de | 4935 | return false; |
ebfd146a | 4936 | |
ebfd146a | 4937 | /* Supportable by target? */ |
00f07b86 RH |
4938 | |
4939 | vec_mode = TYPE_MODE (vectype); | |
4940 | if (code == MULT_HIGHPART_EXPR) | |
523ba738 | 4941 | target_support_p = can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype)); |
00f07b86 RH |
4942 | else |
4943 | { | |
4944 | optab = optab_for_tree_code (code, vectype, optab_default); | |
4945 | if (!optab) | |
5deb57cb | 4946 | { |
73fbfcad | 4947 | if (dump_enabled_p ()) |
78c60e3d | 4948 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4949 | "no optab.\n"); |
00f07b86 | 4950 | return false; |
5deb57cb | 4951 | } |
523ba738 RS |
4952 | target_support_p = (optab_handler (optab, vec_mode) |
4953 | != CODE_FOR_nothing); | |
5deb57cb JJ |
4954 | } |
4955 | ||
523ba738 | 4956 | if (!target_support_p) |
ebfd146a | 4957 | { |
73fbfcad | 4958 | if (dump_enabled_p ()) |
78c60e3d | 4959 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4960 | "op not supported by target.\n"); |
ebfd146a IR |
4961 | /* Check only during analysis. */ |
4962 | if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD | |
5deb57cb | 4963 | || (!vec_stmt && vf < vect_min_worthwhile_factor (code))) |
ebfd146a | 4964 | return false; |
73fbfcad | 4965 | if (dump_enabled_p ()) |
e645e942 TJ |
4966 | dump_printf_loc (MSG_NOTE, vect_location, |
4967 | "proceeding using word mode.\n"); | |
383d9c83 IR |
4968 | } |
4969 | ||
4a00c761 | 4970 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
4971 | if (!VECTOR_MODE_P (vec_mode) |
4972 | && !vec_stmt | |
4973 | && vf < vect_min_worthwhile_factor (code)) | |
7d8930a0 | 4974 | { |
73fbfcad | 4975 | if (dump_enabled_p ()) |
78c60e3d | 4976 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4977 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 4978 | return false; |
7d8930a0 | 4979 | } |
ebfd146a | 4980 | |
ebfd146a IR |
4981 | if (!vec_stmt) /* transformation not required. */ |
4982 | { | |
4a00c761 | 4983 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 4984 | if (dump_enabled_p ()) |
78c60e3d | 4985 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4986 | "=== vectorizable_operation ===\n"); |
c3e7ee41 | 4987 | vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); |
ebfd146a IR |
4988 | return true; |
4989 | } | |
4990 | ||
4991 | /** Transform. **/ | |
4992 | ||
73fbfcad | 4993 | if (dump_enabled_p ()) |
78c60e3d | 4994 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4995 | "transform binary/unary operation.\n"); |
383d9c83 | 4996 | |
ebfd146a | 4997 | /* Handle def. */ |
00f07b86 | 4998 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 4999 | |
ebfd146a IR |
5000 | /* In case the vectorization factor (VF) is bigger than the number |
5001 | of elements that we can fit in a vectype (nunits), we have to generate | |
5002 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
5003 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
5004 | from one copy of the vector stmt to the next, in the field | |
5005 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
5006 | stages to find the correct vector defs to be used when vectorizing | |
5007 | stmts that use the defs of the current stmt. The example below | |
5008 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
5009 | we need to create 4 vectorized stmts): | |
5010 | ||
5011 | before vectorization: | |
5012 | RELATED_STMT VEC_STMT | |
5013 | S1: x = memref - - | |
5014 | S2: z = x + 1 - - | |
5015 | ||
5016 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
5017 | there): | |
5018 | RELATED_STMT VEC_STMT | |
5019 | VS1_0: vx0 = memref0 VS1_1 - | |
5020 | VS1_1: vx1 = memref1 VS1_2 - | |
5021 | VS1_2: vx2 = memref2 VS1_3 - | |
5022 | VS1_3: vx3 = memref3 - - | |
5023 | S1: x = load - VS1_0 | |
5024 | S2: z = x + 1 - - | |
5025 | ||
5026 | step2: vectorize stmt S2 (done here): | |
5027 | To vectorize stmt S2 we first need to find the relevant vector | |
5028 | def for the first operand 'x'. This is, as usual, obtained from | |
5029 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
5030 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
5031 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
5032 | the vector stmt VS2_0, and as usual, record it in the | |
5033 | STMT_VINFO_VEC_STMT of stmt S2. | |
5034 | When creating the second copy (VS2_1), we obtain the relevant vector | |
5035 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
5036 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
5037 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
5038 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
5039 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
5040 | chain of stmts and pointers: | |
5041 | RELATED_STMT VEC_STMT | |
5042 | VS1_0: vx0 = memref0 VS1_1 - | |
5043 | VS1_1: vx1 = memref1 VS1_2 - | |
5044 | VS1_2: vx2 = memref2 VS1_3 - | |
5045 | VS1_3: vx3 = memref3 - - | |
5046 | S1: x = load - VS1_0 | |
5047 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
5048 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
5049 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
5050 | VS2_3: vz3 = vx3 + v1 - - | |
5051 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
5052 | |
5053 | prev_stmt_info = NULL; | |
5054 | for (j = 0; j < ncopies; j++) | |
5055 | { | |
5056 | /* Handle uses. */ | |
5057 | if (j == 0) | |
4a00c761 JJ |
5058 | { |
5059 | if (op_type == binary_op || op_type == ternary_op) | |
5060 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
5061 | slp_node, -1); | |
5062 | else | |
5063 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
5064 | slp_node, -1); | |
5065 | if (op_type == ternary_op) | |
36ba4aae | 5066 | { |
9771b263 DN |
5067 | vec_oprnds2.create (1); |
5068 | vec_oprnds2.quick_push (vect_get_vec_def_for_operand (op2, | |
81c40241 | 5069 | stmt)); |
36ba4aae | 5070 | } |
4a00c761 | 5071 | } |
ebfd146a | 5072 | else |
4a00c761 JJ |
5073 | { |
5074 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5075 | if (op_type == ternary_op) | |
5076 | { | |
9771b263 DN |
5077 | tree vec_oprnd = vec_oprnds2.pop (); |
5078 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
5079 | vec_oprnd)); | |
4a00c761 JJ |
5080 | } |
5081 | } | |
5082 | ||
5083 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 5084 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 5085 | { |
4a00c761 | 5086 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 5087 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 5088 | vop2 = ((op_type == ternary_op) |
9771b263 | 5089 | ? vec_oprnds2[i] : NULL_TREE); |
0d0e4a03 | 5090 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1, vop2); |
4a00c761 JJ |
5091 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5092 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5093 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5094 | if (slp_node) | |
9771b263 | 5095 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
5096 | } |
5097 | ||
4a00c761 JJ |
5098 | if (slp_node) |
5099 | continue; | |
5100 | ||
5101 | if (j == 0) | |
5102 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5103 | else | |
5104 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5105 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
5106 | } |
5107 | ||
9771b263 DN |
5108 | vec_oprnds0.release (); |
5109 | vec_oprnds1.release (); | |
5110 | vec_oprnds2.release (); | |
ebfd146a | 5111 | |
ebfd146a IR |
5112 | return true; |
5113 | } | |
5114 | ||
c716e67f XDL |
5115 | /* A helper function to ensure data reference DR's base alignment |
5116 | for STMT_INFO. */ | |
5117 | ||
5118 | static void | |
5119 | ensure_base_align (stmt_vec_info stmt_info, struct data_reference *dr) | |
5120 | { | |
5121 | if (!dr->aux) | |
5122 | return; | |
5123 | ||
52639a61 | 5124 | if (DR_VECT_AUX (dr)->base_misaligned) |
c716e67f XDL |
5125 | { |
5126 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
52639a61 | 5127 | tree base_decl = DR_VECT_AUX (dr)->base_decl; |
c716e67f | 5128 | |
428f0c67 JH |
5129 | if (decl_in_symtab_p (base_decl)) |
5130 | symtab_node::get (base_decl)->increase_alignment (TYPE_ALIGN (vectype)); | |
5131 | else | |
5132 | { | |
fe37c7af | 5133 | SET_DECL_ALIGN (base_decl, TYPE_ALIGN (vectype)); |
428f0c67 JH |
5134 | DECL_USER_ALIGN (base_decl) = 1; |
5135 | } | |
52639a61 | 5136 | DR_VECT_AUX (dr)->base_misaligned = false; |
c716e67f XDL |
5137 | } |
5138 | } | |
5139 | ||
ebfd146a | 5140 | |
09dfa495 BM |
5141 | /* Given a vector type VECTYPE returns the VECTOR_CST mask that implements |
5142 | reversal of the vector elements. If that is impossible to do, | |
5143 | returns NULL. */ | |
5144 | ||
5145 | static tree | |
5146 | perm_mask_for_reverse (tree vectype) | |
5147 | { | |
5148 | int i, nunits; | |
5149 | unsigned char *sel; | |
5150 | ||
5151 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
5152 | sel = XALLOCAVEC (unsigned char, nunits); | |
5153 | ||
5154 | for (i = 0; i < nunits; ++i) | |
5155 | sel[i] = nunits - 1 - i; | |
5156 | ||
557be5a8 AL |
5157 | if (!can_vec_perm_p (TYPE_MODE (vectype), false, sel)) |
5158 | return NULL_TREE; | |
5159 | return vect_gen_perm_mask_checked (vectype, sel); | |
09dfa495 BM |
5160 | } |
5161 | ||
ebfd146a IR |
5162 | /* Function vectorizable_store. |
5163 | ||
b8698a0f L |
5164 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
5165 | can be vectorized. | |
5166 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
5167 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5168 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5169 | ||
5170 | static bool | |
355fe088 | 5171 | vectorizable_store (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 5172 | slp_tree slp_node) |
ebfd146a IR |
5173 | { |
5174 | tree scalar_dest; | |
5175 | tree data_ref; | |
5176 | tree op; | |
5177 | tree vec_oprnd = NULL_TREE; | |
5178 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5179 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
272c6793 | 5180 | tree elem_type; |
ebfd146a | 5181 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 5182 | struct loop *loop = NULL; |
ef4bddc2 | 5183 | machine_mode vec_mode; |
ebfd146a IR |
5184 | tree dummy; |
5185 | enum dr_alignment_support alignment_support_scheme; | |
355fe088 | 5186 | gimple *def_stmt; |
ebfd146a IR |
5187 | enum vect_def_type dt; |
5188 | stmt_vec_info prev_stmt_info = NULL; | |
5189 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 5190 | tree dataref_offset = NULL_TREE; |
355fe088 | 5191 | gimple *ptr_incr = NULL; |
ebfd146a IR |
5192 | int ncopies; |
5193 | int j; | |
355fe088 | 5194 | gimple *next_stmt, *first_stmt = NULL; |
0d0293ac | 5195 | bool grouped_store = false; |
272c6793 | 5196 | bool store_lanes_p = false; |
ebfd146a | 5197 | unsigned int group_size, i; |
6e1aa848 DN |
5198 | vec<tree> dr_chain = vNULL; |
5199 | vec<tree> oprnds = vNULL; | |
5200 | vec<tree> result_chain = vNULL; | |
ebfd146a | 5201 | bool inv_p; |
09dfa495 BM |
5202 | bool negative = false; |
5203 | tree offset = NULL_TREE; | |
6e1aa848 | 5204 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 5205 | bool slp = (slp_node != NULL); |
ebfd146a | 5206 | unsigned int vec_num; |
a70d6342 | 5207 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5208 | vec_info *vinfo = stmt_info->vinfo; |
272c6793 | 5209 | tree aggr_type; |
3bab6342 AT |
5210 | tree scatter_base = NULL_TREE, scatter_off = NULL_TREE; |
5211 | tree scatter_off_vectype = NULL_TREE, scatter_decl = NULL_TREE; | |
5212 | int scatter_scale = 1; | |
5213 | enum vect_def_type scatter_idx_dt = vect_unknown_def_type; | |
5214 | enum vect_def_type scatter_src_dt = vect_unknown_def_type; | |
355fe088 | 5215 | gimple *new_stmt; |
b17dc4d4 | 5216 | int vf; |
a70d6342 | 5217 | |
a70d6342 | 5218 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5219 | return false; |
5220 | ||
66c16fd9 RB |
5221 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5222 | && ! vec_stmt) | |
ebfd146a IR |
5223 | return false; |
5224 | ||
5225 | /* Is vectorizable store? */ | |
5226 | ||
5227 | if (!is_gimple_assign (stmt)) | |
5228 | return false; | |
5229 | ||
5230 | scalar_dest = gimple_assign_lhs (stmt); | |
ab0ef706 JJ |
5231 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR |
5232 | && is_pattern_stmt_p (stmt_info)) | |
5233 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
ebfd146a | 5234 | if (TREE_CODE (scalar_dest) != ARRAY_REF |
38000232 | 5235 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF |
ebfd146a | 5236 | && TREE_CODE (scalar_dest) != INDIRECT_REF |
e9dbe7bb IR |
5237 | && TREE_CODE (scalar_dest) != COMPONENT_REF |
5238 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
70f34814 RG |
5239 | && TREE_CODE (scalar_dest) != REALPART_EXPR |
5240 | && TREE_CODE (scalar_dest) != MEM_REF) | |
ebfd146a IR |
5241 | return false; |
5242 | ||
fce57248 RS |
5243 | /* Cannot have hybrid store SLP -- that would mean storing to the |
5244 | same location twice. */ | |
5245 | gcc_assert (slp == PURE_SLP_STMT (stmt_info)); | |
5246 | ||
ebfd146a | 5247 | gcc_assert (gimple_assign_single_p (stmt)); |
465c8c19 | 5248 | |
f4d09712 | 5249 | tree vectype = STMT_VINFO_VECTYPE (stmt_info), rhs_vectype = NULL_TREE; |
465c8c19 JJ |
5250 | unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype); |
5251 | ||
5252 | if (loop_vinfo) | |
b17dc4d4 RB |
5253 | { |
5254 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
5255 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
5256 | } | |
5257 | else | |
5258 | vf = 1; | |
465c8c19 JJ |
5259 | |
5260 | /* Multiple types in SLP are handled by creating the appropriate number of | |
5261 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5262 | case of SLP. */ | |
fce57248 | 5263 | if (slp) |
465c8c19 JJ |
5264 | ncopies = 1; |
5265 | else | |
5266 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
5267 | ||
5268 | gcc_assert (ncopies >= 1); | |
5269 | ||
5270 | /* FORNOW. This restriction should be relaxed. */ | |
5271 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) | |
5272 | { | |
5273 | if (dump_enabled_p ()) | |
5274 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5275 | "multiple types in nested loop.\n"); | |
5276 | return false; | |
5277 | } | |
5278 | ||
ebfd146a | 5279 | op = gimple_assign_rhs1 (stmt); |
f4d09712 KY |
5280 | |
5281 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt, &rhs_vectype)) | |
ebfd146a | 5282 | { |
73fbfcad | 5283 | if (dump_enabled_p ()) |
78c60e3d | 5284 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5285 | "use not simple.\n"); |
ebfd146a IR |
5286 | return false; |
5287 | } | |
5288 | ||
f4d09712 KY |
5289 | if (rhs_vectype && !useless_type_conversion_p (vectype, rhs_vectype)) |
5290 | return false; | |
5291 | ||
272c6793 | 5292 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 5293 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 5294 | |
ebfd146a IR |
5295 | /* FORNOW. In some cases can vectorize even if data-type not supported |
5296 | (e.g. - array initialization with 0). */ | |
947131ba | 5297 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) |
ebfd146a IR |
5298 | return false; |
5299 | ||
5300 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
5301 | return false; | |
5302 | ||
f2e2a985 | 5303 | if (!STMT_VINFO_STRIDED_P (stmt_info)) |
09dfa495 | 5304 | { |
f2e2a985 MM |
5305 | negative = |
5306 | tree_int_cst_compare (loop && nested_in_vect_loop_p (loop, stmt) | |
5307 | ? STMT_VINFO_DR_STEP (stmt_info) : DR_STEP (dr), | |
5308 | size_zero_node) < 0; | |
5309 | if (negative && ncopies > 1) | |
09dfa495 BM |
5310 | { |
5311 | if (dump_enabled_p ()) | |
5312 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
f2e2a985 | 5313 | "multiple types with negative step.\n"); |
09dfa495 BM |
5314 | return false; |
5315 | } | |
f2e2a985 | 5316 | if (negative) |
09dfa495 | 5317 | { |
f2e2a985 MM |
5318 | gcc_assert (!grouped_store); |
5319 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
5320 | if (alignment_support_scheme != dr_aligned | |
5321 | && alignment_support_scheme != dr_unaligned_supported) | |
5322 | { | |
5323 | if (dump_enabled_p ()) | |
5324 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5325 | "negative step but alignment required.\n"); | |
5326 | return false; | |
5327 | } | |
5328 | if (dt != vect_constant_def | |
5329 | && dt != vect_external_def | |
5330 | && !perm_mask_for_reverse (vectype)) | |
5331 | { | |
5332 | if (dump_enabled_p ()) | |
5333 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5334 | "negative step and reversing not supported.\n"); | |
5335 | return false; | |
5336 | } | |
09dfa495 BM |
5337 | } |
5338 | } | |
5339 | ||
0d0293ac | 5340 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 5341 | { |
0d0293ac | 5342 | grouped_store = true; |
e14c1050 | 5343 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
cee62fee | 5344 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
fce57248 | 5345 | if (!slp && !STMT_VINFO_STRIDED_P (stmt_info)) |
b602d918 | 5346 | { |
272c6793 RS |
5347 | if (vect_store_lanes_supported (vectype, group_size)) |
5348 | store_lanes_p = true; | |
0d0293ac | 5349 | else if (!vect_grouped_store_supported (vectype, group_size)) |
b602d918 RS |
5350 | return false; |
5351 | } | |
b8698a0f | 5352 | |
ebfd146a IR |
5353 | if (first_stmt == stmt) |
5354 | { | |
5355 | /* STMT is the leader of the group. Check the operands of all the | |
5356 | stmts of the group. */ | |
e14c1050 | 5357 | next_stmt = GROUP_NEXT_ELEMENT (stmt_info); |
ebfd146a IR |
5358 | while (next_stmt) |
5359 | { | |
5360 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
5361 | op = gimple_assign_rhs1 (next_stmt); | |
81c40241 | 5362 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt)) |
ebfd146a | 5363 | { |
73fbfcad | 5364 | if (dump_enabled_p ()) |
78c60e3d | 5365 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5366 | "use not simple.\n"); |
ebfd146a IR |
5367 | return false; |
5368 | } | |
e14c1050 | 5369 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
5370 | } |
5371 | } | |
5372 | } | |
5373 | ||
3bab6342 AT |
5374 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5375 | { | |
355fe088 | 5376 | gimple *def_stmt; |
3bab6342 AT |
5377 | scatter_decl = vect_check_gather_scatter (stmt, loop_vinfo, &scatter_base, |
5378 | &scatter_off, &scatter_scale); | |
5379 | gcc_assert (scatter_decl); | |
81c40241 RB |
5380 | if (!vect_is_simple_use (scatter_off, vinfo, &def_stmt, &scatter_idx_dt, |
5381 | &scatter_off_vectype)) | |
3bab6342 AT |
5382 | { |
5383 | if (dump_enabled_p ()) | |
5384 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5385 | "scatter index use not simple."); | |
5386 | return false; | |
5387 | } | |
5388 | } | |
5389 | ||
ebfd146a IR |
5390 | if (!vec_stmt) /* transformation not required. */ |
5391 | { | |
5392 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; | |
2e8ab70c RB |
5393 | /* The SLP costs are calculated during SLP analysis. */ |
5394 | if (!PURE_SLP_STMT (stmt_info)) | |
5395 | vect_model_store_cost (stmt_info, ncopies, store_lanes_p, dt, | |
5396 | NULL, NULL, NULL); | |
ebfd146a IR |
5397 | return true; |
5398 | } | |
5399 | ||
5400 | /** Transform. **/ | |
5401 | ||
c716e67f XDL |
5402 | ensure_base_align (stmt_info, dr); |
5403 | ||
3bab6342 AT |
5404 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
5405 | { | |
5406 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE, op, src; | |
5407 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (scatter_decl)); | |
5408 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
5409 | tree ptr, mask, var, scale, perm_mask = NULL_TREE; | |
5410 | edge pe = loop_preheader_edge (loop); | |
5411 | gimple_seq seq; | |
5412 | basic_block new_bb; | |
5413 | enum { NARROW, NONE, WIDEN } modifier; | |
5414 | int scatter_off_nunits = TYPE_VECTOR_SUBPARTS (scatter_off_vectype); | |
5415 | ||
5416 | if (nunits == (unsigned int) scatter_off_nunits) | |
5417 | modifier = NONE; | |
5418 | else if (nunits == (unsigned int) scatter_off_nunits / 2) | |
5419 | { | |
5420 | unsigned char *sel = XALLOCAVEC (unsigned char, scatter_off_nunits); | |
5421 | modifier = WIDEN; | |
5422 | ||
5423 | for (i = 0; i < (unsigned int) scatter_off_nunits; ++i) | |
5424 | sel[i] = i | nunits; | |
5425 | ||
5426 | perm_mask = vect_gen_perm_mask_checked (scatter_off_vectype, sel); | |
5427 | gcc_assert (perm_mask != NULL_TREE); | |
5428 | } | |
5429 | else if (nunits == (unsigned int) scatter_off_nunits * 2) | |
5430 | { | |
5431 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
5432 | modifier = NARROW; | |
5433 | ||
5434 | for (i = 0; i < (unsigned int) nunits; ++i) | |
5435 | sel[i] = i | scatter_off_nunits; | |
5436 | ||
5437 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); | |
5438 | gcc_assert (perm_mask != NULL_TREE); | |
5439 | ncopies *= 2; | |
5440 | } | |
5441 | else | |
5442 | gcc_unreachable (); | |
5443 | ||
5444 | rettype = TREE_TYPE (TREE_TYPE (scatter_decl)); | |
5445 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5446 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5447 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5448 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
5449 | scaletype = TREE_VALUE (arglist); | |
5450 | ||
5451 | gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE | |
5452 | && TREE_CODE (rettype) == VOID_TYPE); | |
5453 | ||
5454 | ptr = fold_convert (ptrtype, scatter_base); | |
5455 | if (!is_gimple_min_invariant (ptr)) | |
5456 | { | |
5457 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
5458 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
5459 | gcc_assert (!new_bb); | |
5460 | } | |
5461 | ||
5462 | /* Currently we support only unconditional scatter stores, | |
5463 | so mask should be all ones. */ | |
5464 | mask = build_int_cst (masktype, -1); | |
5465 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
5466 | ||
5467 | scale = build_int_cst (scaletype, scatter_scale); | |
5468 | ||
5469 | prev_stmt_info = NULL; | |
5470 | for (j = 0; j < ncopies; ++j) | |
5471 | { | |
5472 | if (j == 0) | |
5473 | { | |
5474 | src = vec_oprnd1 | |
81c40241 | 5475 | = vect_get_vec_def_for_operand (gimple_assign_rhs1 (stmt), stmt); |
3bab6342 | 5476 | op = vec_oprnd0 |
81c40241 | 5477 | = vect_get_vec_def_for_operand (scatter_off, stmt); |
3bab6342 AT |
5478 | } |
5479 | else if (modifier != NONE && (j & 1)) | |
5480 | { | |
5481 | if (modifier == WIDEN) | |
5482 | { | |
5483 | src = vec_oprnd1 | |
5484 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5485 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, perm_mask, | |
5486 | stmt, gsi); | |
5487 | } | |
5488 | else if (modifier == NARROW) | |
5489 | { | |
5490 | src = permute_vec_elements (vec_oprnd1, vec_oprnd1, perm_mask, | |
5491 | stmt, gsi); | |
5492 | op = vec_oprnd0 | |
5493 | = vect_get_vec_def_for_stmt_copy (scatter_idx_dt, vec_oprnd0); | |
5494 | } | |
5495 | else | |
5496 | gcc_unreachable (); | |
5497 | } | |
5498 | else | |
5499 | { | |
5500 | src = vec_oprnd1 | |
5501 | = vect_get_vec_def_for_stmt_copy (scatter_src_dt, vec_oprnd1); | |
5502 | op = vec_oprnd0 | |
5503 | = vect_get_vec_def_for_stmt_copy (scatter_idx_dt, vec_oprnd0); | |
5504 | } | |
5505 | ||
5506 | if (!useless_type_conversion_p (srctype, TREE_TYPE (src))) | |
5507 | { | |
5508 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)) | |
5509 | == TYPE_VECTOR_SUBPARTS (srctype)); | |
0e22bb5a | 5510 | var = vect_get_new_ssa_name (srctype, vect_simple_var); |
3bab6342 AT |
5511 | src = build1 (VIEW_CONVERT_EXPR, srctype, src); |
5512 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, src); | |
5513 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5514 | src = var; | |
5515 | } | |
5516 | ||
5517 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
5518 | { | |
5519 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
5520 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 5521 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
3bab6342 AT |
5522 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
5523 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
5524 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5525 | op = var; | |
5526 | } | |
5527 | ||
5528 | new_stmt | |
5529 | = gimple_build_call (scatter_decl, 5, ptr, mask, op, src, scale); | |
5530 | ||
5531 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5532 | ||
5533 | if (prev_stmt_info == NULL) | |
5534 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5535 | else | |
5536 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5537 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5538 | } | |
5539 | return true; | |
5540 | } | |
5541 | ||
0d0293ac | 5542 | if (grouped_store) |
ebfd146a IR |
5543 | { |
5544 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 5545 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
ebfd146a | 5546 | |
e14c1050 | 5547 | GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; |
ebfd146a IR |
5548 | |
5549 | /* FORNOW */ | |
a70d6342 | 5550 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
5551 | |
5552 | /* We vectorize all the stmts of the interleaving group when we | |
5553 | reach the last stmt in the group. */ | |
e14c1050 IR |
5554 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
5555 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
5556 | && !slp) |
5557 | { | |
5558 | *vec_stmt = NULL; | |
5559 | return true; | |
5560 | } | |
5561 | ||
5562 | if (slp) | |
4b5caab7 | 5563 | { |
0d0293ac | 5564 | grouped_store = false; |
4b5caab7 IR |
5565 | /* VEC_NUM is the number of vect stmts to be created for this |
5566 | group. */ | |
5567 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 5568 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
52eab378 | 5569 | gcc_assert (GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_stmt)) == first_stmt); |
4b5caab7 | 5570 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
d092494c | 5571 | op = gimple_assign_rhs1 (first_stmt); |
4b5caab7 | 5572 | } |
ebfd146a | 5573 | else |
4b5caab7 IR |
5574 | /* VEC_NUM is the number of vect stmts to be created for this |
5575 | group. */ | |
ebfd146a IR |
5576 | vec_num = group_size; |
5577 | } | |
b8698a0f | 5578 | else |
ebfd146a IR |
5579 | { |
5580 | first_stmt = stmt; | |
5581 | first_dr = dr; | |
5582 | group_size = vec_num = 1; | |
ebfd146a | 5583 | } |
b8698a0f | 5584 | |
73fbfcad | 5585 | if (dump_enabled_p ()) |
78c60e3d | 5586 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5587 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 5588 | |
f2e2a985 MM |
5589 | if (STMT_VINFO_STRIDED_P (stmt_info)) |
5590 | { | |
5591 | gimple_stmt_iterator incr_gsi; | |
5592 | bool insert_after; | |
355fe088 | 5593 | gimple *incr; |
f2e2a985 MM |
5594 | tree offvar; |
5595 | tree ivstep; | |
5596 | tree running_off; | |
5597 | gimple_seq stmts = NULL; | |
5598 | tree stride_base, stride_step, alias_off; | |
5599 | tree vec_oprnd; | |
f502d50e | 5600 | unsigned int g; |
f2e2a985 MM |
5601 | |
5602 | gcc_assert (!nested_in_vect_loop_p (loop, stmt)); | |
5603 | ||
5604 | stride_base | |
5605 | = fold_build_pointer_plus | |
f502d50e | 5606 | (unshare_expr (DR_BASE_ADDRESS (first_dr)), |
f2e2a985 | 5607 | size_binop (PLUS_EXPR, |
f502d50e MM |
5608 | convert_to_ptrofftype (unshare_expr (DR_OFFSET (first_dr))), |
5609 | convert_to_ptrofftype (DR_INIT(first_dr)))); | |
5610 | stride_step = fold_convert (sizetype, unshare_expr (DR_STEP (first_dr))); | |
f2e2a985 MM |
5611 | |
5612 | /* For a store with loop-invariant (but other than power-of-2) | |
5613 | stride (i.e. not a grouped access) like so: | |
5614 | ||
5615 | for (i = 0; i < n; i += stride) | |
5616 | array[i] = ...; | |
5617 | ||
5618 | we generate a new induction variable and new stores from | |
5619 | the components of the (vectorized) rhs: | |
5620 | ||
5621 | for (j = 0; ; j += VF*stride) | |
5622 | vectemp = ...; | |
5623 | tmp1 = vectemp[0]; | |
5624 | array[j] = tmp1; | |
5625 | tmp2 = vectemp[1]; | |
5626 | array[j + stride] = tmp2; | |
5627 | ... | |
5628 | */ | |
5629 | ||
cee62fee | 5630 | unsigned nstores = nunits; |
b17dc4d4 | 5631 | unsigned lnel = 1; |
cee62fee MM |
5632 | tree ltype = elem_type; |
5633 | if (slp) | |
5634 | { | |
b17dc4d4 RB |
5635 | if (group_size < nunits |
5636 | && nunits % group_size == 0) | |
5637 | { | |
5638 | nstores = nunits / group_size; | |
5639 | lnel = group_size; | |
5640 | ltype = build_vector_type (elem_type, group_size); | |
5641 | } | |
5642 | else if (group_size >= nunits | |
5643 | && group_size % nunits == 0) | |
5644 | { | |
5645 | nstores = 1; | |
5646 | lnel = nunits; | |
5647 | ltype = vectype; | |
5648 | } | |
cee62fee MM |
5649 | ltype = build_aligned_type (ltype, TYPE_ALIGN (elem_type)); |
5650 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
5651 | } | |
5652 | ||
f2e2a985 MM |
5653 | ivstep = stride_step; |
5654 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
b17dc4d4 | 5655 | build_int_cst (TREE_TYPE (ivstep), vf)); |
f2e2a985 MM |
5656 | |
5657 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
5658 | ||
5659 | create_iv (stride_base, ivstep, NULL, | |
5660 | loop, &incr_gsi, insert_after, | |
5661 | &offvar, NULL); | |
5662 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 5663 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
f2e2a985 MM |
5664 | |
5665 | stride_step = force_gimple_operand (stride_step, &stmts, true, NULL_TREE); | |
5666 | if (stmts) | |
5667 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
5668 | ||
5669 | prev_stmt_info = NULL; | |
f502d50e MM |
5670 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (first_dr)), 0); |
5671 | next_stmt = first_stmt; | |
5672 | for (g = 0; g < group_size; g++) | |
f2e2a985 | 5673 | { |
f502d50e MM |
5674 | running_off = offvar; |
5675 | if (g) | |
f2e2a985 | 5676 | { |
f502d50e MM |
5677 | tree size = TYPE_SIZE_UNIT (ltype); |
5678 | tree pos = fold_build2 (MULT_EXPR, sizetype, size_int (g), | |
f2e2a985 | 5679 | size); |
f502d50e | 5680 | tree newoff = copy_ssa_name (running_off, NULL); |
f2e2a985 | 5681 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
f502d50e | 5682 | running_off, pos); |
f2e2a985 | 5683 | vect_finish_stmt_generation (stmt, incr, gsi); |
f2e2a985 | 5684 | running_off = newoff; |
f502d50e | 5685 | } |
b17dc4d4 RB |
5686 | unsigned int group_el = 0; |
5687 | unsigned HOST_WIDE_INT | |
5688 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
f502d50e MM |
5689 | for (j = 0; j < ncopies; j++) |
5690 | { | |
5691 | /* We've set op and dt above, from gimple_assign_rhs1(stmt), | |
5692 | and first_stmt == stmt. */ | |
5693 | if (j == 0) | |
5694 | { | |
5695 | if (slp) | |
5696 | { | |
5697 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, NULL, | |
5698 | slp_node, -1); | |
5699 | vec_oprnd = vec_oprnds[0]; | |
5700 | } | |
5701 | else | |
5702 | { | |
5703 | gcc_assert (gimple_assign_single_p (next_stmt)); | |
5704 | op = gimple_assign_rhs1 (next_stmt); | |
81c40241 | 5705 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
f502d50e MM |
5706 | } |
5707 | } | |
f2e2a985 | 5708 | else |
f502d50e MM |
5709 | { |
5710 | if (slp) | |
5711 | vec_oprnd = vec_oprnds[j]; | |
5712 | else | |
c079cbac | 5713 | { |
81c40241 | 5714 | vect_is_simple_use (vec_oprnd, vinfo, &def_stmt, &dt); |
c079cbac RB |
5715 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); |
5716 | } | |
f502d50e MM |
5717 | } |
5718 | ||
5719 | for (i = 0; i < nstores; i++) | |
5720 | { | |
5721 | tree newref, newoff; | |
355fe088 | 5722 | gimple *incr, *assign; |
f502d50e MM |
5723 | tree size = TYPE_SIZE (ltype); |
5724 | /* Extract the i'th component. */ | |
5725 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, | |
5726 | bitsize_int (i), size); | |
5727 | tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd, | |
5728 | size, pos); | |
5729 | ||
5730 | elem = force_gimple_operand_gsi (gsi, elem, true, | |
5731 | NULL_TREE, true, | |
5732 | GSI_SAME_STMT); | |
5733 | ||
b17dc4d4 RB |
5734 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
5735 | group_el * elsz); | |
f502d50e | 5736 | newref = build2 (MEM_REF, ltype, |
b17dc4d4 | 5737 | running_off, this_off); |
f502d50e MM |
5738 | |
5739 | /* And store it to *running_off. */ | |
5740 | assign = gimple_build_assign (newref, elem); | |
5741 | vect_finish_stmt_generation (stmt, assign, gsi); | |
5742 | ||
b17dc4d4 RB |
5743 | group_el += lnel; |
5744 | if (! slp | |
5745 | || group_el == group_size) | |
5746 | { | |
5747 | newoff = copy_ssa_name (running_off, NULL); | |
5748 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
5749 | running_off, stride_step); | |
5750 | vect_finish_stmt_generation (stmt, incr, gsi); | |
f502d50e | 5751 | |
b17dc4d4 RB |
5752 | running_off = newoff; |
5753 | group_el = 0; | |
5754 | } | |
225ce44b RB |
5755 | if (g == group_size - 1 |
5756 | && !slp) | |
f502d50e MM |
5757 | { |
5758 | if (j == 0 && i == 0) | |
225ce44b RB |
5759 | STMT_VINFO_VEC_STMT (stmt_info) |
5760 | = *vec_stmt = assign; | |
f502d50e MM |
5761 | else |
5762 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = assign; | |
5763 | prev_stmt_info = vinfo_for_stmt (assign); | |
5764 | } | |
5765 | } | |
f2e2a985 | 5766 | } |
f502d50e | 5767 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
b17dc4d4 RB |
5768 | if (slp) |
5769 | break; | |
f2e2a985 MM |
5770 | } |
5771 | return true; | |
5772 | } | |
5773 | ||
9771b263 DN |
5774 | dr_chain.create (group_size); |
5775 | oprnds.create (group_size); | |
ebfd146a | 5776 | |
720f5239 | 5777 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 5778 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
5779 | /* Targets with store-lane instructions must not require explicit |
5780 | realignment. */ | |
5781 | gcc_assert (!store_lanes_p | |
5782 | || alignment_support_scheme == dr_aligned | |
5783 | || alignment_support_scheme == dr_unaligned_supported); | |
5784 | ||
09dfa495 BM |
5785 | if (negative) |
5786 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
5787 | ||
272c6793 RS |
5788 | if (store_lanes_p) |
5789 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
5790 | else | |
5791 | aggr_type = vectype; | |
ebfd146a IR |
5792 | |
5793 | /* In case the vectorization factor (VF) is bigger than the number | |
5794 | of elements that we can fit in a vectype (nunits), we have to generate | |
5795 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 5796 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
5797 | vect_get_vec_def_for_copy_stmt. */ |
5798 | ||
0d0293ac | 5799 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
5800 | |
5801 | S1: &base + 2 = x2 | |
5802 | S2: &base = x0 | |
5803 | S3: &base + 1 = x1 | |
5804 | S4: &base + 3 = x3 | |
5805 | ||
5806 | We create vectorized stores starting from base address (the access of the | |
5807 | first stmt in the chain (S2 in the above example), when the last store stmt | |
5808 | of the chain (S4) is reached: | |
5809 | ||
5810 | VS1: &base = vx2 | |
5811 | VS2: &base + vec_size*1 = vx0 | |
5812 | VS3: &base + vec_size*2 = vx1 | |
5813 | VS4: &base + vec_size*3 = vx3 | |
5814 | ||
5815 | Then permutation statements are generated: | |
5816 | ||
3fcc1b55 JJ |
5817 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
5818 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 5819 | ... |
b8698a0f | 5820 | |
ebfd146a IR |
5821 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
5822 | (the order of the data-refs in the output of vect_permute_store_chain | |
5823 | corresponds to the order of scalar stmts in the interleaving chain - see | |
5824 | the documentation of vect_permute_store_chain()). | |
5825 | ||
5826 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 5827 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 5828 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 5829 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
5830 | */ |
5831 | ||
5832 | prev_stmt_info = NULL; | |
5833 | for (j = 0; j < ncopies; j++) | |
5834 | { | |
ebfd146a IR |
5835 | |
5836 | if (j == 0) | |
5837 | { | |
5838 | if (slp) | |
5839 | { | |
5840 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c IR |
5841 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
5842 | NULL, slp_node, -1); | |
ebfd146a | 5843 | |
9771b263 | 5844 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
5845 | } |
5846 | else | |
5847 | { | |
b8698a0f L |
5848 | /* For interleaved stores we collect vectorized defs for all the |
5849 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
5850 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
5851 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
5852 | ||
0d0293ac | 5853 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 5854 | OPRNDS are of size 1. */ |
b8698a0f | 5855 | next_stmt = first_stmt; |
ebfd146a IR |
5856 | for (i = 0; i < group_size; i++) |
5857 | { | |
b8698a0f L |
5858 | /* Since gaps are not supported for interleaved stores, |
5859 | GROUP_SIZE is the exact number of stmts in the chain. | |
5860 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
5861 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a IR |
5862 | iteration of the loop will be executed. */ |
5863 | gcc_assert (next_stmt | |
5864 | && gimple_assign_single_p (next_stmt)); | |
5865 | op = gimple_assign_rhs1 (next_stmt); | |
5866 | ||
81c40241 | 5867 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
9771b263 DN |
5868 | dr_chain.quick_push (vec_oprnd); |
5869 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 5870 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a IR |
5871 | } |
5872 | } | |
5873 | ||
5874 | /* We should have catched mismatched types earlier. */ | |
5875 | gcc_assert (useless_type_conversion_p (vectype, | |
5876 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
5877 | bool simd_lane_access_p |
5878 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
5879 | if (simd_lane_access_p | |
5880 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
5881 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
5882 | && integer_zerop (DR_OFFSET (first_dr)) | |
5883 | && integer_zerop (DR_INIT (first_dr)) | |
5884 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
5885 | get_alias_set (DR_REF (first_dr)))) | |
5886 | { | |
5887 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
5888 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
5889 | (DR_REF (first_dr)), 0); | |
8928eff3 | 5890 | inv_p = false; |
74bf76ed JJ |
5891 | } |
5892 | else | |
5893 | dataref_ptr | |
5894 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
5895 | simd_lane_access_p ? loop : NULL, | |
09dfa495 | 5896 | offset, &dummy, gsi, &ptr_incr, |
74bf76ed | 5897 | simd_lane_access_p, &inv_p); |
a70d6342 | 5898 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 5899 | } |
b8698a0f | 5900 | else |
ebfd146a | 5901 | { |
b8698a0f L |
5902 | /* For interleaved stores we created vectorized defs for all the |
5903 | defs stored in OPRNDS in the previous iteration (previous copy). | |
5904 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
5905 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
5906 | next copy. | |
0d0293ac | 5907 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
5908 | OPRNDS are of size 1. */ |
5909 | for (i = 0; i < group_size; i++) | |
5910 | { | |
9771b263 | 5911 | op = oprnds[i]; |
81c40241 | 5912 | vect_is_simple_use (op, vinfo, &def_stmt, &dt); |
b8698a0f | 5913 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op); |
9771b263 DN |
5914 | dr_chain[i] = vec_oprnd; |
5915 | oprnds[i] = vec_oprnd; | |
ebfd146a | 5916 | } |
74bf76ed JJ |
5917 | if (dataref_offset) |
5918 | dataref_offset | |
5919 | = int_const_binop (PLUS_EXPR, dataref_offset, | |
5920 | TYPE_SIZE_UNIT (aggr_type)); | |
5921 | else | |
5922 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
5923 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a IR |
5924 | } |
5925 | ||
272c6793 | 5926 | if (store_lanes_p) |
ebfd146a | 5927 | { |
272c6793 | 5928 | tree vec_array; |
267d3070 | 5929 | |
272c6793 RS |
5930 | /* Combine all the vectors into an array. */ |
5931 | vec_array = create_vector_array (vectype, vec_num); | |
5932 | for (i = 0; i < vec_num; i++) | |
c2d7ab2a | 5933 | { |
9771b263 | 5934 | vec_oprnd = dr_chain[i]; |
272c6793 | 5935 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 5936 | } |
b8698a0f | 5937 | |
272c6793 RS |
5938 | /* Emit: |
5939 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
5940 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
5941 | new_stmt = gimple_build_call_internal (IFN_STORE_LANES, 1, vec_array); | |
5942 | gimple_call_set_lhs (new_stmt, data_ref); | |
267d3070 | 5943 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
5944 | } |
5945 | else | |
5946 | { | |
5947 | new_stmt = NULL; | |
0d0293ac | 5948 | if (grouped_store) |
272c6793 | 5949 | { |
b6b9227d JJ |
5950 | if (j == 0) |
5951 | result_chain.create (group_size); | |
272c6793 RS |
5952 | /* Permute. */ |
5953 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
5954 | &result_chain); | |
5955 | } | |
c2d7ab2a | 5956 | |
272c6793 RS |
5957 | next_stmt = first_stmt; |
5958 | for (i = 0; i < vec_num; i++) | |
5959 | { | |
644ffefd | 5960 | unsigned align, misalign; |
272c6793 RS |
5961 | |
5962 | if (i > 0) | |
5963 | /* Bump the vector pointer. */ | |
5964 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
5965 | stmt, NULL_TREE); | |
5966 | ||
5967 | if (slp) | |
9771b263 | 5968 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
5969 | else if (grouped_store) |
5970 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 5971 | vect_permute_store_chain(). */ |
9771b263 | 5972 | vec_oprnd = result_chain[i]; |
272c6793 | 5973 | |
aed93b23 RB |
5974 | data_ref = fold_build2 (MEM_REF, TREE_TYPE (vec_oprnd), |
5975 | dataref_ptr, | |
5976 | dataref_offset | |
5977 | ? dataref_offset | |
5978 | : build_int_cst (reference_alias_ptr_type | |
5979 | (DR_REF (first_dr)), 0)); | |
644ffefd | 5980 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 | 5981 | if (aligned_access_p (first_dr)) |
644ffefd | 5982 | misalign = 0; |
272c6793 RS |
5983 | else if (DR_MISALIGNMENT (first_dr) == -1) |
5984 | { | |
52639a61 RB |
5985 | if (DR_VECT_AUX (first_dr)->base_element_aligned) |
5986 | align = TYPE_ALIGN_UNIT (elem_type); | |
5987 | else | |
5988 | align = get_object_alignment (DR_REF (first_dr)) | |
5989 | / BITS_PER_UNIT; | |
5990 | misalign = 0; | |
272c6793 RS |
5991 | TREE_TYPE (data_ref) |
5992 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 5993 | align * BITS_PER_UNIT); |
272c6793 RS |
5994 | } |
5995 | else | |
5996 | { | |
5997 | TREE_TYPE (data_ref) | |
5998 | = build_aligned_type (TREE_TYPE (data_ref), | |
5999 | TYPE_ALIGN (elem_type)); | |
644ffefd | 6000 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 6001 | } |
aed93b23 RB |
6002 | if (dataref_offset == NULL_TREE |
6003 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
6004 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
6005 | misalign); | |
c2d7ab2a | 6006 | |
f234d260 BM |
6007 | if (negative |
6008 | && dt != vect_constant_def | |
6009 | && dt != vect_external_def) | |
09dfa495 BM |
6010 | { |
6011 | tree perm_mask = perm_mask_for_reverse (vectype); | |
6012 | tree perm_dest | |
6013 | = vect_create_destination_var (gimple_assign_rhs1 (stmt), | |
6014 | vectype); | |
b731b390 | 6015 | tree new_temp = make_ssa_name (perm_dest); |
09dfa495 BM |
6016 | |
6017 | /* Generate the permute statement. */ | |
355fe088 | 6018 | gimple *perm_stmt |
0d0e4a03 JJ |
6019 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
6020 | vec_oprnd, perm_mask); | |
09dfa495 BM |
6021 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
6022 | ||
6023 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); | |
6024 | vec_oprnd = new_temp; | |
6025 | } | |
6026 | ||
272c6793 RS |
6027 | /* Arguments are ready. Create the new vector stmt. */ |
6028 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); | |
6029 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
6030 | |
6031 | if (slp) | |
6032 | continue; | |
6033 | ||
e14c1050 | 6034 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
6035 | if (!next_stmt) |
6036 | break; | |
6037 | } | |
ebfd146a | 6038 | } |
1da0876c RS |
6039 | if (!slp) |
6040 | { | |
6041 | if (j == 0) | |
6042 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6043 | else | |
6044 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6045 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6046 | } | |
ebfd146a IR |
6047 | } |
6048 | ||
9771b263 DN |
6049 | dr_chain.release (); |
6050 | oprnds.release (); | |
6051 | result_chain.release (); | |
6052 | vec_oprnds.release (); | |
ebfd146a IR |
6053 | |
6054 | return true; | |
6055 | } | |
6056 | ||
557be5a8 AL |
6057 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
6058 | VECTOR_CST mask. No checks are made that the target platform supports the | |
6059 | mask, so callers may wish to test can_vec_perm_p separately, or use | |
6060 | vect_gen_perm_mask_checked. */ | |
a1e53f3f | 6061 | |
3fcc1b55 | 6062 | tree |
557be5a8 | 6063 | vect_gen_perm_mask_any (tree vectype, const unsigned char *sel) |
a1e53f3f | 6064 | { |
d2a12ae7 | 6065 | tree mask_elt_type, mask_type, mask_vec, *mask_elts; |
2635892a | 6066 | int i, nunits; |
a1e53f3f | 6067 | |
22e4dee7 | 6068 | nunits = TYPE_VECTOR_SUBPARTS (vectype); |
22e4dee7 | 6069 | |
96f9265a RG |
6070 | mask_elt_type = lang_hooks.types.type_for_mode |
6071 | (int_mode_for_mode (TYPE_MODE (TREE_TYPE (vectype))), 1); | |
22e4dee7 | 6072 | mask_type = get_vectype_for_scalar_type (mask_elt_type); |
a1e53f3f | 6073 | |
d2a12ae7 | 6074 | mask_elts = XALLOCAVEC (tree, nunits); |
aec7ae7d | 6075 | for (i = nunits - 1; i >= 0; i--) |
d2a12ae7 RG |
6076 | mask_elts[i] = build_int_cst (mask_elt_type, sel[i]); |
6077 | mask_vec = build_vector (mask_type, mask_elts); | |
a1e53f3f | 6078 | |
2635892a | 6079 | return mask_vec; |
a1e53f3f L |
6080 | } |
6081 | ||
cf7aa6a3 AL |
6082 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_p, |
6083 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ | |
557be5a8 AL |
6084 | |
6085 | tree | |
6086 | vect_gen_perm_mask_checked (tree vectype, const unsigned char *sel) | |
6087 | { | |
6088 | gcc_assert (can_vec_perm_p (TYPE_MODE (vectype), false, sel)); | |
6089 | return vect_gen_perm_mask_any (vectype, sel); | |
6090 | } | |
6091 | ||
aec7ae7d JJ |
6092 | /* Given a vector variable X and Y, that was generated for the scalar |
6093 | STMT, generate instructions to permute the vector elements of X and Y | |
6094 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
6095 | permuted vector variable. */ | |
a1e53f3f L |
6096 | |
6097 | static tree | |
355fe088 | 6098 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple *stmt, |
aec7ae7d | 6099 | gimple_stmt_iterator *gsi) |
a1e53f3f L |
6100 | { |
6101 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 6102 | tree perm_dest, data_ref; |
355fe088 | 6103 | gimple *perm_stmt; |
a1e53f3f | 6104 | |
acdcd61b | 6105 | perm_dest = vect_create_destination_var (gimple_get_lhs (stmt), vectype); |
b731b390 | 6106 | data_ref = make_ssa_name (perm_dest); |
a1e53f3f L |
6107 | |
6108 | /* Generate the permute statement. */ | |
0d0e4a03 | 6109 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
a1e53f3f L |
6110 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
6111 | ||
6112 | return data_ref; | |
6113 | } | |
6114 | ||
6b916b36 RB |
6115 | /* Hoist the definitions of all SSA uses on STMT out of the loop LOOP, |
6116 | inserting them on the loops preheader edge. Returns true if we | |
6117 | were successful in doing so (and thus STMT can be moved then), | |
6118 | otherwise returns false. */ | |
6119 | ||
6120 | static bool | |
355fe088 | 6121 | hoist_defs_of_uses (gimple *stmt, struct loop *loop) |
6b916b36 RB |
6122 | { |
6123 | ssa_op_iter i; | |
6124 | tree op; | |
6125 | bool any = false; | |
6126 | ||
6127 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6128 | { | |
355fe088 | 6129 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6130 | if (!gimple_nop_p (def_stmt) |
6131 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6132 | { | |
6133 | /* Make sure we don't need to recurse. While we could do | |
6134 | so in simple cases when there are more complex use webs | |
6135 | we don't have an easy way to preserve stmt order to fulfil | |
6136 | dependencies within them. */ | |
6137 | tree op2; | |
6138 | ssa_op_iter i2; | |
d1417442 JJ |
6139 | if (gimple_code (def_stmt) == GIMPLE_PHI) |
6140 | return false; | |
6b916b36 RB |
6141 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
6142 | { | |
355fe088 | 6143 | gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2); |
6b916b36 RB |
6144 | if (!gimple_nop_p (def_stmt2) |
6145 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2))) | |
6146 | return false; | |
6147 | } | |
6148 | any = true; | |
6149 | } | |
6150 | } | |
6151 | ||
6152 | if (!any) | |
6153 | return true; | |
6154 | ||
6155 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
6156 | { | |
355fe088 | 6157 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
6158 | if (!gimple_nop_p (def_stmt) |
6159 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
6160 | { | |
6161 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); | |
6162 | gsi_remove (&gsi, false); | |
6163 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); | |
6164 | } | |
6165 | } | |
6166 | ||
6167 | return true; | |
6168 | } | |
6169 | ||
ebfd146a IR |
6170 | /* vectorizable_load. |
6171 | ||
b8698a0f L |
6172 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
6173 | can be vectorized. | |
6174 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
6175 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
6176 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
6177 | ||
6178 | static bool | |
355fe088 | 6179 | vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 6180 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a IR |
6181 | { |
6182 | tree scalar_dest; | |
6183 | tree vec_dest = NULL; | |
6184 | tree data_ref = NULL; | |
6185 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 6186 | stmt_vec_info prev_stmt_info; |
ebfd146a | 6187 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 6188 | struct loop *loop = NULL; |
ebfd146a | 6189 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 6190 | bool nested_in_vect_loop = false; |
c716e67f | 6191 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
272c6793 | 6192 | tree elem_type; |
ebfd146a | 6193 | tree new_temp; |
ef4bddc2 | 6194 | machine_mode mode; |
355fe088 | 6195 | gimple *new_stmt = NULL; |
ebfd146a IR |
6196 | tree dummy; |
6197 | enum dr_alignment_support alignment_support_scheme; | |
6198 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 6199 | tree dataref_offset = NULL_TREE; |
355fe088 | 6200 | gimple *ptr_incr = NULL; |
ebfd146a | 6201 | int ncopies; |
9b999e8c | 6202 | int i, j, group_size = -1, group_gap_adj; |
ebfd146a IR |
6203 | tree msq = NULL_TREE, lsq; |
6204 | tree offset = NULL_TREE; | |
356bbc4c | 6205 | tree byte_offset = NULL_TREE; |
ebfd146a | 6206 | tree realignment_token = NULL_TREE; |
538dd0b7 | 6207 | gphi *phi = NULL; |
6e1aa848 | 6208 | vec<tree> dr_chain = vNULL; |
0d0293ac | 6209 | bool grouped_load = false; |
272c6793 | 6210 | bool load_lanes_p = false; |
355fe088 | 6211 | gimple *first_stmt; |
4f0a0218 | 6212 | gimple *first_stmt_for_drptr = NULL; |
ebfd146a | 6213 | bool inv_p; |
319e6439 | 6214 | bool negative = false; |
ebfd146a IR |
6215 | bool compute_in_loop = false; |
6216 | struct loop *at_loop; | |
6217 | int vec_num; | |
6218 | bool slp = (slp_node != NULL); | |
6219 | bool slp_perm = false; | |
6220 | enum tree_code code; | |
a70d6342 IR |
6221 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
6222 | int vf; | |
272c6793 | 6223 | tree aggr_type; |
aec7ae7d JJ |
6224 | tree gather_base = NULL_TREE, gather_off = NULL_TREE; |
6225 | tree gather_off_vectype = NULL_TREE, gather_decl = NULL_TREE; | |
6226 | int gather_scale = 1; | |
6227 | enum vect_def_type gather_dt = vect_unknown_def_type; | |
310213d4 | 6228 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 6229 | |
465c8c19 JJ |
6230 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
6231 | return false; | |
6232 | ||
66c16fd9 RB |
6233 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
6234 | && ! vec_stmt) | |
465c8c19 JJ |
6235 | return false; |
6236 | ||
6237 | /* Is vectorizable load? */ | |
6238 | if (!is_gimple_assign (stmt)) | |
6239 | return false; | |
6240 | ||
6241 | scalar_dest = gimple_assign_lhs (stmt); | |
6242 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
6243 | return false; | |
6244 | ||
6245 | code = gimple_assign_rhs_code (stmt); | |
6246 | if (code != ARRAY_REF | |
6247 | && code != BIT_FIELD_REF | |
6248 | && code != INDIRECT_REF | |
6249 | && code != COMPONENT_REF | |
6250 | && code != IMAGPART_EXPR | |
6251 | && code != REALPART_EXPR | |
6252 | && code != MEM_REF | |
6253 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
6254 | return false; | |
6255 | ||
6256 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
6257 | return false; | |
6258 | ||
6259 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
6260 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
6261 | ||
a70d6342 IR |
6262 | if (loop_vinfo) |
6263 | { | |
6264 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
6265 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
6266 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
6267 | } | |
6268 | else | |
3533e503 | 6269 | vf = 1; |
ebfd146a IR |
6270 | |
6271 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 6272 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 6273 | case of SLP. */ |
fce57248 | 6274 | if (slp) |
ebfd146a IR |
6275 | ncopies = 1; |
6276 | else | |
6277 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
6278 | ||
6279 | gcc_assert (ncopies >= 1); | |
6280 | ||
6281 | /* FORNOW. This restriction should be relaxed. */ | |
6282 | if (nested_in_vect_loop && ncopies > 1) | |
6283 | { | |
73fbfcad | 6284 | if (dump_enabled_p ()) |
78c60e3d | 6285 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6286 | "multiple types in nested loop.\n"); |
ebfd146a IR |
6287 | return false; |
6288 | } | |
6289 | ||
f2556b68 RB |
6290 | /* Invalidate assumptions made by dependence analysis when vectorization |
6291 | on the unrolled body effectively re-orders stmts. */ | |
6292 | if (ncopies > 1 | |
6293 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
6294 | && ((unsigned)LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
6295 | > STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
6296 | { | |
6297 | if (dump_enabled_p ()) | |
6298 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6299 | "cannot perform implicit CSE when unrolling " | |
6300 | "with negative dependence distance\n"); | |
6301 | return false; | |
6302 | } | |
6303 | ||
7b7b1813 | 6304 | elem_type = TREE_TYPE (vectype); |
947131ba | 6305 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
6306 | |
6307 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
6308 | (e.g. - data copies). */ | |
947131ba | 6309 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 6310 | { |
73fbfcad | 6311 | if (dump_enabled_p ()) |
78c60e3d | 6312 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6313 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
6314 | return false; |
6315 | } | |
6316 | ||
ebfd146a | 6317 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 6318 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 6319 | { |
0d0293ac | 6320 | grouped_load = true; |
ebfd146a | 6321 | /* FORNOW */ |
3bab6342 | 6322 | gcc_assert (!nested_in_vect_loop && !STMT_VINFO_GATHER_SCATTER_P (stmt_info)); |
ebfd146a | 6323 | |
e14c1050 | 6324 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
d3465d72 RS |
6325 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
6326 | ||
fce57248 | 6327 | if (!slp && !STMT_VINFO_STRIDED_P (stmt_info)) |
d3465d72 RS |
6328 | { |
6329 | if (vect_load_lanes_supported (vectype, group_size)) | |
6330 | load_lanes_p = true; | |
6331 | else if (!vect_grouped_load_supported (vectype, group_size)) | |
6332 | return false; | |
6333 | } | |
d5f035ea RB |
6334 | |
6335 | /* If this is single-element interleaving with an element distance | |
6336 | that leaves unused vector loads around punt - we at least create | |
6337 | very sub-optimal code in that case (and blow up memory, | |
6338 | see PR65518). */ | |
6339 | if (first_stmt == stmt | |
72c0f643 RB |
6340 | && !GROUP_NEXT_ELEMENT (stmt_info)) |
6341 | { | |
6342 | if (GROUP_SIZE (stmt_info) > TYPE_VECTOR_SUBPARTS (vectype)) | |
6343 | { | |
6344 | if (dump_enabled_p ()) | |
6345 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6346 | "single-element interleaving not supported " | |
6347 | "for not adjacent vector loads\n"); | |
6348 | return false; | |
6349 | } | |
6350 | ||
6351 | /* Single-element interleaving requires peeling for gaps. */ | |
836dbb1a | 6352 | gcc_assert (GROUP_GAP (stmt_info)); |
72c0f643 RB |
6353 | } |
6354 | ||
c01e092f | 6355 | /* If there is a gap in the end of the group then we access excess |
72c0f643 RB |
6356 | elements in the last iteration and thus need to peel that off. */ |
6357 | if (loop_vinfo | |
6358 | && ! STMT_VINFO_STRIDED_P (stmt_info) | |
c01e092f | 6359 | && GROUP_GAP (vinfo_for_stmt (first_stmt)) != 0) |
d5f035ea RB |
6360 | { |
6361 | if (dump_enabled_p ()) | |
6362 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
72c0f643 RB |
6363 | "Data access with gaps requires scalar " |
6364 | "epilogue loop\n"); | |
6365 | if (loop->inner) | |
6366 | { | |
6367 | if (dump_enabled_p ()) | |
6368 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6369 | "Peeling for outer loop is not supported\n"); | |
6370 | return false; | |
6371 | } | |
6372 | ||
6373 | LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true; | |
d5f035ea RB |
6374 | } |
6375 | ||
b1af7da6 RB |
6376 | if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
6377 | slp_perm = true; | |
6378 | ||
47d3fdb2 RB |
6379 | /* ??? The following is overly pessimistic (as well as the loop |
6380 | case above) in the case we can statically determine the excess | |
6381 | elements loaded are within the bounds of a decl that is accessed. | |
6382 | Likewise for BB vectorizations using masked loads is a possibility. */ | |
6383 | if (bb_vinfo && slp_perm && group_size % nunits != 0) | |
6384 | { | |
6385 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6386 | "BB vectorization with gaps at the end of a load " | |
6387 | "is not supported\n"); | |
6388 | return false; | |
6389 | } | |
6390 | ||
f2556b68 RB |
6391 | /* Invalidate assumptions made by dependence analysis when vectorization |
6392 | on the unrolled body effectively re-orders stmts. */ | |
6393 | if (!PURE_SLP_STMT (stmt_info) | |
6394 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
6395 | && ((unsigned)LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
6396 | > STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
6397 | { | |
6398 | if (dump_enabled_p ()) | |
6399 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6400 | "cannot perform implicit CSE when performing " | |
6401 | "group loads with negative dependence distance\n"); | |
6402 | return false; | |
6403 | } | |
96bb56b2 RB |
6404 | |
6405 | /* Similarly when the stmt is a load that is both part of a SLP | |
6406 | instance and a loop vectorized stmt via the same-dr mechanism | |
6407 | we have to give up. */ | |
6408 | if (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info) | |
6409 | && (STMT_SLP_TYPE (stmt_info) | |
6410 | != STMT_SLP_TYPE (vinfo_for_stmt | |
6411 | (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info))))) | |
6412 | { | |
6413 | if (dump_enabled_p ()) | |
6414 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6415 | "conflicting SLP types for CSEd load\n"); | |
6416 | return false; | |
6417 | } | |
ebfd146a IR |
6418 | } |
6419 | ||
a1e53f3f | 6420 | |
3bab6342 | 6421 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
aec7ae7d | 6422 | { |
355fe088 | 6423 | gimple *def_stmt; |
3bab6342 AT |
6424 | gather_decl = vect_check_gather_scatter (stmt, loop_vinfo, &gather_base, |
6425 | &gather_off, &gather_scale); | |
aec7ae7d | 6426 | gcc_assert (gather_decl); |
81c40241 RB |
6427 | if (!vect_is_simple_use (gather_off, vinfo, &def_stmt, &gather_dt, |
6428 | &gather_off_vectype)) | |
aec7ae7d | 6429 | { |
73fbfcad | 6430 | if (dump_enabled_p ()) |
78c60e3d | 6431 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6432 | "gather index use not simple.\n"); |
aec7ae7d JJ |
6433 | return false; |
6434 | } | |
6435 | } | |
f2e2a985 | 6436 | else if (STMT_VINFO_STRIDED_P (stmt_info)) |
e09b4c37 | 6437 | ; |
319e6439 RG |
6438 | else |
6439 | { | |
6440 | negative = tree_int_cst_compare (nested_in_vect_loop | |
6441 | ? STMT_VINFO_DR_STEP (stmt_info) | |
6442 | : DR_STEP (dr), | |
6443 | size_zero_node) < 0; | |
6444 | if (negative && ncopies > 1) | |
6445 | { | |
73fbfcad | 6446 | if (dump_enabled_p ()) |
78c60e3d | 6447 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6448 | "multiple types with negative step.\n"); |
319e6439 RG |
6449 | return false; |
6450 | } | |
6451 | ||
6452 | if (negative) | |
6453 | { | |
08940f33 RB |
6454 | if (grouped_load) |
6455 | { | |
6456 | if (dump_enabled_p ()) | |
6457 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 TJ |
6458 | "negative step for group load not supported" |
6459 | "\n"); | |
08940f33 RB |
6460 | return false; |
6461 | } | |
319e6439 RG |
6462 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
6463 | if (alignment_support_scheme != dr_aligned | |
6464 | && alignment_support_scheme != dr_unaligned_supported) | |
6465 | { | |
73fbfcad | 6466 | if (dump_enabled_p ()) |
78c60e3d | 6467 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 6468 | "negative step but alignment required.\n"); |
319e6439 RG |
6469 | return false; |
6470 | } | |
6471 | if (!perm_mask_for_reverse (vectype)) | |
6472 | { | |
73fbfcad | 6473 | if (dump_enabled_p ()) |
78c60e3d | 6474 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
6475 | "negative step and reversing not supported." |
6476 | "\n"); | |
319e6439 RG |
6477 | return false; |
6478 | } | |
6479 | } | |
7d75abc8 | 6480 | } |
aec7ae7d | 6481 | |
ebfd146a IR |
6482 | if (!vec_stmt) /* transformation not required. */ |
6483 | { | |
6484 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; | |
2e8ab70c RB |
6485 | /* The SLP costs are calculated during SLP analysis. */ |
6486 | if (!PURE_SLP_STMT (stmt_info)) | |
6487 | vect_model_load_cost (stmt_info, ncopies, load_lanes_p, | |
6488 | NULL, NULL, NULL); | |
ebfd146a IR |
6489 | return true; |
6490 | } | |
6491 | ||
73fbfcad | 6492 | if (dump_enabled_p ()) |
78c60e3d | 6493 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6494 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a IR |
6495 | |
6496 | /** Transform. **/ | |
6497 | ||
c716e67f XDL |
6498 | ensure_base_align (stmt_info, dr); |
6499 | ||
3bab6342 | 6500 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
aec7ae7d JJ |
6501 | { |
6502 | tree vec_oprnd0 = NULL_TREE, op; | |
6503 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gather_decl)); | |
6504 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; | |
d3c2fee0 | 6505 | tree ptr, mask, var, scale, merge, perm_mask = NULL_TREE, prev_res = NULL_TREE; |
aec7ae7d JJ |
6506 | edge pe = loop_preheader_edge (loop); |
6507 | gimple_seq seq; | |
6508 | basic_block new_bb; | |
6509 | enum { NARROW, NONE, WIDEN } modifier; | |
6510 | int gather_off_nunits = TYPE_VECTOR_SUBPARTS (gather_off_vectype); | |
6511 | ||
6512 | if (nunits == gather_off_nunits) | |
6513 | modifier = NONE; | |
6514 | else if (nunits == gather_off_nunits / 2) | |
6515 | { | |
6516 | unsigned char *sel = XALLOCAVEC (unsigned char, gather_off_nunits); | |
6517 | modifier = WIDEN; | |
6518 | ||
6519 | for (i = 0; i < gather_off_nunits; ++i) | |
6520 | sel[i] = i | nunits; | |
6521 | ||
557be5a8 | 6522 | perm_mask = vect_gen_perm_mask_checked (gather_off_vectype, sel); |
aec7ae7d JJ |
6523 | } |
6524 | else if (nunits == gather_off_nunits * 2) | |
6525 | { | |
6526 | unsigned char *sel = XALLOCAVEC (unsigned char, nunits); | |
6527 | modifier = NARROW; | |
6528 | ||
6529 | for (i = 0; i < nunits; ++i) | |
6530 | sel[i] = i < gather_off_nunits | |
6531 | ? i : i + nunits - gather_off_nunits; | |
6532 | ||
557be5a8 | 6533 | perm_mask = vect_gen_perm_mask_checked (vectype, sel); |
aec7ae7d JJ |
6534 | ncopies *= 2; |
6535 | } | |
6536 | else | |
6537 | gcc_unreachable (); | |
6538 | ||
6539 | rettype = TREE_TYPE (TREE_TYPE (gather_decl)); | |
6540 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6541 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6542 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6543 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6544 | scaletype = TREE_VALUE (arglist); | |
d3c2fee0 | 6545 | gcc_checking_assert (types_compatible_p (srctype, rettype)); |
aec7ae7d JJ |
6546 | |
6547 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
6548 | ||
6549 | ptr = fold_convert (ptrtype, gather_base); | |
6550 | if (!is_gimple_min_invariant (ptr)) | |
6551 | { | |
6552 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
6553 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
6554 | gcc_assert (!new_bb); | |
6555 | } | |
6556 | ||
6557 | /* Currently we support only unconditional gather loads, | |
6558 | so mask should be all ones. */ | |
d3c2fee0 AI |
6559 | if (TREE_CODE (masktype) == INTEGER_TYPE) |
6560 | mask = build_int_cst (masktype, -1); | |
6561 | else if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
6562 | { | |
6563 | mask = build_int_cst (TREE_TYPE (masktype), -1); | |
6564 | mask = build_vector_from_val (masktype, mask); | |
03b9e8e4 | 6565 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
d3c2fee0 | 6566 | } |
aec7ae7d JJ |
6567 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) |
6568 | { | |
6569 | REAL_VALUE_TYPE r; | |
6570 | long tmp[6]; | |
6571 | for (j = 0; j < 6; ++j) | |
6572 | tmp[j] = -1; | |
6573 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
6574 | mask = build_real (TREE_TYPE (masktype), r); | |
d3c2fee0 | 6575 | mask = build_vector_from_val (masktype, mask); |
03b9e8e4 | 6576 | mask = vect_init_vector (stmt, mask, masktype, NULL); |
aec7ae7d JJ |
6577 | } |
6578 | else | |
6579 | gcc_unreachable (); | |
aec7ae7d JJ |
6580 | |
6581 | scale = build_int_cst (scaletype, gather_scale); | |
6582 | ||
d3c2fee0 AI |
6583 | if (TREE_CODE (TREE_TYPE (rettype)) == INTEGER_TYPE) |
6584 | merge = build_int_cst (TREE_TYPE (rettype), 0); | |
6585 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (rettype))) | |
6586 | { | |
6587 | REAL_VALUE_TYPE r; | |
6588 | long tmp[6]; | |
6589 | for (j = 0; j < 6; ++j) | |
6590 | tmp[j] = 0; | |
6591 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (rettype))); | |
6592 | merge = build_real (TREE_TYPE (rettype), r); | |
6593 | } | |
6594 | else | |
6595 | gcc_unreachable (); | |
6596 | merge = build_vector_from_val (rettype, merge); | |
6597 | merge = vect_init_vector (stmt, merge, rettype, NULL); | |
6598 | ||
aec7ae7d JJ |
6599 | prev_stmt_info = NULL; |
6600 | for (j = 0; j < ncopies; ++j) | |
6601 | { | |
6602 | if (modifier == WIDEN && (j & 1)) | |
6603 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
6604 | perm_mask, stmt, gsi); | |
6605 | else if (j == 0) | |
6606 | op = vec_oprnd0 | |
81c40241 | 6607 | = vect_get_vec_def_for_operand (gather_off, stmt); |
aec7ae7d JJ |
6608 | else |
6609 | op = vec_oprnd0 | |
6610 | = vect_get_vec_def_for_stmt_copy (gather_dt, vec_oprnd0); | |
6611 | ||
6612 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
6613 | { | |
6614 | gcc_assert (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)) | |
6615 | == TYPE_VECTOR_SUBPARTS (idxtype)); | |
0e22bb5a | 6616 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
aec7ae7d JJ |
6617 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
6618 | new_stmt | |
0d0e4a03 | 6619 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
6620 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
6621 | op = var; | |
6622 | } | |
6623 | ||
6624 | new_stmt | |
d3c2fee0 | 6625 | = gimple_build_call (gather_decl, 5, merge, ptr, op, mask, scale); |
aec7ae7d JJ |
6626 | |
6627 | if (!useless_type_conversion_p (vectype, rettype)) | |
6628 | { | |
6629 | gcc_assert (TYPE_VECTOR_SUBPARTS (vectype) | |
6630 | == TYPE_VECTOR_SUBPARTS (rettype)); | |
0e22bb5a | 6631 | op = vect_get_new_ssa_name (rettype, vect_simple_var); |
aec7ae7d JJ |
6632 | gimple_call_set_lhs (new_stmt, op); |
6633 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
b731b390 | 6634 | var = make_ssa_name (vec_dest); |
aec7ae7d JJ |
6635 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); |
6636 | new_stmt | |
0d0e4a03 | 6637 | = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); |
aec7ae7d JJ |
6638 | } |
6639 | else | |
6640 | { | |
6641 | var = make_ssa_name (vec_dest, new_stmt); | |
6642 | gimple_call_set_lhs (new_stmt, var); | |
6643 | } | |
6644 | ||
6645 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6646 | ||
6647 | if (modifier == NARROW) | |
6648 | { | |
6649 | if ((j & 1) == 0) | |
6650 | { | |
6651 | prev_res = var; | |
6652 | continue; | |
6653 | } | |
6654 | var = permute_vec_elements (prev_res, var, | |
6655 | perm_mask, stmt, gsi); | |
6656 | new_stmt = SSA_NAME_DEF_STMT (var); | |
6657 | } | |
6658 | ||
6659 | if (prev_stmt_info == NULL) | |
6660 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6661 | else | |
6662 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6663 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6664 | } | |
6665 | return true; | |
6666 | } | |
f2e2a985 | 6667 | else if (STMT_VINFO_STRIDED_P (stmt_info)) |
7d75abc8 MM |
6668 | { |
6669 | gimple_stmt_iterator incr_gsi; | |
6670 | bool insert_after; | |
355fe088 | 6671 | gimple *incr; |
7d75abc8 | 6672 | tree offvar; |
7d75abc8 MM |
6673 | tree ivstep; |
6674 | tree running_off; | |
9771b263 | 6675 | vec<constructor_elt, va_gc> *v = NULL; |
7d75abc8 | 6676 | gimple_seq stmts = NULL; |
14ac6aa2 RB |
6677 | tree stride_base, stride_step, alias_off; |
6678 | ||
6679 | gcc_assert (!nested_in_vect_loop); | |
7d75abc8 | 6680 | |
f502d50e | 6681 | if (slp && grouped_load) |
ab313a8c RB |
6682 | first_dr = STMT_VINFO_DATA_REF |
6683 | (vinfo_for_stmt (GROUP_FIRST_ELEMENT (stmt_info))); | |
6684 | else | |
6685 | first_dr = dr; | |
6686 | ||
14ac6aa2 RB |
6687 | stride_base |
6688 | = fold_build_pointer_plus | |
ab313a8c | 6689 | (DR_BASE_ADDRESS (first_dr), |
14ac6aa2 | 6690 | size_binop (PLUS_EXPR, |
ab313a8c RB |
6691 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
6692 | convert_to_ptrofftype (DR_INIT (first_dr)))); | |
6693 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); | |
7d75abc8 MM |
6694 | |
6695 | /* For a load with loop-invariant (but other than power-of-2) | |
6696 | stride (i.e. not a grouped access) like so: | |
6697 | ||
6698 | for (i = 0; i < n; i += stride) | |
6699 | ... = array[i]; | |
6700 | ||
6701 | we generate a new induction variable and new accesses to | |
6702 | form a new vector (or vectors, depending on ncopies): | |
6703 | ||
6704 | for (j = 0; ; j += VF*stride) | |
6705 | tmp1 = array[j]; | |
6706 | tmp2 = array[j + stride]; | |
6707 | ... | |
6708 | vectemp = {tmp1, tmp2, ...} | |
6709 | */ | |
6710 | ||
ab313a8c RB |
6711 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step, |
6712 | build_int_cst (TREE_TYPE (stride_step), vf)); | |
7d75abc8 MM |
6713 | |
6714 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
6715 | ||
ab313a8c | 6716 | create_iv (unshare_expr (stride_base), unshare_expr (ivstep), NULL, |
7d75abc8 MM |
6717 | loop, &incr_gsi, insert_after, |
6718 | &offvar, NULL); | |
6719 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 6720 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
7d75abc8 | 6721 | |
ab313a8c RB |
6722 | stride_step = force_gimple_operand (unshare_expr (stride_step), |
6723 | &stmts, true, NULL_TREE); | |
7d75abc8 MM |
6724 | if (stmts) |
6725 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
6726 | ||
6727 | prev_stmt_info = NULL; | |
6728 | running_off = offvar; | |
ab313a8c | 6729 | alias_off = build_int_cst (reference_alias_ptr_type (DR_REF (first_dr)), 0); |
7b5fc413 | 6730 | int nloads = nunits; |
e09b4c37 | 6731 | int lnel = 1; |
7b5fc413 | 6732 | tree ltype = TREE_TYPE (vectype); |
b266b968 | 6733 | auto_vec<tree> dr_chain; |
7b5fc413 RB |
6734 | if (slp) |
6735 | { | |
e09b4c37 RB |
6736 | if (group_size < nunits |
6737 | && nunits % group_size == 0) | |
6738 | { | |
6739 | nloads = nunits / group_size; | |
6740 | lnel = group_size; | |
6741 | ltype = build_vector_type (TREE_TYPE (vectype), group_size); | |
6742 | ltype = build_aligned_type (ltype, | |
6743 | TYPE_ALIGN (TREE_TYPE (vectype))); | |
6744 | } | |
6745 | else if (group_size >= nunits | |
6746 | && group_size % nunits == 0) | |
6747 | { | |
6748 | nloads = 1; | |
6749 | lnel = nunits; | |
6750 | ltype = vectype; | |
6751 | ltype = build_aligned_type (ltype, | |
6752 | TYPE_ALIGN (TREE_TYPE (vectype))); | |
6753 | } | |
66c16fd9 RB |
6754 | /* For SLP permutation support we need to load the whole group, |
6755 | not only the number of vector stmts the permutation result | |
6756 | fits in. */ | |
b266b968 | 6757 | if (slp_perm) |
66c16fd9 RB |
6758 | { |
6759 | ncopies = (group_size * vf + nunits - 1) / nunits; | |
6760 | dr_chain.create (ncopies); | |
6761 | } | |
6762 | else | |
6763 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
7b5fc413 | 6764 | } |
e09b4c37 RB |
6765 | int group_el = 0; |
6766 | unsigned HOST_WIDE_INT | |
6767 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
7d75abc8 MM |
6768 | for (j = 0; j < ncopies; j++) |
6769 | { | |
7b5fc413 | 6770 | if (nloads > 1) |
e09b4c37 RB |
6771 | vec_alloc (v, nloads); |
6772 | for (i = 0; i < nloads; i++) | |
7b5fc413 | 6773 | { |
e09b4c37 RB |
6774 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
6775 | group_el * elsz); | |
6776 | new_stmt = gimple_build_assign (make_ssa_name (ltype), | |
6777 | build2 (MEM_REF, ltype, | |
6778 | running_off, this_off)); | |
6779 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6780 | if (nloads > 1) | |
6781 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, | |
6782 | gimple_assign_lhs (new_stmt)); | |
6783 | ||
6784 | group_el += lnel; | |
6785 | if (! slp | |
6786 | || group_el == group_size) | |
7b5fc413 | 6787 | { |
e09b4c37 RB |
6788 | tree newoff = copy_ssa_name (running_off); |
6789 | gimple *incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
6790 | running_off, stride_step); | |
7b5fc413 RB |
6791 | vect_finish_stmt_generation (stmt, incr, gsi); |
6792 | ||
6793 | running_off = newoff; | |
e09b4c37 | 6794 | group_el = 0; |
7b5fc413 | 6795 | } |
7b5fc413 | 6796 | } |
e09b4c37 | 6797 | if (nloads > 1) |
7d75abc8 | 6798 | { |
e09b4c37 RB |
6799 | tree vec_inv = build_constructor (vectype, v); |
6800 | new_temp = vect_init_vector (stmt, vec_inv, vectype, gsi); | |
6801 | new_stmt = SSA_NAME_DEF_STMT (new_temp); | |
7d75abc8 MM |
6802 | } |
6803 | ||
7b5fc413 | 6804 | if (slp) |
b266b968 | 6805 | { |
b266b968 RB |
6806 | if (slp_perm) |
6807 | dr_chain.quick_push (gimple_assign_lhs (new_stmt)); | |
66c16fd9 RB |
6808 | else |
6809 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
b266b968 | 6810 | } |
7d75abc8 | 6811 | else |
225ce44b RB |
6812 | { |
6813 | if (j == 0) | |
6814 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6815 | else | |
6816 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6817 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6818 | } | |
7d75abc8 | 6819 | } |
b266b968 RB |
6820 | if (slp_perm) |
6821 | vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, | |
6822 | slp_node_instance, false); | |
7d75abc8 MM |
6823 | return true; |
6824 | } | |
aec7ae7d | 6825 | |
0d0293ac | 6826 | if (grouped_load) |
ebfd146a | 6827 | { |
e14c1050 | 6828 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
4f0a0218 | 6829 | /* For SLP vectorization we directly vectorize a subchain |
52eab378 RB |
6830 | without permutation. */ |
6831 | if (slp && ! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
4f0a0218 RB |
6832 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
6833 | /* For BB vectorization always use the first stmt to base | |
6834 | the data ref pointer on. */ | |
6835 | if (bb_vinfo) | |
6836 | first_stmt_for_drptr = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 6837 | |
ebfd146a | 6838 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
6839 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
6840 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
6841 | ??? But we can only do so if there is exactly one | |
6842 | as we have no way to get at the rest. Leave the CSE | |
6843 | opportunity alone. | |
6844 | ??? With the group load eventually participating | |
6845 | in multiple different permutations (having multiple | |
6846 | slp nodes which refer to the same group) the CSE | |
6847 | is even wrong code. See PR56270. */ | |
6848 | && !slp) | |
ebfd146a IR |
6849 | { |
6850 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
6851 | return true; | |
6852 | } | |
6853 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
e14c1050 | 6854 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
9b999e8c | 6855 | group_gap_adj = 0; |
ebfd146a IR |
6856 | |
6857 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
6858 | if (slp) | |
6859 | { | |
0d0293ac | 6860 | grouped_load = false; |
91ff1504 RB |
6861 | /* For SLP permutation support we need to load the whole group, |
6862 | not only the number of vector stmts the permutation result | |
6863 | fits in. */ | |
6864 | if (slp_perm) | |
6865 | vec_num = (group_size * vf + nunits - 1) / nunits; | |
6866 | else | |
6867 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9b999e8c | 6868 | group_gap_adj = vf * group_size - nunits * vec_num; |
a70d6342 | 6869 | } |
ebfd146a | 6870 | else |
9b999e8c | 6871 | vec_num = group_size; |
ebfd146a IR |
6872 | } |
6873 | else | |
6874 | { | |
6875 | first_stmt = stmt; | |
6876 | first_dr = dr; | |
6877 | group_size = vec_num = 1; | |
9b999e8c | 6878 | group_gap_adj = 0; |
ebfd146a IR |
6879 | } |
6880 | ||
720f5239 | 6881 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 6882 | gcc_assert (alignment_support_scheme); |
272c6793 RS |
6883 | /* Targets with load-lane instructions must not require explicit |
6884 | realignment. */ | |
6885 | gcc_assert (!load_lanes_p | |
6886 | || alignment_support_scheme == dr_aligned | |
6887 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
6888 | |
6889 | /* In case the vectorization factor (VF) is bigger than the number | |
6890 | of elements that we can fit in a vectype (nunits), we have to generate | |
6891 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 6892 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 6893 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 6894 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 6895 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
6896 | stmts that use the defs of the current stmt. The example below |
6897 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
6898 | need to create 4 vectorized stmts): | |
ebfd146a IR |
6899 | |
6900 | before vectorization: | |
6901 | RELATED_STMT VEC_STMT | |
6902 | S1: x = memref - - | |
6903 | S2: z = x + 1 - - | |
6904 | ||
6905 | step 1: vectorize stmt S1: | |
6906 | We first create the vector stmt VS1_0, and, as usual, record a | |
6907 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
6908 | Next, we create the vector stmt VS1_1, and record a pointer to | |
6909 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 6910 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
6911 | stmts and pointers: |
6912 | RELATED_STMT VEC_STMT | |
6913 | VS1_0: vx0 = memref0 VS1_1 - | |
6914 | VS1_1: vx1 = memref1 VS1_2 - | |
6915 | VS1_2: vx2 = memref2 VS1_3 - | |
6916 | VS1_3: vx3 = memref3 - - | |
6917 | S1: x = load - VS1_0 | |
6918 | S2: z = x + 1 - - | |
6919 | ||
b8698a0f L |
6920 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
6921 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
6922 | stmt S2. */ |
6923 | ||
0d0293ac | 6924 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
6925 | |
6926 | S1: x2 = &base + 2 | |
6927 | S2: x0 = &base | |
6928 | S3: x1 = &base + 1 | |
6929 | S4: x3 = &base + 3 | |
6930 | ||
b8698a0f | 6931 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
6932 | starting from the access of the first stmt of the chain: |
6933 | ||
6934 | VS1: vx0 = &base | |
6935 | VS2: vx1 = &base + vec_size*1 | |
6936 | VS3: vx3 = &base + vec_size*2 | |
6937 | VS4: vx4 = &base + vec_size*3 | |
6938 | ||
6939 | Then permutation statements are generated: | |
6940 | ||
e2c83630 RH |
6941 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
6942 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
6943 | ... |
6944 | ||
6945 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
6946 | (the order of the data-refs in the output of vect_permute_load_chain | |
6947 | corresponds to the order of scalar stmts in the interleaving chain - see | |
6948 | the documentation of vect_permute_load_chain()). | |
6949 | The generation of permutation stmts and recording them in | |
0d0293ac | 6950 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 6951 | |
b8698a0f | 6952 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
6953 | permutation stmts above are created for every copy. The result vector |
6954 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
6955 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
6956 | |
6957 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
6958 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
6959 | we generate the following code: | |
6960 | p = initial_addr; | |
6961 | indx = 0; | |
6962 | loop { | |
6963 | p = p + indx * vectype_size; | |
6964 | vec_dest = *(p); | |
6965 | indx = indx + 1; | |
6966 | } | |
6967 | ||
6968 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 6969 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
6970 | then generate the following code, in which the data in each iteration is |
6971 | obtained by two vector loads, one from the previous iteration, and one | |
6972 | from the current iteration: | |
6973 | p1 = initial_addr; | |
6974 | msq_init = *(floor(p1)) | |
6975 | p2 = initial_addr + VS - 1; | |
6976 | realignment_token = call target_builtin; | |
6977 | indx = 0; | |
6978 | loop { | |
6979 | p2 = p2 + indx * vectype_size | |
6980 | lsq = *(floor(p2)) | |
6981 | vec_dest = realign_load (msq, lsq, realignment_token) | |
6982 | indx = indx + 1; | |
6983 | msq = lsq; | |
6984 | } */ | |
6985 | ||
6986 | /* If the misalignment remains the same throughout the execution of the | |
6987 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 6988 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
6989 | This can only occur when vectorizing memory accesses in the inner-loop |
6990 | nested within an outer-loop that is being vectorized. */ | |
6991 | ||
d1e4b493 | 6992 | if (nested_in_vect_loop |
211bea38 | 6993 | && (TREE_INT_CST_LOW (DR_STEP (dr)) |
ebfd146a IR |
6994 | % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)) |
6995 | { | |
6996 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
6997 | compute_in_loop = true; | |
6998 | } | |
6999 | ||
7000 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
7001 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 7002 | && !compute_in_loop) |
ebfd146a IR |
7003 | { |
7004 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
7005 | alignment_support_scheme, NULL_TREE, | |
7006 | &at_loop); | |
7007 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7008 | { | |
538dd0b7 | 7009 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (msq)); |
356bbc4c JJ |
7010 | byte_offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
7011 | size_one_node); | |
ebfd146a IR |
7012 | } |
7013 | } | |
7014 | else | |
7015 | at_loop = loop; | |
7016 | ||
a1e53f3f L |
7017 | if (negative) |
7018 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); | |
7019 | ||
272c6793 RS |
7020 | if (load_lanes_p) |
7021 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
7022 | else | |
7023 | aggr_type = vectype; | |
7024 | ||
ebfd146a IR |
7025 | prev_stmt_info = NULL; |
7026 | for (j = 0; j < ncopies; j++) | |
b8698a0f | 7027 | { |
272c6793 | 7028 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 7029 | if (j == 0) |
74bf76ed JJ |
7030 | { |
7031 | bool simd_lane_access_p | |
7032 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
7033 | if (simd_lane_access_p | |
7034 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
7035 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
7036 | && integer_zerop (DR_OFFSET (first_dr)) | |
7037 | && integer_zerop (DR_INIT (first_dr)) | |
7038 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
7039 | get_alias_set (DR_REF (first_dr))) | |
7040 | && (alignment_support_scheme == dr_aligned | |
7041 | || alignment_support_scheme == dr_unaligned_supported)) | |
7042 | { | |
7043 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
7044 | dataref_offset = build_int_cst (reference_alias_ptr_type | |
7045 | (DR_REF (first_dr)), 0); | |
8928eff3 | 7046 | inv_p = false; |
74bf76ed | 7047 | } |
4f0a0218 RB |
7048 | else if (first_stmt_for_drptr |
7049 | && first_stmt != first_stmt_for_drptr) | |
7050 | { | |
7051 | dataref_ptr | |
7052 | = vect_create_data_ref_ptr (first_stmt_for_drptr, aggr_type, | |
7053 | at_loop, offset, &dummy, gsi, | |
7054 | &ptr_incr, simd_lane_access_p, | |
7055 | &inv_p, byte_offset); | |
7056 | /* Adjust the pointer by the difference to first_stmt. */ | |
7057 | data_reference_p ptrdr | |
7058 | = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt_for_drptr)); | |
7059 | tree diff = fold_convert (sizetype, | |
7060 | size_binop (MINUS_EXPR, | |
7061 | DR_INIT (first_dr), | |
7062 | DR_INIT (ptrdr))); | |
7063 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
7064 | stmt, diff); | |
7065 | } | |
74bf76ed JJ |
7066 | else |
7067 | dataref_ptr | |
7068 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
7069 | offset, &dummy, gsi, &ptr_incr, | |
356bbc4c JJ |
7070 | simd_lane_access_p, &inv_p, |
7071 | byte_offset); | |
74bf76ed JJ |
7072 | } |
7073 | else if (dataref_offset) | |
7074 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
7075 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 7076 | else |
272c6793 RS |
7077 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, |
7078 | TYPE_SIZE_UNIT (aggr_type)); | |
ebfd146a | 7079 | |
0d0293ac | 7080 | if (grouped_load || slp_perm) |
9771b263 | 7081 | dr_chain.create (vec_num); |
5ce1ee7f | 7082 | |
272c6793 | 7083 | if (load_lanes_p) |
ebfd146a | 7084 | { |
272c6793 RS |
7085 | tree vec_array; |
7086 | ||
7087 | vec_array = create_vector_array (vectype, vec_num); | |
7088 | ||
7089 | /* Emit: | |
7090 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
7091 | data_ref = create_array_ref (aggr_type, dataref_ptr, first_dr); | |
7092 | new_stmt = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
7093 | gimple_call_set_lhs (new_stmt, vec_array); | |
7094 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
ebfd146a | 7095 | |
272c6793 RS |
7096 | /* Extract each vector into an SSA_NAME. */ |
7097 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 7098 | { |
272c6793 RS |
7099 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
7100 | vec_array, i); | |
9771b263 | 7101 | dr_chain.quick_push (new_temp); |
272c6793 RS |
7102 | } |
7103 | ||
7104 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 7105 | vect_record_grouped_load_vectors (stmt, dr_chain); |
272c6793 RS |
7106 | } |
7107 | else | |
7108 | { | |
7109 | for (i = 0; i < vec_num; i++) | |
7110 | { | |
7111 | if (i > 0) | |
7112 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
7113 | stmt, NULL_TREE); | |
7114 | ||
7115 | /* 2. Create the vector-load in the loop. */ | |
7116 | switch (alignment_support_scheme) | |
7117 | { | |
7118 | case dr_aligned: | |
7119 | case dr_unaligned_supported: | |
be1ac4ec | 7120 | { |
644ffefd MJ |
7121 | unsigned int align, misalign; |
7122 | ||
272c6793 | 7123 | data_ref |
aed93b23 RB |
7124 | = fold_build2 (MEM_REF, vectype, dataref_ptr, |
7125 | dataref_offset | |
7126 | ? dataref_offset | |
7127 | : build_int_cst (reference_alias_ptr_type | |
7128 | (DR_REF (first_dr)), 0)); | |
644ffefd | 7129 | align = TYPE_ALIGN_UNIT (vectype); |
272c6793 RS |
7130 | if (alignment_support_scheme == dr_aligned) |
7131 | { | |
7132 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 7133 | misalign = 0; |
272c6793 RS |
7134 | } |
7135 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
7136 | { | |
52639a61 RB |
7137 | if (DR_VECT_AUX (first_dr)->base_element_aligned) |
7138 | align = TYPE_ALIGN_UNIT (elem_type); | |
7139 | else | |
7140 | align = (get_object_alignment (DR_REF (first_dr)) | |
7141 | / BITS_PER_UNIT); | |
7142 | misalign = 0; | |
272c6793 RS |
7143 | TREE_TYPE (data_ref) |
7144 | = build_aligned_type (TREE_TYPE (data_ref), | |
52639a61 | 7145 | align * BITS_PER_UNIT); |
272c6793 RS |
7146 | } |
7147 | else | |
7148 | { | |
7149 | TREE_TYPE (data_ref) | |
7150 | = build_aligned_type (TREE_TYPE (data_ref), | |
7151 | TYPE_ALIGN (elem_type)); | |
644ffefd | 7152 | misalign = DR_MISALIGNMENT (first_dr); |
272c6793 | 7153 | } |
aed93b23 RB |
7154 | if (dataref_offset == NULL_TREE |
7155 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
7156 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), |
7157 | align, misalign); | |
272c6793 | 7158 | break; |
be1ac4ec | 7159 | } |
272c6793 | 7160 | case dr_explicit_realign: |
267d3070 | 7161 | { |
272c6793 | 7162 | tree ptr, bump; |
272c6793 | 7163 | |
d88981fc | 7164 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
272c6793 RS |
7165 | |
7166 | if (compute_in_loop) | |
7167 | msq = vect_setup_realignment (first_stmt, gsi, | |
7168 | &realignment_token, | |
7169 | dr_explicit_realign, | |
7170 | dataref_ptr, NULL); | |
7171 | ||
aed93b23 RB |
7172 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
7173 | ptr = copy_ssa_name (dataref_ptr); | |
7174 | else | |
7175 | ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
0d0e4a03 JJ |
7176 | new_stmt = gimple_build_assign |
7177 | (ptr, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
7178 | build_int_cst |
7179 | (TREE_TYPE (dataref_ptr), | |
7180 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
7181 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7182 | data_ref | |
7183 | = build2 (MEM_REF, vectype, ptr, | |
7184 | build_int_cst (reference_alias_ptr_type | |
7185 | (DR_REF (first_dr)), 0)); | |
7186 | vec_dest = vect_create_destination_var (scalar_dest, | |
7187 | vectype); | |
7188 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
7189 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
7190 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7191 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
7192 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
7193 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7194 | msq = new_temp; | |
7195 | ||
d88981fc | 7196 | bump = size_binop (MULT_EXPR, vs, |
7b7b1813 | 7197 | TYPE_SIZE_UNIT (elem_type)); |
d88981fc | 7198 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
272c6793 | 7199 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
0d0e4a03 JJ |
7200 | new_stmt = gimple_build_assign |
7201 | (NULL_TREE, BIT_AND_EXPR, ptr, | |
272c6793 RS |
7202 | build_int_cst |
7203 | (TREE_TYPE (ptr), | |
7204 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
aed93b23 | 7205 | ptr = copy_ssa_name (ptr, new_stmt); |
272c6793 RS |
7206 | gimple_assign_set_lhs (new_stmt, ptr); |
7207 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7208 | data_ref | |
7209 | = build2 (MEM_REF, vectype, ptr, | |
7210 | build_int_cst (reference_alias_ptr_type | |
7211 | (DR_REF (first_dr)), 0)); | |
7212 | break; | |
267d3070 | 7213 | } |
272c6793 | 7214 | case dr_explicit_realign_optimized: |
aed93b23 RB |
7215 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
7216 | new_temp = copy_ssa_name (dataref_ptr); | |
7217 | else | |
7218 | new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
0d0e4a03 JJ |
7219 | new_stmt = gimple_build_assign |
7220 | (new_temp, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
7221 | build_int_cst |
7222 | (TREE_TYPE (dataref_ptr), | |
7223 | -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); | |
272c6793 RS |
7224 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7225 | data_ref | |
7226 | = build2 (MEM_REF, vectype, new_temp, | |
7227 | build_int_cst (reference_alias_ptr_type | |
7228 | (DR_REF (first_dr)), 0)); | |
7229 | break; | |
7230 | default: | |
7231 | gcc_unreachable (); | |
7232 | } | |
ebfd146a | 7233 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
272c6793 | 7234 | new_stmt = gimple_build_assign (vec_dest, data_ref); |
ebfd146a IR |
7235 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7236 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7237 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7238 | ||
272c6793 RS |
7239 | /* 3. Handle explicit realignment if necessary/supported. |
7240 | Create in loop: | |
7241 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
7242 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
7243 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 7244 | { |
272c6793 RS |
7245 | lsq = gimple_assign_lhs (new_stmt); |
7246 | if (!realignment_token) | |
7247 | realignment_token = dataref_ptr; | |
7248 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
0d0e4a03 JJ |
7249 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, |
7250 | msq, lsq, realignment_token); | |
272c6793 RS |
7251 | new_temp = make_ssa_name (vec_dest, new_stmt); |
7252 | gimple_assign_set_lhs (new_stmt, new_temp); | |
7253 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7254 | ||
7255 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7256 | { | |
7257 | gcc_assert (phi); | |
7258 | if (i == vec_num - 1 && j == ncopies - 1) | |
7259 | add_phi_arg (phi, lsq, | |
7260 | loop_latch_edge (containing_loop), | |
9e227d60 | 7261 | UNKNOWN_LOCATION); |
272c6793 RS |
7262 | msq = lsq; |
7263 | } | |
ebfd146a | 7264 | } |
ebfd146a | 7265 | |
59fd17e3 RB |
7266 | /* 4. Handle invariant-load. */ |
7267 | if (inv_p && !bb_vinfo) | |
7268 | { | |
59fd17e3 | 7269 | gcc_assert (!grouped_load); |
d1417442 JJ |
7270 | /* If we have versioned for aliasing or the loop doesn't |
7271 | have any data dependencies that would preclude this, | |
7272 | then we are sure this is a loop invariant load and | |
7273 | thus we can insert it on the preheader edge. */ | |
7274 | if (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) | |
7275 | && !nested_in_vect_loop | |
6b916b36 | 7276 | && hoist_defs_of_uses (stmt, loop)) |
a0e35eb0 RB |
7277 | { |
7278 | if (dump_enabled_p ()) | |
7279 | { | |
7280 | dump_printf_loc (MSG_NOTE, vect_location, | |
7281 | "hoisting out of the vectorized " | |
7282 | "loop: "); | |
7283 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
a0e35eb0 | 7284 | } |
b731b390 | 7285 | tree tem = copy_ssa_name (scalar_dest); |
a0e35eb0 RB |
7286 | gsi_insert_on_edge_immediate |
7287 | (loop_preheader_edge (loop), | |
7288 | gimple_build_assign (tem, | |
7289 | unshare_expr | |
7290 | (gimple_assign_rhs1 (stmt)))); | |
7291 | new_temp = vect_init_vector (stmt, tem, vectype, NULL); | |
34cd48e5 RB |
7292 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
7293 | set_vinfo_for_stmt (new_stmt, | |
7294 | new_stmt_vec_info (new_stmt, vinfo)); | |
a0e35eb0 RB |
7295 | } |
7296 | else | |
7297 | { | |
7298 | gimple_stmt_iterator gsi2 = *gsi; | |
7299 | gsi_next (&gsi2); | |
7300 | new_temp = vect_init_vector (stmt, scalar_dest, | |
7301 | vectype, &gsi2); | |
34cd48e5 | 7302 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
a0e35eb0 | 7303 | } |
59fd17e3 RB |
7304 | } |
7305 | ||
272c6793 RS |
7306 | if (negative) |
7307 | { | |
aec7ae7d JJ |
7308 | tree perm_mask = perm_mask_for_reverse (vectype); |
7309 | new_temp = permute_vec_elements (new_temp, new_temp, | |
7310 | perm_mask, stmt, gsi); | |
ebfd146a IR |
7311 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
7312 | } | |
267d3070 | 7313 | |
272c6793 | 7314 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
7315 | vect_transform_grouped_load (). */ |
7316 | if (grouped_load || slp_perm) | |
9771b263 | 7317 | dr_chain.quick_push (new_temp); |
267d3070 | 7318 | |
272c6793 RS |
7319 | /* Store vector loads in the corresponding SLP_NODE. */ |
7320 | if (slp && !slp_perm) | |
9771b263 | 7321 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
272c6793 | 7322 | } |
9b999e8c RB |
7323 | /* Bump the vector pointer to account for a gap or for excess |
7324 | elements loaded for a permuted SLP load. */ | |
7325 | if (group_gap_adj != 0) | |
a64b9c26 | 7326 | { |
9b999e8c RB |
7327 | bool ovf; |
7328 | tree bump | |
7329 | = wide_int_to_tree (sizetype, | |
7330 | wi::smul (TYPE_SIZE_UNIT (elem_type), | |
7331 | group_gap_adj, &ovf)); | |
a64b9c26 RB |
7332 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
7333 | stmt, bump); | |
7334 | } | |
ebfd146a IR |
7335 | } |
7336 | ||
7337 | if (slp && !slp_perm) | |
7338 | continue; | |
7339 | ||
7340 | if (slp_perm) | |
7341 | { | |
01d8bf07 | 7342 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
ebfd146a IR |
7343 | slp_node_instance, false)) |
7344 | { | |
9771b263 | 7345 | dr_chain.release (); |
ebfd146a IR |
7346 | return false; |
7347 | } | |
7348 | } | |
7349 | else | |
7350 | { | |
0d0293ac | 7351 | if (grouped_load) |
ebfd146a | 7352 | { |
272c6793 | 7353 | if (!load_lanes_p) |
0d0293ac | 7354 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 7355 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
7356 | } |
7357 | else | |
7358 | { | |
7359 | if (j == 0) | |
7360 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7361 | else | |
7362 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7363 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7364 | } | |
7365 | } | |
9771b263 | 7366 | dr_chain.release (); |
ebfd146a IR |
7367 | } |
7368 | ||
ebfd146a IR |
7369 | return true; |
7370 | } | |
7371 | ||
7372 | /* Function vect_is_simple_cond. | |
b8698a0f | 7373 | |
ebfd146a IR |
7374 | Input: |
7375 | LOOP - the loop that is being vectorized. | |
7376 | COND - Condition that is checked for simple use. | |
7377 | ||
e9e1d143 RG |
7378 | Output: |
7379 | *COMP_VECTYPE - the vector type for the comparison. | |
7380 | ||
ebfd146a IR |
7381 | Returns whether a COND can be vectorized. Checks whether |
7382 | condition operands are supportable using vec_is_simple_use. */ | |
7383 | ||
87aab9b2 | 7384 | static bool |
81c40241 | 7385 | vect_is_simple_cond (tree cond, vec_info *vinfo, tree *comp_vectype) |
ebfd146a IR |
7386 | { |
7387 | tree lhs, rhs; | |
ebfd146a | 7388 | enum vect_def_type dt; |
e9e1d143 | 7389 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a | 7390 | |
a414c77f IE |
7391 | /* Mask case. */ |
7392 | if (TREE_CODE (cond) == SSA_NAME | |
7393 | && TREE_CODE (TREE_TYPE (cond)) == BOOLEAN_TYPE) | |
7394 | { | |
7395 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (cond); | |
7396 | if (!vect_is_simple_use (cond, vinfo, &lhs_def_stmt, | |
7397 | &dt, comp_vectype) | |
7398 | || !*comp_vectype | |
7399 | || !VECTOR_BOOLEAN_TYPE_P (*comp_vectype)) | |
7400 | return false; | |
7401 | return true; | |
7402 | } | |
7403 | ||
ebfd146a IR |
7404 | if (!COMPARISON_CLASS_P (cond)) |
7405 | return false; | |
7406 | ||
7407 | lhs = TREE_OPERAND (cond, 0); | |
7408 | rhs = TREE_OPERAND (cond, 1); | |
7409 | ||
7410 | if (TREE_CODE (lhs) == SSA_NAME) | |
7411 | { | |
355fe088 | 7412 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); |
81c40241 | 7413 | if (!vect_is_simple_use (lhs, vinfo, &lhs_def_stmt, &dt, &vectype1)) |
ebfd146a IR |
7414 | return false; |
7415 | } | |
7416 | else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST | |
7417 | && TREE_CODE (lhs) != FIXED_CST) | |
7418 | return false; | |
7419 | ||
7420 | if (TREE_CODE (rhs) == SSA_NAME) | |
7421 | { | |
355fe088 | 7422 | gimple *rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); |
81c40241 | 7423 | if (!vect_is_simple_use (rhs, vinfo, &rhs_def_stmt, &dt, &vectype2)) |
ebfd146a IR |
7424 | return false; |
7425 | } | |
f7e531cf | 7426 | else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST |
ebfd146a IR |
7427 | && TREE_CODE (rhs) != FIXED_CST) |
7428 | return false; | |
7429 | ||
28b33016 IE |
7430 | if (vectype1 && vectype2 |
7431 | && TYPE_VECTOR_SUBPARTS (vectype1) != TYPE_VECTOR_SUBPARTS (vectype2)) | |
7432 | return false; | |
7433 | ||
e9e1d143 | 7434 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
ebfd146a IR |
7435 | return true; |
7436 | } | |
7437 | ||
7438 | /* vectorizable_condition. | |
7439 | ||
b8698a0f L |
7440 | Check if STMT is conditional modify expression that can be vectorized. |
7441 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
7442 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
7443 | at GSI. |
7444 | ||
7445 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
7446 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
0ad23163 | 7447 | else clause if it is 2). |
ebfd146a IR |
7448 | |
7449 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7450 | ||
4bbe8262 | 7451 | bool |
355fe088 TS |
7452 | vectorizable_condition (gimple *stmt, gimple_stmt_iterator *gsi, |
7453 | gimple **vec_stmt, tree reduc_def, int reduc_index, | |
f7e531cf | 7454 | slp_tree slp_node) |
ebfd146a IR |
7455 | { |
7456 | tree scalar_dest = NULL_TREE; | |
7457 | tree vec_dest = NULL_TREE; | |
ebfd146a IR |
7458 | tree cond_expr, then_clause, else_clause; |
7459 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
df11cc78 | 7460 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
7461 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
7462 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
5958f9e2 | 7463 | tree vec_compare; |
ebfd146a IR |
7464 | tree new_temp; |
7465 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a855b1b1 | 7466 | enum vect_def_type dt, dts[4]; |
f7e531cf | 7467 | int ncopies; |
ebfd146a | 7468 | enum tree_code code; |
a855b1b1 | 7469 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
7470 | int i, j; |
7471 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
7472 | vec<tree> vec_oprnds0 = vNULL; |
7473 | vec<tree> vec_oprnds1 = vNULL; | |
7474 | vec<tree> vec_oprnds2 = vNULL; | |
7475 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 7476 | tree vec_cmp_type; |
a414c77f | 7477 | bool masked = false; |
b8698a0f | 7478 | |
f7e531cf IR |
7479 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
7480 | return false; | |
7481 | ||
af29617a AH |
7482 | if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) == TREE_CODE_REDUCTION) |
7483 | { | |
7484 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
7485 | return false; | |
ebfd146a | 7486 | |
af29617a AH |
7487 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7488 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
7489 | && reduc_def)) | |
7490 | return false; | |
ebfd146a | 7491 | |
af29617a AH |
7492 | /* FORNOW: not yet supported. */ |
7493 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
7494 | { | |
7495 | if (dump_enabled_p ()) | |
7496 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7497 | "value used after loop.\n"); | |
7498 | return false; | |
7499 | } | |
ebfd146a IR |
7500 | } |
7501 | ||
7502 | /* Is vectorizable conditional operation? */ | |
7503 | if (!is_gimple_assign (stmt)) | |
7504 | return false; | |
7505 | ||
7506 | code = gimple_assign_rhs_code (stmt); | |
7507 | ||
7508 | if (code != COND_EXPR) | |
7509 | return false; | |
7510 | ||
465c8c19 JJ |
7511 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
7512 | int nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
2947d3b2 | 7513 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
465c8c19 | 7514 | |
fce57248 | 7515 | if (slp_node) |
465c8c19 JJ |
7516 | ncopies = 1; |
7517 | else | |
7518 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
7519 | ||
7520 | gcc_assert (ncopies >= 1); | |
7521 | if (reduc_index && ncopies > 1) | |
7522 | return false; /* FORNOW */ | |
7523 | ||
4e71066d RG |
7524 | cond_expr = gimple_assign_rhs1 (stmt); |
7525 | then_clause = gimple_assign_rhs2 (stmt); | |
7526 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 7527 | |
81c40241 | 7528 | if (!vect_is_simple_cond (cond_expr, stmt_info->vinfo, &comp_vectype) |
e9e1d143 | 7529 | || !comp_vectype) |
ebfd146a IR |
7530 | return false; |
7531 | ||
81c40241 | 7532 | gimple *def_stmt; |
2947d3b2 IE |
7533 | if (!vect_is_simple_use (then_clause, stmt_info->vinfo, &def_stmt, &dt, |
7534 | &vectype1)) | |
7535 | return false; | |
7536 | if (!vect_is_simple_use (else_clause, stmt_info->vinfo, &def_stmt, &dt, | |
7537 | &vectype2)) | |
ebfd146a | 7538 | return false; |
2947d3b2 IE |
7539 | |
7540 | if (vectype1 && !useless_type_conversion_p (vectype, vectype1)) | |
7541 | return false; | |
7542 | ||
7543 | if (vectype2 && !useless_type_conversion_p (vectype, vectype2)) | |
ebfd146a IR |
7544 | return false; |
7545 | ||
28b33016 IE |
7546 | masked = !COMPARISON_CLASS_P (cond_expr); |
7547 | vec_cmp_type = build_same_sized_truth_vector_type (comp_vectype); | |
7548 | ||
74946978 MP |
7549 | if (vec_cmp_type == NULL_TREE) |
7550 | return false; | |
784fb9b3 | 7551 | |
b8698a0f | 7552 | if (!vec_stmt) |
ebfd146a IR |
7553 | { |
7554 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
e9e1d143 | 7555 | return expand_vec_cond_expr_p (vectype, comp_vectype); |
ebfd146a IR |
7556 | } |
7557 | ||
f7e531cf IR |
7558 | /* Transform. */ |
7559 | ||
7560 | if (!slp_node) | |
7561 | { | |
9771b263 DN |
7562 | vec_oprnds0.create (1); |
7563 | vec_oprnds1.create (1); | |
7564 | vec_oprnds2.create (1); | |
7565 | vec_oprnds3.create (1); | |
f7e531cf | 7566 | } |
ebfd146a IR |
7567 | |
7568 | /* Handle def. */ | |
7569 | scalar_dest = gimple_assign_lhs (stmt); | |
7570 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
7571 | ||
7572 | /* Handle cond expr. */ | |
a855b1b1 MM |
7573 | for (j = 0; j < ncopies; j++) |
7574 | { | |
538dd0b7 | 7575 | gassign *new_stmt = NULL; |
a855b1b1 MM |
7576 | if (j == 0) |
7577 | { | |
f7e531cf IR |
7578 | if (slp_node) |
7579 | { | |
00f96dc9 TS |
7580 | auto_vec<tree, 4> ops; |
7581 | auto_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 7582 | |
a414c77f IE |
7583 | if (masked) |
7584 | ops.safe_push (cond_expr); | |
7585 | else | |
7586 | { | |
7587 | ops.safe_push (TREE_OPERAND (cond_expr, 0)); | |
7588 | ops.safe_push (TREE_OPERAND (cond_expr, 1)); | |
7589 | } | |
9771b263 DN |
7590 | ops.safe_push (then_clause); |
7591 | ops.safe_push (else_clause); | |
f7e531cf | 7592 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); |
37b5ec8f JJ |
7593 | vec_oprnds3 = vec_defs.pop (); |
7594 | vec_oprnds2 = vec_defs.pop (); | |
a414c77f IE |
7595 | if (!masked) |
7596 | vec_oprnds1 = vec_defs.pop (); | |
37b5ec8f | 7597 | vec_oprnds0 = vec_defs.pop (); |
f7e531cf | 7598 | |
9771b263 DN |
7599 | ops.release (); |
7600 | vec_defs.release (); | |
f7e531cf IR |
7601 | } |
7602 | else | |
7603 | { | |
355fe088 | 7604 | gimple *gtemp; |
a414c77f IE |
7605 | if (masked) |
7606 | { | |
7607 | vec_cond_lhs | |
7608 | = vect_get_vec_def_for_operand (cond_expr, stmt, | |
7609 | comp_vectype); | |
7610 | vect_is_simple_use (cond_expr, stmt_info->vinfo, | |
7611 | >emp, &dts[0]); | |
7612 | } | |
7613 | else | |
7614 | { | |
7615 | vec_cond_lhs = | |
7616 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), | |
7617 | stmt, comp_vectype); | |
7618 | vect_is_simple_use (TREE_OPERAND (cond_expr, 0), | |
7619 | loop_vinfo, >emp, &dts[0]); | |
7620 | ||
7621 | vec_cond_rhs = | |
7622 | vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), | |
7623 | stmt, comp_vectype); | |
7624 | vect_is_simple_use (TREE_OPERAND (cond_expr, 1), | |
7625 | loop_vinfo, >emp, &dts[1]); | |
7626 | } | |
f7e531cf IR |
7627 | if (reduc_index == 1) |
7628 | vec_then_clause = reduc_def; | |
7629 | else | |
7630 | { | |
7631 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
81c40241 RB |
7632 | stmt); |
7633 | vect_is_simple_use (then_clause, loop_vinfo, | |
7634 | >emp, &dts[2]); | |
f7e531cf IR |
7635 | } |
7636 | if (reduc_index == 2) | |
7637 | vec_else_clause = reduc_def; | |
7638 | else | |
7639 | { | |
7640 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
81c40241 RB |
7641 | stmt); |
7642 | vect_is_simple_use (else_clause, loop_vinfo, >emp, &dts[3]); | |
f7e531cf | 7643 | } |
a855b1b1 MM |
7644 | } |
7645 | } | |
7646 | else | |
7647 | { | |
a414c77f IE |
7648 | vec_cond_lhs |
7649 | = vect_get_vec_def_for_stmt_copy (dts[0], | |
7650 | vec_oprnds0.pop ()); | |
7651 | if (!masked) | |
7652 | vec_cond_rhs | |
7653 | = vect_get_vec_def_for_stmt_copy (dts[1], | |
7654 | vec_oprnds1.pop ()); | |
7655 | ||
a855b1b1 | 7656 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 7657 | vec_oprnds2.pop ()); |
a855b1b1 | 7658 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 7659 | vec_oprnds3.pop ()); |
f7e531cf IR |
7660 | } |
7661 | ||
7662 | if (!slp_node) | |
7663 | { | |
9771b263 | 7664 | vec_oprnds0.quick_push (vec_cond_lhs); |
a414c77f IE |
7665 | if (!masked) |
7666 | vec_oprnds1.quick_push (vec_cond_rhs); | |
9771b263 DN |
7667 | vec_oprnds2.quick_push (vec_then_clause); |
7668 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
7669 | } |
7670 | ||
9dc3f7de | 7671 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 7672 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 7673 | { |
9771b263 DN |
7674 | vec_then_clause = vec_oprnds2[i]; |
7675 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 7676 | |
a414c77f IE |
7677 | if (masked) |
7678 | vec_compare = vec_cond_lhs; | |
7679 | else | |
7680 | { | |
7681 | vec_cond_rhs = vec_oprnds1[i]; | |
7682 | vec_compare = build2 (TREE_CODE (cond_expr), vec_cmp_type, | |
7683 | vec_cond_lhs, vec_cond_rhs); | |
7684 | } | |
5958f9e2 JJ |
7685 | new_temp = make_ssa_name (vec_dest); |
7686 | new_stmt = gimple_build_assign (new_temp, VEC_COND_EXPR, | |
7687 | vec_compare, vec_then_clause, | |
7688 | vec_else_clause); | |
f7e531cf IR |
7689 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7690 | if (slp_node) | |
9771b263 | 7691 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
7692 | } |
7693 | ||
7694 | if (slp_node) | |
7695 | continue; | |
7696 | ||
7697 | if (j == 0) | |
7698 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7699 | else | |
7700 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7701 | ||
7702 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 7703 | } |
b8698a0f | 7704 | |
9771b263 DN |
7705 | vec_oprnds0.release (); |
7706 | vec_oprnds1.release (); | |
7707 | vec_oprnds2.release (); | |
7708 | vec_oprnds3.release (); | |
f7e531cf | 7709 | |
ebfd146a IR |
7710 | return true; |
7711 | } | |
7712 | ||
42fd8198 IE |
7713 | /* vectorizable_comparison. |
7714 | ||
7715 | Check if STMT is comparison expression that can be vectorized. | |
7716 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
7717 | comparison, put it in VEC_STMT, and insert it at GSI. | |
7718 | ||
7719 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7720 | ||
fce57248 | 7721 | static bool |
42fd8198 IE |
7722 | vectorizable_comparison (gimple *stmt, gimple_stmt_iterator *gsi, |
7723 | gimple **vec_stmt, tree reduc_def, | |
7724 | slp_tree slp_node) | |
7725 | { | |
7726 | tree lhs, rhs1, rhs2; | |
7727 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
7728 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; | |
7729 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
7730 | tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE; | |
7731 | tree new_temp; | |
7732 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
7733 | enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
7734 | unsigned nunits; | |
7735 | int ncopies; | |
49e76ff1 | 7736 | enum tree_code code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
42fd8198 IE |
7737 | stmt_vec_info prev_stmt_info = NULL; |
7738 | int i, j; | |
7739 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
7740 | vec<tree> vec_oprnds0 = vNULL; | |
7741 | vec<tree> vec_oprnds1 = vNULL; | |
7742 | gimple *def_stmt; | |
7743 | tree mask_type; | |
7744 | tree mask; | |
7745 | ||
c245362b IE |
7746 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
7747 | return false; | |
7748 | ||
30480bcd | 7749 | if (!vectype || !VECTOR_BOOLEAN_TYPE_P (vectype)) |
42fd8198 IE |
7750 | return false; |
7751 | ||
7752 | mask_type = vectype; | |
7753 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
7754 | ||
fce57248 | 7755 | if (slp_node) |
42fd8198 IE |
7756 | ncopies = 1; |
7757 | else | |
7758 | ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; | |
7759 | ||
7760 | gcc_assert (ncopies >= 1); | |
42fd8198 IE |
7761 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7762 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
7763 | && reduc_def)) | |
7764 | return false; | |
7765 | ||
7766 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
7767 | { | |
7768 | if (dump_enabled_p ()) | |
7769 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7770 | "value used after loop.\n"); | |
7771 | return false; | |
7772 | } | |
7773 | ||
7774 | if (!is_gimple_assign (stmt)) | |
7775 | return false; | |
7776 | ||
7777 | code = gimple_assign_rhs_code (stmt); | |
7778 | ||
7779 | if (TREE_CODE_CLASS (code) != tcc_comparison) | |
7780 | return false; | |
7781 | ||
7782 | rhs1 = gimple_assign_rhs1 (stmt); | |
7783 | rhs2 = gimple_assign_rhs2 (stmt); | |
7784 | ||
7785 | if (!vect_is_simple_use (rhs1, stmt_info->vinfo, &def_stmt, | |
7786 | &dts[0], &vectype1)) | |
7787 | return false; | |
7788 | ||
7789 | if (!vect_is_simple_use (rhs2, stmt_info->vinfo, &def_stmt, | |
7790 | &dts[1], &vectype2)) | |
7791 | return false; | |
7792 | ||
7793 | if (vectype1 && vectype2 | |
7794 | && TYPE_VECTOR_SUBPARTS (vectype1) != TYPE_VECTOR_SUBPARTS (vectype2)) | |
7795 | return false; | |
7796 | ||
7797 | vectype = vectype1 ? vectype1 : vectype2; | |
7798 | ||
7799 | /* Invariant comparison. */ | |
7800 | if (!vectype) | |
7801 | { | |
69a9a66f RB |
7802 | vectype = get_vectype_for_scalar_type (TREE_TYPE (rhs1)); |
7803 | if (TYPE_VECTOR_SUBPARTS (vectype) != nunits) | |
42fd8198 IE |
7804 | return false; |
7805 | } | |
7806 | else if (nunits != TYPE_VECTOR_SUBPARTS (vectype)) | |
7807 | return false; | |
7808 | ||
49e76ff1 IE |
7809 | /* Can't compare mask and non-mask types. */ |
7810 | if (vectype1 && vectype2 | |
7811 | && (VECTOR_BOOLEAN_TYPE_P (vectype1) ^ VECTOR_BOOLEAN_TYPE_P (vectype2))) | |
7812 | return false; | |
7813 | ||
7814 | /* Boolean values may have another representation in vectors | |
7815 | and therefore we prefer bit operations over comparison for | |
7816 | them (which also works for scalar masks). We store opcodes | |
7817 | to use in bitop1 and bitop2. Statement is vectorized as | |
7818 | BITOP2 (rhs1 BITOP1 rhs2) or | |
7819 | rhs1 BITOP2 (BITOP1 rhs2) | |
7820 | depending on bitop1 and bitop2 arity. */ | |
7821 | if (VECTOR_BOOLEAN_TYPE_P (vectype)) | |
7822 | { | |
7823 | if (code == GT_EXPR) | |
7824 | { | |
7825 | bitop1 = BIT_NOT_EXPR; | |
7826 | bitop2 = BIT_AND_EXPR; | |
7827 | } | |
7828 | else if (code == GE_EXPR) | |
7829 | { | |
7830 | bitop1 = BIT_NOT_EXPR; | |
7831 | bitop2 = BIT_IOR_EXPR; | |
7832 | } | |
7833 | else if (code == LT_EXPR) | |
7834 | { | |
7835 | bitop1 = BIT_NOT_EXPR; | |
7836 | bitop2 = BIT_AND_EXPR; | |
7837 | std::swap (rhs1, rhs2); | |
264d951a | 7838 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
7839 | } |
7840 | else if (code == LE_EXPR) | |
7841 | { | |
7842 | bitop1 = BIT_NOT_EXPR; | |
7843 | bitop2 = BIT_IOR_EXPR; | |
7844 | std::swap (rhs1, rhs2); | |
264d951a | 7845 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
7846 | } |
7847 | else | |
7848 | { | |
7849 | bitop1 = BIT_XOR_EXPR; | |
7850 | if (code == EQ_EXPR) | |
7851 | bitop2 = BIT_NOT_EXPR; | |
7852 | } | |
7853 | } | |
7854 | ||
42fd8198 IE |
7855 | if (!vec_stmt) |
7856 | { | |
7857 | STMT_VINFO_TYPE (stmt_info) = comparison_vec_info_type; | |
49e76ff1 IE |
7858 | vect_model_simple_cost (stmt_info, ncopies * (1 + (bitop2 != NOP_EXPR)), |
7859 | dts, NULL, NULL); | |
7860 | if (bitop1 == NOP_EXPR) | |
7861 | return expand_vec_cmp_expr_p (vectype, mask_type); | |
7862 | else | |
7863 | { | |
7864 | machine_mode mode = TYPE_MODE (vectype); | |
7865 | optab optab; | |
7866 | ||
7867 | optab = optab_for_tree_code (bitop1, vectype, optab_default); | |
7868 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
7869 | return false; | |
7870 | ||
7871 | if (bitop2 != NOP_EXPR) | |
7872 | { | |
7873 | optab = optab_for_tree_code (bitop2, vectype, optab_default); | |
7874 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
7875 | return false; | |
7876 | } | |
7877 | return true; | |
7878 | } | |
42fd8198 IE |
7879 | } |
7880 | ||
7881 | /* Transform. */ | |
7882 | if (!slp_node) | |
7883 | { | |
7884 | vec_oprnds0.create (1); | |
7885 | vec_oprnds1.create (1); | |
7886 | } | |
7887 | ||
7888 | /* Handle def. */ | |
7889 | lhs = gimple_assign_lhs (stmt); | |
7890 | mask = vect_create_destination_var (lhs, mask_type); | |
7891 | ||
7892 | /* Handle cmp expr. */ | |
7893 | for (j = 0; j < ncopies; j++) | |
7894 | { | |
7895 | gassign *new_stmt = NULL; | |
7896 | if (j == 0) | |
7897 | { | |
7898 | if (slp_node) | |
7899 | { | |
7900 | auto_vec<tree, 2> ops; | |
7901 | auto_vec<vec<tree>, 2> vec_defs; | |
7902 | ||
7903 | ops.safe_push (rhs1); | |
7904 | ops.safe_push (rhs2); | |
7905 | vect_get_slp_defs (ops, slp_node, &vec_defs, -1); | |
7906 | vec_oprnds1 = vec_defs.pop (); | |
7907 | vec_oprnds0 = vec_defs.pop (); | |
7908 | } | |
7909 | else | |
7910 | { | |
e4af0bc4 IE |
7911 | vec_rhs1 = vect_get_vec_def_for_operand (rhs1, stmt, vectype); |
7912 | vec_rhs2 = vect_get_vec_def_for_operand (rhs2, stmt, vectype); | |
42fd8198 IE |
7913 | } |
7914 | } | |
7915 | else | |
7916 | { | |
7917 | vec_rhs1 = vect_get_vec_def_for_stmt_copy (dts[0], | |
7918 | vec_oprnds0.pop ()); | |
7919 | vec_rhs2 = vect_get_vec_def_for_stmt_copy (dts[1], | |
7920 | vec_oprnds1.pop ()); | |
7921 | } | |
7922 | ||
7923 | if (!slp_node) | |
7924 | { | |
7925 | vec_oprnds0.quick_push (vec_rhs1); | |
7926 | vec_oprnds1.quick_push (vec_rhs2); | |
7927 | } | |
7928 | ||
7929 | /* Arguments are ready. Create the new vector stmt. */ | |
7930 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1) | |
7931 | { | |
7932 | vec_rhs2 = vec_oprnds1[i]; | |
7933 | ||
7934 | new_temp = make_ssa_name (mask); | |
49e76ff1 IE |
7935 | if (bitop1 == NOP_EXPR) |
7936 | { | |
7937 | new_stmt = gimple_build_assign (new_temp, code, | |
7938 | vec_rhs1, vec_rhs2); | |
7939 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7940 | } | |
7941 | else | |
7942 | { | |
7943 | if (bitop1 == BIT_NOT_EXPR) | |
7944 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs2); | |
7945 | else | |
7946 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs1, | |
7947 | vec_rhs2); | |
7948 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7949 | if (bitop2 != NOP_EXPR) | |
7950 | { | |
7951 | tree res = make_ssa_name (mask); | |
7952 | if (bitop2 == BIT_NOT_EXPR) | |
7953 | new_stmt = gimple_build_assign (res, bitop2, new_temp); | |
7954 | else | |
7955 | new_stmt = gimple_build_assign (res, bitop2, vec_rhs1, | |
7956 | new_temp); | |
7957 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7958 | } | |
7959 | } | |
42fd8198 IE |
7960 | if (slp_node) |
7961 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
7962 | } | |
7963 | ||
7964 | if (slp_node) | |
7965 | continue; | |
7966 | ||
7967 | if (j == 0) | |
7968 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7969 | else | |
7970 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7971 | ||
7972 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7973 | } | |
7974 | ||
7975 | vec_oprnds0.release (); | |
7976 | vec_oprnds1.release (); | |
7977 | ||
7978 | return true; | |
7979 | } | |
ebfd146a | 7980 | |
8644a673 | 7981 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
7982 | |
7983 | bool | |
355fe088 | 7984 | vect_analyze_stmt (gimple *stmt, bool *need_to_vectorize, slp_tree node) |
ebfd146a | 7985 | { |
8644a673 | 7986 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 7987 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 7988 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 7989 | bool ok; |
a70d6342 | 7990 | tree scalar_type, vectype; |
355fe088 | 7991 | gimple *pattern_stmt; |
363477c0 | 7992 | gimple_seq pattern_def_seq; |
ebfd146a | 7993 | |
73fbfcad | 7994 | if (dump_enabled_p ()) |
ebfd146a | 7995 | { |
78c60e3d SS |
7996 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
7997 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
8644a673 | 7998 | } |
ebfd146a | 7999 | |
1825a1f3 | 8000 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 8001 | { |
73fbfcad | 8002 | if (dump_enabled_p ()) |
78c60e3d | 8003 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8004 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
8005 | |
8006 | return false; | |
8007 | } | |
b8698a0f L |
8008 | |
8009 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
8010 | to include: |
8011 | - the COND_EXPR which is the loop exit condition | |
8012 | - any LABEL_EXPRs in the loop | |
b8698a0f | 8013 | - computations that are used only for array indexing or loop control. |
8644a673 | 8014 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 8015 | instance, therefore, all the statements are relevant. |
ebfd146a | 8016 | |
d092494c | 8017 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 8018 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
8019 | statements. In basic blocks we are called from some SLP instance |
8020 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
8021 | already will be part of SLP instance. */ | |
83197f37 IR |
8022 | |
8023 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 8024 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 8025 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 8026 | { |
9d5e7640 | 8027 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 8028 | && pattern_stmt |
9d5e7640 IR |
8029 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
8030 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
8031 | { | |
83197f37 | 8032 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
8033 | stmt = pattern_stmt; |
8034 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 8035 | if (dump_enabled_p ()) |
9d5e7640 | 8036 | { |
78c60e3d SS |
8037 | dump_printf_loc (MSG_NOTE, vect_location, |
8038 | "==> examining pattern statement: "); | |
8039 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
9d5e7640 IR |
8040 | } |
8041 | } | |
8042 | else | |
8043 | { | |
73fbfcad | 8044 | if (dump_enabled_p ()) |
e645e942 | 8045 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 8046 | |
9d5e7640 IR |
8047 | return true; |
8048 | } | |
8644a673 | 8049 | } |
83197f37 | 8050 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 8051 | && node == NULL |
83197f37 IR |
8052 | && pattern_stmt |
8053 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
8054 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
8055 | { | |
8056 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 8057 | if (dump_enabled_p ()) |
83197f37 | 8058 | { |
78c60e3d SS |
8059 | dump_printf_loc (MSG_NOTE, vect_location, |
8060 | "==> examining pattern statement: "); | |
8061 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
83197f37 IR |
8062 | } |
8063 | ||
8064 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node)) | |
8065 | return false; | |
8066 | } | |
ebfd146a | 8067 | |
1107f3ae | 8068 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 8069 | && node == NULL |
363477c0 | 8070 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 8071 | { |
363477c0 | 8072 | gimple_stmt_iterator si; |
1107f3ae | 8073 | |
363477c0 JJ |
8074 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
8075 | { | |
355fe088 | 8076 | gimple *pattern_def_stmt = gsi_stmt (si); |
363477c0 JJ |
8077 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) |
8078 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
8079 | { | |
8080 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 8081 | if (dump_enabled_p ()) |
363477c0 | 8082 | { |
78c60e3d SS |
8083 | dump_printf_loc (MSG_NOTE, vect_location, |
8084 | "==> examining pattern def statement: "); | |
8085 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
363477c0 | 8086 | } |
1107f3ae | 8087 | |
363477c0 JJ |
8088 | if (!vect_analyze_stmt (pattern_def_stmt, |
8089 | need_to_vectorize, node)) | |
8090 | return false; | |
8091 | } | |
8092 | } | |
8093 | } | |
1107f3ae | 8094 | |
8644a673 IR |
8095 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
8096 | { | |
8097 | case vect_internal_def: | |
8098 | break; | |
ebfd146a | 8099 | |
8644a673 | 8100 | case vect_reduction_def: |
7c5222ff | 8101 | case vect_nested_cycle: |
14a61437 RB |
8102 | gcc_assert (!bb_vinfo |
8103 | && (relevance == vect_used_in_outer | |
8104 | || relevance == vect_used_in_outer_by_reduction | |
8105 | || relevance == vect_used_by_reduction | |
b28ead45 AH |
8106 | || relevance == vect_unused_in_scope |
8107 | || relevance == vect_used_only_live)); | |
8644a673 IR |
8108 | break; |
8109 | ||
8110 | case vect_induction_def: | |
8111 | case vect_constant_def: | |
8112 | case vect_external_def: | |
8113 | case vect_unknown_def_type: | |
8114 | default: | |
8115 | gcc_unreachable (); | |
8116 | } | |
ebfd146a | 8117 | |
a70d6342 IR |
8118 | if (bb_vinfo) |
8119 | { | |
8120 | gcc_assert (PURE_SLP_STMT (stmt_info)); | |
8121 | ||
b690cc0f | 8122 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); |
73fbfcad | 8123 | if (dump_enabled_p ()) |
a70d6342 | 8124 | { |
78c60e3d SS |
8125 | dump_printf_loc (MSG_NOTE, vect_location, |
8126 | "get vectype for scalar type: "); | |
8127 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
e645e942 | 8128 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
8129 | } |
8130 | ||
8131 | vectype = get_vectype_for_scalar_type (scalar_type); | |
8132 | if (!vectype) | |
8133 | { | |
73fbfcad | 8134 | if (dump_enabled_p ()) |
a70d6342 | 8135 | { |
78c60e3d SS |
8136 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8137 | "not SLPed: unsupported data-type "); | |
8138 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
8139 | scalar_type); | |
e645e942 | 8140 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
a70d6342 IR |
8141 | } |
8142 | return false; | |
8143 | } | |
8144 | ||
73fbfcad | 8145 | if (dump_enabled_p ()) |
a70d6342 | 8146 | { |
78c60e3d SS |
8147 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); |
8148 | dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); | |
e645e942 | 8149 | dump_printf (MSG_NOTE, "\n"); |
a70d6342 IR |
8150 | } |
8151 | ||
8152 | STMT_VINFO_VECTYPE (stmt_info) = vectype; | |
8153 | } | |
8154 | ||
8644a673 | 8155 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 8156 | { |
8644a673 | 8157 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
0136f8f0 AH |
8158 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
8159 | || (is_gimple_call (stmt) | |
8160 | && gimple_call_lhs (stmt) == NULL_TREE)); | |
8644a673 | 8161 | *need_to_vectorize = true; |
ebfd146a IR |
8162 | } |
8163 | ||
b1af7da6 RB |
8164 | if (PURE_SLP_STMT (stmt_info) && !node) |
8165 | { | |
8166 | dump_printf_loc (MSG_NOTE, vect_location, | |
8167 | "handled only by SLP analysis\n"); | |
8168 | return true; | |
8169 | } | |
8170 | ||
8171 | ok = true; | |
8172 | if (!bb_vinfo | |
8173 | && (STMT_VINFO_RELEVANT_P (stmt_info) | |
8174 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
8175 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
8176 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
8177 | || vectorizable_shift (stmt, NULL, NULL, node) | |
8178 | || vectorizable_operation (stmt, NULL, NULL, node) | |
8179 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
8180 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
8181 | || vectorizable_call (stmt, NULL, NULL, node) | |
8182 | || vectorizable_store (stmt, NULL, NULL, node) | |
8183 | || vectorizable_reduction (stmt, NULL, NULL, node) | |
42fd8198 IE |
8184 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
8185 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 RB |
8186 | else |
8187 | { | |
8188 | if (bb_vinfo) | |
8189 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
8190 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
8191 | || vectorizable_shift (stmt, NULL, NULL, node) | |
8192 | || vectorizable_operation (stmt, NULL, NULL, node) | |
8193 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
8194 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
8195 | || vectorizable_call (stmt, NULL, NULL, node) | |
8196 | || vectorizable_store (stmt, NULL, NULL, node) | |
42fd8198 IE |
8197 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
8198 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 | 8199 | } |
8644a673 IR |
8200 | |
8201 | if (!ok) | |
ebfd146a | 8202 | { |
73fbfcad | 8203 | if (dump_enabled_p ()) |
8644a673 | 8204 | { |
78c60e3d SS |
8205 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8206 | "not vectorized: relevant stmt not "); | |
8207 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
8208 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 8209 | } |
b8698a0f | 8210 | |
ebfd146a IR |
8211 | return false; |
8212 | } | |
8213 | ||
a70d6342 IR |
8214 | if (bb_vinfo) |
8215 | return true; | |
8216 | ||
8644a673 IR |
8217 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
8218 | need extra handling, except for vectorizable reductions. */ | |
8219 | if (STMT_VINFO_LIVE_P (stmt_info) | |
8220 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) | |
b28ead45 | 8221 | ok = vectorizable_live_operation (stmt, NULL, NULL, -1, NULL); |
ebfd146a | 8222 | |
8644a673 | 8223 | if (!ok) |
ebfd146a | 8224 | { |
73fbfcad | 8225 | if (dump_enabled_p ()) |
8644a673 | 8226 | { |
78c60e3d SS |
8227 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8228 | "not vectorized: live stmt not "); | |
8229 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
8230 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 8231 | } |
b8698a0f | 8232 | |
8644a673 | 8233 | return false; |
ebfd146a IR |
8234 | } |
8235 | ||
ebfd146a IR |
8236 | return true; |
8237 | } | |
8238 | ||
8239 | ||
8240 | /* Function vect_transform_stmt. | |
8241 | ||
8242 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
8243 | ||
8244 | bool | |
355fe088 | 8245 | vect_transform_stmt (gimple *stmt, gimple_stmt_iterator *gsi, |
0d0293ac | 8246 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
8247 | slp_instance slp_node_instance) |
8248 | { | |
8249 | bool is_store = false; | |
355fe088 | 8250 | gimple *vec_stmt = NULL; |
ebfd146a | 8251 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
ebfd146a | 8252 | bool done; |
ebfd146a | 8253 | |
fce57248 | 8254 | gcc_assert (slp_node || !PURE_SLP_STMT (stmt_info)); |
355fe088 | 8255 | gimple *old_vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
225ce44b | 8256 | |
ebfd146a IR |
8257 | switch (STMT_VINFO_TYPE (stmt_info)) |
8258 | { | |
8259 | case type_demotion_vec_info_type: | |
ebfd146a | 8260 | case type_promotion_vec_info_type: |
ebfd146a IR |
8261 | case type_conversion_vec_info_type: |
8262 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
8263 | gcc_assert (done); | |
8264 | break; | |
8265 | ||
8266 | case induc_vec_info_type: | |
8267 | gcc_assert (!slp_node); | |
8268 | done = vectorizable_induction (stmt, gsi, &vec_stmt); | |
8269 | gcc_assert (done); | |
8270 | break; | |
8271 | ||
9dc3f7de IR |
8272 | case shift_vec_info_type: |
8273 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
8274 | gcc_assert (done); | |
8275 | break; | |
8276 | ||
ebfd146a IR |
8277 | case op_vec_info_type: |
8278 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
8279 | gcc_assert (done); | |
8280 | break; | |
8281 | ||
8282 | case assignment_vec_info_type: | |
8283 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
8284 | gcc_assert (done); | |
8285 | break; | |
8286 | ||
8287 | case load_vec_info_type: | |
b8698a0f | 8288 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
8289 | slp_node_instance); |
8290 | gcc_assert (done); | |
8291 | break; | |
8292 | ||
8293 | case store_vec_info_type: | |
8294 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
8295 | gcc_assert (done); | |
0d0293ac | 8296 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
8297 | { |
8298 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 8299 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
8300 | one are skipped, and there vec_stmt_info shouldn't be freed |
8301 | meanwhile. */ | |
0d0293ac | 8302 | *grouped_store = true; |
ebfd146a IR |
8303 | if (STMT_VINFO_VEC_STMT (stmt_info)) |
8304 | is_store = true; | |
8305 | } | |
8306 | else | |
8307 | is_store = true; | |
8308 | break; | |
8309 | ||
8310 | case condition_vec_info_type: | |
f7e531cf | 8311 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
8312 | gcc_assert (done); |
8313 | break; | |
8314 | ||
42fd8198 IE |
8315 | case comparison_vec_info_type: |
8316 | done = vectorizable_comparison (stmt, gsi, &vec_stmt, NULL, slp_node); | |
8317 | gcc_assert (done); | |
8318 | break; | |
8319 | ||
ebfd146a | 8320 | case call_vec_info_type: |
190c2236 | 8321 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 8322 | stmt = gsi_stmt (*gsi); |
5ce9450f JJ |
8323 | if (is_gimple_call (stmt) |
8324 | && gimple_call_internal_p (stmt) | |
8325 | && gimple_call_internal_fn (stmt) == IFN_MASK_STORE) | |
8326 | is_store = true; | |
ebfd146a IR |
8327 | break; |
8328 | ||
0136f8f0 AH |
8329 | case call_simd_clone_vec_info_type: |
8330 | done = vectorizable_simd_clone_call (stmt, gsi, &vec_stmt, slp_node); | |
8331 | stmt = gsi_stmt (*gsi); | |
8332 | break; | |
8333 | ||
ebfd146a | 8334 | case reduc_vec_info_type: |
b5aeb3bb | 8335 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
8336 | gcc_assert (done); |
8337 | break; | |
8338 | ||
8339 | default: | |
8340 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
8341 | { | |
73fbfcad | 8342 | if (dump_enabled_p ()) |
78c60e3d | 8343 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8344 | "stmt not supported.\n"); |
ebfd146a IR |
8345 | gcc_unreachable (); |
8346 | } | |
8347 | } | |
8348 | ||
225ce44b RB |
8349 | /* Verify SLP vectorization doesn't mess with STMT_VINFO_VEC_STMT. |
8350 | This would break hybrid SLP vectorization. */ | |
8351 | if (slp_node) | |
d90f8440 RB |
8352 | gcc_assert (!vec_stmt |
8353 | && STMT_VINFO_VEC_STMT (stmt_info) == old_vec_stmt); | |
225ce44b | 8354 | |
ebfd146a IR |
8355 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that |
8356 | is being vectorized, but outside the immediately enclosing loop. */ | |
8357 | if (vec_stmt | |
a70d6342 IR |
8358 | && STMT_VINFO_LOOP_VINFO (stmt_info) |
8359 | && nested_in_vect_loop_p (LOOP_VINFO_LOOP ( | |
8360 | STMT_VINFO_LOOP_VINFO (stmt_info)), stmt) | |
ebfd146a IR |
8361 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
8362 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 8363 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 8364 | vect_used_in_outer_by_reduction)) |
ebfd146a | 8365 | { |
a70d6342 IR |
8366 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
8367 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
8368 | imm_use_iterator imm_iter; |
8369 | use_operand_p use_p; | |
8370 | tree scalar_dest; | |
355fe088 | 8371 | gimple *exit_phi; |
ebfd146a | 8372 | |
73fbfcad | 8373 | if (dump_enabled_p ()) |
78c60e3d | 8374 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 8375 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
8376 | |
8377 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
8378 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
8379 | STMT). */ | |
8380 | if (gimple_code (stmt) == GIMPLE_PHI) | |
8381 | scalar_dest = PHI_RESULT (stmt); | |
8382 | else | |
8383 | scalar_dest = gimple_assign_lhs (stmt); | |
8384 | ||
8385 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
8386 | { | |
8387 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
8388 | { | |
8389 | exit_phi = USE_STMT (use_p); | |
8390 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
8391 | } | |
8392 | } | |
8393 | } | |
8394 | ||
8395 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
8396 | being vectorized. */ | |
b28ead45 AH |
8397 | if (slp_node) |
8398 | { | |
8399 | gimple *slp_stmt; | |
8400 | int i; | |
8401 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (slp_node), i, slp_stmt) | |
8402 | { | |
8403 | stmt_vec_info slp_stmt_info = vinfo_for_stmt (slp_stmt); | |
8404 | if (STMT_VINFO_LIVE_P (slp_stmt_info) | |
8405 | && STMT_VINFO_TYPE (slp_stmt_info) != reduc_vec_info_type) | |
8406 | { | |
8407 | done = vectorizable_live_operation (slp_stmt, gsi, slp_node, i, | |
8408 | &vec_stmt); | |
8409 | gcc_assert (done); | |
8410 | } | |
8411 | } | |
8412 | } | |
8413 | else if (STMT_VINFO_LIVE_P (stmt_info) | |
ebfd146a IR |
8414 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) |
8415 | { | |
b28ead45 | 8416 | done = vectorizable_live_operation (stmt, gsi, slp_node, -1, &vec_stmt); |
ebfd146a IR |
8417 | gcc_assert (done); |
8418 | } | |
8419 | ||
8420 | if (vec_stmt) | |
83197f37 | 8421 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 8422 | |
b8698a0f | 8423 | return is_store; |
ebfd146a IR |
8424 | } |
8425 | ||
8426 | ||
b8698a0f | 8427 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
8428 | stmt_vec_info. */ |
8429 | ||
8430 | void | |
355fe088 | 8431 | vect_remove_stores (gimple *first_stmt) |
ebfd146a | 8432 | { |
355fe088 TS |
8433 | gimple *next = first_stmt; |
8434 | gimple *tmp; | |
ebfd146a IR |
8435 | gimple_stmt_iterator next_si; |
8436 | ||
8437 | while (next) | |
8438 | { | |
78048b1c JJ |
8439 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
8440 | ||
8441 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
8442 | if (is_pattern_stmt_p (stmt_info)) | |
8443 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
8444 | /* Free the attached stmt_vec_info and remove the stmt. */ |
8445 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 8446 | unlink_stmt_vdef (next); |
ebfd146a | 8447 | gsi_remove (&next_si, true); |
3d3f2249 | 8448 | release_defs (next); |
ebfd146a IR |
8449 | free_stmt_vec_info (next); |
8450 | next = tmp; | |
8451 | } | |
8452 | } | |
8453 | ||
8454 | ||
8455 | /* Function new_stmt_vec_info. | |
8456 | ||
8457 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
8458 | ||
8459 | stmt_vec_info | |
310213d4 | 8460 | new_stmt_vec_info (gimple *stmt, vec_info *vinfo) |
ebfd146a IR |
8461 | { |
8462 | stmt_vec_info res; | |
8463 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
8464 | ||
8465 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
8466 | STMT_VINFO_STMT (res) = stmt; | |
310213d4 | 8467 | res->vinfo = vinfo; |
8644a673 | 8468 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
8469 | STMT_VINFO_LIVE_P (res) = false; |
8470 | STMT_VINFO_VECTYPE (res) = NULL; | |
8471 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 8472 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
8473 | STMT_VINFO_IN_PATTERN_P (res) = false; |
8474 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 8475 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a | 8476 | STMT_VINFO_DATA_REF (res) = NULL; |
af29617a | 8477 | STMT_VINFO_VEC_REDUCTION_TYPE (res) = TREE_CODE_REDUCTION; |
ebfd146a IR |
8478 | |
8479 | STMT_VINFO_DR_BASE_ADDRESS (res) = NULL; | |
8480 | STMT_VINFO_DR_OFFSET (res) = NULL; | |
8481 | STMT_VINFO_DR_INIT (res) = NULL; | |
8482 | STMT_VINFO_DR_STEP (res) = NULL; | |
8483 | STMT_VINFO_DR_ALIGNED_TO (res) = NULL; | |
8484 | ||
8485 | if (gimple_code (stmt) == GIMPLE_PHI | |
8486 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
8487 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
8488 | else | |
8644a673 IR |
8489 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
8490 | ||
9771b263 | 8491 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 8492 | STMT_SLP_TYPE (res) = loop_vect; |
78810bd3 RB |
8493 | STMT_VINFO_NUM_SLP_USES (res) = 0; |
8494 | ||
e14c1050 IR |
8495 | GROUP_FIRST_ELEMENT (res) = NULL; |
8496 | GROUP_NEXT_ELEMENT (res) = NULL; | |
8497 | GROUP_SIZE (res) = 0; | |
8498 | GROUP_STORE_COUNT (res) = 0; | |
8499 | GROUP_GAP (res) = 0; | |
8500 | GROUP_SAME_DR_STMT (res) = NULL; | |
ebfd146a IR |
8501 | |
8502 | return res; | |
8503 | } | |
8504 | ||
8505 | ||
8506 | /* Create a hash table for stmt_vec_info. */ | |
8507 | ||
8508 | void | |
8509 | init_stmt_vec_info_vec (void) | |
8510 | { | |
9771b263 DN |
8511 | gcc_assert (!stmt_vec_info_vec.exists ()); |
8512 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
8513 | } |
8514 | ||
8515 | ||
8516 | /* Free hash table for stmt_vec_info. */ | |
8517 | ||
8518 | void | |
8519 | free_stmt_vec_info_vec (void) | |
8520 | { | |
93675444 | 8521 | unsigned int i; |
3161455c | 8522 | stmt_vec_info info; |
93675444 JJ |
8523 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) |
8524 | if (info != NULL) | |
3161455c | 8525 | free_stmt_vec_info (STMT_VINFO_STMT (info)); |
9771b263 DN |
8526 | gcc_assert (stmt_vec_info_vec.exists ()); |
8527 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
8528 | } |
8529 | ||
8530 | ||
8531 | /* Free stmt vectorization related info. */ | |
8532 | ||
8533 | void | |
355fe088 | 8534 | free_stmt_vec_info (gimple *stmt) |
ebfd146a IR |
8535 | { |
8536 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8537 | ||
8538 | if (!stmt_info) | |
8539 | return; | |
8540 | ||
78048b1c JJ |
8541 | /* Check if this statement has a related "pattern stmt" |
8542 | (introduced by the vectorizer during the pattern recognition | |
8543 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
8544 | too. */ | |
8545 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
8546 | { | |
8547 | stmt_vec_info patt_info | |
8548 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
8549 | if (patt_info) | |
8550 | { | |
363477c0 | 8551 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
355fe088 | 8552 | gimple *patt_stmt = STMT_VINFO_STMT (patt_info); |
f0281fde RB |
8553 | gimple_set_bb (patt_stmt, NULL); |
8554 | tree lhs = gimple_get_lhs (patt_stmt); | |
e6f5c25d | 8555 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde | 8556 | release_ssa_name (lhs); |
363477c0 JJ |
8557 | if (seq) |
8558 | { | |
8559 | gimple_stmt_iterator si; | |
8560 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
f0281fde | 8561 | { |
355fe088 | 8562 | gimple *seq_stmt = gsi_stmt (si); |
f0281fde | 8563 | gimple_set_bb (seq_stmt, NULL); |
7532abf2 | 8564 | lhs = gimple_get_lhs (seq_stmt); |
e6f5c25d | 8565 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde RB |
8566 | release_ssa_name (lhs); |
8567 | free_stmt_vec_info (seq_stmt); | |
8568 | } | |
363477c0 | 8569 | } |
f0281fde | 8570 | free_stmt_vec_info (patt_stmt); |
78048b1c JJ |
8571 | } |
8572 | } | |
8573 | ||
9771b263 | 8574 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
6c9e85fb | 8575 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).release (); |
ebfd146a IR |
8576 | set_vinfo_for_stmt (stmt, NULL); |
8577 | free (stmt_info); | |
8578 | } | |
8579 | ||
8580 | ||
bb67d9c7 | 8581 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 8582 | |
bb67d9c7 | 8583 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
8584 | by the target. */ |
8585 | ||
bb67d9c7 RG |
8586 | static tree |
8587 | get_vectype_for_scalar_type_and_size (tree scalar_type, unsigned size) | |
ebfd146a | 8588 | { |
ef4bddc2 RS |
8589 | machine_mode inner_mode = TYPE_MODE (scalar_type); |
8590 | machine_mode simd_mode; | |
2f816591 | 8591 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
ebfd146a IR |
8592 | int nunits; |
8593 | tree vectype; | |
8594 | ||
cc4b5170 | 8595 | if (nbytes == 0) |
ebfd146a IR |
8596 | return NULL_TREE; |
8597 | ||
48f2e373 RB |
8598 | if (GET_MODE_CLASS (inner_mode) != MODE_INT |
8599 | && GET_MODE_CLASS (inner_mode) != MODE_FLOAT) | |
8600 | return NULL_TREE; | |
8601 | ||
7b7b1813 RG |
8602 | /* For vector types of elements whose mode precision doesn't |
8603 | match their types precision we use a element type of mode | |
8604 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
8605 | they support the proper result truncation/extension. |
8606 | We also make sure to build vector types with INTEGER_TYPE | |
8607 | component type only. */ | |
6d7971b8 | 8608 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
8609 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
8610 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
8611 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
8612 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 8613 | |
ccbf5bb4 RG |
8614 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
8615 | When the component mode passes the above test simply use a type | |
8616 | corresponding to that mode. The theory is that any use that | |
8617 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 8618 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 8619 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
8620 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
8621 | ||
8622 | /* We can't build a vector type of elements with alignment bigger than | |
8623 | their size. */ | |
dfc2e2ac | 8624 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
8625 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
8626 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 8627 | |
dfc2e2ac RB |
8628 | /* If we felt back to using the mode fail if there was |
8629 | no scalar type for it. */ | |
8630 | if (scalar_type == NULL_TREE) | |
8631 | return NULL_TREE; | |
8632 | ||
bb67d9c7 RG |
8633 | /* If no size was supplied use the mode the target prefers. Otherwise |
8634 | lookup a vector mode of the specified size. */ | |
8635 | if (size == 0) | |
8636 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); | |
8637 | else | |
8638 | simd_mode = mode_for_vector (inner_mode, size / nbytes); | |
cc4b5170 RG |
8639 | nunits = GET_MODE_SIZE (simd_mode) / nbytes; |
8640 | if (nunits <= 1) | |
8641 | return NULL_TREE; | |
ebfd146a IR |
8642 | |
8643 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
8644 | |
8645 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
8646 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 8647 | return NULL_TREE; |
ebfd146a IR |
8648 | |
8649 | return vectype; | |
8650 | } | |
8651 | ||
bb67d9c7 RG |
8652 | unsigned int current_vector_size; |
8653 | ||
8654 | /* Function get_vectype_for_scalar_type. | |
8655 | ||
8656 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
8657 | by the target. */ | |
8658 | ||
8659 | tree | |
8660 | get_vectype_for_scalar_type (tree scalar_type) | |
8661 | { | |
8662 | tree vectype; | |
8663 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
8664 | current_vector_size); | |
8665 | if (vectype | |
8666 | && current_vector_size == 0) | |
8667 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
8668 | return vectype; | |
8669 | } | |
8670 | ||
42fd8198 IE |
8671 | /* Function get_mask_type_for_scalar_type. |
8672 | ||
8673 | Returns the mask type corresponding to a result of comparison | |
8674 | of vectors of specified SCALAR_TYPE as supported by target. */ | |
8675 | ||
8676 | tree | |
8677 | get_mask_type_for_scalar_type (tree scalar_type) | |
8678 | { | |
8679 | tree vectype = get_vectype_for_scalar_type (scalar_type); | |
8680 | ||
8681 | if (!vectype) | |
8682 | return NULL; | |
8683 | ||
8684 | return build_truth_vector_type (TYPE_VECTOR_SUBPARTS (vectype), | |
8685 | current_vector_size); | |
8686 | } | |
8687 | ||
b690cc0f RG |
8688 | /* Function get_same_sized_vectype |
8689 | ||
8690 | Returns a vector type corresponding to SCALAR_TYPE of size | |
8691 | VECTOR_TYPE if supported by the target. */ | |
8692 | ||
8693 | tree | |
bb67d9c7 | 8694 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 8695 | { |
9f47c7e5 IE |
8696 | if (TREE_CODE (scalar_type) == BOOLEAN_TYPE) |
8697 | return build_same_sized_truth_vector_type (vector_type); | |
8698 | ||
bb67d9c7 RG |
8699 | return get_vectype_for_scalar_type_and_size |
8700 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
8701 | } |
8702 | ||
ebfd146a IR |
8703 | /* Function vect_is_simple_use. |
8704 | ||
8705 | Input: | |
81c40241 RB |
8706 | VINFO - the vect info of the loop or basic block that is being vectorized. |
8707 | OPERAND - operand in the loop or bb. | |
8708 | Output: | |
8709 | DEF_STMT - the defining stmt in case OPERAND is an SSA_NAME. | |
8710 | DT - the type of definition | |
ebfd146a IR |
8711 | |
8712 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 8713 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 8714 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 8715 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
8716 | is the case in reduction/induction computations). |
8717 | For basic blocks, supportable operands are constants and bb invariants. | |
8718 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
8719 | |
8720 | bool | |
81c40241 RB |
8721 | vect_is_simple_use (tree operand, vec_info *vinfo, |
8722 | gimple **def_stmt, enum vect_def_type *dt) | |
b8698a0f | 8723 | { |
ebfd146a | 8724 | *def_stmt = NULL; |
3fc356dc | 8725 | *dt = vect_unknown_def_type; |
b8698a0f | 8726 | |
73fbfcad | 8727 | if (dump_enabled_p ()) |
ebfd146a | 8728 | { |
78c60e3d SS |
8729 | dump_printf_loc (MSG_NOTE, vect_location, |
8730 | "vect_is_simple_use: operand "); | |
8731 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
e645e942 | 8732 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 8733 | } |
b8698a0f | 8734 | |
b758f602 | 8735 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
8736 | { |
8737 | *dt = vect_constant_def; | |
8738 | return true; | |
8739 | } | |
b8698a0f | 8740 | |
ebfd146a IR |
8741 | if (is_gimple_min_invariant (operand)) |
8742 | { | |
8644a673 | 8743 | *dt = vect_external_def; |
ebfd146a IR |
8744 | return true; |
8745 | } | |
8746 | ||
ebfd146a IR |
8747 | if (TREE_CODE (operand) != SSA_NAME) |
8748 | { | |
73fbfcad | 8749 | if (dump_enabled_p ()) |
af29617a AH |
8750 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
8751 | "not ssa-name.\n"); | |
ebfd146a IR |
8752 | return false; |
8753 | } | |
b8698a0f | 8754 | |
3fc356dc | 8755 | if (SSA_NAME_IS_DEFAULT_DEF (operand)) |
ebfd146a | 8756 | { |
3fc356dc RB |
8757 | *dt = vect_external_def; |
8758 | return true; | |
ebfd146a IR |
8759 | } |
8760 | ||
3fc356dc | 8761 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
73fbfcad | 8762 | if (dump_enabled_p ()) |
ebfd146a | 8763 | { |
78c60e3d SS |
8764 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
8765 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
ebfd146a IR |
8766 | } |
8767 | ||
61d371eb | 8768 | if (! vect_stmt_in_region_p (vinfo, *def_stmt)) |
8644a673 | 8769 | *dt = vect_external_def; |
ebfd146a IR |
8770 | else |
8771 | { | |
3fc356dc | 8772 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (*def_stmt); |
603cca93 | 8773 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
ebfd146a IR |
8774 | } |
8775 | ||
2e8ab70c RB |
8776 | if (dump_enabled_p ()) |
8777 | { | |
8778 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: "); | |
8779 | switch (*dt) | |
8780 | { | |
8781 | case vect_uninitialized_def: | |
8782 | dump_printf (MSG_NOTE, "uninitialized\n"); | |
8783 | break; | |
8784 | case vect_constant_def: | |
8785 | dump_printf (MSG_NOTE, "constant\n"); | |
8786 | break; | |
8787 | case vect_external_def: | |
8788 | dump_printf (MSG_NOTE, "external\n"); | |
8789 | break; | |
8790 | case vect_internal_def: | |
8791 | dump_printf (MSG_NOTE, "internal\n"); | |
8792 | break; | |
8793 | case vect_induction_def: | |
8794 | dump_printf (MSG_NOTE, "induction\n"); | |
8795 | break; | |
8796 | case vect_reduction_def: | |
8797 | dump_printf (MSG_NOTE, "reduction\n"); | |
8798 | break; | |
8799 | case vect_double_reduction_def: | |
8800 | dump_printf (MSG_NOTE, "double reduction\n"); | |
8801 | break; | |
8802 | case vect_nested_cycle: | |
8803 | dump_printf (MSG_NOTE, "nested cycle\n"); | |
8804 | break; | |
8805 | case vect_unknown_def_type: | |
8806 | dump_printf (MSG_NOTE, "unknown\n"); | |
8807 | break; | |
8808 | } | |
8809 | } | |
8810 | ||
81c40241 | 8811 | if (*dt == vect_unknown_def_type) |
ebfd146a | 8812 | { |
73fbfcad | 8813 | if (dump_enabled_p ()) |
78c60e3d | 8814 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8815 | "Unsupported pattern.\n"); |
ebfd146a IR |
8816 | return false; |
8817 | } | |
8818 | ||
ebfd146a IR |
8819 | switch (gimple_code (*def_stmt)) |
8820 | { | |
8821 | case GIMPLE_PHI: | |
ebfd146a | 8822 | case GIMPLE_ASSIGN: |
ebfd146a | 8823 | case GIMPLE_CALL: |
81c40241 | 8824 | break; |
ebfd146a | 8825 | default: |
73fbfcad | 8826 | if (dump_enabled_p ()) |
78c60e3d | 8827 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 8828 | "unsupported defining stmt:\n"); |
ebfd146a IR |
8829 | return false; |
8830 | } | |
8831 | ||
8832 | return true; | |
8833 | } | |
8834 | ||
81c40241 | 8835 | /* Function vect_is_simple_use. |
b690cc0f | 8836 | |
81c40241 | 8837 | Same as vect_is_simple_use but also determines the vector operand |
b690cc0f RG |
8838 | type of OPERAND and stores it to *VECTYPE. If the definition of |
8839 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
8840 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
8841 | is responsible to compute the best suited vector type for the | |
8842 | scalar operand. */ | |
8843 | ||
8844 | bool | |
81c40241 RB |
8845 | vect_is_simple_use (tree operand, vec_info *vinfo, |
8846 | gimple **def_stmt, enum vect_def_type *dt, tree *vectype) | |
b690cc0f | 8847 | { |
81c40241 | 8848 | if (!vect_is_simple_use (operand, vinfo, def_stmt, dt)) |
b690cc0f RG |
8849 | return false; |
8850 | ||
8851 | /* Now get a vector type if the def is internal, otherwise supply | |
8852 | NULL_TREE and leave it up to the caller to figure out a proper | |
8853 | type for the use stmt. */ | |
8854 | if (*dt == vect_internal_def | |
8855 | || *dt == vect_induction_def | |
8856 | || *dt == vect_reduction_def | |
8857 | || *dt == vect_double_reduction_def | |
8858 | || *dt == vect_nested_cycle) | |
8859 | { | |
8860 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
8861 | |
8862 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
8863 | && !STMT_VINFO_RELEVANT (stmt_info) | |
8864 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 8865 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 8866 | |
b690cc0f RG |
8867 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
8868 | gcc_assert (*vectype != NULL_TREE); | |
8869 | } | |
8870 | else if (*dt == vect_uninitialized_def | |
8871 | || *dt == vect_constant_def | |
8872 | || *dt == vect_external_def) | |
8873 | *vectype = NULL_TREE; | |
8874 | else | |
8875 | gcc_unreachable (); | |
8876 | ||
8877 | return true; | |
8878 | } | |
8879 | ||
ebfd146a IR |
8880 | |
8881 | /* Function supportable_widening_operation | |
8882 | ||
b8698a0f L |
8883 | Check whether an operation represented by the code CODE is a |
8884 | widening operation that is supported by the target platform in | |
b690cc0f RG |
8885 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
8886 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 8887 | |
ebfd146a IR |
8888 | Widening operations we currently support are NOP (CONVERT), FLOAT |
8889 | and WIDEN_MULT. This function checks if these operations are supported | |
8890 | by the target platform either directly (via vector tree-codes), or via | |
8891 | target builtins. | |
8892 | ||
8893 | Output: | |
b8698a0f L |
8894 | - CODE1 and CODE2 are codes of vector operations to be used when |
8895 | vectorizing the operation, if available. | |
ebfd146a IR |
8896 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
8897 | case of multi-step conversion (like char->short->int - in that case | |
8898 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
8899 | - INTERM_TYPES contains the intermediate type required to perform the |
8900 | widening operation (short in the above example). */ | |
ebfd146a IR |
8901 | |
8902 | bool | |
355fe088 | 8903 | supportable_widening_operation (enum tree_code code, gimple *stmt, |
b690cc0f | 8904 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
8905 | enum tree_code *code1, enum tree_code *code2, |
8906 | int *multi_step_cvt, | |
9771b263 | 8907 | vec<tree> *interm_types) |
ebfd146a IR |
8908 | { |
8909 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8910 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 8911 | struct loop *vect_loop = NULL; |
ef4bddc2 | 8912 | machine_mode vec_mode; |
81f40b79 | 8913 | enum insn_code icode1, icode2; |
ebfd146a | 8914 | optab optab1, optab2; |
b690cc0f RG |
8915 | tree vectype = vectype_in; |
8916 | tree wide_vectype = vectype_out; | |
ebfd146a | 8917 | enum tree_code c1, c2; |
4a00c761 JJ |
8918 | int i; |
8919 | tree prev_type, intermediate_type; | |
ef4bddc2 | 8920 | machine_mode intermediate_mode, prev_mode; |
4a00c761 | 8921 | optab optab3, optab4; |
ebfd146a | 8922 | |
4a00c761 | 8923 | *multi_step_cvt = 0; |
4ef69dfc IR |
8924 | if (loop_info) |
8925 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
8926 | ||
ebfd146a IR |
8927 | switch (code) |
8928 | { | |
8929 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
8930 | /* The result of a vectorized widening operation usually requires |
8931 | two vectors (because the widened results do not fit into one vector). | |
8932 | The generated vector results would normally be expected to be | |
8933 | generated in the same order as in the original scalar computation, | |
8934 | i.e. if 8 results are generated in each vector iteration, they are | |
8935 | to be organized as follows: | |
8936 | vect1: [res1,res2,res3,res4], | |
8937 | vect2: [res5,res6,res7,res8]. | |
8938 | ||
8939 | However, in the special case that the result of the widening | |
8940 | operation is used in a reduction computation only, the order doesn't | |
8941 | matter (because when vectorizing a reduction we change the order of | |
8942 | the computation). Some targets can take advantage of this and | |
8943 | generate more efficient code. For example, targets like Altivec, | |
8944 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
8945 | generate the following vectors: | |
8946 | vect1: [res1,res3,res5,res7], | |
8947 | vect2: [res2,res4,res6,res8]. | |
8948 | ||
8949 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
8950 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
8951 | iterations in parallel). We therefore don't allow to change the | |
8952 | order of the computation in the inner-loop during outer-loop | |
8953 | vectorization. */ | |
8954 | /* TODO: Another case in which order doesn't *really* matter is when we | |
8955 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
8956 | Normally, pack_trunc performs an even/odd permute, whereas the | |
8957 | repack from an even/odd expansion would be an interleave, which | |
8958 | would be significantly simpler for e.g. AVX2. */ | |
8959 | /* In any case, in order to avoid duplicating the code below, recurse | |
8960 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
8961 | are properly set up for the caller. If we fail, we'll continue with | |
8962 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
8963 | if (vect_loop | |
8964 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
8965 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
8966 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
8967 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
8968 | code1, code2, multi_step_cvt, |
8969 | interm_types)) | |
ebc047a2 CH |
8970 | { |
8971 | /* Elements in a vector with vect_used_by_reduction property cannot | |
8972 | be reordered if the use chain with this property does not have the | |
8973 | same operation. One such an example is s += a * b, where elements | |
8974 | in a and b cannot be reordered. Here we check if the vector defined | |
8975 | by STMT is only directly used in the reduction statement. */ | |
8976 | tree lhs = gimple_assign_lhs (stmt); | |
8977 | use_operand_p dummy; | |
355fe088 | 8978 | gimple *use_stmt; |
ebc047a2 CH |
8979 | stmt_vec_info use_stmt_info = NULL; |
8980 | if (single_imm_use (lhs, &dummy, &use_stmt) | |
8981 | && (use_stmt_info = vinfo_for_stmt (use_stmt)) | |
8982 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) | |
8983 | return true; | |
8984 | } | |
4a00c761 JJ |
8985 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
8986 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
8987 | break; |
8988 | ||
81c40241 RB |
8989 | case DOT_PROD_EXPR: |
8990 | c1 = DOT_PROD_EXPR; | |
8991 | c2 = DOT_PROD_EXPR; | |
8992 | break; | |
8993 | ||
8994 | case SAD_EXPR: | |
8995 | c1 = SAD_EXPR; | |
8996 | c2 = SAD_EXPR; | |
8997 | break; | |
8998 | ||
6ae6116f RH |
8999 | case VEC_WIDEN_MULT_EVEN_EXPR: |
9000 | /* Support the recursion induced just above. */ | |
9001 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
9002 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
9003 | break; | |
9004 | ||
36ba4aae | 9005 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
9006 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
9007 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
9008 | break; |
9009 | ||
ebfd146a | 9010 | CASE_CONVERT: |
4a00c761 JJ |
9011 | c1 = VEC_UNPACK_LO_EXPR; |
9012 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
9013 | break; |
9014 | ||
9015 | case FLOAT_EXPR: | |
4a00c761 JJ |
9016 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
9017 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
9018 | break; |
9019 | ||
9020 | case FIX_TRUNC_EXPR: | |
9021 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
9022 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
9023 | computing the operation. */ | |
9024 | return false; | |
9025 | ||
9026 | default: | |
9027 | gcc_unreachable (); | |
9028 | } | |
9029 | ||
6ae6116f | 9030 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
6b4db501 | 9031 | std::swap (c1, c2); |
4a00c761 | 9032 | |
ebfd146a IR |
9033 | if (code == FIX_TRUNC_EXPR) |
9034 | { | |
9035 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
9036 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
9037 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
9038 | } |
9039 | else | |
9040 | { | |
9041 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
9042 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
9043 | } | |
9044 | ||
9045 | if (!optab1 || !optab2) | |
9046 | return false; | |
9047 | ||
9048 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
9049 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
9050 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
9051 | return false; |
9052 | ||
4a00c761 JJ |
9053 | *code1 = c1; |
9054 | *code2 = c2; | |
9055 | ||
9056 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
9057 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff IE |
9058 | /* For scalar masks we may have different boolean |
9059 | vector types having the same QImode. Thus we | |
9060 | add additional check for elements number. */ | |
9061 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
9062 | || (TYPE_VECTOR_SUBPARTS (vectype) / 2 | |
9063 | == TYPE_VECTOR_SUBPARTS (wide_vectype))); | |
4a00c761 | 9064 | |
b8698a0f | 9065 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 9066 | types. */ |
ebfd146a | 9067 | |
4a00c761 JJ |
9068 | prev_type = vectype; |
9069 | prev_mode = vec_mode; | |
b8698a0f | 9070 | |
4a00c761 JJ |
9071 | if (!CONVERT_EXPR_CODE_P (code)) |
9072 | return false; | |
b8698a0f | 9073 | |
4a00c761 JJ |
9074 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
9075 | intermediate steps in promotion sequence. We try | |
9076 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
9077 | not. */ | |
9771b263 | 9078 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
9079 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
9080 | { | |
9081 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
9082 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
9083 | { | |
9084 | intermediate_type | |
9085 | = build_truth_vector_type (TYPE_VECTOR_SUBPARTS (prev_type) / 2, | |
9086 | current_vector_size); | |
9087 | if (intermediate_mode != TYPE_MODE (intermediate_type)) | |
9088 | return false; | |
9089 | } | |
9090 | else | |
9091 | intermediate_type | |
9092 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
9093 | TYPE_UNSIGNED (prev_type)); | |
9094 | ||
4a00c761 JJ |
9095 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); |
9096 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
9097 | ||
9098 | if (!optab3 || !optab4 | |
9099 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
9100 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
9101 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
9102 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
9103 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
9104 | == CODE_FOR_nothing) | |
9105 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
9106 | == CODE_FOR_nothing)) | |
9107 | break; | |
ebfd146a | 9108 | |
9771b263 | 9109 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
9110 | (*multi_step_cvt)++; |
9111 | ||
9112 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
9113 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff IE |
9114 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
9115 | || (TYPE_VECTOR_SUBPARTS (intermediate_type) / 2 | |
9116 | == TYPE_VECTOR_SUBPARTS (wide_vectype))); | |
4a00c761 JJ |
9117 | |
9118 | prev_type = intermediate_type; | |
9119 | prev_mode = intermediate_mode; | |
ebfd146a IR |
9120 | } |
9121 | ||
9771b263 | 9122 | interm_types->release (); |
4a00c761 | 9123 | return false; |
ebfd146a IR |
9124 | } |
9125 | ||
9126 | ||
9127 | /* Function supportable_narrowing_operation | |
9128 | ||
b8698a0f L |
9129 | Check whether an operation represented by the code CODE is a |
9130 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
9131 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
9132 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 9133 | |
ebfd146a | 9134 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 9135 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
9136 | the target platform directly via vector tree-codes. |
9137 | ||
9138 | Output: | |
b8698a0f L |
9139 | - CODE1 is the code of a vector operation to be used when |
9140 | vectorizing the operation, if available. | |
ebfd146a IR |
9141 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
9142 | case of multi-step conversion (like int->short->char - in that case | |
9143 | MULTI_STEP_CVT will be 1). | |
9144 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 9145 | narrowing operation (short in the above example). */ |
ebfd146a IR |
9146 | |
9147 | bool | |
9148 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 9149 | tree vectype_out, tree vectype_in, |
ebfd146a | 9150 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 9151 | vec<tree> *interm_types) |
ebfd146a | 9152 | { |
ef4bddc2 | 9153 | machine_mode vec_mode; |
ebfd146a IR |
9154 | enum insn_code icode1; |
9155 | optab optab1, interm_optab; | |
b690cc0f RG |
9156 | tree vectype = vectype_in; |
9157 | tree narrow_vectype = vectype_out; | |
ebfd146a | 9158 | enum tree_code c1; |
3ae0661a | 9159 | tree intermediate_type, prev_type; |
ef4bddc2 | 9160 | machine_mode intermediate_mode, prev_mode; |
ebfd146a | 9161 | int i; |
4a00c761 | 9162 | bool uns; |
ebfd146a | 9163 | |
4a00c761 | 9164 | *multi_step_cvt = 0; |
ebfd146a IR |
9165 | switch (code) |
9166 | { | |
9167 | CASE_CONVERT: | |
9168 | c1 = VEC_PACK_TRUNC_EXPR; | |
9169 | break; | |
9170 | ||
9171 | case FIX_TRUNC_EXPR: | |
9172 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
9173 | break; | |
9174 | ||
9175 | case FLOAT_EXPR: | |
9176 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
9177 | tree code and optabs used for computing the operation. */ | |
9178 | return false; | |
9179 | ||
9180 | default: | |
9181 | gcc_unreachable (); | |
9182 | } | |
9183 | ||
9184 | if (code == FIX_TRUNC_EXPR) | |
9185 | /* The signedness is determined from output operand. */ | |
b690cc0f | 9186 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
9187 | else |
9188 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
9189 | ||
9190 | if (!optab1) | |
9191 | return false; | |
9192 | ||
9193 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 9194 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
9195 | return false; |
9196 | ||
4a00c761 JJ |
9197 | *code1 = c1; |
9198 | ||
9199 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff IE |
9200 | /* For scalar masks we may have different boolean |
9201 | vector types having the same QImode. Thus we | |
9202 | add additional check for elements number. */ | |
9203 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
9204 | || (TYPE_VECTOR_SUBPARTS (vectype) * 2 | |
9205 | == TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 | 9206 | |
ebfd146a IR |
9207 | /* Check if it's a multi-step conversion that can be done using intermediate |
9208 | types. */ | |
4a00c761 | 9209 | prev_mode = vec_mode; |
3ae0661a | 9210 | prev_type = vectype; |
4a00c761 JJ |
9211 | if (code == FIX_TRUNC_EXPR) |
9212 | uns = TYPE_UNSIGNED (vectype_out); | |
9213 | else | |
9214 | uns = TYPE_UNSIGNED (vectype); | |
9215 | ||
9216 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
9217 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
9218 | costly than signed. */ | |
9219 | if (code == FIX_TRUNC_EXPR && uns) | |
9220 | { | |
9221 | enum insn_code icode2; | |
9222 | ||
9223 | intermediate_type | |
9224 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
9225 | interm_optab | |
9226 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 9227 | if (interm_optab != unknown_optab |
4a00c761 JJ |
9228 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
9229 | && insn_data[icode1].operand[0].mode | |
9230 | == insn_data[icode2].operand[0].mode) | |
9231 | { | |
9232 | uns = false; | |
9233 | optab1 = interm_optab; | |
9234 | icode1 = icode2; | |
9235 | } | |
9236 | } | |
ebfd146a | 9237 | |
4a00c761 JJ |
9238 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
9239 | intermediate steps in promotion sequence. We try | |
9240 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 9241 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
9242 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
9243 | { | |
9244 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
9245 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
9246 | { | |
9247 | intermediate_type | |
9248 | = build_truth_vector_type (TYPE_VECTOR_SUBPARTS (prev_type) * 2, | |
9249 | current_vector_size); | |
9250 | if (intermediate_mode != TYPE_MODE (intermediate_type)) | |
9251 | return false; | |
9252 | } | |
9253 | else | |
9254 | intermediate_type | |
9255 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
4a00c761 JJ |
9256 | interm_optab |
9257 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
9258 | optab_default); | |
9259 | if (!interm_optab | |
9260 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
9261 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
9262 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
9263 | == CODE_FOR_nothing)) | |
9264 | break; | |
9265 | ||
9771b263 | 9266 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
9267 | (*multi_step_cvt)++; |
9268 | ||
9269 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff IE |
9270 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
9271 | || (TYPE_VECTOR_SUBPARTS (intermediate_type) * 2 | |
9272 | == TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 JJ |
9273 | |
9274 | prev_mode = intermediate_mode; | |
3ae0661a | 9275 | prev_type = intermediate_type; |
4a00c761 | 9276 | optab1 = interm_optab; |
ebfd146a IR |
9277 | } |
9278 | ||
9771b263 | 9279 | interm_types->release (); |
4a00c761 | 9280 | return false; |
ebfd146a | 9281 | } |