]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
85ec4feb | 2 | Copyright (C) 2003-2018 Free Software Foundation, Inc. |
b8698a0f | 3 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
4 | and Ira Rosen <irar@il.ibm.com> |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 3, or (at your option) any later | |
11 | version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
c7131fb2 | 25 | #include "backend.h" |
957060b5 AM |
26 | #include "target.h" |
27 | #include "rtl.h" | |
ebfd146a | 28 | #include "tree.h" |
c7131fb2 | 29 | #include "gimple.h" |
c7131fb2 | 30 | #include "ssa.h" |
957060b5 AM |
31 | #include "optabs-tree.h" |
32 | #include "insn-config.h" | |
33 | #include "recog.h" /* FIXME: for insn_data */ | |
34 | #include "cgraph.h" | |
957060b5 | 35 | #include "dumpfile.h" |
c7131fb2 | 36 | #include "alias.h" |
40e23961 | 37 | #include "fold-const.h" |
d8a2d370 | 38 | #include "stor-layout.h" |
2fb9a547 | 39 | #include "tree-eh.h" |
45b0be94 | 40 | #include "gimplify.h" |
5be5c238 | 41 | #include "gimple-iterator.h" |
18f429e2 | 42 | #include "gimplify-me.h" |
442b4905 | 43 | #include "tree-cfg.h" |
e28030cf | 44 | #include "tree-ssa-loop-manip.h" |
ebfd146a | 45 | #include "cfgloop.h" |
0136f8f0 AH |
46 | #include "tree-ssa-loop.h" |
47 | #include "tree-scalar-evolution.h" | |
ebfd146a | 48 | #include "tree-vectorizer.h" |
9b2b7279 | 49 | #include "builtins.h" |
70439f0d | 50 | #include "internal-fn.h" |
5ebaa477 | 51 | #include "tree-vector-builder.h" |
f151c9e1 | 52 | #include "vec-perm-indices.h" |
7cfb4d93 RS |
53 | #include "tree-ssa-loop-niter.h" |
54 | #include "gimple-fold.h" | |
ebfd146a | 55 | |
7ee2468b SB |
56 | /* For lang_hooks.types.type_for_mode. */ |
57 | #include "langhooks.h" | |
ebfd146a | 58 | |
c3e7ee41 BS |
59 | /* Return the vectorized type for the given statement. */ |
60 | ||
61 | tree | |
62 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
63 | { | |
64 | return STMT_VINFO_VECTYPE (stmt_info); | |
65 | } | |
66 | ||
67 | /* Return TRUE iff the given statement is in an inner loop relative to | |
68 | the loop being vectorized. */ | |
69 | bool | |
70 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
71 | { | |
355fe088 | 72 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
c3e7ee41 BS |
73 | basic_block bb = gimple_bb (stmt); |
74 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
75 | struct loop* loop; | |
76 | ||
77 | if (!loop_vinfo) | |
78 | return false; | |
79 | ||
80 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
81 | ||
82 | return (bb->loop_father == loop->inner); | |
83 | } | |
84 | ||
85 | /* Record the cost of a statement, either by directly informing the | |
86 | target model or by saving it in a vector for later processing. | |
87 | Return a preliminary estimate of the statement's cost. */ | |
88 | ||
89 | unsigned | |
92345349 | 90 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 91 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 92 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 93 | { |
cc9fe6bb JH |
94 | if ((kind == vector_load || kind == unaligned_load) |
95 | && STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
96 | kind = vector_gather_load; | |
97 | if ((kind == vector_store || kind == unaligned_store) | |
98 | && STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
99 | kind = vector_scatter_store; | |
68435eb2 RB |
100 | |
101 | stmt_info_for_cost si = { count, kind, where, | |
102 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
103 | misalign }; | |
104 | body_cost_vec->safe_push (si); | |
105 | ||
106 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; | |
107 | return (unsigned) | |
108 | (builtin_vectorization_cost (kind, vectype, misalign) * count); | |
c3e7ee41 BS |
109 | } |
110 | ||
272c6793 RS |
111 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
112 | ||
113 | static tree | |
114 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
115 | { | |
116 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
117 | "vect_array"); | |
118 | } | |
119 | ||
120 | /* ARRAY is an array of vectors created by create_vector_array. | |
121 | Return an SSA_NAME for the vector in index N. The reference | |
122 | is part of the vectorization of STMT and the vector is associated | |
123 | with scalar destination SCALAR_DEST. */ | |
124 | ||
125 | static tree | |
355fe088 | 126 | read_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree scalar_dest, |
272c6793 RS |
127 | tree array, unsigned HOST_WIDE_INT n) |
128 | { | |
129 | tree vect_type, vect, vect_name, array_ref; | |
355fe088 | 130 | gimple *new_stmt; |
272c6793 RS |
131 | |
132 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
133 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
134 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
135 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
136 | build_int_cst (size_type_node, n), | |
137 | NULL_TREE, NULL_TREE); | |
138 | ||
139 | new_stmt = gimple_build_assign (vect, array_ref); | |
140 | vect_name = make_ssa_name (vect, new_stmt); | |
141 | gimple_assign_set_lhs (new_stmt, vect_name); | |
142 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
143 | |
144 | return vect_name; | |
145 | } | |
146 | ||
147 | /* ARRAY is an array of vectors created by create_vector_array. | |
148 | Emit code to store SSA_NAME VECT in index N of the array. | |
149 | The store is part of the vectorization of STMT. */ | |
150 | ||
151 | static void | |
355fe088 | 152 | write_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree vect, |
272c6793 RS |
153 | tree array, unsigned HOST_WIDE_INT n) |
154 | { | |
155 | tree array_ref; | |
355fe088 | 156 | gimple *new_stmt; |
272c6793 RS |
157 | |
158 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
159 | build_int_cst (size_type_node, n), | |
160 | NULL_TREE, NULL_TREE); | |
161 | ||
162 | new_stmt = gimple_build_assign (array_ref, vect); | |
163 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
164 | } |
165 | ||
166 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
167 | of *PTR. The memory reference replaces those in FIRST_DR | |
168 | (and its group). */ | |
169 | ||
170 | static tree | |
44fc7854 | 171 | create_array_ref (tree type, tree ptr, tree alias_ptr_type) |
272c6793 | 172 | { |
44fc7854 | 173 | tree mem_ref; |
272c6793 | 174 | |
272c6793 RS |
175 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); |
176 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 177 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
178 | return mem_ref; |
179 | } | |
180 | ||
3ba4ff41 RS |
181 | /* Add a clobber of variable VAR to the vectorization of STMT. |
182 | Emit the clobber before *GSI. */ | |
183 | ||
184 | static void | |
185 | vect_clobber_variable (gimple *stmt, gimple_stmt_iterator *gsi, tree var) | |
186 | { | |
187 | tree clobber = build_clobber (TREE_TYPE (var)); | |
188 | gimple *new_stmt = gimple_build_assign (var, clobber); | |
189 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
190 | } | |
191 | ||
ebfd146a IR |
192 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
193 | ||
194 | /* Function vect_mark_relevant. | |
195 | ||
196 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
197 | ||
198 | static void | |
355fe088 | 199 | vect_mark_relevant (vec<gimple *> *worklist, gimple *stmt, |
97ecdb46 | 200 | enum vect_relevant relevant, bool live_p) |
ebfd146a IR |
201 | { |
202 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
203 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
204 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
355fe088 | 205 | gimple *pattern_stmt; |
ebfd146a | 206 | |
73fbfcad | 207 | if (dump_enabled_p ()) |
66c16fd9 RB |
208 | { |
209 | dump_printf_loc (MSG_NOTE, vect_location, | |
210 | "mark relevant %d, live %d: ", relevant, live_p); | |
211 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
212 | } | |
ebfd146a | 213 | |
83197f37 IR |
214 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
215 | related pattern stmt instead of the original stmt. However, such stmts | |
216 | may have their own uses that are not in any pattern, in such cases the | |
217 | stmt itself should be marked. */ | |
ebfd146a IR |
218 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
219 | { | |
97ecdb46 JJ |
220 | /* This is the last stmt in a sequence that was detected as a |
221 | pattern that can potentially be vectorized. Don't mark the stmt | |
222 | as relevant/live because it's not going to be vectorized. | |
223 | Instead mark the pattern-stmt that replaces it. */ | |
83197f37 | 224 | |
97ecdb46 JJ |
225 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); |
226 | ||
227 | if (dump_enabled_p ()) | |
228 | dump_printf_loc (MSG_NOTE, vect_location, | |
229 | "last stmt in pattern. don't mark" | |
230 | " relevant/live.\n"); | |
231 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
232 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
233 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
234 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
235 | stmt = pattern_stmt; | |
ebfd146a IR |
236 | } |
237 | ||
238 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
239 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
240 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
241 | ||
242 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
243 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
244 | { | |
73fbfcad | 245 | if (dump_enabled_p ()) |
78c60e3d | 246 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 247 | "already marked relevant/live.\n"); |
ebfd146a IR |
248 | return; |
249 | } | |
250 | ||
9771b263 | 251 | worklist->safe_push (stmt); |
ebfd146a IR |
252 | } |
253 | ||
254 | ||
b28ead45 AH |
255 | /* Function is_simple_and_all_uses_invariant |
256 | ||
257 | Return true if STMT is simple and all uses of it are invariant. */ | |
258 | ||
259 | bool | |
260 | is_simple_and_all_uses_invariant (gimple *stmt, loop_vec_info loop_vinfo) | |
261 | { | |
262 | tree op; | |
263 | gimple *def_stmt; | |
264 | ssa_op_iter iter; | |
265 | ||
266 | if (!is_gimple_assign (stmt)) | |
267 | return false; | |
268 | ||
269 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) | |
270 | { | |
271 | enum vect_def_type dt = vect_uninitialized_def; | |
272 | ||
273 | if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt)) | |
274 | { | |
275 | if (dump_enabled_p ()) | |
276 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
277 | "use not simple.\n"); | |
278 | return false; | |
279 | } | |
280 | ||
281 | if (dt != vect_external_def && dt != vect_constant_def) | |
282 | return false; | |
283 | } | |
284 | return true; | |
285 | } | |
286 | ||
ebfd146a IR |
287 | /* Function vect_stmt_relevant_p. |
288 | ||
289 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
290 | "relevant for vectorization". | |
291 | ||
292 | A stmt is considered "relevant for vectorization" if: | |
293 | - it has uses outside the loop. | |
294 | - it has vdefs (it alters memory). | |
295 | - control stmts in the loop (except for the exit condition). | |
296 | ||
297 | CHECKME: what other side effects would the vectorizer allow? */ | |
298 | ||
299 | static bool | |
355fe088 | 300 | vect_stmt_relevant_p (gimple *stmt, loop_vec_info loop_vinfo, |
ebfd146a IR |
301 | enum vect_relevant *relevant, bool *live_p) |
302 | { | |
303 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
304 | ssa_op_iter op_iter; | |
305 | imm_use_iterator imm_iter; | |
306 | use_operand_p use_p; | |
307 | def_operand_p def_p; | |
308 | ||
8644a673 | 309 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
310 | *live_p = false; |
311 | ||
312 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
313 | if (is_ctrl_stmt (stmt) |
314 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
315 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 316 | *relevant = vect_used_in_scope; |
ebfd146a IR |
317 | |
318 | /* changing memory. */ | |
319 | if (gimple_code (stmt) != GIMPLE_PHI) | |
ac6aeab4 RB |
320 | if (gimple_vdef (stmt) |
321 | && !gimple_clobber_p (stmt)) | |
ebfd146a | 322 | { |
73fbfcad | 323 | if (dump_enabled_p ()) |
78c60e3d | 324 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 325 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 326 | *relevant = vect_used_in_scope; |
ebfd146a IR |
327 | } |
328 | ||
329 | /* uses outside the loop. */ | |
330 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
331 | { | |
332 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
333 | { | |
334 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
335 | if (!flow_bb_inside_loop_p (loop, bb)) | |
336 | { | |
73fbfcad | 337 | if (dump_enabled_p ()) |
78c60e3d | 338 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 339 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 340 | |
3157b0c2 AO |
341 | if (is_gimple_debug (USE_STMT (use_p))) |
342 | continue; | |
343 | ||
ebfd146a IR |
344 | /* We expect all such uses to be in the loop exit phis |
345 | (because of loop closed form) */ | |
346 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
347 | gcc_assert (bb == single_exit (loop)->dest); | |
348 | ||
349 | *live_p = true; | |
350 | } | |
351 | } | |
352 | } | |
353 | ||
3a2edf4c AH |
354 | if (*live_p && *relevant == vect_unused_in_scope |
355 | && !is_simple_and_all_uses_invariant (stmt, loop_vinfo)) | |
b28ead45 AH |
356 | { |
357 | if (dump_enabled_p ()) | |
358 | dump_printf_loc (MSG_NOTE, vect_location, | |
359 | "vec_stmt_relevant_p: stmt live but not relevant.\n"); | |
360 | *relevant = vect_used_only_live; | |
361 | } | |
362 | ||
ebfd146a IR |
363 | return (*live_p || *relevant); |
364 | } | |
365 | ||
366 | ||
b8698a0f | 367 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 368 | |
ff802fa1 | 369 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
370 | used in STMT for anything other than indexing an array. */ |
371 | ||
372 | static bool | |
355fe088 | 373 | exist_non_indexing_operands_for_use_p (tree use, gimple *stmt) |
ebfd146a IR |
374 | { |
375 | tree operand; | |
376 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 377 | |
ff802fa1 | 378 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
379 | reference in STMT, then any operand that corresponds to USE |
380 | is not indexing an array. */ | |
381 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
382 | return true; | |
59a05b0c | 383 | |
ebfd146a IR |
384 | /* STMT has a data_ref. FORNOW this means that its of one of |
385 | the following forms: | |
386 | -1- ARRAY_REF = var | |
387 | -2- var = ARRAY_REF | |
388 | (This should have been verified in analyze_data_refs). | |
389 | ||
390 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 391 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
392 | for array indexing. |
393 | ||
394 | Therefore, all we need to check is if STMT falls into the | |
395 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
396 | |
397 | if (!gimple_assign_copy_p (stmt)) | |
5ce9450f JJ |
398 | { |
399 | if (is_gimple_call (stmt) | |
400 | && gimple_call_internal_p (stmt)) | |
bfaa08b7 RS |
401 | { |
402 | internal_fn ifn = gimple_call_internal_fn (stmt); | |
403 | int mask_index = internal_fn_mask_index (ifn); | |
404 | if (mask_index >= 0 | |
405 | && use == gimple_call_arg (stmt, mask_index)) | |
406 | return true; | |
f307441a RS |
407 | int stored_value_index = internal_fn_stored_value_index (ifn); |
408 | if (stored_value_index >= 0 | |
409 | && use == gimple_call_arg (stmt, stored_value_index)) | |
410 | return true; | |
bfaa08b7 RS |
411 | if (internal_gather_scatter_fn_p (ifn) |
412 | && use == gimple_call_arg (stmt, 1)) | |
413 | return true; | |
bfaa08b7 | 414 | } |
5ce9450f JJ |
415 | return false; |
416 | } | |
417 | ||
59a05b0c EB |
418 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
419 | return false; | |
ebfd146a | 420 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
421 | if (TREE_CODE (operand) != SSA_NAME) |
422 | return false; | |
423 | ||
424 | if (operand == use) | |
425 | return true; | |
426 | ||
427 | return false; | |
428 | } | |
429 | ||
430 | ||
b8698a0f | 431 | /* |
ebfd146a IR |
432 | Function process_use. |
433 | ||
434 | Inputs: | |
435 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b28ead45 | 436 | - RELEVANT - enum value to be set in the STMT_VINFO of the stmt |
ff802fa1 | 437 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 438 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
439 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
440 | be performed. | |
ebfd146a IR |
441 | |
442 | Outputs: | |
443 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
444 | relevance info of the DEF_STMT of this USE: | |
445 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
446 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
447 | Exceptions: | |
448 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 449 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 450 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
451 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
452 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
453 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
454 | be modified accordingly. | |
455 | ||
456 | Return true if everything is as expected. Return false otherwise. */ | |
457 | ||
458 | static bool | |
b28ead45 | 459 | process_use (gimple *stmt, tree use, loop_vec_info loop_vinfo, |
355fe088 | 460 | enum vect_relevant relevant, vec<gimple *> *worklist, |
aec7ae7d | 461 | bool force) |
ebfd146a IR |
462 | { |
463 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
464 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
465 | stmt_vec_info dstmt_vinfo; | |
466 | basic_block bb, def_bb; | |
355fe088 | 467 | gimple *def_stmt; |
ebfd146a IR |
468 | enum vect_def_type dt; |
469 | ||
b8698a0f | 470 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 471 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 472 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
473 | return true; |
474 | ||
81c40241 | 475 | if (!vect_is_simple_use (use, loop_vinfo, &def_stmt, &dt)) |
b8698a0f | 476 | { |
73fbfcad | 477 | if (dump_enabled_p ()) |
78c60e3d | 478 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 479 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
480 | return false; |
481 | } | |
482 | ||
483 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
484 | return true; | |
485 | ||
486 | def_bb = gimple_bb (def_stmt); | |
487 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
488 | { | |
73fbfcad | 489 | if (dump_enabled_p ()) |
e645e942 | 490 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n"); |
ebfd146a IR |
491 | return true; |
492 | } | |
493 | ||
b8698a0f L |
494 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
495 | DEF_STMT must have already been processed, because this should be the | |
496 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
497 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
498 | check that everything is as expected, and we are done. */ |
499 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
500 | bb = gimple_bb (stmt); | |
501 | if (gimple_code (stmt) == GIMPLE_PHI | |
502 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
503 | && gimple_code (def_stmt) != GIMPLE_PHI | |
504 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
505 | && bb->loop_father == def_bb->loop_father) | |
506 | { | |
73fbfcad | 507 | if (dump_enabled_p ()) |
78c60e3d | 508 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 509 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a IR |
510 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
511 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
512 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 513 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 514 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
515 | return true; |
516 | } | |
517 | ||
518 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
519 | outer-loop-header-bb: | |
520 | d = def_stmt | |
521 | inner-loop: | |
522 | stmt # use (d) | |
523 | outer-loop-tail-bb: | |
524 | ... */ | |
525 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
526 | { | |
73fbfcad | 527 | if (dump_enabled_p ()) |
78c60e3d | 528 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 529 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 530 | |
ebfd146a IR |
531 | switch (relevant) |
532 | { | |
8644a673 | 533 | case vect_unused_in_scope: |
7c5222ff IR |
534 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
535 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 536 | break; |
7c5222ff | 537 | |
ebfd146a | 538 | case vect_used_in_outer_by_reduction: |
7c5222ff | 539 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
540 | relevant = vect_used_by_reduction; |
541 | break; | |
7c5222ff | 542 | |
ebfd146a | 543 | case vect_used_in_outer: |
7c5222ff | 544 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 545 | relevant = vect_used_in_scope; |
ebfd146a | 546 | break; |
7c5222ff | 547 | |
8644a673 | 548 | case vect_used_in_scope: |
ebfd146a IR |
549 | break; |
550 | ||
551 | default: | |
552 | gcc_unreachable (); | |
b8698a0f | 553 | } |
ebfd146a IR |
554 | } |
555 | ||
556 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
557 | outer-loop-header-bb: | |
558 | ... | |
559 | inner-loop: | |
560 | d = def_stmt | |
06066f92 | 561 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
562 | stmt # use (d) */ |
563 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
564 | { | |
73fbfcad | 565 | if (dump_enabled_p ()) |
78c60e3d | 566 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 567 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 568 | |
ebfd146a IR |
569 | switch (relevant) |
570 | { | |
8644a673 | 571 | case vect_unused_in_scope: |
b8698a0f | 572 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 573 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 574 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
575 | break; |
576 | ||
ebfd146a | 577 | case vect_used_by_reduction: |
b28ead45 | 578 | case vect_used_only_live: |
ebfd146a IR |
579 | relevant = vect_used_in_outer_by_reduction; |
580 | break; | |
581 | ||
8644a673 | 582 | case vect_used_in_scope: |
ebfd146a IR |
583 | relevant = vect_used_in_outer; |
584 | break; | |
585 | ||
586 | default: | |
587 | gcc_unreachable (); | |
588 | } | |
589 | } | |
643a9684 RB |
590 | /* We are also not interested in uses on loop PHI backedges that are |
591 | inductions. Otherwise we'll needlessly vectorize the IV increment | |
e294f495 RB |
592 | and cause hybrid SLP for SLP inductions. Unless the PHI is live |
593 | of course. */ | |
643a9684 RB |
594 | else if (gimple_code (stmt) == GIMPLE_PHI |
595 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_induction_def | |
e294f495 | 596 | && ! STMT_VINFO_LIVE_P (stmt_vinfo) |
643a9684 RB |
597 | && (PHI_ARG_DEF_FROM_EDGE (stmt, loop_latch_edge (bb->loop_father)) |
598 | == use)) | |
599 | { | |
600 | if (dump_enabled_p ()) | |
601 | dump_printf_loc (MSG_NOTE, vect_location, | |
602 | "induction value on backedge.\n"); | |
603 | return true; | |
604 | } | |
605 | ||
ebfd146a | 606 | |
b28ead45 | 607 | vect_mark_relevant (worklist, def_stmt, relevant, false); |
ebfd146a IR |
608 | return true; |
609 | } | |
610 | ||
611 | ||
612 | /* Function vect_mark_stmts_to_be_vectorized. | |
613 | ||
614 | Not all stmts in the loop need to be vectorized. For example: | |
615 | ||
616 | for i... | |
617 | for j... | |
618 | 1. T0 = i + j | |
619 | 2. T1 = a[T0] | |
620 | ||
621 | 3. j = j + 1 | |
622 | ||
623 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
624 | addressing of vectorized data-refs are handled differently. | |
625 | ||
626 | This pass detects such stmts. */ | |
627 | ||
628 | bool | |
629 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
630 | { | |
ebfd146a IR |
631 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
632 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
633 | unsigned int nbbs = loop->num_nodes; | |
634 | gimple_stmt_iterator si; | |
355fe088 | 635 | gimple *stmt; |
ebfd146a IR |
636 | unsigned int i; |
637 | stmt_vec_info stmt_vinfo; | |
638 | basic_block bb; | |
355fe088 | 639 | gimple *phi; |
ebfd146a | 640 | bool live_p; |
b28ead45 | 641 | enum vect_relevant relevant; |
ebfd146a | 642 | |
73fbfcad | 643 | if (dump_enabled_p ()) |
78c60e3d | 644 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 645 | "=== vect_mark_stmts_to_be_vectorized ===\n"); |
ebfd146a | 646 | |
355fe088 | 647 | auto_vec<gimple *, 64> worklist; |
ebfd146a IR |
648 | |
649 | /* 1. Init worklist. */ | |
650 | for (i = 0; i < nbbs; i++) | |
651 | { | |
652 | bb = bbs[i]; | |
653 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 654 | { |
ebfd146a | 655 | phi = gsi_stmt (si); |
73fbfcad | 656 | if (dump_enabled_p ()) |
ebfd146a | 657 | { |
78c60e3d SS |
658 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
659 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
660 | } |
661 | ||
662 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 663 | vect_mark_relevant (&worklist, phi, relevant, live_p); |
ebfd146a IR |
664 | } |
665 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
666 | { | |
667 | stmt = gsi_stmt (si); | |
73fbfcad | 668 | if (dump_enabled_p ()) |
ebfd146a | 669 | { |
78c60e3d SS |
670 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
671 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
b8698a0f | 672 | } |
ebfd146a IR |
673 | |
674 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 675 | vect_mark_relevant (&worklist, stmt, relevant, live_p); |
ebfd146a IR |
676 | } |
677 | } | |
678 | ||
679 | /* 2. Process_worklist */ | |
9771b263 | 680 | while (worklist.length () > 0) |
ebfd146a IR |
681 | { |
682 | use_operand_p use_p; | |
683 | ssa_op_iter iter; | |
684 | ||
9771b263 | 685 | stmt = worklist.pop (); |
73fbfcad | 686 | if (dump_enabled_p ()) |
ebfd146a | 687 | { |
78c60e3d SS |
688 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
689 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
ebfd146a IR |
690 | } |
691 | ||
b8698a0f | 692 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
b28ead45 AH |
693 | (DEF_STMT) as relevant/irrelevant according to the relevance property |
694 | of STMT. */ | |
ebfd146a IR |
695 | stmt_vinfo = vinfo_for_stmt (stmt); |
696 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
ebfd146a | 697 | |
b28ead45 AH |
698 | /* Generally, the relevance property of STMT (in STMT_VINFO_RELEVANT) is |
699 | propagated as is to the DEF_STMTs of its USEs. | |
ebfd146a IR |
700 | |
701 | One exception is when STMT has been identified as defining a reduction | |
b28ead45 | 702 | variable; in this case we set the relevance to vect_used_by_reduction. |
ebfd146a | 703 | This is because we distinguish between two kinds of relevant stmts - |
b8698a0f | 704 | those that are used by a reduction computation, and those that are |
ff802fa1 | 705 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 706 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 707 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 708 | |
b28ead45 | 709 | switch (STMT_VINFO_DEF_TYPE (stmt_vinfo)) |
ebfd146a | 710 | { |
06066f92 | 711 | case vect_reduction_def: |
b28ead45 AH |
712 | gcc_assert (relevant != vect_unused_in_scope); |
713 | if (relevant != vect_unused_in_scope | |
714 | && relevant != vect_used_in_scope | |
715 | && relevant != vect_used_by_reduction | |
716 | && relevant != vect_used_only_live) | |
06066f92 | 717 | { |
b28ead45 AH |
718 | if (dump_enabled_p ()) |
719 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
720 | "unsupported use of reduction.\n"); | |
721 | return false; | |
06066f92 | 722 | } |
06066f92 | 723 | break; |
b8698a0f | 724 | |
06066f92 | 725 | case vect_nested_cycle: |
b28ead45 AH |
726 | if (relevant != vect_unused_in_scope |
727 | && relevant != vect_used_in_outer_by_reduction | |
728 | && relevant != vect_used_in_outer) | |
06066f92 | 729 | { |
73fbfcad | 730 | if (dump_enabled_p ()) |
78c60e3d | 731 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 732 | "unsupported use of nested cycle.\n"); |
7c5222ff | 733 | |
06066f92 IR |
734 | return false; |
735 | } | |
b8698a0f L |
736 | break; |
737 | ||
06066f92 | 738 | case vect_double_reduction_def: |
b28ead45 AH |
739 | if (relevant != vect_unused_in_scope |
740 | && relevant != vect_used_by_reduction | |
741 | && relevant != vect_used_only_live) | |
06066f92 | 742 | { |
73fbfcad | 743 | if (dump_enabled_p ()) |
78c60e3d | 744 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 745 | "unsupported use of double reduction.\n"); |
7c5222ff | 746 | |
7c5222ff | 747 | return false; |
06066f92 | 748 | } |
b8698a0f | 749 | break; |
7c5222ff | 750 | |
06066f92 IR |
751 | default: |
752 | break; | |
7c5222ff | 753 | } |
b8698a0f | 754 | |
aec7ae7d | 755 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
756 | { |
757 | /* Pattern statements are not inserted into the code, so | |
758 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
759 | have to scan the RHS or function arguments instead. */ | |
760 | if (is_gimple_assign (stmt)) | |
761 | { | |
69d2aade JJ |
762 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
763 | tree op = gimple_assign_rhs1 (stmt); | |
764 | ||
765 | i = 1; | |
766 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
767 | { | |
768 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
b28ead45 | 769 | relevant, &worklist, false) |
69d2aade | 770 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
b28ead45 | 771 | relevant, &worklist, false)) |
566d377a | 772 | return false; |
69d2aade JJ |
773 | i = 2; |
774 | } | |
775 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 776 | { |
69d2aade | 777 | op = gimple_op (stmt, i); |
afbe6325 | 778 | if (TREE_CODE (op) == SSA_NAME |
b28ead45 | 779 | && !process_use (stmt, op, loop_vinfo, relevant, |
afbe6325 | 780 | &worklist, false)) |
07687835 | 781 | return false; |
9d5e7640 IR |
782 | } |
783 | } | |
784 | else if (is_gimple_call (stmt)) | |
785 | { | |
786 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
787 | { | |
788 | tree arg = gimple_call_arg (stmt, i); | |
b28ead45 | 789 | if (!process_use (stmt, arg, loop_vinfo, relevant, |
aec7ae7d | 790 | &worklist, false)) |
07687835 | 791 | return false; |
9d5e7640 IR |
792 | } |
793 | } | |
794 | } | |
795 | else | |
796 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
797 | { | |
798 | tree op = USE_FROM_PTR (use_p); | |
b28ead45 | 799 | if (!process_use (stmt, op, loop_vinfo, relevant, |
aec7ae7d | 800 | &worklist, false)) |
07687835 | 801 | return false; |
9d5e7640 | 802 | } |
aec7ae7d | 803 | |
3bab6342 | 804 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo)) |
aec7ae7d | 805 | { |
134c85ca RS |
806 | gather_scatter_info gs_info; |
807 | if (!vect_check_gather_scatter (stmt, loop_vinfo, &gs_info)) | |
808 | gcc_unreachable (); | |
809 | if (!process_use (stmt, gs_info.offset, loop_vinfo, relevant, | |
810 | &worklist, true)) | |
566d377a | 811 | return false; |
aec7ae7d | 812 | } |
ebfd146a IR |
813 | } /* while worklist */ |
814 | ||
ebfd146a IR |
815 | return true; |
816 | } | |
817 | ||
68435eb2 RB |
818 | /* Compute the prologue cost for invariant or constant operands. */ |
819 | ||
820 | static unsigned | |
821 | vect_prologue_cost_for_slp_op (slp_tree node, stmt_vec_info stmt_info, | |
822 | unsigned opno, enum vect_def_type dt, | |
823 | stmt_vector_for_cost *cost_vec) | |
824 | { | |
825 | gimple *stmt = SLP_TREE_SCALAR_STMTS (node)[0]; | |
826 | tree op = gimple_op (stmt, opno); | |
827 | unsigned prologue_cost = 0; | |
828 | ||
829 | /* Without looking at the actual initializer a vector of | |
830 | constants can be implemented as load from the constant pool. | |
831 | When all elements are the same we can use a splat. */ | |
832 | tree vectype = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
833 | unsigned group_size = SLP_TREE_SCALAR_STMTS (node).length (); | |
834 | unsigned num_vects_to_check; | |
835 | unsigned HOST_WIDE_INT const_nunits; | |
836 | unsigned nelt_limit; | |
837 | if (TYPE_VECTOR_SUBPARTS (vectype).is_constant (&const_nunits) | |
838 | && ! multiple_p (const_nunits, group_size)) | |
839 | { | |
840 | num_vects_to_check = SLP_TREE_NUMBER_OF_VEC_STMTS (node); | |
841 | nelt_limit = const_nunits; | |
842 | } | |
843 | else | |
844 | { | |
845 | /* If either the vector has variable length or the vectors | |
846 | are composed of repeated whole groups we only need to | |
847 | cost construction once. All vectors will be the same. */ | |
848 | num_vects_to_check = 1; | |
849 | nelt_limit = group_size; | |
850 | } | |
851 | tree elt = NULL_TREE; | |
852 | unsigned nelt = 0; | |
853 | for (unsigned j = 0; j < num_vects_to_check * nelt_limit; ++j) | |
854 | { | |
855 | unsigned si = j % group_size; | |
856 | if (nelt == 0) | |
857 | elt = gimple_op (SLP_TREE_SCALAR_STMTS (node)[si], opno); | |
858 | /* ??? We're just tracking whether all operands of a single | |
859 | vector initializer are the same, ideally we'd check if | |
860 | we emitted the same one already. */ | |
861 | else if (elt != gimple_op (SLP_TREE_SCALAR_STMTS (node)[si], | |
862 | opno)) | |
863 | elt = NULL_TREE; | |
864 | nelt++; | |
865 | if (nelt == nelt_limit) | |
866 | { | |
867 | /* ??? We need to pass down stmt_info for a vector type | |
868 | even if it points to the wrong stmt. */ | |
869 | prologue_cost += record_stmt_cost | |
870 | (cost_vec, 1, | |
871 | dt == vect_external_def | |
872 | ? (elt ? scalar_to_vec : vec_construct) | |
873 | : vector_load, | |
874 | stmt_info, 0, vect_prologue); | |
875 | nelt = 0; | |
876 | } | |
877 | } | |
878 | ||
879 | return prologue_cost; | |
880 | } | |
ebfd146a | 881 | |
b8698a0f | 882 | /* Function vect_model_simple_cost. |
ebfd146a | 883 | |
b8698a0f | 884 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
885 | single op. Right now, this does not account for multiple insns that could |
886 | be generated for the single vector op. We will handle that shortly. */ | |
887 | ||
68435eb2 | 888 | static void |
b8698a0f | 889 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 | 890 | enum vect_def_type *dt, |
4fc5ebf1 | 891 | int ndts, |
68435eb2 RB |
892 | slp_tree node, |
893 | stmt_vector_for_cost *cost_vec) | |
ebfd146a | 894 | { |
92345349 | 895 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a | 896 | |
68435eb2 | 897 | gcc_assert (cost_vec != NULL); |
ebfd146a | 898 | |
68435eb2 RB |
899 | /* ??? Somehow we need to fix this at the callers. */ |
900 | if (node) | |
901 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (node); | |
902 | ||
903 | if (node) | |
904 | { | |
905 | /* Scan operands and account for prologue cost of constants/externals. | |
906 | ??? This over-estimates cost for multiple uses and should be | |
907 | re-engineered. */ | |
908 | gimple *stmt = SLP_TREE_SCALAR_STMTS (node)[0]; | |
909 | tree lhs = gimple_get_lhs (stmt); | |
910 | for (unsigned i = 0; i < gimple_num_ops (stmt); ++i) | |
911 | { | |
912 | tree op = gimple_op (stmt, i); | |
913 | gimple *def_stmt; | |
914 | enum vect_def_type dt; | |
915 | if (!op || op == lhs) | |
916 | continue; | |
917 | if (vect_is_simple_use (op, stmt_info->vinfo, &def_stmt, &dt) | |
918 | && (dt == vect_constant_def || dt == vect_external_def)) | |
919 | prologue_cost += vect_prologue_cost_for_slp_op (node, stmt_info, | |
920 | i, dt, cost_vec); | |
921 | } | |
922 | } | |
923 | else | |
924 | /* Cost the "broadcast" of a scalar operand in to a vector operand. | |
925 | Use scalar_to_vec to cost the broadcast, as elsewhere in the vector | |
926 | cost model. */ | |
927 | for (int i = 0; i < ndts; i++) | |
928 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) | |
929 | prologue_cost += record_stmt_cost (cost_vec, 1, scalar_to_vec, | |
930 | stmt_info, 0, vect_prologue); | |
931 | ||
932 | /* Adjust for two-operator SLP nodes. */ | |
933 | if (node && SLP_TREE_TWO_OPERATORS (node)) | |
934 | { | |
935 | ncopies *= 2; | |
936 | inside_cost += record_stmt_cost (cost_vec, ncopies, vec_perm, | |
937 | stmt_info, 0, vect_body); | |
938 | } | |
c3e7ee41 BS |
939 | |
940 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
68435eb2 RB |
941 | inside_cost += record_stmt_cost (cost_vec, ncopies, vector_stmt, |
942 | stmt_info, 0, vect_body); | |
c3e7ee41 | 943 | |
73fbfcad | 944 | if (dump_enabled_p ()) |
78c60e3d SS |
945 | dump_printf_loc (MSG_NOTE, vect_location, |
946 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 947 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
948 | } |
949 | ||
950 | ||
8bd37302 BS |
951 | /* Model cost for type demotion and promotion operations. PWR is normally |
952 | zero for single-step promotions and demotions. It will be one if | |
953 | two-step promotion/demotion is required, and so on. Each additional | |
954 | step doubles the number of instructions required. */ | |
955 | ||
956 | static void | |
957 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
68435eb2 RB |
958 | enum vect_def_type *dt, int pwr, |
959 | stmt_vector_for_cost *cost_vec) | |
8bd37302 BS |
960 | { |
961 | int i, tmp; | |
92345349 | 962 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 | 963 | |
8bd37302 BS |
964 | for (i = 0; i < pwr + 1; i++) |
965 | { | |
966 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
967 | (i + 1) : i; | |
68435eb2 RB |
968 | inside_cost += record_stmt_cost (cost_vec, vect_pow2 (tmp), |
969 | vec_promote_demote, stmt_info, 0, | |
970 | vect_body); | |
8bd37302 BS |
971 | } |
972 | ||
973 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
974 | for (i = 0; i < 2; i++) | |
92345349 | 975 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
68435eb2 RB |
976 | prologue_cost += record_stmt_cost (cost_vec, 1, vector_stmt, |
977 | stmt_info, 0, vect_prologue); | |
8bd37302 | 978 | |
73fbfcad | 979 | if (dump_enabled_p ()) |
78c60e3d SS |
980 | dump_printf_loc (MSG_NOTE, vect_location, |
981 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 982 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
983 | } |
984 | ||
ebfd146a IR |
985 | /* Function vect_model_store_cost |
986 | ||
0d0293ac MM |
987 | Models cost for stores. In the case of grouped accesses, one access |
988 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a | 989 | |
68435eb2 | 990 | static void |
b8698a0f | 991 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
68435eb2 | 992 | enum vect_def_type dt, |
2de001ee | 993 | vect_memory_access_type memory_access_type, |
9ce4345a | 994 | vec_load_store_type vls_type, slp_tree slp_node, |
68435eb2 | 995 | stmt_vector_for_cost *cost_vec) |
ebfd146a | 996 | { |
92345349 | 997 | unsigned int inside_cost = 0, prologue_cost = 0; |
892a981f RS |
998 | gimple *first_stmt = STMT_VINFO_STMT (stmt_info); |
999 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); | |
ebfd146a | 1000 | |
68435eb2 RB |
1001 | /* ??? Somehow we need to fix this at the callers. */ |
1002 | if (slp_node) | |
1003 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
1004 | ||
9ce4345a | 1005 | if (vls_type == VLS_STORE_INVARIANT) |
68435eb2 RB |
1006 | { |
1007 | if (slp_node) | |
1008 | prologue_cost += vect_prologue_cost_for_slp_op (slp_node, stmt_info, | |
1009 | 1, dt, cost_vec); | |
1010 | else | |
1011 | prologue_cost += record_stmt_cost (cost_vec, 1, scalar_to_vec, | |
1012 | stmt_info, 0, vect_prologue); | |
1013 | } | |
ebfd146a | 1014 | |
892a981f RS |
1015 | /* Grouped stores update all elements in the group at once, |
1016 | so we want the DR for the first statement. */ | |
1017 | if (!slp_node && grouped_access_p) | |
57c454d2 | 1018 | first_stmt = DR_GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a | 1019 | |
892a981f RS |
1020 | /* True if we should include any once-per-group costs as well as |
1021 | the cost of the statement itself. For SLP we only get called | |
1022 | once per group anyhow. */ | |
1023 | bool first_stmt_p = (first_stmt == STMT_VINFO_STMT (stmt_info)); | |
1024 | ||
272c6793 | 1025 | /* We assume that the cost of a single store-lanes instruction is |
2c53b149 | 1026 | equivalent to the cost of DR_GROUP_SIZE separate stores. If a grouped |
272c6793 | 1027 | access is instead being provided by a permute-and-store operation, |
2de001ee RS |
1028 | include the cost of the permutes. */ |
1029 | if (first_stmt_p | |
1030 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 1031 | { |
e1377713 ES |
1032 | /* Uses a high and low interleave or shuffle operations for each |
1033 | needed permute. */ | |
2c53b149 | 1034 | int group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
e1377713 | 1035 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
68435eb2 | 1036 | inside_cost = record_stmt_cost (cost_vec, nstmts, vec_perm, |
92345349 | 1037 | stmt_info, 0, vect_body); |
ebfd146a | 1038 | |
73fbfcad | 1039 | if (dump_enabled_p ()) |
78c60e3d | 1040 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1041 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 1042 | group_size); |
ebfd146a IR |
1043 | } |
1044 | ||
cee62fee | 1045 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a | 1046 | /* Costs of the stores. */ |
067bc855 RB |
1047 | if (memory_access_type == VMAT_ELEMENTWISE |
1048 | || memory_access_type == VMAT_GATHER_SCATTER) | |
c5126ce8 RS |
1049 | { |
1050 | /* N scalar stores plus extracting the elements. */ | |
1051 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); | |
68435eb2 | 1052 | inside_cost += record_stmt_cost (cost_vec, |
c5126ce8 RS |
1053 | ncopies * assumed_nunits, |
1054 | scalar_store, stmt_info, 0, vect_body); | |
1055 | } | |
f2e2a985 | 1056 | else |
57c454d2 | 1057 | vect_get_store_cost (stmt_info, ncopies, &inside_cost, cost_vec); |
ebfd146a | 1058 | |
2de001ee RS |
1059 | if (memory_access_type == VMAT_ELEMENTWISE |
1060 | || memory_access_type == VMAT_STRIDED_SLP) | |
c5126ce8 RS |
1061 | { |
1062 | /* N scalar stores plus extracting the elements. */ | |
1063 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); | |
68435eb2 | 1064 | inside_cost += record_stmt_cost (cost_vec, |
c5126ce8 RS |
1065 | ncopies * assumed_nunits, |
1066 | vec_to_scalar, stmt_info, 0, vect_body); | |
1067 | } | |
cee62fee | 1068 | |
73fbfcad | 1069 | if (dump_enabled_p ()) |
78c60e3d SS |
1070 | dump_printf_loc (MSG_NOTE, vect_location, |
1071 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 1072 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
1073 | } |
1074 | ||
1075 | ||
720f5239 IR |
1076 | /* Calculate cost of DR's memory access. */ |
1077 | void | |
57c454d2 | 1078 | vect_get_store_cost (stmt_vec_info stmt_info, int ncopies, |
c3e7ee41 | 1079 | unsigned int *inside_cost, |
92345349 | 1080 | stmt_vector_for_cost *body_cost_vec) |
720f5239 | 1081 | { |
57c454d2 | 1082 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
720f5239 IR |
1083 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
1084 | ||
1085 | switch (alignment_support_scheme) | |
1086 | { | |
1087 | case dr_aligned: | |
1088 | { | |
92345349 BS |
1089 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
1090 | vector_store, stmt_info, 0, | |
1091 | vect_body); | |
720f5239 | 1092 | |
73fbfcad | 1093 | if (dump_enabled_p ()) |
78c60e3d | 1094 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1095 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
1096 | break; |
1097 | } | |
1098 | ||
1099 | case dr_unaligned_supported: | |
1100 | { | |
720f5239 | 1101 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1102 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1103 | unaligned_store, stmt_info, |
92345349 | 1104 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1105 | if (dump_enabled_p ()) |
78c60e3d SS |
1106 | dump_printf_loc (MSG_NOTE, vect_location, |
1107 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1108 | "hardware.\n"); |
720f5239 IR |
1109 | break; |
1110 | } | |
1111 | ||
38eec4c6 UW |
1112 | case dr_unaligned_unsupported: |
1113 | { | |
1114 | *inside_cost = VECT_MAX_COST; | |
1115 | ||
73fbfcad | 1116 | if (dump_enabled_p ()) |
78c60e3d | 1117 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1118 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1119 | break; |
1120 | } | |
1121 | ||
720f5239 IR |
1122 | default: |
1123 | gcc_unreachable (); | |
1124 | } | |
1125 | } | |
1126 | ||
1127 | ||
ebfd146a IR |
1128 | /* Function vect_model_load_cost |
1129 | ||
892a981f RS |
1130 | Models cost for loads. In the case of grouped accesses, one access has |
1131 | the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1132 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1133 | access scheme chosen. */ |
1134 | ||
68435eb2 RB |
1135 | static void |
1136 | vect_model_load_cost (stmt_vec_info stmt_info, unsigned ncopies, | |
2de001ee | 1137 | vect_memory_access_type memory_access_type, |
68435eb2 | 1138 | slp_instance instance, |
2de001ee | 1139 | slp_tree slp_node, |
68435eb2 | 1140 | stmt_vector_for_cost *cost_vec) |
ebfd146a | 1141 | { |
892a981f | 1142 | gimple *first_stmt = STMT_VINFO_STMT (stmt_info); |
92345349 | 1143 | unsigned int inside_cost = 0, prologue_cost = 0; |
892a981f | 1144 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); |
ebfd146a | 1145 | |
68435eb2 RB |
1146 | gcc_assert (cost_vec); |
1147 | ||
1148 | /* ??? Somehow we need to fix this at the callers. */ | |
1149 | if (slp_node) | |
1150 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
1151 | ||
1152 | if (slp_node && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
1153 | { | |
1154 | /* If the load is permuted then the alignment is determined by | |
1155 | the first group element not by the first scalar stmt DR. */ | |
2c53b149 | 1156 | gimple *stmt = DR_GROUP_FIRST_ELEMENT (stmt_info); |
68435eb2 RB |
1157 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
1158 | /* Record the cost for the permutation. */ | |
1159 | unsigned n_perms; | |
1160 | unsigned assumed_nunits | |
1161 | = vect_nunits_for_cost (STMT_VINFO_VECTYPE (stmt_info)); | |
1162 | unsigned slp_vf = (ncopies * assumed_nunits) / instance->group_size; | |
1163 | vect_transform_slp_perm_load (slp_node, vNULL, NULL, | |
1164 | slp_vf, instance, true, | |
1165 | &n_perms); | |
1166 | inside_cost += record_stmt_cost (cost_vec, n_perms, vec_perm, | |
1167 | stmt_info, 0, vect_body); | |
1168 | /* And adjust the number of loads performed. This handles | |
1169 | redundancies as well as loads that are later dead. */ | |
2c53b149 | 1170 | auto_sbitmap perm (DR_GROUP_SIZE (stmt_info)); |
68435eb2 RB |
1171 | bitmap_clear (perm); |
1172 | for (unsigned i = 0; | |
1173 | i < SLP_TREE_LOAD_PERMUTATION (slp_node).length (); ++i) | |
1174 | bitmap_set_bit (perm, SLP_TREE_LOAD_PERMUTATION (slp_node)[i]); | |
1175 | ncopies = 0; | |
1176 | bool load_seen = false; | |
2c53b149 | 1177 | for (unsigned i = 0; i < DR_GROUP_SIZE (stmt_info); ++i) |
68435eb2 RB |
1178 | { |
1179 | if (i % assumed_nunits == 0) | |
1180 | { | |
1181 | if (load_seen) | |
1182 | ncopies++; | |
1183 | load_seen = false; | |
1184 | } | |
1185 | if (bitmap_bit_p (perm, i)) | |
1186 | load_seen = true; | |
1187 | } | |
1188 | if (load_seen) | |
1189 | ncopies++; | |
1190 | gcc_assert (ncopies | |
2c53b149 | 1191 | <= (DR_GROUP_SIZE (stmt_info) - DR_GROUP_GAP (stmt_info) |
68435eb2 RB |
1192 | + assumed_nunits - 1) / assumed_nunits); |
1193 | } | |
1194 | ||
892a981f RS |
1195 | /* Grouped loads read all elements in the group at once, |
1196 | so we want the DR for the first statement. */ | |
1197 | if (!slp_node && grouped_access_p) | |
57c454d2 | 1198 | first_stmt = DR_GROUP_FIRST_ELEMENT (stmt_info); |
ebfd146a | 1199 | |
892a981f RS |
1200 | /* True if we should include any once-per-group costs as well as |
1201 | the cost of the statement itself. For SLP we only get called | |
1202 | once per group anyhow. */ | |
1203 | bool first_stmt_p = (first_stmt == STMT_VINFO_STMT (stmt_info)); | |
1204 | ||
272c6793 | 1205 | /* We assume that the cost of a single load-lanes instruction is |
2c53b149 | 1206 | equivalent to the cost of DR_GROUP_SIZE separate loads. If a grouped |
272c6793 | 1207 | access is instead being provided by a load-and-permute operation, |
2de001ee RS |
1208 | include the cost of the permutes. */ |
1209 | if (first_stmt_p | |
1210 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 1211 | { |
2c23db6d ES |
1212 | /* Uses an even and odd extract operations or shuffle operations |
1213 | for each needed permute. */ | |
2c53b149 | 1214 | int group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
2c23db6d | 1215 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
68435eb2 RB |
1216 | inside_cost += record_stmt_cost (cost_vec, nstmts, vec_perm, |
1217 | stmt_info, 0, vect_body); | |
ebfd146a | 1218 | |
73fbfcad | 1219 | if (dump_enabled_p ()) |
e645e942 TJ |
1220 | dump_printf_loc (MSG_NOTE, vect_location, |
1221 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1222 | group_size); |
ebfd146a IR |
1223 | } |
1224 | ||
1225 | /* The loads themselves. */ | |
067bc855 RB |
1226 | if (memory_access_type == VMAT_ELEMENTWISE |
1227 | || memory_access_type == VMAT_GATHER_SCATTER) | |
a82960aa | 1228 | { |
a21892ad BS |
1229 | /* N scalar loads plus gathering them into a vector. */ |
1230 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
c5126ce8 | 1231 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); |
68435eb2 | 1232 | inside_cost += record_stmt_cost (cost_vec, |
c5126ce8 | 1233 | ncopies * assumed_nunits, |
92345349 | 1234 | scalar_load, stmt_info, 0, vect_body); |
a82960aa RG |
1235 | } |
1236 | else | |
57c454d2 | 1237 | vect_get_load_cost (stmt_info, ncopies, first_stmt_p, |
92345349 | 1238 | &inside_cost, &prologue_cost, |
68435eb2 | 1239 | cost_vec, cost_vec, true); |
2de001ee RS |
1240 | if (memory_access_type == VMAT_ELEMENTWISE |
1241 | || memory_access_type == VMAT_STRIDED_SLP) | |
68435eb2 | 1242 | inside_cost += record_stmt_cost (cost_vec, ncopies, vec_construct, |
892a981f | 1243 | stmt_info, 0, vect_body); |
720f5239 | 1244 | |
73fbfcad | 1245 | if (dump_enabled_p ()) |
78c60e3d SS |
1246 | dump_printf_loc (MSG_NOTE, vect_location, |
1247 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1248 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1249 | } |
1250 | ||
1251 | ||
1252 | /* Calculate cost of DR's memory access. */ | |
1253 | void | |
57c454d2 | 1254 | vect_get_load_cost (stmt_vec_info stmt_info, int ncopies, |
c3e7ee41 | 1255 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1256 | unsigned int *prologue_cost, |
1257 | stmt_vector_for_cost *prologue_cost_vec, | |
1258 | stmt_vector_for_cost *body_cost_vec, | |
1259 | bool record_prologue_costs) | |
720f5239 | 1260 | { |
57c454d2 | 1261 | data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
720f5239 IR |
1262 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); |
1263 | ||
1264 | switch (alignment_support_scheme) | |
ebfd146a IR |
1265 | { |
1266 | case dr_aligned: | |
1267 | { | |
92345349 BS |
1268 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1269 | stmt_info, 0, vect_body); | |
ebfd146a | 1270 | |
73fbfcad | 1271 | if (dump_enabled_p ()) |
78c60e3d | 1272 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1273 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1274 | |
1275 | break; | |
1276 | } | |
1277 | case dr_unaligned_supported: | |
1278 | { | |
720f5239 | 1279 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1280 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1281 | unaligned_load, stmt_info, |
92345349 | 1282 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1283 | |
73fbfcad | 1284 | if (dump_enabled_p ()) |
78c60e3d SS |
1285 | dump_printf_loc (MSG_NOTE, vect_location, |
1286 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1287 | "hardware.\n"); |
ebfd146a IR |
1288 | |
1289 | break; | |
1290 | } | |
1291 | case dr_explicit_realign: | |
1292 | { | |
92345349 BS |
1293 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1294 | vector_load, stmt_info, 0, vect_body); | |
1295 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1296 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1297 | |
1298 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1299 | the containing loop, the following cost should be added to the | |
92345349 | 1300 | prologue costs. */ |
ebfd146a | 1301 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1302 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1303 | stmt_info, 0, vect_body); | |
ebfd146a | 1304 | |
73fbfcad | 1305 | if (dump_enabled_p ()) |
e645e942 TJ |
1306 | dump_printf_loc (MSG_NOTE, vect_location, |
1307 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1308 | |
ebfd146a IR |
1309 | break; |
1310 | } | |
1311 | case dr_explicit_realign_optimized: | |
1312 | { | |
73fbfcad | 1313 | if (dump_enabled_p ()) |
e645e942 | 1314 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1315 | "vect_model_load_cost: unaligned software " |
e645e942 | 1316 | "pipelined.\n"); |
ebfd146a IR |
1317 | |
1318 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1319 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1320 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1321 | access, then the above cost should only be considered for one |
ff802fa1 | 1322 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1323 | and a realignment op. */ |
1324 | ||
92345349 | 1325 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1326 | { |
92345349 BS |
1327 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1328 | vector_stmt, stmt_info, | |
1329 | 0, vect_prologue); | |
ebfd146a | 1330 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1331 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1332 | vector_stmt, stmt_info, | |
1333 | 0, vect_prologue); | |
ebfd146a IR |
1334 | } |
1335 | ||
92345349 BS |
1336 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1337 | stmt_info, 0, vect_body); | |
1338 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1339 | stmt_info, 0, vect_body); | |
8bd37302 | 1340 | |
73fbfcad | 1341 | if (dump_enabled_p ()) |
78c60e3d | 1342 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1343 | "vect_model_load_cost: explicit realign optimized" |
1344 | "\n"); | |
8bd37302 | 1345 | |
ebfd146a IR |
1346 | break; |
1347 | } | |
1348 | ||
38eec4c6 UW |
1349 | case dr_unaligned_unsupported: |
1350 | { | |
1351 | *inside_cost = VECT_MAX_COST; | |
1352 | ||
73fbfcad | 1353 | if (dump_enabled_p ()) |
78c60e3d | 1354 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1355 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1356 | break; |
1357 | } | |
1358 | ||
ebfd146a IR |
1359 | default: |
1360 | gcc_unreachable (); | |
1361 | } | |
ebfd146a IR |
1362 | } |
1363 | ||
418b7df3 RG |
1364 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1365 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1366 | |
418b7df3 | 1367 | static void |
355fe088 | 1368 | vect_init_vector_1 (gimple *stmt, gimple *new_stmt, gimple_stmt_iterator *gsi) |
ebfd146a | 1369 | { |
ebfd146a | 1370 | if (gsi) |
418b7df3 | 1371 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1372 | else |
1373 | { | |
418b7df3 | 1374 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1375 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1376 | |
a70d6342 IR |
1377 | if (loop_vinfo) |
1378 | { | |
1379 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1380 | basic_block new_bb; |
1381 | edge pe; | |
a70d6342 IR |
1382 | |
1383 | if (nested_in_vect_loop_p (loop, stmt)) | |
1384 | loop = loop->inner; | |
b8698a0f | 1385 | |
a70d6342 | 1386 | pe = loop_preheader_edge (loop); |
418b7df3 | 1387 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1388 | gcc_assert (!new_bb); |
1389 | } | |
1390 | else | |
1391 | { | |
1392 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1393 | basic_block bb; | |
1394 | gimple_stmt_iterator gsi_bb_start; | |
1395 | ||
1396 | gcc_assert (bb_vinfo); | |
1397 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1398 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1399 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1400 | } |
ebfd146a IR |
1401 | } |
1402 | ||
73fbfcad | 1403 | if (dump_enabled_p ()) |
ebfd146a | 1404 | { |
78c60e3d SS |
1405 | dump_printf_loc (MSG_NOTE, vect_location, |
1406 | "created new init_stmt: "); | |
1407 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
ebfd146a | 1408 | } |
418b7df3 RG |
1409 | } |
1410 | ||
1411 | /* Function vect_init_vector. | |
ebfd146a | 1412 | |
5467ee52 RG |
1413 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1414 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1415 | vector type a vector with all elements equal to VAL is created first. | |
1416 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1417 | initialization at the loop preheader. | |
418b7df3 RG |
1418 | Return the DEF of INIT_STMT. |
1419 | It will be used in the vectorization of STMT. */ | |
1420 | ||
1421 | tree | |
355fe088 | 1422 | vect_init_vector (gimple *stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 | 1423 | { |
355fe088 | 1424 | gimple *init_stmt; |
418b7df3 RG |
1425 | tree new_temp; |
1426 | ||
e412ece4 RB |
1427 | /* We abuse this function to push sth to a SSA name with initial 'val'. */ |
1428 | if (! useless_type_conversion_p (type, TREE_TYPE (val))) | |
418b7df3 | 1429 | { |
e412ece4 RB |
1430 | gcc_assert (TREE_CODE (type) == VECTOR_TYPE); |
1431 | if (! types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) | |
418b7df3 | 1432 | { |
5a308cf1 IE |
1433 | /* Scalar boolean value should be transformed into |
1434 | all zeros or all ones value before building a vector. */ | |
1435 | if (VECTOR_BOOLEAN_TYPE_P (type)) | |
1436 | { | |
b3d51f23 IE |
1437 | tree true_val = build_all_ones_cst (TREE_TYPE (type)); |
1438 | tree false_val = build_zero_cst (TREE_TYPE (type)); | |
5a308cf1 IE |
1439 | |
1440 | if (CONSTANT_CLASS_P (val)) | |
1441 | val = integer_zerop (val) ? false_val : true_val; | |
1442 | else | |
1443 | { | |
1444 | new_temp = make_ssa_name (TREE_TYPE (type)); | |
1445 | init_stmt = gimple_build_assign (new_temp, COND_EXPR, | |
1446 | val, true_val, false_val); | |
1447 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
1448 | val = new_temp; | |
1449 | } | |
1450 | } | |
1451 | else if (CONSTANT_CLASS_P (val)) | |
42fd8198 | 1452 | val = fold_convert (TREE_TYPE (type), val); |
418b7df3 RG |
1453 | else |
1454 | { | |
b731b390 | 1455 | new_temp = make_ssa_name (TREE_TYPE (type)); |
e412ece4 RB |
1456 | if (! INTEGRAL_TYPE_P (TREE_TYPE (val))) |
1457 | init_stmt = gimple_build_assign (new_temp, | |
1458 | fold_build1 (VIEW_CONVERT_EXPR, | |
1459 | TREE_TYPE (type), | |
1460 | val)); | |
1461 | else | |
1462 | init_stmt = gimple_build_assign (new_temp, NOP_EXPR, val); | |
418b7df3 | 1463 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1464 | val = new_temp; |
418b7df3 RG |
1465 | } |
1466 | } | |
5467ee52 | 1467 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1468 | } |
1469 | ||
0e22bb5a RB |
1470 | new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_"); |
1471 | init_stmt = gimple_build_assign (new_temp, val); | |
418b7df3 | 1472 | vect_init_vector_1 (stmt, init_stmt, gsi); |
0e22bb5a | 1473 | return new_temp; |
ebfd146a IR |
1474 | } |
1475 | ||
c83a894c | 1476 | /* Function vect_get_vec_def_for_operand_1. |
a70d6342 | 1477 | |
c83a894c AH |
1478 | For a defining stmt DEF_STMT of a scalar stmt, return a vector def with type |
1479 | DT that will be used in the vectorized stmt. */ | |
ebfd146a IR |
1480 | |
1481 | tree | |
c83a894c | 1482 | vect_get_vec_def_for_operand_1 (gimple *def_stmt, enum vect_def_type dt) |
ebfd146a IR |
1483 | { |
1484 | tree vec_oprnd; | |
355fe088 | 1485 | gimple *vec_stmt; |
ebfd146a | 1486 | stmt_vec_info def_stmt_info = NULL; |
ebfd146a IR |
1487 | |
1488 | switch (dt) | |
1489 | { | |
81c40241 | 1490 | /* operand is a constant or a loop invariant. */ |
ebfd146a | 1491 | case vect_constant_def: |
81c40241 | 1492 | case vect_external_def: |
c83a894c AH |
1493 | /* Code should use vect_get_vec_def_for_operand. */ |
1494 | gcc_unreachable (); | |
ebfd146a | 1495 | |
81c40241 | 1496 | /* operand is defined inside the loop. */ |
8644a673 | 1497 | case vect_internal_def: |
ebfd146a | 1498 | { |
ebfd146a IR |
1499 | /* Get the def from the vectorized stmt. */ |
1500 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1501 | |
ebfd146a | 1502 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1503 | /* Get vectorized pattern statement. */ |
1504 | if (!vec_stmt | |
1505 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1506 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1507 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1508 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1509 | gcc_assert (vec_stmt); |
1510 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1511 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1512 | else if (is_gimple_call (vec_stmt)) | |
1513 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1514 | else | |
1515 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1516 | return vec_oprnd; | |
1517 | } | |
1518 | ||
c78e3652 | 1519 | /* operand is defined by a loop header phi. */ |
ebfd146a | 1520 | case vect_reduction_def: |
06066f92 | 1521 | case vect_double_reduction_def: |
7c5222ff | 1522 | case vect_nested_cycle: |
ebfd146a IR |
1523 | case vect_induction_def: |
1524 | { | |
1525 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1526 | ||
1527 | /* Get the def from the vectorized stmt. */ | |
1528 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1529 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1530 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1531 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1532 | else | |
1533 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1534 | return vec_oprnd; |
1535 | } | |
1536 | ||
1537 | default: | |
1538 | gcc_unreachable (); | |
1539 | } | |
1540 | } | |
1541 | ||
1542 | ||
c83a894c AH |
1543 | /* Function vect_get_vec_def_for_operand. |
1544 | ||
1545 | OP is an operand in STMT. This function returns a (vector) def that will be | |
1546 | used in the vectorized stmt for STMT. | |
1547 | ||
1548 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1549 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1550 | ||
1551 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1552 | needs to be introduced. VECTYPE may be used to specify a required type for | |
1553 | vector invariant. */ | |
1554 | ||
1555 | tree | |
1556 | vect_get_vec_def_for_operand (tree op, gimple *stmt, tree vectype) | |
1557 | { | |
1558 | gimple *def_stmt; | |
1559 | enum vect_def_type dt; | |
1560 | bool is_simple_use; | |
1561 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
1562 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
1563 | ||
1564 | if (dump_enabled_p ()) | |
1565 | { | |
1566 | dump_printf_loc (MSG_NOTE, vect_location, | |
1567 | "vect_get_vec_def_for_operand: "); | |
1568 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
1569 | dump_printf (MSG_NOTE, "\n"); | |
1570 | } | |
1571 | ||
1572 | is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt); | |
1573 | gcc_assert (is_simple_use); | |
1574 | if (def_stmt && dump_enabled_p ()) | |
1575 | { | |
1576 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1577 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
1578 | } | |
1579 | ||
1580 | if (dt == vect_constant_def || dt == vect_external_def) | |
1581 | { | |
1582 | tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo); | |
1583 | tree vector_type; | |
1584 | ||
1585 | if (vectype) | |
1586 | vector_type = vectype; | |
2568d8a1 | 1587 | else if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op)) |
c83a894c AH |
1588 | && VECTOR_BOOLEAN_TYPE_P (stmt_vectype)) |
1589 | vector_type = build_same_sized_truth_vector_type (stmt_vectype); | |
1590 | else | |
1591 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
1592 | ||
1593 | gcc_assert (vector_type); | |
1594 | return vect_init_vector (stmt, op, vector_type, NULL); | |
1595 | } | |
1596 | else | |
1597 | return vect_get_vec_def_for_operand_1 (def_stmt, dt); | |
1598 | } | |
1599 | ||
1600 | ||
ebfd146a IR |
1601 | /* Function vect_get_vec_def_for_stmt_copy |
1602 | ||
ff802fa1 | 1603 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1604 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1605 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1606 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1607 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1608 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1609 | DT is the type of the vector def VEC_OPRND. |
1610 | ||
1611 | Context: | |
1612 | In case the vectorization factor (VF) is bigger than the number | |
1613 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1614 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1615 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1616 | smallest data-type determines the VF, and as a result, when vectorizing |
1617 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1618 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1619 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1620 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1621 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1622 | ||
1623 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1624 | |
ebfd146a IR |
1625 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1626 | VS1.1: vx.1 = memref1 VS1.2 | |
1627 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1628 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1629 | |
1630 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1631 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1632 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1633 | VSnew.3: vz3 = vx.3 + ... | |
1634 | ||
1635 | The vectorization of S1 is explained in vectorizable_load. | |
1636 | The vectorization of S2: | |
b8698a0f L |
1637 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1638 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1639 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1640 | returns the vector-def 'vx.0'. |
1641 | ||
b8698a0f L |
1642 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1643 | function is called to get the relevant vector-def for each operand. It is | |
1644 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1645 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1646 | ||
b8698a0f L |
1647 | For example, to obtain the vector-def 'vx.1' in order to create the |
1648 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1649 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1650 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1651 | and return its def ('vx.1'). | |
1652 | Overall, to create the above sequence this function will be called 3 times: | |
1653 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1654 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1655 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1656 | ||
1657 | tree | |
1658 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1659 | { | |
355fe088 | 1660 | gimple *vec_stmt_for_operand; |
ebfd146a IR |
1661 | stmt_vec_info def_stmt_info; |
1662 | ||
1663 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1664 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1665 | return vec_oprnd; |
1666 | ||
1667 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1668 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1669 | gcc_assert (def_stmt_info); | |
1670 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1671 | gcc_assert (vec_stmt_for_operand); | |
ebfd146a IR |
1672 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) |
1673 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1674 | else | |
1675 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1676 | return vec_oprnd; | |
1677 | } | |
1678 | ||
1679 | ||
1680 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1681 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a | 1682 | |
c78e3652 | 1683 | void |
b8698a0f | 1684 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1685 | vec<tree> *vec_oprnds0, |
1686 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1687 | { |
9771b263 | 1688 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1689 | |
1690 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1691 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1692 | |
9771b263 | 1693 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1694 | { |
9771b263 | 1695 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1696 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1697 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1698 | } |
1699 | } | |
1700 | ||
1701 | ||
c78e3652 | 1702 | /* Get vectorized definitions for OP0 and OP1. */ |
ebfd146a | 1703 | |
c78e3652 | 1704 | void |
355fe088 | 1705 | vect_get_vec_defs (tree op0, tree op1, gimple *stmt, |
9771b263 DN |
1706 | vec<tree> *vec_oprnds0, |
1707 | vec<tree> *vec_oprnds1, | |
306b0c92 | 1708 | slp_tree slp_node) |
ebfd146a IR |
1709 | { |
1710 | if (slp_node) | |
d092494c IR |
1711 | { |
1712 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1713 | auto_vec<tree> ops (nops); |
1714 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1715 | |
9771b263 | 1716 | ops.quick_push (op0); |
d092494c | 1717 | if (op1) |
9771b263 | 1718 | ops.quick_push (op1); |
d092494c | 1719 | |
306b0c92 | 1720 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
d092494c | 1721 | |
37b5ec8f | 1722 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1723 | if (op1) |
37b5ec8f | 1724 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1725 | } |
ebfd146a IR |
1726 | else |
1727 | { | |
1728 | tree vec_oprnd; | |
1729 | ||
9771b263 | 1730 | vec_oprnds0->create (1); |
81c40241 | 1731 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 1732 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1733 | |
1734 | if (op1) | |
1735 | { | |
9771b263 | 1736 | vec_oprnds1->create (1); |
81c40241 | 1737 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 1738 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1739 | } |
1740 | } | |
1741 | } | |
1742 | ||
bb6c2b68 RS |
1743 | /* Helper function called by vect_finish_replace_stmt and |
1744 | vect_finish_stmt_generation. Set the location of the new | |
1745 | statement and create a stmt_vec_info for it. */ | |
1746 | ||
1747 | static void | |
1748 | vect_finish_stmt_generation_1 (gimple *stmt, gimple *vec_stmt) | |
1749 | { | |
1750 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1751 | vec_info *vinfo = stmt_info->vinfo; | |
1752 | ||
1753 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, vinfo)); | |
1754 | ||
1755 | if (dump_enabled_p ()) | |
1756 | { | |
1757 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); | |
1758 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
1759 | } | |
1760 | ||
1761 | gimple_set_location (vec_stmt, gimple_location (stmt)); | |
1762 | ||
1763 | /* While EH edges will generally prevent vectorization, stmt might | |
1764 | e.g. be in a must-not-throw region. Ensure newly created stmts | |
1765 | that could throw are part of the same region. */ | |
1766 | int lp_nr = lookup_stmt_eh_lp (stmt); | |
1767 | if (lp_nr != 0 && stmt_could_throw_p (vec_stmt)) | |
1768 | add_stmt_to_eh_lp (vec_stmt, lp_nr); | |
1769 | } | |
1770 | ||
1771 | /* Replace the scalar statement STMT with a new vector statement VEC_STMT, | |
1772 | which sets the same scalar result as STMT did. */ | |
1773 | ||
1774 | void | |
1775 | vect_finish_replace_stmt (gimple *stmt, gimple *vec_stmt) | |
1776 | { | |
1777 | gcc_assert (gimple_get_lhs (stmt) == gimple_get_lhs (vec_stmt)); | |
1778 | ||
1779 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); | |
1780 | gsi_replace (&gsi, vec_stmt, false); | |
1781 | ||
1782 | vect_finish_stmt_generation_1 (stmt, vec_stmt); | |
1783 | } | |
ebfd146a IR |
1784 | |
1785 | /* Function vect_finish_stmt_generation. | |
1786 | ||
1787 | Insert a new stmt. */ | |
1788 | ||
1789 | void | |
355fe088 | 1790 | vect_finish_stmt_generation (gimple *stmt, gimple *vec_stmt, |
ebfd146a IR |
1791 | gimple_stmt_iterator *gsi) |
1792 | { | |
ebfd146a IR |
1793 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); |
1794 | ||
54e8e2c3 RG |
1795 | if (!gsi_end_p (*gsi) |
1796 | && gimple_has_mem_ops (vec_stmt)) | |
1797 | { | |
355fe088 | 1798 | gimple *at_stmt = gsi_stmt (*gsi); |
54e8e2c3 RG |
1799 | tree vuse = gimple_vuse (at_stmt); |
1800 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1801 | { | |
1802 | tree vdef = gimple_vdef (at_stmt); | |
1803 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1804 | /* If we have an SSA vuse and insert a store, update virtual | |
1805 | SSA form to avoid triggering the renamer. Do so only | |
1806 | if we can easily see all uses - which is what almost always | |
1807 | happens with the way vectorized stmts are inserted. */ | |
1808 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1809 | && ((is_gimple_assign (vec_stmt) | |
1810 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1811 | || (is_gimple_call (vec_stmt) | |
1812 | && !(gimple_call_flags (vec_stmt) | |
1813 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1814 | { | |
1815 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1816 | gimple_set_vdef (vec_stmt, new_vdef); | |
1817 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1818 | } | |
1819 | } | |
1820 | } | |
ebfd146a | 1821 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
bb6c2b68 | 1822 | vect_finish_stmt_generation_1 (stmt, vec_stmt); |
ebfd146a IR |
1823 | } |
1824 | ||
70439f0d RS |
1825 | /* We want to vectorize a call to combined function CFN with function |
1826 | decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN | |
1827 | as the types of all inputs. Check whether this is possible using | |
1828 | an internal function, returning its code if so or IFN_LAST if not. */ | |
ebfd146a | 1829 | |
70439f0d RS |
1830 | static internal_fn |
1831 | vectorizable_internal_function (combined_fn cfn, tree fndecl, | |
1832 | tree vectype_out, tree vectype_in) | |
ebfd146a | 1833 | { |
70439f0d RS |
1834 | internal_fn ifn; |
1835 | if (internal_fn_p (cfn)) | |
1836 | ifn = as_internal_fn (cfn); | |
1837 | else | |
1838 | ifn = associated_internal_fn (fndecl); | |
1839 | if (ifn != IFN_LAST && direct_internal_fn_p (ifn)) | |
1840 | { | |
1841 | const direct_internal_fn_info &info = direct_internal_fn (ifn); | |
1842 | if (info.vectorizable) | |
1843 | { | |
1844 | tree type0 = (info.type0 < 0 ? vectype_out : vectype_in); | |
1845 | tree type1 = (info.type1 < 0 ? vectype_out : vectype_in); | |
d95ab70a RS |
1846 | if (direct_internal_fn_supported_p (ifn, tree_pair (type0, type1), |
1847 | OPTIMIZE_FOR_SPEED)) | |
70439f0d RS |
1848 | return ifn; |
1849 | } | |
1850 | } | |
1851 | return IFN_LAST; | |
ebfd146a IR |
1852 | } |
1853 | ||
5ce9450f | 1854 | |
355fe088 | 1855 | static tree permute_vec_elements (tree, tree, tree, gimple *, |
5ce9450f JJ |
1856 | gimple_stmt_iterator *); |
1857 | ||
7cfb4d93 RS |
1858 | /* Check whether a load or store statement in the loop described by |
1859 | LOOP_VINFO is possible in a fully-masked loop. This is testing | |
1860 | whether the vectorizer pass has the appropriate support, as well as | |
1861 | whether the target does. | |
1862 | ||
1863 | VLS_TYPE says whether the statement is a load or store and VECTYPE | |
1864 | is the type of the vector being loaded or stored. MEMORY_ACCESS_TYPE | |
1865 | says how the load or store is going to be implemented and GROUP_SIZE | |
1866 | is the number of load or store statements in the containing group. | |
bfaa08b7 RS |
1867 | If the access is a gather load or scatter store, GS_INFO describes |
1868 | its arguments. | |
7cfb4d93 RS |
1869 | |
1870 | Clear LOOP_VINFO_CAN_FULLY_MASK_P if a fully-masked loop is not | |
1871 | supported, otherwise record the required mask types. */ | |
1872 | ||
1873 | static void | |
1874 | check_load_store_masking (loop_vec_info loop_vinfo, tree vectype, | |
1875 | vec_load_store_type vls_type, int group_size, | |
bfaa08b7 RS |
1876 | vect_memory_access_type memory_access_type, |
1877 | gather_scatter_info *gs_info) | |
7cfb4d93 RS |
1878 | { |
1879 | /* Invariant loads need no special support. */ | |
1880 | if (memory_access_type == VMAT_INVARIANT) | |
1881 | return; | |
1882 | ||
1883 | vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); | |
1884 | machine_mode vecmode = TYPE_MODE (vectype); | |
1885 | bool is_load = (vls_type == VLS_LOAD); | |
1886 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
1887 | { | |
1888 | if (is_load | |
1889 | ? !vect_load_lanes_supported (vectype, group_size, true) | |
1890 | : !vect_store_lanes_supported (vectype, group_size, true)) | |
1891 | { | |
1892 | if (dump_enabled_p ()) | |
1893 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1894 | "can't use a fully-masked loop because the" | |
1895 | " target doesn't have an appropriate masked" | |
1896 | " load/store-lanes instruction.\n"); | |
1897 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1898 | return; | |
1899 | } | |
1900 | unsigned int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
1901 | vect_record_loop_mask (loop_vinfo, masks, ncopies, vectype); | |
1902 | return; | |
1903 | } | |
1904 | ||
bfaa08b7 RS |
1905 | if (memory_access_type == VMAT_GATHER_SCATTER) |
1906 | { | |
f307441a RS |
1907 | internal_fn ifn = (is_load |
1908 | ? IFN_MASK_GATHER_LOAD | |
1909 | : IFN_MASK_SCATTER_STORE); | |
bfaa08b7 | 1910 | tree offset_type = TREE_TYPE (gs_info->offset); |
f307441a | 1911 | if (!internal_gather_scatter_fn_supported_p (ifn, vectype, |
bfaa08b7 RS |
1912 | gs_info->memory_type, |
1913 | TYPE_SIGN (offset_type), | |
1914 | gs_info->scale)) | |
1915 | { | |
1916 | if (dump_enabled_p ()) | |
1917 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1918 | "can't use a fully-masked loop because the" | |
1919 | " target doesn't have an appropriate masked" | |
f307441a | 1920 | " gather load or scatter store instruction.\n"); |
bfaa08b7 RS |
1921 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; |
1922 | return; | |
1923 | } | |
1924 | unsigned int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
1925 | vect_record_loop_mask (loop_vinfo, masks, ncopies, vectype); | |
1926 | return; | |
1927 | } | |
1928 | ||
7cfb4d93 RS |
1929 | if (memory_access_type != VMAT_CONTIGUOUS |
1930 | && memory_access_type != VMAT_CONTIGUOUS_PERMUTE) | |
1931 | { | |
1932 | /* Element X of the data must come from iteration i * VF + X of the | |
1933 | scalar loop. We need more work to support other mappings. */ | |
1934 | if (dump_enabled_p ()) | |
1935 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1936 | "can't use a fully-masked loop because an access" | |
1937 | " isn't contiguous.\n"); | |
1938 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1939 | return; | |
1940 | } | |
1941 | ||
1942 | machine_mode mask_mode; | |
1943 | if (!(targetm.vectorize.get_mask_mode | |
1944 | (GET_MODE_NUNITS (vecmode), | |
1945 | GET_MODE_SIZE (vecmode)).exists (&mask_mode)) | |
1946 | || !can_vec_mask_load_store_p (vecmode, mask_mode, is_load)) | |
1947 | { | |
1948 | if (dump_enabled_p ()) | |
1949 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1950 | "can't use a fully-masked loop because the target" | |
1951 | " doesn't have the appropriate masked load or" | |
1952 | " store.\n"); | |
1953 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1954 | return; | |
1955 | } | |
1956 | /* We might load more scalars than we need for permuting SLP loads. | |
1957 | We checked in get_group_load_store_type that the extra elements | |
1958 | don't leak into a new vector. */ | |
1959 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
1960 | poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
1961 | unsigned int nvectors; | |
1962 | if (can_div_away_from_zero_p (group_size * vf, nunits, &nvectors)) | |
1963 | vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype); | |
1964 | else | |
1965 | gcc_unreachable (); | |
1966 | } | |
1967 | ||
1968 | /* Return the mask input to a masked load or store. VEC_MASK is the vectorized | |
1969 | form of the scalar mask condition and LOOP_MASK, if nonnull, is the mask | |
1970 | that needs to be applied to all loads and stores in a vectorized loop. | |
1971 | Return VEC_MASK if LOOP_MASK is null, otherwise return VEC_MASK & LOOP_MASK. | |
1972 | ||
1973 | MASK_TYPE is the type of both masks. If new statements are needed, | |
1974 | insert them before GSI. */ | |
1975 | ||
1976 | static tree | |
1977 | prepare_load_store_mask (tree mask_type, tree loop_mask, tree vec_mask, | |
1978 | gimple_stmt_iterator *gsi) | |
1979 | { | |
1980 | gcc_assert (useless_type_conversion_p (mask_type, TREE_TYPE (vec_mask))); | |
1981 | if (!loop_mask) | |
1982 | return vec_mask; | |
1983 | ||
1984 | gcc_assert (TREE_TYPE (loop_mask) == mask_type); | |
1985 | tree and_res = make_temp_ssa_name (mask_type, NULL, "vec_mask_and"); | |
1986 | gimple *and_stmt = gimple_build_assign (and_res, BIT_AND_EXPR, | |
1987 | vec_mask, loop_mask); | |
1988 | gsi_insert_before (gsi, and_stmt, GSI_SAME_STMT); | |
1989 | return and_res; | |
1990 | } | |
1991 | ||
429ef523 RS |
1992 | /* Determine whether we can use a gather load or scatter store to vectorize |
1993 | strided load or store STMT by truncating the current offset to a smaller | |
1994 | width. We need to be able to construct an offset vector: | |
1995 | ||
1996 | { 0, X, X*2, X*3, ... } | |
1997 | ||
1998 | without loss of precision, where X is STMT's DR_STEP. | |
1999 | ||
2000 | Return true if this is possible, describing the gather load or scatter | |
2001 | store in GS_INFO. MASKED_P is true if the load or store is conditional. */ | |
2002 | ||
2003 | static bool | |
2004 | vect_truncate_gather_scatter_offset (gimple *stmt, loop_vec_info loop_vinfo, | |
2005 | bool masked_p, | |
2006 | gather_scatter_info *gs_info) | |
2007 | { | |
2008 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2009 | data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
2010 | tree step = DR_STEP (dr); | |
2011 | if (TREE_CODE (step) != INTEGER_CST) | |
2012 | { | |
2013 | /* ??? Perhaps we could use range information here? */ | |
2014 | if (dump_enabled_p ()) | |
2015 | dump_printf_loc (MSG_NOTE, vect_location, | |
2016 | "cannot truncate variable step.\n"); | |
2017 | return false; | |
2018 | } | |
2019 | ||
2020 | /* Get the number of bits in an element. */ | |
2021 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2022 | scalar_mode element_mode = SCALAR_TYPE_MODE (TREE_TYPE (vectype)); | |
2023 | unsigned int element_bits = GET_MODE_BITSIZE (element_mode); | |
2024 | ||
2025 | /* Set COUNT to the upper limit on the number of elements - 1. | |
2026 | Start with the maximum vectorization factor. */ | |
2027 | unsigned HOST_WIDE_INT count = vect_max_vf (loop_vinfo) - 1; | |
2028 | ||
2029 | /* Try lowering COUNT to the number of scalar latch iterations. */ | |
2030 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2031 | widest_int max_iters; | |
2032 | if (max_loop_iterations (loop, &max_iters) | |
2033 | && max_iters < count) | |
2034 | count = max_iters.to_shwi (); | |
2035 | ||
2036 | /* Try scales of 1 and the element size. */ | |
2037 | int scales[] = { 1, vect_get_scalar_dr_size (dr) }; | |
2038 | bool overflow_p = false; | |
2039 | for (int i = 0; i < 2; ++i) | |
2040 | { | |
2041 | int scale = scales[i]; | |
2042 | widest_int factor; | |
2043 | if (!wi::multiple_of_p (wi::to_widest (step), scale, SIGNED, &factor)) | |
2044 | continue; | |
2045 | ||
2046 | /* See whether we can calculate (COUNT - 1) * STEP / SCALE | |
2047 | in OFFSET_BITS bits. */ | |
2048 | widest_int range = wi::mul (count, factor, SIGNED, &overflow_p); | |
2049 | if (overflow_p) | |
2050 | continue; | |
2051 | signop sign = range >= 0 ? UNSIGNED : SIGNED; | |
2052 | if (wi::min_precision (range, sign) > element_bits) | |
2053 | { | |
2054 | overflow_p = true; | |
2055 | continue; | |
2056 | } | |
2057 | ||
2058 | /* See whether the target supports the operation. */ | |
2059 | tree memory_type = TREE_TYPE (DR_REF (dr)); | |
2060 | if (!vect_gather_scatter_fn_p (DR_IS_READ (dr), masked_p, vectype, | |
2061 | memory_type, element_bits, sign, scale, | |
2062 | &gs_info->ifn, &gs_info->element_type)) | |
2063 | continue; | |
2064 | ||
2065 | tree offset_type = build_nonstandard_integer_type (element_bits, | |
2066 | sign == UNSIGNED); | |
2067 | ||
2068 | gs_info->decl = NULL_TREE; | |
2069 | /* Logically the sum of DR_BASE_ADDRESS, DR_INIT and DR_OFFSET, | |
2070 | but we don't need to store that here. */ | |
2071 | gs_info->base = NULL_TREE; | |
2072 | gs_info->offset = fold_convert (offset_type, step); | |
929b4411 | 2073 | gs_info->offset_dt = vect_constant_def; |
429ef523 RS |
2074 | gs_info->offset_vectype = NULL_TREE; |
2075 | gs_info->scale = scale; | |
2076 | gs_info->memory_type = memory_type; | |
2077 | return true; | |
2078 | } | |
2079 | ||
2080 | if (overflow_p && dump_enabled_p ()) | |
2081 | dump_printf_loc (MSG_NOTE, vect_location, | |
2082 | "truncating gather/scatter offset to %d bits" | |
2083 | " might change its value.\n", element_bits); | |
2084 | ||
2085 | return false; | |
2086 | } | |
2087 | ||
ab2fc782 RS |
2088 | /* Return true if we can use gather/scatter internal functions to |
2089 | vectorize STMT, which is a grouped or strided load or store. | |
429ef523 RS |
2090 | MASKED_P is true if load or store is conditional. When returning |
2091 | true, fill in GS_INFO with the information required to perform the | |
2092 | operation. */ | |
ab2fc782 RS |
2093 | |
2094 | static bool | |
2095 | vect_use_strided_gather_scatters_p (gimple *stmt, loop_vec_info loop_vinfo, | |
429ef523 | 2096 | bool masked_p, |
ab2fc782 RS |
2097 | gather_scatter_info *gs_info) |
2098 | { | |
2099 | if (!vect_check_gather_scatter (stmt, loop_vinfo, gs_info) | |
2100 | || gs_info->decl) | |
429ef523 RS |
2101 | return vect_truncate_gather_scatter_offset (stmt, loop_vinfo, |
2102 | masked_p, gs_info); | |
ab2fc782 RS |
2103 | |
2104 | scalar_mode element_mode = SCALAR_TYPE_MODE (gs_info->element_type); | |
2105 | unsigned int element_bits = GET_MODE_BITSIZE (element_mode); | |
2106 | tree offset_type = TREE_TYPE (gs_info->offset); | |
2107 | unsigned int offset_bits = TYPE_PRECISION (offset_type); | |
2108 | ||
2109 | /* Enforced by vect_check_gather_scatter. */ | |
2110 | gcc_assert (element_bits >= offset_bits); | |
2111 | ||
2112 | /* If the elements are wider than the offset, convert the offset to the | |
2113 | same width, without changing its sign. */ | |
2114 | if (element_bits > offset_bits) | |
2115 | { | |
2116 | bool unsigned_p = TYPE_UNSIGNED (offset_type); | |
2117 | offset_type = build_nonstandard_integer_type (element_bits, unsigned_p); | |
2118 | gs_info->offset = fold_convert (offset_type, gs_info->offset); | |
2119 | } | |
2120 | ||
2121 | if (dump_enabled_p ()) | |
2122 | dump_printf_loc (MSG_NOTE, vect_location, | |
2123 | "using gather/scatter for strided/grouped access," | |
2124 | " scale = %d\n", gs_info->scale); | |
2125 | ||
2126 | return true; | |
2127 | } | |
2128 | ||
62da9e14 RS |
2129 | /* STMT is a non-strided load or store, meaning that it accesses |
2130 | elements with a known constant step. Return -1 if that step | |
2131 | is negative, 0 if it is zero, and 1 if it is greater than zero. */ | |
2132 | ||
2133 | static int | |
2134 | compare_step_with_zero (gimple *stmt) | |
2135 | { | |
2136 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3f5e8a76 RS |
2137 | data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
2138 | return tree_int_cst_compare (vect_dr_behavior (dr)->step, | |
2139 | size_zero_node); | |
62da9e14 RS |
2140 | } |
2141 | ||
2142 | /* If the target supports a permute mask that reverses the elements in | |
2143 | a vector of type VECTYPE, return that mask, otherwise return null. */ | |
2144 | ||
2145 | static tree | |
2146 | perm_mask_for_reverse (tree vectype) | |
2147 | { | |
928686b1 | 2148 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
62da9e14 | 2149 | |
d980067b RS |
2150 | /* The encoding has a single stepped pattern. */ |
2151 | vec_perm_builder sel (nunits, 1, 3); | |
928686b1 | 2152 | for (int i = 0; i < 3; ++i) |
908a1a16 | 2153 | sel.quick_push (nunits - 1 - i); |
62da9e14 | 2154 | |
e3342de4 RS |
2155 | vec_perm_indices indices (sel, 1, nunits); |
2156 | if (!can_vec_perm_const_p (TYPE_MODE (vectype), indices)) | |
62da9e14 | 2157 | return NULL_TREE; |
e3342de4 | 2158 | return vect_gen_perm_mask_checked (vectype, indices); |
62da9e14 | 2159 | } |
5ce9450f | 2160 | |
c3a8f964 RS |
2161 | /* STMT is either a masked or unconditional store. Return the value |
2162 | being stored. */ | |
2163 | ||
f307441a | 2164 | tree |
c3a8f964 RS |
2165 | vect_get_store_rhs (gimple *stmt) |
2166 | { | |
2167 | if (gassign *assign = dyn_cast <gassign *> (stmt)) | |
2168 | { | |
2169 | gcc_assert (gimple_assign_single_p (assign)); | |
2170 | return gimple_assign_rhs1 (assign); | |
2171 | } | |
2172 | if (gcall *call = dyn_cast <gcall *> (stmt)) | |
2173 | { | |
2174 | internal_fn ifn = gimple_call_internal_fn (call); | |
f307441a RS |
2175 | int index = internal_fn_stored_value_index (ifn); |
2176 | gcc_assert (index >= 0); | |
2177 | return gimple_call_arg (stmt, index); | |
c3a8f964 RS |
2178 | } |
2179 | gcc_unreachable (); | |
2180 | } | |
2181 | ||
2de001ee RS |
2182 | /* A subroutine of get_load_store_type, with a subset of the same |
2183 | arguments. Handle the case where STMT is part of a grouped load | |
2184 | or store. | |
2185 | ||
2186 | For stores, the statements in the group are all consecutive | |
2187 | and there is no gap at the end. For loads, the statements in the | |
2188 | group might not be consecutive; there can be gaps between statements | |
2189 | as well as at the end. */ | |
2190 | ||
2191 | static bool | |
2192 | get_group_load_store_type (gimple *stmt, tree vectype, bool slp, | |
7e11fc7f | 2193 | bool masked_p, vec_load_store_type vls_type, |
429ef523 RS |
2194 | vect_memory_access_type *memory_access_type, |
2195 | gather_scatter_info *gs_info) | |
2de001ee RS |
2196 | { |
2197 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2198 | vec_info *vinfo = stmt_info->vinfo; | |
2199 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2200 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; | |
2c53b149 | 2201 | gimple *first_stmt = DR_GROUP_FIRST_ELEMENT (stmt_info); |
f702e7d4 | 2202 | data_reference *first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
2c53b149 | 2203 | unsigned int group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
2de001ee | 2204 | bool single_element_p = (stmt == first_stmt |
2c53b149 RB |
2205 | && !DR_GROUP_NEXT_ELEMENT (stmt_info)); |
2206 | unsigned HOST_WIDE_INT gap = DR_GROUP_GAP (vinfo_for_stmt (first_stmt)); | |
928686b1 | 2207 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2de001ee RS |
2208 | |
2209 | /* True if the vectorized statements would access beyond the last | |
2210 | statement in the group. */ | |
2211 | bool overrun_p = false; | |
2212 | ||
2213 | /* True if we can cope with such overrun by peeling for gaps, so that | |
2214 | there is at least one final scalar iteration after the vector loop. */ | |
7e11fc7f RS |
2215 | bool can_overrun_p = (!masked_p |
2216 | && vls_type == VLS_LOAD | |
2217 | && loop_vinfo | |
2218 | && !loop->inner); | |
2de001ee RS |
2219 | |
2220 | /* There can only be a gap at the end of the group if the stride is | |
2221 | known at compile time. */ | |
2222 | gcc_assert (!STMT_VINFO_STRIDED_P (stmt_info) || gap == 0); | |
2223 | ||
2224 | /* Stores can't yet have gaps. */ | |
2225 | gcc_assert (slp || vls_type == VLS_LOAD || gap == 0); | |
2226 | ||
2227 | if (slp) | |
2228 | { | |
2229 | if (STMT_VINFO_STRIDED_P (stmt_info)) | |
2230 | { | |
2c53b149 | 2231 | /* Try to use consecutive accesses of DR_GROUP_SIZE elements, |
2de001ee RS |
2232 | separated by the stride, until we have a complete vector. |
2233 | Fall back to scalar accesses if that isn't possible. */ | |
928686b1 | 2234 | if (multiple_p (nunits, group_size)) |
2de001ee RS |
2235 | *memory_access_type = VMAT_STRIDED_SLP; |
2236 | else | |
2237 | *memory_access_type = VMAT_ELEMENTWISE; | |
2238 | } | |
2239 | else | |
2240 | { | |
2241 | overrun_p = loop_vinfo && gap != 0; | |
2242 | if (overrun_p && vls_type != VLS_LOAD) | |
2243 | { | |
2244 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2245 | "Grouped store with gaps requires" | |
2246 | " non-consecutive accesses\n"); | |
2247 | return false; | |
2248 | } | |
f702e7d4 RS |
2249 | /* An overrun is fine if the trailing elements are smaller |
2250 | than the alignment boundary B. Every vector access will | |
2251 | be a multiple of B and so we are guaranteed to access a | |
2252 | non-gap element in the same B-sized block. */ | |
f9ef2c76 | 2253 | if (overrun_p |
f702e7d4 RS |
2254 | && gap < (vect_known_alignment_in_bytes (first_dr) |
2255 | / vect_get_scalar_dr_size (first_dr))) | |
f9ef2c76 | 2256 | overrun_p = false; |
2de001ee RS |
2257 | if (overrun_p && !can_overrun_p) |
2258 | { | |
2259 | if (dump_enabled_p ()) | |
2260 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2261 | "Peeling for outer loop is not supported\n"); | |
2262 | return false; | |
2263 | } | |
2264 | *memory_access_type = VMAT_CONTIGUOUS; | |
2265 | } | |
2266 | } | |
2267 | else | |
2268 | { | |
2269 | /* We can always handle this case using elementwise accesses, | |
2270 | but see if something more efficient is available. */ | |
2271 | *memory_access_type = VMAT_ELEMENTWISE; | |
2272 | ||
2273 | /* If there is a gap at the end of the group then these optimizations | |
2274 | would access excess elements in the last iteration. */ | |
2275 | bool would_overrun_p = (gap != 0); | |
f702e7d4 RS |
2276 | /* An overrun is fine if the trailing elements are smaller than the |
2277 | alignment boundary B. Every vector access will be a multiple of B | |
2278 | and so we are guaranteed to access a non-gap element in the | |
2279 | same B-sized block. */ | |
f9ef2c76 | 2280 | if (would_overrun_p |
7e11fc7f | 2281 | && !masked_p |
f702e7d4 RS |
2282 | && gap < (vect_known_alignment_in_bytes (first_dr) |
2283 | / vect_get_scalar_dr_size (first_dr))) | |
f9ef2c76 | 2284 | would_overrun_p = false; |
f702e7d4 | 2285 | |
2de001ee | 2286 | if (!STMT_VINFO_STRIDED_P (stmt_info) |
62da9e14 RS |
2287 | && (can_overrun_p || !would_overrun_p) |
2288 | && compare_step_with_zero (stmt) > 0) | |
2de001ee | 2289 | { |
6737facb RS |
2290 | /* First cope with the degenerate case of a single-element |
2291 | vector. */ | |
2292 | if (known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U)) | |
2293 | *memory_access_type = VMAT_CONTIGUOUS; | |
2294 | ||
2295 | /* Otherwise try using LOAD/STORE_LANES. */ | |
2296 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2297 | && (vls_type == VLS_LOAD | |
7e11fc7f RS |
2298 | ? vect_load_lanes_supported (vectype, group_size, masked_p) |
2299 | : vect_store_lanes_supported (vectype, group_size, | |
2300 | masked_p))) | |
2de001ee RS |
2301 | { |
2302 | *memory_access_type = VMAT_LOAD_STORE_LANES; | |
2303 | overrun_p = would_overrun_p; | |
2304 | } | |
2305 | ||
2306 | /* If that fails, try using permuting loads. */ | |
2307 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2308 | && (vls_type == VLS_LOAD | |
2309 | ? vect_grouped_load_supported (vectype, single_element_p, | |
2310 | group_size) | |
2311 | : vect_grouped_store_supported (vectype, group_size))) | |
2312 | { | |
2313 | *memory_access_type = VMAT_CONTIGUOUS_PERMUTE; | |
2314 | overrun_p = would_overrun_p; | |
2315 | } | |
2316 | } | |
429ef523 RS |
2317 | |
2318 | /* As a last resort, trying using a gather load or scatter store. | |
2319 | ||
2320 | ??? Although the code can handle all group sizes correctly, | |
2321 | it probably isn't a win to use separate strided accesses based | |
2322 | on nearby locations. Or, even if it's a win over scalar code, | |
2323 | it might not be a win over vectorizing at a lower VF, if that | |
2324 | allows us to use contiguous accesses. */ | |
2325 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2326 | && single_element_p | |
2327 | && loop_vinfo | |
2328 | && vect_use_strided_gather_scatters_p (stmt, loop_vinfo, | |
2329 | masked_p, gs_info)) | |
2330 | *memory_access_type = VMAT_GATHER_SCATTER; | |
2de001ee RS |
2331 | } |
2332 | ||
2333 | if (vls_type != VLS_LOAD && first_stmt == stmt) | |
2334 | { | |
2335 | /* STMT is the leader of the group. Check the operands of all the | |
2336 | stmts of the group. */ | |
2c53b149 | 2337 | gimple *next_stmt = DR_GROUP_NEXT_ELEMENT (stmt_info); |
2de001ee RS |
2338 | while (next_stmt) |
2339 | { | |
7e11fc7f | 2340 | tree op = vect_get_store_rhs (next_stmt); |
2de001ee RS |
2341 | gimple *def_stmt; |
2342 | enum vect_def_type dt; | |
2343 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt)) | |
2344 | { | |
2345 | if (dump_enabled_p ()) | |
2346 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2347 | "use not simple.\n"); | |
2348 | return false; | |
2349 | } | |
2c53b149 | 2350 | next_stmt = DR_GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
2de001ee RS |
2351 | } |
2352 | } | |
2353 | ||
2354 | if (overrun_p) | |
2355 | { | |
2356 | gcc_assert (can_overrun_p); | |
2357 | if (dump_enabled_p ()) | |
2358 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2359 | "Data access with gaps requires scalar " | |
2360 | "epilogue loop\n"); | |
2361 | LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true; | |
2362 | } | |
2363 | ||
2364 | return true; | |
2365 | } | |
2366 | ||
62da9e14 RS |
2367 | /* A subroutine of get_load_store_type, with a subset of the same |
2368 | arguments. Handle the case where STMT is a load or store that | |
2369 | accesses consecutive elements with a negative step. */ | |
2370 | ||
2371 | static vect_memory_access_type | |
2372 | get_negative_load_store_type (gimple *stmt, tree vectype, | |
2373 | vec_load_store_type vls_type, | |
2374 | unsigned int ncopies) | |
2375 | { | |
2376 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2377 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
2378 | dr_alignment_support alignment_support_scheme; | |
2379 | ||
2380 | if (ncopies > 1) | |
2381 | { | |
2382 | if (dump_enabled_p ()) | |
2383 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2384 | "multiple types with negative step.\n"); | |
2385 | return VMAT_ELEMENTWISE; | |
2386 | } | |
2387 | ||
2388 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
2389 | if (alignment_support_scheme != dr_aligned | |
2390 | && alignment_support_scheme != dr_unaligned_supported) | |
2391 | { | |
2392 | if (dump_enabled_p ()) | |
2393 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2394 | "negative step but alignment required.\n"); | |
2395 | return VMAT_ELEMENTWISE; | |
2396 | } | |
2397 | ||
2398 | if (vls_type == VLS_STORE_INVARIANT) | |
2399 | { | |
2400 | if (dump_enabled_p ()) | |
2401 | dump_printf_loc (MSG_NOTE, vect_location, | |
2402 | "negative step with invariant source;" | |
2403 | " no permute needed.\n"); | |
2404 | return VMAT_CONTIGUOUS_DOWN; | |
2405 | } | |
2406 | ||
2407 | if (!perm_mask_for_reverse (vectype)) | |
2408 | { | |
2409 | if (dump_enabled_p ()) | |
2410 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2411 | "negative step and reversing not supported.\n"); | |
2412 | return VMAT_ELEMENTWISE; | |
2413 | } | |
2414 | ||
2415 | return VMAT_CONTIGUOUS_REVERSE; | |
2416 | } | |
2417 | ||
2de001ee RS |
2418 | /* Analyze load or store statement STMT of type VLS_TYPE. Return true |
2419 | if there is a memory access type that the vectorized form can use, | |
2420 | storing it in *MEMORY_ACCESS_TYPE if so. If we decide to use gathers | |
2421 | or scatters, fill in GS_INFO accordingly. | |
2422 | ||
2423 | SLP says whether we're performing SLP rather than loop vectorization. | |
7e11fc7f | 2424 | MASKED_P is true if the statement is conditional on a vectorized mask. |
62da9e14 RS |
2425 | VECTYPE is the vector type that the vectorized statements will use. |
2426 | NCOPIES is the number of vector statements that will be needed. */ | |
2de001ee RS |
2427 | |
2428 | static bool | |
7e11fc7f | 2429 | get_load_store_type (gimple *stmt, tree vectype, bool slp, bool masked_p, |
62da9e14 | 2430 | vec_load_store_type vls_type, unsigned int ncopies, |
2de001ee RS |
2431 | vect_memory_access_type *memory_access_type, |
2432 | gather_scatter_info *gs_info) | |
2433 | { | |
2434 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2435 | vec_info *vinfo = stmt_info->vinfo; | |
2436 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4d694b27 | 2437 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2de001ee RS |
2438 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
2439 | { | |
2440 | *memory_access_type = VMAT_GATHER_SCATTER; | |
2441 | gimple *def_stmt; | |
2442 | if (!vect_check_gather_scatter (stmt, loop_vinfo, gs_info)) | |
2443 | gcc_unreachable (); | |
2444 | else if (!vect_is_simple_use (gs_info->offset, vinfo, &def_stmt, | |
2445 | &gs_info->offset_dt, | |
2446 | &gs_info->offset_vectype)) | |
2447 | { | |
2448 | if (dump_enabled_p ()) | |
2449 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2450 | "%s index use not simple.\n", | |
2451 | vls_type == VLS_LOAD ? "gather" : "scatter"); | |
2452 | return false; | |
2453 | } | |
2454 | } | |
2455 | else if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
2456 | { | |
7e11fc7f | 2457 | if (!get_group_load_store_type (stmt, vectype, slp, masked_p, vls_type, |
429ef523 | 2458 | memory_access_type, gs_info)) |
2de001ee RS |
2459 | return false; |
2460 | } | |
2461 | else if (STMT_VINFO_STRIDED_P (stmt_info)) | |
2462 | { | |
2463 | gcc_assert (!slp); | |
ab2fc782 | 2464 | if (loop_vinfo |
429ef523 RS |
2465 | && vect_use_strided_gather_scatters_p (stmt, loop_vinfo, |
2466 | masked_p, gs_info)) | |
ab2fc782 RS |
2467 | *memory_access_type = VMAT_GATHER_SCATTER; |
2468 | else | |
2469 | *memory_access_type = VMAT_ELEMENTWISE; | |
2de001ee RS |
2470 | } |
2471 | else | |
62da9e14 RS |
2472 | { |
2473 | int cmp = compare_step_with_zero (stmt); | |
2474 | if (cmp < 0) | |
2475 | *memory_access_type = get_negative_load_store_type | |
2476 | (stmt, vectype, vls_type, ncopies); | |
2477 | else if (cmp == 0) | |
2478 | { | |
2479 | gcc_assert (vls_type == VLS_LOAD); | |
2480 | *memory_access_type = VMAT_INVARIANT; | |
2481 | } | |
2482 | else | |
2483 | *memory_access_type = VMAT_CONTIGUOUS; | |
2484 | } | |
2de001ee | 2485 | |
4d694b27 RS |
2486 | if ((*memory_access_type == VMAT_ELEMENTWISE |
2487 | || *memory_access_type == VMAT_STRIDED_SLP) | |
2488 | && !nunits.is_constant ()) | |
2489 | { | |
2490 | if (dump_enabled_p ()) | |
2491 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2492 | "Not using elementwise accesses due to variable " | |
2493 | "vectorization factor.\n"); | |
2494 | return false; | |
2495 | } | |
2496 | ||
2de001ee RS |
2497 | /* FIXME: At the moment the cost model seems to underestimate the |
2498 | cost of using elementwise accesses. This check preserves the | |
2499 | traditional behavior until that can be fixed. */ | |
2500 | if (*memory_access_type == VMAT_ELEMENTWISE | |
4aa157e8 | 2501 | && !STMT_VINFO_STRIDED_P (stmt_info) |
2c53b149 RB |
2502 | && !(stmt == DR_GROUP_FIRST_ELEMENT (stmt_info) |
2503 | && !DR_GROUP_NEXT_ELEMENT (stmt_info) | |
2504 | && !pow2p_hwi (DR_GROUP_SIZE (stmt_info)))) | |
2de001ee RS |
2505 | { |
2506 | if (dump_enabled_p ()) | |
2507 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2508 | "not falling back to elementwise accesses\n"); | |
2509 | return false; | |
2510 | } | |
2511 | return true; | |
2512 | } | |
2513 | ||
aaeefd88 | 2514 | /* Return true if boolean argument MASK is suitable for vectorizing |
929b4411 RS |
2515 | conditional load or store STMT. When returning true, store the type |
2516 | of the definition in *MASK_DT_OUT and the type of the vectorized mask | |
2517 | in *MASK_VECTYPE_OUT. */ | |
aaeefd88 RS |
2518 | |
2519 | static bool | |
929b4411 RS |
2520 | vect_check_load_store_mask (gimple *stmt, tree mask, |
2521 | vect_def_type *mask_dt_out, | |
2522 | tree *mask_vectype_out) | |
aaeefd88 RS |
2523 | { |
2524 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (mask))) | |
2525 | { | |
2526 | if (dump_enabled_p ()) | |
2527 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2528 | "mask argument is not a boolean.\n"); | |
2529 | return false; | |
2530 | } | |
2531 | ||
2532 | if (TREE_CODE (mask) != SSA_NAME) | |
2533 | { | |
2534 | if (dump_enabled_p ()) | |
2535 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2536 | "mask argument is not an SSA name.\n"); | |
2537 | return false; | |
2538 | } | |
2539 | ||
2540 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2541 | gimple *def_stmt; | |
929b4411 | 2542 | enum vect_def_type mask_dt; |
aaeefd88 | 2543 | tree mask_vectype; |
929b4411 | 2544 | if (!vect_is_simple_use (mask, stmt_info->vinfo, &def_stmt, &mask_dt, |
aaeefd88 RS |
2545 | &mask_vectype)) |
2546 | { | |
2547 | if (dump_enabled_p ()) | |
2548 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2549 | "mask use not simple.\n"); | |
2550 | return false; | |
2551 | } | |
2552 | ||
2553 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2554 | if (!mask_vectype) | |
2555 | mask_vectype = get_mask_type_for_scalar_type (TREE_TYPE (vectype)); | |
2556 | ||
2557 | if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype)) | |
2558 | { | |
2559 | if (dump_enabled_p ()) | |
2560 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2561 | "could not find an appropriate vector mask type.\n"); | |
2562 | return false; | |
2563 | } | |
2564 | ||
2565 | if (maybe_ne (TYPE_VECTOR_SUBPARTS (mask_vectype), | |
2566 | TYPE_VECTOR_SUBPARTS (vectype))) | |
2567 | { | |
2568 | if (dump_enabled_p ()) | |
2569 | { | |
2570 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2571 | "vector mask type "); | |
2572 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, mask_vectype); | |
2573 | dump_printf (MSG_MISSED_OPTIMIZATION, | |
2574 | " does not match vector data type "); | |
2575 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, vectype); | |
2576 | dump_printf (MSG_MISSED_OPTIMIZATION, ".\n"); | |
2577 | } | |
2578 | return false; | |
2579 | } | |
2580 | ||
929b4411 | 2581 | *mask_dt_out = mask_dt; |
aaeefd88 RS |
2582 | *mask_vectype_out = mask_vectype; |
2583 | return true; | |
2584 | } | |
2585 | ||
3133c3b6 RS |
2586 | /* Return true if stored value RHS is suitable for vectorizing store |
2587 | statement STMT. When returning true, store the type of the | |
929b4411 RS |
2588 | definition in *RHS_DT_OUT, the type of the vectorized store value in |
2589 | *RHS_VECTYPE_OUT and the type of the store in *VLS_TYPE_OUT. */ | |
3133c3b6 RS |
2590 | |
2591 | static bool | |
929b4411 RS |
2592 | vect_check_store_rhs (gimple *stmt, tree rhs, vect_def_type *rhs_dt_out, |
2593 | tree *rhs_vectype_out, vec_load_store_type *vls_type_out) | |
3133c3b6 RS |
2594 | { |
2595 | /* In the case this is a store from a constant make sure | |
2596 | native_encode_expr can handle it. */ | |
2597 | if (CONSTANT_CLASS_P (rhs) && native_encode_expr (rhs, NULL, 64) == 0) | |
2598 | { | |
2599 | if (dump_enabled_p ()) | |
2600 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2601 | "cannot encode constant as a byte sequence.\n"); | |
2602 | return false; | |
2603 | } | |
2604 | ||
2605 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2606 | gimple *def_stmt; | |
929b4411 | 2607 | enum vect_def_type rhs_dt; |
3133c3b6 | 2608 | tree rhs_vectype; |
929b4411 | 2609 | if (!vect_is_simple_use (rhs, stmt_info->vinfo, &def_stmt, &rhs_dt, |
3133c3b6 RS |
2610 | &rhs_vectype)) |
2611 | { | |
2612 | if (dump_enabled_p ()) | |
2613 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2614 | "use not simple.\n"); | |
2615 | return false; | |
2616 | } | |
2617 | ||
2618 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2619 | if (rhs_vectype && !useless_type_conversion_p (vectype, rhs_vectype)) | |
2620 | { | |
2621 | if (dump_enabled_p ()) | |
2622 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2623 | "incompatible vector types.\n"); | |
2624 | return false; | |
2625 | } | |
2626 | ||
929b4411 | 2627 | *rhs_dt_out = rhs_dt; |
3133c3b6 | 2628 | *rhs_vectype_out = rhs_vectype; |
929b4411 | 2629 | if (rhs_dt == vect_constant_def || rhs_dt == vect_external_def) |
3133c3b6 RS |
2630 | *vls_type_out = VLS_STORE_INVARIANT; |
2631 | else | |
2632 | *vls_type_out = VLS_STORE; | |
2633 | return true; | |
2634 | } | |
2635 | ||
bc9587eb RS |
2636 | /* Build an all-ones vector mask of type MASKTYPE while vectorizing STMT. |
2637 | Note that we support masks with floating-point type, in which case the | |
2638 | floats are interpreted as a bitmask. */ | |
2639 | ||
2640 | static tree | |
2641 | vect_build_all_ones_mask (gimple *stmt, tree masktype) | |
2642 | { | |
2643 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
2644 | return build_int_cst (masktype, -1); | |
2645 | else if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
2646 | { | |
2647 | tree mask = build_int_cst (TREE_TYPE (masktype), -1); | |
2648 | mask = build_vector_from_val (masktype, mask); | |
2649 | return vect_init_vector (stmt, mask, masktype, NULL); | |
2650 | } | |
2651 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) | |
2652 | { | |
2653 | REAL_VALUE_TYPE r; | |
2654 | long tmp[6]; | |
2655 | for (int j = 0; j < 6; ++j) | |
2656 | tmp[j] = -1; | |
2657 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
2658 | tree mask = build_real (TREE_TYPE (masktype), r); | |
2659 | mask = build_vector_from_val (masktype, mask); | |
2660 | return vect_init_vector (stmt, mask, masktype, NULL); | |
2661 | } | |
2662 | gcc_unreachable (); | |
2663 | } | |
2664 | ||
2665 | /* Build an all-zero merge value of type VECTYPE while vectorizing | |
2666 | STMT as a gather load. */ | |
2667 | ||
2668 | static tree | |
2669 | vect_build_zero_merge_argument (gimple *stmt, tree vectype) | |
2670 | { | |
2671 | tree merge; | |
2672 | if (TREE_CODE (TREE_TYPE (vectype)) == INTEGER_TYPE) | |
2673 | merge = build_int_cst (TREE_TYPE (vectype), 0); | |
2674 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (vectype))) | |
2675 | { | |
2676 | REAL_VALUE_TYPE r; | |
2677 | long tmp[6]; | |
2678 | for (int j = 0; j < 6; ++j) | |
2679 | tmp[j] = 0; | |
2680 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (vectype))); | |
2681 | merge = build_real (TREE_TYPE (vectype), r); | |
2682 | } | |
2683 | else | |
2684 | gcc_unreachable (); | |
2685 | merge = build_vector_from_val (vectype, merge); | |
2686 | return vect_init_vector (stmt, merge, vectype, NULL); | |
2687 | } | |
2688 | ||
c48d2d35 RS |
2689 | /* Build a gather load call while vectorizing STMT. Insert new instructions |
2690 | before GSI and add them to VEC_STMT. GS_INFO describes the gather load | |
2691 | operation. If the load is conditional, MASK is the unvectorized | |
929b4411 | 2692 | condition and MASK_DT is its definition type, otherwise MASK is null. */ |
c48d2d35 RS |
2693 | |
2694 | static void | |
2695 | vect_build_gather_load_calls (gimple *stmt, gimple_stmt_iterator *gsi, | |
2696 | gimple **vec_stmt, gather_scatter_info *gs_info, | |
929b4411 | 2697 | tree mask, vect_def_type mask_dt) |
c48d2d35 RS |
2698 | { |
2699 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2700 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2701 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2702 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2703 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
2704 | int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
2705 | edge pe = loop_preheader_edge (loop); | |
2706 | enum { NARROW, NONE, WIDEN } modifier; | |
2707 | poly_uint64 gather_off_nunits | |
2708 | = TYPE_VECTOR_SUBPARTS (gs_info->offset_vectype); | |
2709 | ||
2710 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info->decl)); | |
2711 | tree rettype = TREE_TYPE (TREE_TYPE (gs_info->decl)); | |
2712 | tree srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2713 | tree ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2714 | tree idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2715 | tree masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2716 | tree scaletype = TREE_VALUE (arglist); | |
2717 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
2718 | && (!mask || types_compatible_p (srctype, masktype))); | |
2719 | ||
2720 | tree perm_mask = NULL_TREE; | |
2721 | tree mask_perm_mask = NULL_TREE; | |
2722 | if (known_eq (nunits, gather_off_nunits)) | |
2723 | modifier = NONE; | |
2724 | else if (known_eq (nunits * 2, gather_off_nunits)) | |
2725 | { | |
2726 | modifier = WIDEN; | |
2727 | ||
2728 | /* Currently widening gathers and scatters are only supported for | |
2729 | fixed-length vectors. */ | |
2730 | int count = gather_off_nunits.to_constant (); | |
2731 | vec_perm_builder sel (count, count, 1); | |
2732 | for (int i = 0; i < count; ++i) | |
2733 | sel.quick_push (i | (count / 2)); | |
2734 | ||
2735 | vec_perm_indices indices (sel, 1, count); | |
2736 | perm_mask = vect_gen_perm_mask_checked (gs_info->offset_vectype, | |
2737 | indices); | |
2738 | } | |
2739 | else if (known_eq (nunits, gather_off_nunits * 2)) | |
2740 | { | |
2741 | modifier = NARROW; | |
2742 | ||
2743 | /* Currently narrowing gathers and scatters are only supported for | |
2744 | fixed-length vectors. */ | |
2745 | int count = nunits.to_constant (); | |
2746 | vec_perm_builder sel (count, count, 1); | |
2747 | sel.quick_grow (count); | |
2748 | for (int i = 0; i < count; ++i) | |
2749 | sel[i] = i < count / 2 ? i : i + count / 2; | |
2750 | vec_perm_indices indices (sel, 2, count); | |
2751 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); | |
2752 | ||
2753 | ncopies *= 2; | |
2754 | ||
2755 | if (mask) | |
2756 | { | |
2757 | for (int i = 0; i < count; ++i) | |
2758 | sel[i] = i | (count / 2); | |
2759 | indices.new_vector (sel, 2, count); | |
2760 | mask_perm_mask = vect_gen_perm_mask_checked (masktype, indices); | |
2761 | } | |
2762 | } | |
2763 | else | |
2764 | gcc_unreachable (); | |
2765 | ||
2766 | tree vec_dest = vect_create_destination_var (gimple_get_lhs (stmt), | |
2767 | vectype); | |
2768 | ||
2769 | tree ptr = fold_convert (ptrtype, gs_info->base); | |
2770 | if (!is_gimple_min_invariant (ptr)) | |
2771 | { | |
2772 | gimple_seq seq; | |
2773 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
2774 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
2775 | gcc_assert (!new_bb); | |
2776 | } | |
2777 | ||
2778 | tree scale = build_int_cst (scaletype, gs_info->scale); | |
2779 | ||
2780 | tree vec_oprnd0 = NULL_TREE; | |
2781 | tree vec_mask = NULL_TREE; | |
2782 | tree src_op = NULL_TREE; | |
2783 | tree mask_op = NULL_TREE; | |
2784 | tree prev_res = NULL_TREE; | |
2785 | stmt_vec_info prev_stmt_info = NULL; | |
2786 | ||
2787 | if (!mask) | |
2788 | { | |
2789 | src_op = vect_build_zero_merge_argument (stmt, rettype); | |
2790 | mask_op = vect_build_all_ones_mask (stmt, masktype); | |
2791 | } | |
2792 | ||
2793 | for (int j = 0; j < ncopies; ++j) | |
2794 | { | |
2795 | tree op, var; | |
2796 | gimple *new_stmt; | |
2797 | if (modifier == WIDEN && (j & 1)) | |
2798 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
2799 | perm_mask, stmt, gsi); | |
2800 | else if (j == 0) | |
2801 | op = vec_oprnd0 | |
2802 | = vect_get_vec_def_for_operand (gs_info->offset, stmt); | |
2803 | else | |
2804 | op = vec_oprnd0 | |
2805 | = vect_get_vec_def_for_stmt_copy (gs_info->offset_dt, vec_oprnd0); | |
2806 | ||
2807 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
2808 | { | |
2809 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)), | |
2810 | TYPE_VECTOR_SUBPARTS (idxtype))); | |
2811 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); | |
2812 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); | |
2813 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
2814 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2815 | op = var; | |
2816 | } | |
2817 | ||
2818 | if (mask) | |
2819 | { | |
2820 | if (mask_perm_mask && (j & 1)) | |
2821 | mask_op = permute_vec_elements (mask_op, mask_op, | |
2822 | mask_perm_mask, stmt, gsi); | |
2823 | else | |
2824 | { | |
2825 | if (j == 0) | |
2826 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); | |
2827 | else | |
929b4411 | 2828 | vec_mask = vect_get_vec_def_for_stmt_copy (mask_dt, vec_mask); |
c48d2d35 RS |
2829 | |
2830 | mask_op = vec_mask; | |
2831 | if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask))) | |
2832 | { | |
2833 | gcc_assert | |
2834 | (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)), | |
2835 | TYPE_VECTOR_SUBPARTS (masktype))); | |
2836 | var = vect_get_new_ssa_name (masktype, vect_simple_var); | |
2837 | mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op); | |
2838 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, | |
2839 | mask_op); | |
2840 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2841 | mask_op = var; | |
2842 | } | |
2843 | } | |
2844 | src_op = mask_op; | |
2845 | } | |
2846 | ||
2847 | new_stmt = gimple_build_call (gs_info->decl, 5, src_op, ptr, op, | |
2848 | mask_op, scale); | |
2849 | ||
2850 | if (!useless_type_conversion_p (vectype, rettype)) | |
2851 | { | |
2852 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (vectype), | |
2853 | TYPE_VECTOR_SUBPARTS (rettype))); | |
2854 | op = vect_get_new_ssa_name (rettype, vect_simple_var); | |
2855 | gimple_call_set_lhs (new_stmt, op); | |
2856 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2857 | var = make_ssa_name (vec_dest); | |
2858 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); | |
2859 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
2860 | } | |
2861 | else | |
2862 | { | |
2863 | var = make_ssa_name (vec_dest, new_stmt); | |
2864 | gimple_call_set_lhs (new_stmt, var); | |
2865 | } | |
2866 | ||
2867 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2868 | ||
2869 | if (modifier == NARROW) | |
2870 | { | |
2871 | if ((j & 1) == 0) | |
2872 | { | |
2873 | prev_res = var; | |
2874 | continue; | |
2875 | } | |
2876 | var = permute_vec_elements (prev_res, var, perm_mask, stmt, gsi); | |
2877 | new_stmt = SSA_NAME_DEF_STMT (var); | |
2878 | } | |
2879 | ||
2880 | if (prev_stmt_info == NULL) | |
2881 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2882 | else | |
2883 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2884 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2885 | } | |
2886 | } | |
2887 | ||
bfaa08b7 RS |
2888 | /* Prepare the base and offset in GS_INFO for vectorization. |
2889 | Set *DATAREF_PTR to the loop-invariant base address and *VEC_OFFSET | |
2890 | to the vectorized offset argument for the first copy of STMT. STMT | |
2891 | is the statement described by GS_INFO and LOOP is the containing loop. */ | |
2892 | ||
2893 | static void | |
2894 | vect_get_gather_scatter_ops (struct loop *loop, gimple *stmt, | |
2895 | gather_scatter_info *gs_info, | |
2896 | tree *dataref_ptr, tree *vec_offset) | |
2897 | { | |
2898 | gimple_seq stmts = NULL; | |
2899 | *dataref_ptr = force_gimple_operand (gs_info->base, &stmts, true, NULL_TREE); | |
2900 | if (stmts != NULL) | |
2901 | { | |
2902 | basic_block new_bb; | |
2903 | edge pe = loop_preheader_edge (loop); | |
2904 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
2905 | gcc_assert (!new_bb); | |
2906 | } | |
2907 | tree offset_type = TREE_TYPE (gs_info->offset); | |
2908 | tree offset_vectype = get_vectype_for_scalar_type (offset_type); | |
2909 | *vec_offset = vect_get_vec_def_for_operand (gs_info->offset, stmt, | |
2910 | offset_vectype); | |
2911 | } | |
2912 | ||
ab2fc782 RS |
2913 | /* Prepare to implement a grouped or strided load or store using |
2914 | the gather load or scatter store operation described by GS_INFO. | |
2915 | STMT is the load or store statement. | |
2916 | ||
2917 | Set *DATAREF_BUMP to the amount that should be added to the base | |
2918 | address after each copy of the vectorized statement. Set *VEC_OFFSET | |
2919 | to an invariant offset vector in which element I has the value | |
2920 | I * DR_STEP / SCALE. */ | |
2921 | ||
2922 | static void | |
2923 | vect_get_strided_load_store_ops (gimple *stmt, loop_vec_info loop_vinfo, | |
2924 | gather_scatter_info *gs_info, | |
2925 | tree *dataref_bump, tree *vec_offset) | |
2926 | { | |
2927 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2928 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
2929 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2930 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2931 | gimple_seq stmts; | |
2932 | ||
2933 | tree bump = size_binop (MULT_EXPR, | |
2934 | fold_convert (sizetype, DR_STEP (dr)), | |
2935 | size_int (TYPE_VECTOR_SUBPARTS (vectype))); | |
2936 | *dataref_bump = force_gimple_operand (bump, &stmts, true, NULL_TREE); | |
2937 | if (stmts) | |
2938 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
2939 | ||
2940 | /* The offset given in GS_INFO can have pointer type, so use the element | |
2941 | type of the vector instead. */ | |
2942 | tree offset_type = TREE_TYPE (gs_info->offset); | |
2943 | tree offset_vectype = get_vectype_for_scalar_type (offset_type); | |
2944 | offset_type = TREE_TYPE (offset_vectype); | |
2945 | ||
2946 | /* Calculate X = DR_STEP / SCALE and convert it to the appropriate type. */ | |
2947 | tree step = size_binop (EXACT_DIV_EXPR, DR_STEP (dr), | |
2948 | ssize_int (gs_info->scale)); | |
2949 | step = fold_convert (offset_type, step); | |
2950 | step = force_gimple_operand (step, &stmts, true, NULL_TREE); | |
2951 | ||
2952 | /* Create {0, X, X*2, X*3, ...}. */ | |
2953 | *vec_offset = gimple_build (&stmts, VEC_SERIES_EXPR, offset_vectype, | |
2954 | build_zero_cst (offset_type), step); | |
2955 | if (stmts) | |
2956 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
2957 | } | |
2958 | ||
2959 | /* Return the amount that should be added to a vector pointer to move | |
2960 | to the next or previous copy of AGGR_TYPE. DR is the data reference | |
2961 | being vectorized and MEMORY_ACCESS_TYPE describes the type of | |
2962 | vectorization. */ | |
2963 | ||
2964 | static tree | |
2965 | vect_get_data_ptr_increment (data_reference *dr, tree aggr_type, | |
2966 | vect_memory_access_type memory_access_type) | |
2967 | { | |
2968 | if (memory_access_type == VMAT_INVARIANT) | |
2969 | return size_zero_node; | |
2970 | ||
2971 | tree iv_step = TYPE_SIZE_UNIT (aggr_type); | |
2972 | tree step = vect_dr_behavior (dr)->step; | |
2973 | if (tree_int_cst_sgn (step) == -1) | |
2974 | iv_step = fold_build1 (NEGATE_EXPR, TREE_TYPE (iv_step), iv_step); | |
2975 | return iv_step; | |
2976 | } | |
2977 | ||
37b14185 RB |
2978 | /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64}. */ |
2979 | ||
2980 | static bool | |
2981 | vectorizable_bswap (gimple *stmt, gimple_stmt_iterator *gsi, | |
2982 | gimple **vec_stmt, slp_tree slp_node, | |
68435eb2 RB |
2983 | tree vectype_in, enum vect_def_type *dt, |
2984 | stmt_vector_for_cost *cost_vec) | |
37b14185 RB |
2985 | { |
2986 | tree op, vectype; | |
2987 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2988 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
928686b1 RS |
2989 | unsigned ncopies; |
2990 | unsigned HOST_WIDE_INT nunits, num_bytes; | |
37b14185 RB |
2991 | |
2992 | op = gimple_call_arg (stmt, 0); | |
2993 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
928686b1 RS |
2994 | |
2995 | if (!TYPE_VECTOR_SUBPARTS (vectype).is_constant (&nunits)) | |
2996 | return false; | |
37b14185 RB |
2997 | |
2998 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2999 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
3000 | case of SLP. */ | |
3001 | if (slp_node) | |
3002 | ncopies = 1; | |
3003 | else | |
e8f142e2 | 3004 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
37b14185 RB |
3005 | |
3006 | gcc_assert (ncopies >= 1); | |
3007 | ||
3008 | tree char_vectype = get_same_sized_vectype (char_type_node, vectype_in); | |
3009 | if (! char_vectype) | |
3010 | return false; | |
3011 | ||
928686b1 RS |
3012 | if (!TYPE_VECTOR_SUBPARTS (char_vectype).is_constant (&num_bytes)) |
3013 | return false; | |
3014 | ||
794e3180 | 3015 | unsigned word_bytes = num_bytes / nunits; |
908a1a16 | 3016 | |
d980067b RS |
3017 | /* The encoding uses one stepped pattern for each byte in the word. */ |
3018 | vec_perm_builder elts (num_bytes, word_bytes, 3); | |
3019 | for (unsigned i = 0; i < 3; ++i) | |
37b14185 | 3020 | for (unsigned j = 0; j < word_bytes; ++j) |
908a1a16 | 3021 | elts.quick_push ((i + 1) * word_bytes - j - 1); |
37b14185 | 3022 | |
e3342de4 RS |
3023 | vec_perm_indices indices (elts, 1, num_bytes); |
3024 | if (!can_vec_perm_const_p (TYPE_MODE (char_vectype), indices)) | |
37b14185 RB |
3025 | return false; |
3026 | ||
3027 | if (! vec_stmt) | |
3028 | { | |
3029 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
3030 | if (dump_enabled_p ()) | |
3031 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_bswap ===" | |
3032 | "\n"); | |
78604de0 | 3033 | if (! slp_node) |
37b14185 | 3034 | { |
68435eb2 RB |
3035 | record_stmt_cost (cost_vec, |
3036 | 1, vector_stmt, stmt_info, 0, vect_prologue); | |
3037 | record_stmt_cost (cost_vec, | |
3038 | ncopies, vec_perm, stmt_info, 0, vect_body); | |
37b14185 RB |
3039 | } |
3040 | return true; | |
3041 | } | |
3042 | ||
736d0f28 | 3043 | tree bswap_vconst = vec_perm_indices_to_tree (char_vectype, indices); |
37b14185 RB |
3044 | |
3045 | /* Transform. */ | |
3046 | vec<tree> vec_oprnds = vNULL; | |
3047 | gimple *new_stmt = NULL; | |
3048 | stmt_vec_info prev_stmt_info = NULL; | |
3049 | for (unsigned j = 0; j < ncopies; j++) | |
3050 | { | |
3051 | /* Handle uses. */ | |
3052 | if (j == 0) | |
306b0c92 | 3053 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); |
37b14185 RB |
3054 | else |
3055 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
3056 | ||
3057 | /* Arguments are ready. create the new vector stmt. */ | |
3058 | unsigned i; | |
3059 | tree vop; | |
3060 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) | |
3061 | { | |
3062 | tree tem = make_ssa_name (char_vectype); | |
3063 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
3064 | char_vectype, vop)); | |
3065 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3066 | tree tem2 = make_ssa_name (char_vectype); | |
3067 | new_stmt = gimple_build_assign (tem2, VEC_PERM_EXPR, | |
3068 | tem, tem, bswap_vconst); | |
3069 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3070 | tem = make_ssa_name (vectype); | |
3071 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
3072 | vectype, tem2)); | |
3073 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3074 | if (slp_node) | |
3075 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
3076 | } | |
3077 | ||
3078 | if (slp_node) | |
3079 | continue; | |
3080 | ||
3081 | if (j == 0) | |
3082 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
3083 | else | |
3084 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3085 | ||
3086 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3087 | } | |
3088 | ||
3089 | vec_oprnds.release (); | |
3090 | return true; | |
3091 | } | |
3092 | ||
b1b6836e RS |
3093 | /* Return true if vector types VECTYPE_IN and VECTYPE_OUT have |
3094 | integer elements and if we can narrow VECTYPE_IN to VECTYPE_OUT | |
3095 | in a single step. On success, store the binary pack code in | |
3096 | *CONVERT_CODE. */ | |
3097 | ||
3098 | static bool | |
3099 | simple_integer_narrowing (tree vectype_out, tree vectype_in, | |
3100 | tree_code *convert_code) | |
3101 | { | |
3102 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_out)) | |
3103 | || !INTEGRAL_TYPE_P (TREE_TYPE (vectype_in))) | |
3104 | return false; | |
3105 | ||
3106 | tree_code code; | |
3107 | int multi_step_cvt = 0; | |
3108 | auto_vec <tree, 8> interm_types; | |
3109 | if (!supportable_narrowing_operation (NOP_EXPR, vectype_out, vectype_in, | |
3110 | &code, &multi_step_cvt, | |
3111 | &interm_types) | |
3112 | || multi_step_cvt) | |
3113 | return false; | |
3114 | ||
3115 | *convert_code = code; | |
3116 | return true; | |
3117 | } | |
5ce9450f | 3118 | |
ebfd146a IR |
3119 | /* Function vectorizable_call. |
3120 | ||
538dd0b7 | 3121 | Check if GS performs a function call that can be vectorized. |
b8698a0f | 3122 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
3123 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
3124 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3125 | ||
3126 | static bool | |
355fe088 | 3127 | vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
68435eb2 | 3128 | slp_tree slp_node, stmt_vector_for_cost *cost_vec) |
ebfd146a | 3129 | { |
538dd0b7 | 3130 | gcall *stmt; |
ebfd146a IR |
3131 | tree vec_dest; |
3132 | tree scalar_dest; | |
3133 | tree op, type; | |
3134 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
538dd0b7 | 3135 | stmt_vec_info stmt_info = vinfo_for_stmt (gs), prev_stmt_info; |
ebfd146a | 3136 | tree vectype_out, vectype_in; |
c7bda0f4 RS |
3137 | poly_uint64 nunits_in; |
3138 | poly_uint64 nunits_out; | |
ebfd146a | 3139 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
190c2236 | 3140 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 3141 | vec_info *vinfo = stmt_info->vinfo; |
81c40241 | 3142 | tree fndecl, new_temp, rhs_type; |
355fe088 | 3143 | gimple *def_stmt; |
0502fb85 UB |
3144 | enum vect_def_type dt[3] |
3145 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 3146 | int ndts = 3; |
355fe088 | 3147 | gimple *new_stmt = NULL; |
ebfd146a | 3148 | int ncopies, j; |
6e1aa848 | 3149 | vec<tree> vargs = vNULL; |
ebfd146a IR |
3150 | enum { NARROW, NONE, WIDEN } modifier; |
3151 | size_t i, nargs; | |
9d5e7640 | 3152 | tree lhs; |
ebfd146a | 3153 | |
190c2236 | 3154 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3155 | return false; |
3156 | ||
66c16fd9 RB |
3157 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3158 | && ! vec_stmt) | |
ebfd146a IR |
3159 | return false; |
3160 | ||
538dd0b7 DM |
3161 | /* Is GS a vectorizable call? */ |
3162 | stmt = dyn_cast <gcall *> (gs); | |
3163 | if (!stmt) | |
ebfd146a IR |
3164 | return false; |
3165 | ||
5ce9450f | 3166 | if (gimple_call_internal_p (stmt) |
bfaa08b7 | 3167 | && (internal_load_fn_p (gimple_call_internal_fn (stmt)) |
f307441a | 3168 | || internal_store_fn_p (gimple_call_internal_fn (stmt)))) |
c3a8f964 RS |
3169 | /* Handled by vectorizable_load and vectorizable_store. */ |
3170 | return false; | |
5ce9450f | 3171 | |
0136f8f0 AH |
3172 | if (gimple_call_lhs (stmt) == NULL_TREE |
3173 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
ebfd146a IR |
3174 | return false; |
3175 | ||
0136f8f0 | 3176 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); |
5a2c1986 | 3177 | |
b690cc0f RG |
3178 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
3179 | ||
ebfd146a IR |
3180 | /* Process function arguments. */ |
3181 | rhs_type = NULL_TREE; | |
b690cc0f | 3182 | vectype_in = NULL_TREE; |
ebfd146a IR |
3183 | nargs = gimple_call_num_args (stmt); |
3184 | ||
1b1562a5 MM |
3185 | /* Bail out if the function has more than three arguments, we do not have |
3186 | interesting builtin functions to vectorize with more than two arguments | |
3187 | except for fma. No arguments is also not good. */ | |
3188 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
3189 | return false; |
3190 | ||
74bf76ed JJ |
3191 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
3192 | if (gimple_call_internal_p (stmt) | |
3193 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
3194 | { | |
3195 | nargs = 0; | |
3196 | rhs_type = unsigned_type_node; | |
3197 | } | |
3198 | ||
ebfd146a IR |
3199 | for (i = 0; i < nargs; i++) |
3200 | { | |
b690cc0f RG |
3201 | tree opvectype; |
3202 | ||
ebfd146a IR |
3203 | op = gimple_call_arg (stmt, i); |
3204 | ||
3205 | /* We can only handle calls with arguments of the same type. */ | |
3206 | if (rhs_type | |
8533c9d8 | 3207 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 3208 | { |
73fbfcad | 3209 | if (dump_enabled_p ()) |
78c60e3d | 3210 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3211 | "argument types differ.\n"); |
ebfd146a IR |
3212 | return false; |
3213 | } | |
b690cc0f RG |
3214 | if (!rhs_type) |
3215 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 3216 | |
81c40241 | 3217 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[i], &opvectype)) |
ebfd146a | 3218 | { |
73fbfcad | 3219 | if (dump_enabled_p ()) |
78c60e3d | 3220 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3221 | "use not simple.\n"); |
ebfd146a IR |
3222 | return false; |
3223 | } | |
ebfd146a | 3224 | |
b690cc0f RG |
3225 | if (!vectype_in) |
3226 | vectype_in = opvectype; | |
3227 | else if (opvectype | |
3228 | && opvectype != vectype_in) | |
3229 | { | |
73fbfcad | 3230 | if (dump_enabled_p ()) |
78c60e3d | 3231 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3232 | "argument vector types differ.\n"); |
b690cc0f RG |
3233 | return false; |
3234 | } | |
3235 | } | |
3236 | /* If all arguments are external or constant defs use a vector type with | |
3237 | the same size as the output vector type. */ | |
ebfd146a | 3238 | if (!vectype_in) |
b690cc0f | 3239 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
3240 | if (vec_stmt) |
3241 | gcc_assert (vectype_in); | |
3242 | if (!vectype_in) | |
3243 | { | |
73fbfcad | 3244 | if (dump_enabled_p ()) |
7d8930a0 | 3245 | { |
78c60e3d SS |
3246 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3247 | "no vectype for scalar type "); | |
3248 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 3249 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
3250 | } |
3251 | ||
3252 | return false; | |
3253 | } | |
ebfd146a IR |
3254 | |
3255 | /* FORNOW */ | |
b690cc0f RG |
3256 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
3257 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
c7bda0f4 | 3258 | if (known_eq (nunits_in * 2, nunits_out)) |
ebfd146a | 3259 | modifier = NARROW; |
c7bda0f4 | 3260 | else if (known_eq (nunits_out, nunits_in)) |
ebfd146a | 3261 | modifier = NONE; |
c7bda0f4 | 3262 | else if (known_eq (nunits_out * 2, nunits_in)) |
ebfd146a IR |
3263 | modifier = WIDEN; |
3264 | else | |
3265 | return false; | |
3266 | ||
70439f0d RS |
3267 | /* We only handle functions that do not read or clobber memory. */ |
3268 | if (gimple_vuse (stmt)) | |
3269 | { | |
3270 | if (dump_enabled_p ()) | |
3271 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3272 | "function reads from or writes to memory.\n"); | |
3273 | return false; | |
3274 | } | |
3275 | ||
ebfd146a IR |
3276 | /* For now, we only vectorize functions if a target specific builtin |
3277 | is available. TODO -- in some cases, it might be profitable to | |
3278 | insert the calls for pieces of the vector, in order to be able | |
3279 | to vectorize other operations in the loop. */ | |
70439f0d RS |
3280 | fndecl = NULL_TREE; |
3281 | internal_fn ifn = IFN_LAST; | |
3282 | combined_fn cfn = gimple_call_combined_fn (stmt); | |
3283 | tree callee = gimple_call_fndecl (stmt); | |
3284 | ||
3285 | /* First try using an internal function. */ | |
b1b6836e RS |
3286 | tree_code convert_code = ERROR_MARK; |
3287 | if (cfn != CFN_LAST | |
3288 | && (modifier == NONE | |
3289 | || (modifier == NARROW | |
3290 | && simple_integer_narrowing (vectype_out, vectype_in, | |
3291 | &convert_code)))) | |
70439f0d RS |
3292 | ifn = vectorizable_internal_function (cfn, callee, vectype_out, |
3293 | vectype_in); | |
3294 | ||
3295 | /* If that fails, try asking for a target-specific built-in function. */ | |
3296 | if (ifn == IFN_LAST) | |
3297 | { | |
3298 | if (cfn != CFN_LAST) | |
3299 | fndecl = targetm.vectorize.builtin_vectorized_function | |
3300 | (cfn, vectype_out, vectype_in); | |
7672aa9b | 3301 | else if (callee) |
70439f0d RS |
3302 | fndecl = targetm.vectorize.builtin_md_vectorized_function |
3303 | (callee, vectype_out, vectype_in); | |
3304 | } | |
3305 | ||
3306 | if (ifn == IFN_LAST && !fndecl) | |
ebfd146a | 3307 | { |
70439f0d | 3308 | if (cfn == CFN_GOMP_SIMD_LANE |
74bf76ed JJ |
3309 | && !slp_node |
3310 | && loop_vinfo | |
3311 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
3312 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
3313 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
3314 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
3315 | { | |
3316 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
3317 | { 0, 1, 2, ... vf - 1 } vector. */ | |
3318 | gcc_assert (nargs == 0); | |
3319 | } | |
37b14185 RB |
3320 | else if (modifier == NONE |
3321 | && (gimple_call_builtin_p (stmt, BUILT_IN_BSWAP16) | |
3322 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP32) | |
3323 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP64))) | |
3324 | return vectorizable_bswap (stmt, gsi, vec_stmt, slp_node, | |
68435eb2 | 3325 | vectype_in, dt, cost_vec); |
74bf76ed JJ |
3326 | else |
3327 | { | |
3328 | if (dump_enabled_p ()) | |
3329 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 3330 | "function is not vectorizable.\n"); |
74bf76ed JJ |
3331 | return false; |
3332 | } | |
ebfd146a IR |
3333 | } |
3334 | ||
fce57248 | 3335 | if (slp_node) |
190c2236 | 3336 | ncopies = 1; |
b1b6836e | 3337 | else if (modifier == NARROW && ifn == IFN_LAST) |
e8f142e2 | 3338 | ncopies = vect_get_num_copies (loop_vinfo, vectype_out); |
ebfd146a | 3339 | else |
e8f142e2 | 3340 | ncopies = vect_get_num_copies (loop_vinfo, vectype_in); |
ebfd146a IR |
3341 | |
3342 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
3343 | needs to be generated. */ | |
3344 | gcc_assert (ncopies >= 1); | |
3345 | ||
3346 | if (!vec_stmt) /* transformation not required. */ | |
3347 | { | |
3348 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 3349 | if (dump_enabled_p ()) |
e645e942 TJ |
3350 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ===" |
3351 | "\n"); | |
68435eb2 RB |
3352 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, cost_vec); |
3353 | if (ifn != IFN_LAST && modifier == NARROW && !slp_node) | |
3354 | record_stmt_cost (cost_vec, ncopies / 2, | |
3355 | vec_promote_demote, stmt_info, 0, vect_body); | |
b1b6836e | 3356 | |
ebfd146a IR |
3357 | return true; |
3358 | } | |
3359 | ||
67b8dbac | 3360 | /* Transform. */ |
ebfd146a | 3361 | |
73fbfcad | 3362 | if (dump_enabled_p ()) |
e645e942 | 3363 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
3364 | |
3365 | /* Handle def. */ | |
3366 | scalar_dest = gimple_call_lhs (stmt); | |
3367 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
3368 | ||
3369 | prev_stmt_info = NULL; | |
b1b6836e | 3370 | if (modifier == NONE || ifn != IFN_LAST) |
ebfd146a | 3371 | { |
b1b6836e | 3372 | tree prev_res = NULL_TREE; |
ebfd146a IR |
3373 | for (j = 0; j < ncopies; ++j) |
3374 | { | |
3375 | /* Build argument list for the vectorized call. */ | |
3376 | if (j == 0) | |
9771b263 | 3377 | vargs.create (nargs); |
ebfd146a | 3378 | else |
9771b263 | 3379 | vargs.truncate (0); |
ebfd146a | 3380 | |
190c2236 JJ |
3381 | if (slp_node) |
3382 | { | |
ef062b13 | 3383 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 3384 | vec<tree> vec_oprnds0; |
190c2236 JJ |
3385 | |
3386 | for (i = 0; i < nargs; i++) | |
9771b263 | 3387 | vargs.quick_push (gimple_call_arg (stmt, i)); |
306b0c92 | 3388 | vect_get_slp_defs (vargs, slp_node, &vec_defs); |
37b5ec8f | 3389 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
3390 | |
3391 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3392 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
3393 | { |
3394 | size_t k; | |
3395 | for (k = 0; k < nargs; k++) | |
3396 | { | |
37b5ec8f | 3397 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 3398 | vargs[k] = vec_oprndsk[i]; |
190c2236 | 3399 | } |
b1b6836e RS |
3400 | if (modifier == NARROW) |
3401 | { | |
3402 | tree half_res = make_ssa_name (vectype_in); | |
a844293d RS |
3403 | gcall *call |
3404 | = gimple_build_call_internal_vec (ifn, vargs); | |
3405 | gimple_call_set_lhs (call, half_res); | |
3406 | gimple_call_set_nothrow (call, true); | |
3407 | new_stmt = call; | |
b1b6836e RS |
3408 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3409 | if ((i & 1) == 0) | |
3410 | { | |
3411 | prev_res = half_res; | |
3412 | continue; | |
3413 | } | |
3414 | new_temp = make_ssa_name (vec_dest); | |
3415 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
3416 | prev_res, half_res); | |
3417 | } | |
70439f0d | 3418 | else |
b1b6836e | 3419 | { |
a844293d | 3420 | gcall *call; |
b1b6836e | 3421 | if (ifn != IFN_LAST) |
a844293d | 3422 | call = gimple_build_call_internal_vec (ifn, vargs); |
b1b6836e | 3423 | else |
a844293d RS |
3424 | call = gimple_build_call_vec (fndecl, vargs); |
3425 | new_temp = make_ssa_name (vec_dest, call); | |
3426 | gimple_call_set_lhs (call, new_temp); | |
3427 | gimple_call_set_nothrow (call, true); | |
3428 | new_stmt = call; | |
b1b6836e | 3429 | } |
190c2236 | 3430 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 3431 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
3432 | } |
3433 | ||
3434 | for (i = 0; i < nargs; i++) | |
3435 | { | |
37b5ec8f | 3436 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 3437 | vec_oprndsi.release (); |
190c2236 | 3438 | } |
190c2236 JJ |
3439 | continue; |
3440 | } | |
3441 | ||
ebfd146a IR |
3442 | for (i = 0; i < nargs; i++) |
3443 | { | |
3444 | op = gimple_call_arg (stmt, i); | |
3445 | if (j == 0) | |
3446 | vec_oprnd0 | |
81c40241 | 3447 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 3448 | else |
63827fb8 IR |
3449 | { |
3450 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
3451 | vec_oprnd0 | |
3452 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
3453 | } | |
ebfd146a | 3454 | |
9771b263 | 3455 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
3456 | } |
3457 | ||
74bf76ed JJ |
3458 | if (gimple_call_internal_p (stmt) |
3459 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
3460 | { | |
c7bda0f4 | 3461 | tree cst = build_index_vector (vectype_out, j * nunits_out, 1); |
74bf76ed | 3462 | tree new_var |
0e22bb5a | 3463 | = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_"); |
355fe088 | 3464 | gimple *init_stmt = gimple_build_assign (new_var, cst); |
74bf76ed | 3465 | vect_init_vector_1 (stmt, init_stmt, NULL); |
b731b390 | 3466 | new_temp = make_ssa_name (vec_dest); |
0e22bb5a | 3467 | new_stmt = gimple_build_assign (new_temp, new_var); |
74bf76ed | 3468 | } |
b1b6836e RS |
3469 | else if (modifier == NARROW) |
3470 | { | |
3471 | tree half_res = make_ssa_name (vectype_in); | |
a844293d RS |
3472 | gcall *call = gimple_build_call_internal_vec (ifn, vargs); |
3473 | gimple_call_set_lhs (call, half_res); | |
3474 | gimple_call_set_nothrow (call, true); | |
3475 | new_stmt = call; | |
b1b6836e RS |
3476 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3477 | if ((j & 1) == 0) | |
3478 | { | |
3479 | prev_res = half_res; | |
3480 | continue; | |
3481 | } | |
3482 | new_temp = make_ssa_name (vec_dest); | |
3483 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
3484 | prev_res, half_res); | |
3485 | } | |
74bf76ed JJ |
3486 | else |
3487 | { | |
a844293d | 3488 | gcall *call; |
70439f0d | 3489 | if (ifn != IFN_LAST) |
a844293d | 3490 | call = gimple_build_call_internal_vec (ifn, vargs); |
70439f0d | 3491 | else |
a844293d | 3492 | call = gimple_build_call_vec (fndecl, vargs); |
74bf76ed | 3493 | new_temp = make_ssa_name (vec_dest, new_stmt); |
a844293d RS |
3494 | gimple_call_set_lhs (call, new_temp); |
3495 | gimple_call_set_nothrow (call, true); | |
3496 | new_stmt = call; | |
74bf76ed | 3497 | } |
ebfd146a IR |
3498 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3499 | ||
b1b6836e | 3500 | if (j == (modifier == NARROW ? 1 : 0)) |
ebfd146a IR |
3501 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; |
3502 | else | |
3503 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3504 | ||
3505 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3506 | } | |
b1b6836e RS |
3507 | } |
3508 | else if (modifier == NARROW) | |
3509 | { | |
ebfd146a IR |
3510 | for (j = 0; j < ncopies; ++j) |
3511 | { | |
3512 | /* Build argument list for the vectorized call. */ | |
3513 | if (j == 0) | |
9771b263 | 3514 | vargs.create (nargs * 2); |
ebfd146a | 3515 | else |
9771b263 | 3516 | vargs.truncate (0); |
ebfd146a | 3517 | |
190c2236 JJ |
3518 | if (slp_node) |
3519 | { | |
ef062b13 | 3520 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 3521 | vec<tree> vec_oprnds0; |
190c2236 JJ |
3522 | |
3523 | for (i = 0; i < nargs; i++) | |
9771b263 | 3524 | vargs.quick_push (gimple_call_arg (stmt, i)); |
306b0c92 | 3525 | vect_get_slp_defs (vargs, slp_node, &vec_defs); |
37b5ec8f | 3526 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
3527 | |
3528 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3529 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
3530 | { |
3531 | size_t k; | |
9771b263 | 3532 | vargs.truncate (0); |
190c2236 JJ |
3533 | for (k = 0; k < nargs; k++) |
3534 | { | |
37b5ec8f | 3535 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
3536 | vargs.quick_push (vec_oprndsk[i]); |
3537 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 | 3538 | } |
a844293d | 3539 | gcall *call; |
70439f0d | 3540 | if (ifn != IFN_LAST) |
a844293d | 3541 | call = gimple_build_call_internal_vec (ifn, vargs); |
70439f0d | 3542 | else |
a844293d RS |
3543 | call = gimple_build_call_vec (fndecl, vargs); |
3544 | new_temp = make_ssa_name (vec_dest, call); | |
3545 | gimple_call_set_lhs (call, new_temp); | |
3546 | gimple_call_set_nothrow (call, true); | |
3547 | new_stmt = call; | |
190c2236 | 3548 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 3549 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
3550 | } |
3551 | ||
3552 | for (i = 0; i < nargs; i++) | |
3553 | { | |
37b5ec8f | 3554 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 3555 | vec_oprndsi.release (); |
190c2236 | 3556 | } |
190c2236 JJ |
3557 | continue; |
3558 | } | |
3559 | ||
ebfd146a IR |
3560 | for (i = 0; i < nargs; i++) |
3561 | { | |
3562 | op = gimple_call_arg (stmt, i); | |
3563 | if (j == 0) | |
3564 | { | |
3565 | vec_oprnd0 | |
81c40241 | 3566 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 3567 | vec_oprnd1 |
63827fb8 | 3568 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
3569 | } |
3570 | else | |
3571 | { | |
336ecb65 | 3572 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 3573 | vec_oprnd0 |
63827fb8 | 3574 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 3575 | vec_oprnd1 |
63827fb8 | 3576 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
3577 | } |
3578 | ||
9771b263 DN |
3579 | vargs.quick_push (vec_oprnd0); |
3580 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
3581 | } |
3582 | ||
b1b6836e | 3583 | new_stmt = gimple_build_call_vec (fndecl, vargs); |
ebfd146a IR |
3584 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3585 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
3586 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3587 | ||
3588 | if (j == 0) | |
3589 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3590 | else | |
3591 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3592 | ||
3593 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3594 | } | |
3595 | ||
3596 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a | 3597 | } |
b1b6836e RS |
3598 | else |
3599 | /* No current target implements this case. */ | |
3600 | return false; | |
ebfd146a | 3601 | |
9771b263 | 3602 | vargs.release (); |
ebfd146a | 3603 | |
ebfd146a IR |
3604 | /* The call in STMT might prevent it from being removed in dce. |
3605 | We however cannot remove it here, due to the way the ssa name | |
3606 | it defines is mapped to the new definition. So just replace | |
3607 | rhs of the statement with something harmless. */ | |
3608 | ||
dd34c087 JJ |
3609 | if (slp_node) |
3610 | return true; | |
3611 | ||
ebfd146a | 3612 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
3613 | if (is_pattern_stmt_p (stmt_info)) |
3614 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
3615 | else | |
3616 | lhs = gimple_call_lhs (stmt); | |
3cc2fa2a | 3617 | |
9d5e7640 | 3618 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); |
ebfd146a | 3619 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 3620 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
3621 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
3622 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
3623 | |
3624 | return true; | |
3625 | } | |
3626 | ||
3627 | ||
0136f8f0 AH |
3628 | struct simd_call_arg_info |
3629 | { | |
3630 | tree vectype; | |
3631 | tree op; | |
0136f8f0 | 3632 | HOST_WIDE_INT linear_step; |
34e82342 | 3633 | enum vect_def_type dt; |
0136f8f0 | 3634 | unsigned int align; |
17b658af | 3635 | bool simd_lane_linear; |
0136f8f0 AH |
3636 | }; |
3637 | ||
17b658af JJ |
3638 | /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME, |
3639 | is linear within simd lane (but not within whole loop), note it in | |
3640 | *ARGINFO. */ | |
3641 | ||
3642 | static void | |
3643 | vect_simd_lane_linear (tree op, struct loop *loop, | |
3644 | struct simd_call_arg_info *arginfo) | |
3645 | { | |
355fe088 | 3646 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
17b658af JJ |
3647 | |
3648 | if (!is_gimple_assign (def_stmt) | |
3649 | || gimple_assign_rhs_code (def_stmt) != POINTER_PLUS_EXPR | |
3650 | || !is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))) | |
3651 | return; | |
3652 | ||
3653 | tree base = gimple_assign_rhs1 (def_stmt); | |
3654 | HOST_WIDE_INT linear_step = 0; | |
3655 | tree v = gimple_assign_rhs2 (def_stmt); | |
3656 | while (TREE_CODE (v) == SSA_NAME) | |
3657 | { | |
3658 | tree t; | |
3659 | def_stmt = SSA_NAME_DEF_STMT (v); | |
3660 | if (is_gimple_assign (def_stmt)) | |
3661 | switch (gimple_assign_rhs_code (def_stmt)) | |
3662 | { | |
3663 | case PLUS_EXPR: | |
3664 | t = gimple_assign_rhs2 (def_stmt); | |
3665 | if (linear_step || TREE_CODE (t) != INTEGER_CST) | |
3666 | return; | |
3667 | base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t); | |
3668 | v = gimple_assign_rhs1 (def_stmt); | |
3669 | continue; | |
3670 | case MULT_EXPR: | |
3671 | t = gimple_assign_rhs2 (def_stmt); | |
3672 | if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t)) | |
3673 | return; | |
3674 | linear_step = tree_to_shwi (t); | |
3675 | v = gimple_assign_rhs1 (def_stmt); | |
3676 | continue; | |
3677 | CASE_CONVERT: | |
3678 | t = gimple_assign_rhs1 (def_stmt); | |
3679 | if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE | |
3680 | || (TYPE_PRECISION (TREE_TYPE (v)) | |
3681 | < TYPE_PRECISION (TREE_TYPE (t)))) | |
3682 | return; | |
3683 | if (!linear_step) | |
3684 | linear_step = 1; | |
3685 | v = t; | |
3686 | continue; | |
3687 | default: | |
3688 | return; | |
3689 | } | |
8e4284d0 | 3690 | else if (gimple_call_internal_p (def_stmt, IFN_GOMP_SIMD_LANE) |
17b658af JJ |
3691 | && loop->simduid |
3692 | && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME | |
3693 | && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0)) | |
3694 | == loop->simduid)) | |
3695 | { | |
3696 | if (!linear_step) | |
3697 | linear_step = 1; | |
3698 | arginfo->linear_step = linear_step; | |
3699 | arginfo->op = base; | |
3700 | arginfo->simd_lane_linear = true; | |
3701 | return; | |
3702 | } | |
3703 | } | |
3704 | } | |
3705 | ||
cf1b2ba4 RS |
3706 | /* Return the number of elements in vector type VECTYPE, which is associated |
3707 | with a SIMD clone. At present these vectors always have a constant | |
3708 | length. */ | |
3709 | ||
3710 | static unsigned HOST_WIDE_INT | |
3711 | simd_clone_subparts (tree vectype) | |
3712 | { | |
928686b1 | 3713 | return TYPE_VECTOR_SUBPARTS (vectype).to_constant (); |
cf1b2ba4 RS |
3714 | } |
3715 | ||
0136f8f0 AH |
3716 | /* Function vectorizable_simd_clone_call. |
3717 | ||
3718 | Check if STMT performs a function call that can be vectorized | |
3719 | by calling a simd clone of the function. | |
3720 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3721 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3722 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3723 | ||
3724 | static bool | |
355fe088 | 3725 | vectorizable_simd_clone_call (gimple *stmt, gimple_stmt_iterator *gsi, |
68435eb2 RB |
3726 | gimple **vec_stmt, slp_tree slp_node, |
3727 | stmt_vector_for_cost *) | |
0136f8f0 AH |
3728 | { |
3729 | tree vec_dest; | |
3730 | tree scalar_dest; | |
3731 | tree op, type; | |
3732 | tree vec_oprnd0 = NULL_TREE; | |
3733 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
3734 | tree vectype; | |
3735 | unsigned int nunits; | |
3736 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3737 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
310213d4 | 3738 | vec_info *vinfo = stmt_info->vinfo; |
0136f8f0 | 3739 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
81c40241 | 3740 | tree fndecl, new_temp; |
355fe088 TS |
3741 | gimple *def_stmt; |
3742 | gimple *new_stmt = NULL; | |
0136f8f0 | 3743 | int ncopies, j; |
00426f9a | 3744 | auto_vec<simd_call_arg_info> arginfo; |
0136f8f0 AH |
3745 | vec<tree> vargs = vNULL; |
3746 | size_t i, nargs; | |
3747 | tree lhs, rtype, ratype; | |
e7a74006 | 3748 | vec<constructor_elt, va_gc> *ret_ctor_elts = NULL; |
0136f8f0 AH |
3749 | |
3750 | /* Is STMT a vectorizable call? */ | |
3751 | if (!is_gimple_call (stmt)) | |
3752 | return false; | |
3753 | ||
3754 | fndecl = gimple_call_fndecl (stmt); | |
3755 | if (fndecl == NULL_TREE) | |
3756 | return false; | |
3757 | ||
d52f5295 | 3758 | struct cgraph_node *node = cgraph_node::get (fndecl); |
0136f8f0 AH |
3759 | if (node == NULL || node->simd_clones == NULL) |
3760 | return false; | |
3761 | ||
3762 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
3763 | return false; | |
3764 | ||
66c16fd9 RB |
3765 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3766 | && ! vec_stmt) | |
0136f8f0 AH |
3767 | return false; |
3768 | ||
3769 | if (gimple_call_lhs (stmt) | |
3770 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
3771 | return false; | |
3772 | ||
3773 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); | |
3774 | ||
3775 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3776 | ||
3777 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt)) | |
3778 | return false; | |
3779 | ||
3780 | /* FORNOW */ | |
fce57248 | 3781 | if (slp_node) |
0136f8f0 AH |
3782 | return false; |
3783 | ||
3784 | /* Process function arguments. */ | |
3785 | nargs = gimple_call_num_args (stmt); | |
3786 | ||
3787 | /* Bail out if the function has zero arguments. */ | |
3788 | if (nargs == 0) | |
3789 | return false; | |
3790 | ||
00426f9a | 3791 | arginfo.reserve (nargs, true); |
0136f8f0 AH |
3792 | |
3793 | for (i = 0; i < nargs; i++) | |
3794 | { | |
3795 | simd_call_arg_info thisarginfo; | |
3796 | affine_iv iv; | |
3797 | ||
3798 | thisarginfo.linear_step = 0; | |
3799 | thisarginfo.align = 0; | |
3800 | thisarginfo.op = NULL_TREE; | |
17b658af | 3801 | thisarginfo.simd_lane_linear = false; |
0136f8f0 AH |
3802 | |
3803 | op = gimple_call_arg (stmt, i); | |
81c40241 RB |
3804 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &thisarginfo.dt, |
3805 | &thisarginfo.vectype) | |
0136f8f0 AH |
3806 | || thisarginfo.dt == vect_uninitialized_def) |
3807 | { | |
3808 | if (dump_enabled_p ()) | |
3809 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3810 | "use not simple.\n"); | |
0136f8f0 AH |
3811 | return false; |
3812 | } | |
3813 | ||
3814 | if (thisarginfo.dt == vect_constant_def | |
3815 | || thisarginfo.dt == vect_external_def) | |
3816 | gcc_assert (thisarginfo.vectype == NULL_TREE); | |
3817 | else | |
3818 | gcc_assert (thisarginfo.vectype != NULL_TREE); | |
3819 | ||
6c9e85fb JJ |
3820 | /* For linear arguments, the analyze phase should have saved |
3821 | the base and step in STMT_VINFO_SIMD_CLONE_INFO. */ | |
17b658af JJ |
3822 | if (i * 3 + 4 <= STMT_VINFO_SIMD_CLONE_INFO (stmt_info).length () |
3823 | && STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]) | |
6c9e85fb JJ |
3824 | { |
3825 | gcc_assert (vec_stmt); | |
3826 | thisarginfo.linear_step | |
17b658af | 3827 | = tree_to_shwi (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]); |
6c9e85fb | 3828 | thisarginfo.op |
17b658af JJ |
3829 | = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 1]; |
3830 | thisarginfo.simd_lane_linear | |
3831 | = (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 3] | |
3832 | == boolean_true_node); | |
6c9e85fb JJ |
3833 | /* If loop has been peeled for alignment, we need to adjust it. */ |
3834 | tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo); | |
3835 | tree n2 = LOOP_VINFO_NITERS (loop_vinfo); | |
17b658af | 3836 | if (n1 != n2 && !thisarginfo.simd_lane_linear) |
6c9e85fb JJ |
3837 | { |
3838 | tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2); | |
17b658af | 3839 | tree step = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]; |
6c9e85fb JJ |
3840 | tree opt = TREE_TYPE (thisarginfo.op); |
3841 | bias = fold_convert (TREE_TYPE (step), bias); | |
3842 | bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step); | |
3843 | thisarginfo.op | |
3844 | = fold_build2 (POINTER_TYPE_P (opt) | |
3845 | ? POINTER_PLUS_EXPR : PLUS_EXPR, opt, | |
3846 | thisarginfo.op, bias); | |
3847 | } | |
3848 | } | |
3849 | else if (!vec_stmt | |
3850 | && thisarginfo.dt != vect_constant_def | |
3851 | && thisarginfo.dt != vect_external_def | |
3852 | && loop_vinfo | |
3853 | && TREE_CODE (op) == SSA_NAME | |
3854 | && simple_iv (loop, loop_containing_stmt (stmt), op, | |
3855 | &iv, false) | |
3856 | && tree_fits_shwi_p (iv.step)) | |
0136f8f0 AH |
3857 | { |
3858 | thisarginfo.linear_step = tree_to_shwi (iv.step); | |
3859 | thisarginfo.op = iv.base; | |
3860 | } | |
3861 | else if ((thisarginfo.dt == vect_constant_def | |
3862 | || thisarginfo.dt == vect_external_def) | |
3863 | && POINTER_TYPE_P (TREE_TYPE (op))) | |
3864 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; | |
17b658af JJ |
3865 | /* Addresses of array elements indexed by GOMP_SIMD_LANE are |
3866 | linear too. */ | |
3867 | if (POINTER_TYPE_P (TREE_TYPE (op)) | |
3868 | && !thisarginfo.linear_step | |
3869 | && !vec_stmt | |
3870 | && thisarginfo.dt != vect_constant_def | |
3871 | && thisarginfo.dt != vect_external_def | |
3872 | && loop_vinfo | |
3873 | && !slp_node | |
3874 | && TREE_CODE (op) == SSA_NAME) | |
3875 | vect_simd_lane_linear (op, loop, &thisarginfo); | |
0136f8f0 AH |
3876 | |
3877 | arginfo.quick_push (thisarginfo); | |
3878 | } | |
3879 | ||
d9f21f6a RS |
3880 | unsigned HOST_WIDE_INT vf; |
3881 | if (!LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (&vf)) | |
3882 | { | |
3883 | if (dump_enabled_p ()) | |
3884 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3885 | "not considering SIMD clones; not yet supported" | |
3886 | " for variable-width vectors.\n"); | |
3887 | return NULL; | |
3888 | } | |
3889 | ||
0136f8f0 AH |
3890 | unsigned int badness = 0; |
3891 | struct cgraph_node *bestn = NULL; | |
6c9e85fb JJ |
3892 | if (STMT_VINFO_SIMD_CLONE_INFO (stmt_info).exists ()) |
3893 | bestn = cgraph_node::get (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[0]); | |
0136f8f0 AH |
3894 | else |
3895 | for (struct cgraph_node *n = node->simd_clones; n != NULL; | |
3896 | n = n->simdclone->next_clone) | |
3897 | { | |
3898 | unsigned int this_badness = 0; | |
d9f21f6a | 3899 | if (n->simdclone->simdlen > vf |
0136f8f0 AH |
3900 | || n->simdclone->nargs != nargs) |
3901 | continue; | |
d9f21f6a RS |
3902 | if (n->simdclone->simdlen < vf) |
3903 | this_badness += (exact_log2 (vf) | |
0136f8f0 AH |
3904 | - exact_log2 (n->simdclone->simdlen)) * 1024; |
3905 | if (n->simdclone->inbranch) | |
3906 | this_badness += 2048; | |
3907 | int target_badness = targetm.simd_clone.usable (n); | |
3908 | if (target_badness < 0) | |
3909 | continue; | |
3910 | this_badness += target_badness * 512; | |
3911 | /* FORNOW: Have to add code to add the mask argument. */ | |
3912 | if (n->simdclone->inbranch) | |
3913 | continue; | |
3914 | for (i = 0; i < nargs; i++) | |
3915 | { | |
3916 | switch (n->simdclone->args[i].arg_type) | |
3917 | { | |
3918 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3919 | if (!useless_type_conversion_p | |
3920 | (n->simdclone->args[i].orig_type, | |
3921 | TREE_TYPE (gimple_call_arg (stmt, i)))) | |
3922 | i = -1; | |
3923 | else if (arginfo[i].dt == vect_constant_def | |
3924 | || arginfo[i].dt == vect_external_def | |
3925 | || arginfo[i].linear_step) | |
3926 | this_badness += 64; | |
3927 | break; | |
3928 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3929 | if (arginfo[i].dt != vect_constant_def | |
3930 | && arginfo[i].dt != vect_external_def) | |
3931 | i = -1; | |
3932 | break; | |
3933 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
d9a6bd32 | 3934 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
3935 | if (arginfo[i].dt == vect_constant_def |
3936 | || arginfo[i].dt == vect_external_def | |
3937 | || (arginfo[i].linear_step | |
3938 | != n->simdclone->args[i].linear_step)) | |
3939 | i = -1; | |
3940 | break; | |
3941 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
d9a6bd32 JJ |
3942 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
3943 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
e01d41e5 JJ |
3944 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3945 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3946 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3947 | /* FORNOW */ |
3948 | i = -1; | |
3949 | break; | |
3950 | case SIMD_CLONE_ARG_TYPE_MASK: | |
3951 | gcc_unreachable (); | |
3952 | } | |
3953 | if (i == (size_t) -1) | |
3954 | break; | |
3955 | if (n->simdclone->args[i].alignment > arginfo[i].align) | |
3956 | { | |
3957 | i = -1; | |
3958 | break; | |
3959 | } | |
3960 | if (arginfo[i].align) | |
3961 | this_badness += (exact_log2 (arginfo[i].align) | |
3962 | - exact_log2 (n->simdclone->args[i].alignment)); | |
3963 | } | |
3964 | if (i == (size_t) -1) | |
3965 | continue; | |
3966 | if (bestn == NULL || this_badness < badness) | |
3967 | { | |
3968 | bestn = n; | |
3969 | badness = this_badness; | |
3970 | } | |
3971 | } | |
3972 | ||
3973 | if (bestn == NULL) | |
00426f9a | 3974 | return false; |
0136f8f0 AH |
3975 | |
3976 | for (i = 0; i < nargs; i++) | |
3977 | if ((arginfo[i].dt == vect_constant_def | |
3978 | || arginfo[i].dt == vect_external_def) | |
3979 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
3980 | { | |
3981 | arginfo[i].vectype | |
3982 | = get_vectype_for_scalar_type (TREE_TYPE (gimple_call_arg (stmt, | |
3983 | i))); | |
3984 | if (arginfo[i].vectype == NULL | |
cf1b2ba4 | 3985 | || (simd_clone_subparts (arginfo[i].vectype) |
0136f8f0 | 3986 | > bestn->simdclone->simdlen)) |
00426f9a | 3987 | return false; |
0136f8f0 AH |
3988 | } |
3989 | ||
3990 | fndecl = bestn->decl; | |
3991 | nunits = bestn->simdclone->simdlen; | |
d9f21f6a | 3992 | ncopies = vf / nunits; |
0136f8f0 AH |
3993 | |
3994 | /* If the function isn't const, only allow it in simd loops where user | |
3995 | has asserted that at least nunits consecutive iterations can be | |
3996 | performed using SIMD instructions. */ | |
3997 | if ((loop == NULL || (unsigned) loop->safelen < nunits) | |
3998 | && gimple_vuse (stmt)) | |
00426f9a | 3999 | return false; |
0136f8f0 AH |
4000 | |
4001 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
4002 | needs to be generated. */ | |
4003 | gcc_assert (ncopies >= 1); | |
4004 | ||
4005 | if (!vec_stmt) /* transformation not required. */ | |
4006 | { | |
6c9e85fb JJ |
4007 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (bestn->decl); |
4008 | for (i = 0; i < nargs; i++) | |
7adb26f2 JJ |
4009 | if ((bestn->simdclone->args[i].arg_type |
4010 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
4011 | || (bestn->simdclone->args[i].arg_type | |
4012 | == SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP)) | |
6c9e85fb | 4013 | { |
17b658af | 4014 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_grow_cleared (i * 3 |
6c9e85fb JJ |
4015 | + 1); |
4016 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (arginfo[i].op); | |
4017 | tree lst = POINTER_TYPE_P (TREE_TYPE (arginfo[i].op)) | |
4018 | ? size_type_node : TREE_TYPE (arginfo[i].op); | |
4019 | tree ls = build_int_cst (lst, arginfo[i].linear_step); | |
4020 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (ls); | |
17b658af JJ |
4021 | tree sll = arginfo[i].simd_lane_linear |
4022 | ? boolean_true_node : boolean_false_node; | |
4023 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (sll); | |
6c9e85fb | 4024 | } |
0136f8f0 AH |
4025 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; |
4026 | if (dump_enabled_p ()) | |
4027 | dump_printf_loc (MSG_NOTE, vect_location, | |
4028 | "=== vectorizable_simd_clone_call ===\n"); | |
68435eb2 | 4029 | /* vect_model_simple_cost (stmt_info, ncopies, dt, slp_node, cost_vec); */ |
0136f8f0 AH |
4030 | return true; |
4031 | } | |
4032 | ||
67b8dbac | 4033 | /* Transform. */ |
0136f8f0 AH |
4034 | |
4035 | if (dump_enabled_p ()) | |
4036 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); | |
4037 | ||
4038 | /* Handle def. */ | |
4039 | scalar_dest = gimple_call_lhs (stmt); | |
4040 | vec_dest = NULL_TREE; | |
4041 | rtype = NULL_TREE; | |
4042 | ratype = NULL_TREE; | |
4043 | if (scalar_dest) | |
4044 | { | |
4045 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
4046 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); | |
4047 | if (TREE_CODE (rtype) == ARRAY_TYPE) | |
4048 | { | |
4049 | ratype = rtype; | |
4050 | rtype = TREE_TYPE (ratype); | |
4051 | } | |
4052 | } | |
4053 | ||
4054 | prev_stmt_info = NULL; | |
4055 | for (j = 0; j < ncopies; ++j) | |
4056 | { | |
4057 | /* Build argument list for the vectorized call. */ | |
4058 | if (j == 0) | |
4059 | vargs.create (nargs); | |
4060 | else | |
4061 | vargs.truncate (0); | |
4062 | ||
4063 | for (i = 0; i < nargs; i++) | |
4064 | { | |
4065 | unsigned int k, l, m, o; | |
4066 | tree atype; | |
4067 | op = gimple_call_arg (stmt, i); | |
4068 | switch (bestn->simdclone->args[i].arg_type) | |
4069 | { | |
4070 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
4071 | atype = bestn->simdclone->args[i].vector_type; | |
cf1b2ba4 | 4072 | o = nunits / simd_clone_subparts (atype); |
0136f8f0 AH |
4073 | for (m = j * o; m < (j + 1) * o; m++) |
4074 | { | |
cf1b2ba4 RS |
4075 | if (simd_clone_subparts (atype) |
4076 | < simd_clone_subparts (arginfo[i].vectype)) | |
0136f8f0 | 4077 | { |
73a699ae | 4078 | poly_uint64 prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); |
cf1b2ba4 RS |
4079 | k = (simd_clone_subparts (arginfo[i].vectype) |
4080 | / simd_clone_subparts (atype)); | |
0136f8f0 AH |
4081 | gcc_assert ((k & (k - 1)) == 0); |
4082 | if (m == 0) | |
4083 | vec_oprnd0 | |
81c40241 | 4084 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
4085 | else |
4086 | { | |
4087 | vec_oprnd0 = arginfo[i].op; | |
4088 | if ((m & (k - 1)) == 0) | |
4089 | vec_oprnd0 | |
4090 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
4091 | vec_oprnd0); | |
4092 | } | |
4093 | arginfo[i].op = vec_oprnd0; | |
4094 | vec_oprnd0 | |
4095 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, | |
92e29a5e | 4096 | bitsize_int (prec), |
0136f8f0 AH |
4097 | bitsize_int ((m & (k - 1)) * prec)); |
4098 | new_stmt | |
b731b390 | 4099 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
4100 | vec_oprnd0); |
4101 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4102 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
4103 | } | |
4104 | else | |
4105 | { | |
cf1b2ba4 RS |
4106 | k = (simd_clone_subparts (atype) |
4107 | / simd_clone_subparts (arginfo[i].vectype)); | |
0136f8f0 AH |
4108 | gcc_assert ((k & (k - 1)) == 0); |
4109 | vec<constructor_elt, va_gc> *ctor_elts; | |
4110 | if (k != 1) | |
4111 | vec_alloc (ctor_elts, k); | |
4112 | else | |
4113 | ctor_elts = NULL; | |
4114 | for (l = 0; l < k; l++) | |
4115 | { | |
4116 | if (m == 0 && l == 0) | |
4117 | vec_oprnd0 | |
81c40241 | 4118 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
4119 | else |
4120 | vec_oprnd0 | |
4121 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
4122 | arginfo[i].op); | |
4123 | arginfo[i].op = vec_oprnd0; | |
4124 | if (k == 1) | |
4125 | break; | |
4126 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, | |
4127 | vec_oprnd0); | |
4128 | } | |
4129 | if (k == 1) | |
4130 | vargs.safe_push (vec_oprnd0); | |
4131 | else | |
4132 | { | |
4133 | vec_oprnd0 = build_constructor (atype, ctor_elts); | |
4134 | new_stmt | |
b731b390 | 4135 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
4136 | vec_oprnd0); |
4137 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4138 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
4139 | } | |
4140 | } | |
4141 | } | |
4142 | break; | |
4143 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
4144 | vargs.safe_push (op); | |
4145 | break; | |
4146 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
7adb26f2 | 4147 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
4148 | if (j == 0) |
4149 | { | |
4150 | gimple_seq stmts; | |
4151 | arginfo[i].op | |
4152 | = force_gimple_operand (arginfo[i].op, &stmts, true, | |
4153 | NULL_TREE); | |
4154 | if (stmts != NULL) | |
4155 | { | |
4156 | basic_block new_bb; | |
4157 | edge pe = loop_preheader_edge (loop); | |
4158 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
4159 | gcc_assert (!new_bb); | |
4160 | } | |
17b658af JJ |
4161 | if (arginfo[i].simd_lane_linear) |
4162 | { | |
4163 | vargs.safe_push (arginfo[i].op); | |
4164 | break; | |
4165 | } | |
b731b390 | 4166 | tree phi_res = copy_ssa_name (op); |
538dd0b7 | 4167 | gphi *new_phi = create_phi_node (phi_res, loop->header); |
0136f8f0 | 4168 | set_vinfo_for_stmt (new_phi, |
310213d4 | 4169 | new_stmt_vec_info (new_phi, loop_vinfo)); |
0136f8f0 AH |
4170 | add_phi_arg (new_phi, arginfo[i].op, |
4171 | loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
4172 | enum tree_code code | |
4173 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
4174 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
4175 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
4176 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
4177 | widest_int cst |
4178 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
4179 | ncopies * nunits); | |
4180 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 4181 | tree phi_arg = copy_ssa_name (op); |
0d0e4a03 JJ |
4182 | new_stmt |
4183 | = gimple_build_assign (phi_arg, code, phi_res, tcst); | |
0136f8f0 AH |
4184 | gimple_stmt_iterator si = gsi_after_labels (loop->header); |
4185 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); | |
4186 | set_vinfo_for_stmt (new_stmt, | |
310213d4 | 4187 | new_stmt_vec_info (new_stmt, loop_vinfo)); |
0136f8f0 AH |
4188 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), |
4189 | UNKNOWN_LOCATION); | |
4190 | arginfo[i].op = phi_res; | |
4191 | vargs.safe_push (phi_res); | |
4192 | } | |
4193 | else | |
4194 | { | |
4195 | enum tree_code code | |
4196 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
4197 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
4198 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
4199 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
4200 | widest_int cst |
4201 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
4202 | j * nunits); | |
4203 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 4204 | new_temp = make_ssa_name (TREE_TYPE (op)); |
0d0e4a03 JJ |
4205 | new_stmt = gimple_build_assign (new_temp, code, |
4206 | arginfo[i].op, tcst); | |
0136f8f0 AH |
4207 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4208 | vargs.safe_push (new_temp); | |
4209 | } | |
4210 | break; | |
7adb26f2 JJ |
4211 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
4212 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
0136f8f0 | 4213 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: |
e01d41e5 JJ |
4214 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
4215 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
4216 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
4217 | default: |
4218 | gcc_unreachable (); | |
4219 | } | |
4220 | } | |
4221 | ||
4222 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
4223 | if (vec_dest) | |
4224 | { | |
cf1b2ba4 | 4225 | gcc_assert (ratype || simd_clone_subparts (rtype) == nunits); |
0136f8f0 | 4226 | if (ratype) |
b731b390 | 4227 | new_temp = create_tmp_var (ratype); |
cf1b2ba4 RS |
4228 | else if (simd_clone_subparts (vectype) |
4229 | == simd_clone_subparts (rtype)) | |
0136f8f0 AH |
4230 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4231 | else | |
4232 | new_temp = make_ssa_name (rtype, new_stmt); | |
4233 | gimple_call_set_lhs (new_stmt, new_temp); | |
4234 | } | |
4235 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4236 | ||
4237 | if (vec_dest) | |
4238 | { | |
cf1b2ba4 | 4239 | if (simd_clone_subparts (vectype) < nunits) |
0136f8f0 AH |
4240 | { |
4241 | unsigned int k, l; | |
73a699ae RS |
4242 | poly_uint64 prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); |
4243 | poly_uint64 bytes = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
cf1b2ba4 | 4244 | k = nunits / simd_clone_subparts (vectype); |
0136f8f0 AH |
4245 | gcc_assert ((k & (k - 1)) == 0); |
4246 | for (l = 0; l < k; l++) | |
4247 | { | |
4248 | tree t; | |
4249 | if (ratype) | |
4250 | { | |
4251 | t = build_fold_addr_expr (new_temp); | |
4252 | t = build2 (MEM_REF, vectype, t, | |
73a699ae | 4253 | build_int_cst (TREE_TYPE (t), l * bytes)); |
0136f8f0 AH |
4254 | } |
4255 | else | |
4256 | t = build3 (BIT_FIELD_REF, vectype, new_temp, | |
92e29a5e | 4257 | bitsize_int (prec), bitsize_int (l * prec)); |
0136f8f0 | 4258 | new_stmt |
b731b390 | 4259 | = gimple_build_assign (make_ssa_name (vectype), t); |
0136f8f0 AH |
4260 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4261 | if (j == 0 && l == 0) | |
4262 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4263 | else | |
4264 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4265 | ||
4266 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4267 | } | |
4268 | ||
4269 | if (ratype) | |
3ba4ff41 | 4270 | vect_clobber_variable (stmt, gsi, new_temp); |
0136f8f0 AH |
4271 | continue; |
4272 | } | |
cf1b2ba4 | 4273 | else if (simd_clone_subparts (vectype) > nunits) |
0136f8f0 | 4274 | { |
cf1b2ba4 RS |
4275 | unsigned int k = (simd_clone_subparts (vectype) |
4276 | / simd_clone_subparts (rtype)); | |
0136f8f0 AH |
4277 | gcc_assert ((k & (k - 1)) == 0); |
4278 | if ((j & (k - 1)) == 0) | |
4279 | vec_alloc (ret_ctor_elts, k); | |
4280 | if (ratype) | |
4281 | { | |
cf1b2ba4 | 4282 | unsigned int m, o = nunits / simd_clone_subparts (rtype); |
0136f8f0 AH |
4283 | for (m = 0; m < o; m++) |
4284 | { | |
4285 | tree tem = build4 (ARRAY_REF, rtype, new_temp, | |
4286 | size_int (m), NULL_TREE, NULL_TREE); | |
4287 | new_stmt | |
b731b390 | 4288 | = gimple_build_assign (make_ssa_name (rtype), tem); |
0136f8f0 AH |
4289 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4290 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, | |
4291 | gimple_assign_lhs (new_stmt)); | |
4292 | } | |
3ba4ff41 | 4293 | vect_clobber_variable (stmt, gsi, new_temp); |
0136f8f0 AH |
4294 | } |
4295 | else | |
4296 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); | |
4297 | if ((j & (k - 1)) != k - 1) | |
4298 | continue; | |
4299 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); | |
4300 | new_stmt | |
b731b390 | 4301 | = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0); |
0136f8f0 AH |
4302 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4303 | ||
4304 | if ((unsigned) j == k - 1) | |
4305 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4306 | else | |
4307 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4308 | ||
4309 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4310 | continue; | |
4311 | } | |
4312 | else if (ratype) | |
4313 | { | |
4314 | tree t = build_fold_addr_expr (new_temp); | |
4315 | t = build2 (MEM_REF, vectype, t, | |
4316 | build_int_cst (TREE_TYPE (t), 0)); | |
4317 | new_stmt | |
b731b390 | 4318 | = gimple_build_assign (make_ssa_name (vec_dest), t); |
0136f8f0 | 4319 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3ba4ff41 | 4320 | vect_clobber_variable (stmt, gsi, new_temp); |
0136f8f0 AH |
4321 | } |
4322 | } | |
4323 | ||
4324 | if (j == 0) | |
4325 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4326 | else | |
4327 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4328 | ||
4329 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4330 | } | |
4331 | ||
4332 | vargs.release (); | |
4333 | ||
4334 | /* The call in STMT might prevent it from being removed in dce. | |
4335 | We however cannot remove it here, due to the way the ssa name | |
4336 | it defines is mapped to the new definition. So just replace | |
4337 | rhs of the statement with something harmless. */ | |
4338 | ||
4339 | if (slp_node) | |
4340 | return true; | |
4341 | ||
4342 | if (scalar_dest) | |
4343 | { | |
4344 | type = TREE_TYPE (scalar_dest); | |
4345 | if (is_pattern_stmt_p (stmt_info)) | |
4346 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
4347 | else | |
4348 | lhs = gimple_call_lhs (stmt); | |
4349 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
4350 | } | |
4351 | else | |
4352 | new_stmt = gimple_build_nop (); | |
4353 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
4354 | set_vinfo_for_stmt (stmt, NULL); | |
4355 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2865f32a | 4356 | gsi_replace (gsi, new_stmt, true); |
0136f8f0 AH |
4357 | unlink_stmt_vdef (stmt); |
4358 | ||
4359 | return true; | |
4360 | } | |
4361 | ||
4362 | ||
ebfd146a IR |
4363 | /* Function vect_gen_widened_results_half |
4364 | ||
4365 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 4366 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 4367 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
4368 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
4369 | needs to be created (DECL is a function-decl of a target-builtin). | |
4370 | STMT is the original scalar stmt that we are vectorizing. */ | |
4371 | ||
355fe088 | 4372 | static gimple * |
ebfd146a IR |
4373 | vect_gen_widened_results_half (enum tree_code code, |
4374 | tree decl, | |
4375 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
4376 | tree vec_dest, gimple_stmt_iterator *gsi, | |
355fe088 | 4377 | gimple *stmt) |
b8698a0f | 4378 | { |
355fe088 | 4379 | gimple *new_stmt; |
b8698a0f L |
4380 | tree new_temp; |
4381 | ||
4382 | /* Generate half of the widened result: */ | |
4383 | if (code == CALL_EXPR) | |
4384 | { | |
4385 | /* Target specific support */ | |
ebfd146a IR |
4386 | if (op_type == binary_op) |
4387 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
4388 | else | |
4389 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
4390 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4391 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
4392 | } |
4393 | else | |
ebfd146a | 4394 | { |
b8698a0f L |
4395 | /* Generic support */ |
4396 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
4397 | if (op_type != binary_op) |
4398 | vec_oprnd1 = NULL; | |
0d0e4a03 | 4399 | new_stmt = gimple_build_assign (vec_dest, code, vec_oprnd0, vec_oprnd1); |
ebfd146a IR |
4400 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4401 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 4402 | } |
ebfd146a IR |
4403 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4404 | ||
ebfd146a IR |
4405 | return new_stmt; |
4406 | } | |
4407 | ||
4a00c761 JJ |
4408 | |
4409 | /* Get vectorized definitions for loop-based vectorization. For the first | |
4410 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
4411 | scalar operand), and for the rest we get a copy with | |
4412 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
4413 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
4414 | The vectors are collected into VEC_OPRNDS. */ | |
4415 | ||
4416 | static void | |
355fe088 | 4417 | vect_get_loop_based_defs (tree *oprnd, gimple *stmt, enum vect_def_type dt, |
9771b263 | 4418 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
4419 | { |
4420 | tree vec_oprnd; | |
4421 | ||
4422 | /* Get first vector operand. */ | |
4423 | /* All the vector operands except the very first one (that is scalar oprnd) | |
4424 | are stmt copies. */ | |
4425 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
81c40241 | 4426 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt); |
4a00c761 JJ |
4427 | else |
4428 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
4429 | ||
9771b263 | 4430 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
4431 | |
4432 | /* Get second vector operand. */ | |
4433 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 4434 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
4435 | |
4436 | *oprnd = vec_oprnd; | |
4437 | ||
4438 | /* For conversion in multiple steps, continue to get operands | |
4439 | recursively. */ | |
4440 | if (multi_step_cvt) | |
4441 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
4442 | } | |
4443 | ||
4444 | ||
4445 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
4446 | For multi-step conversions store the resulting vectors and call the function | |
4447 | recursively. */ | |
4448 | ||
4449 | static void | |
9771b263 | 4450 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
355fe088 | 4451 | int multi_step_cvt, gimple *stmt, |
9771b263 | 4452 | vec<tree> vec_dsts, |
4a00c761 JJ |
4453 | gimple_stmt_iterator *gsi, |
4454 | slp_tree slp_node, enum tree_code code, | |
4455 | stmt_vec_info *prev_stmt_info) | |
4456 | { | |
4457 | unsigned int i; | |
4458 | tree vop0, vop1, new_tmp, vec_dest; | |
355fe088 | 4459 | gimple *new_stmt; |
4a00c761 JJ |
4460 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
4461 | ||
9771b263 | 4462 | vec_dest = vec_dsts.pop (); |
4a00c761 | 4463 | |
9771b263 | 4464 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
4465 | { |
4466 | /* Create demotion operation. */ | |
9771b263 DN |
4467 | vop0 = (*vec_oprnds)[i]; |
4468 | vop1 = (*vec_oprnds)[i + 1]; | |
0d0e4a03 | 4469 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
4a00c761 JJ |
4470 | new_tmp = make_ssa_name (vec_dest, new_stmt); |
4471 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
4472 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4473 | ||
4474 | if (multi_step_cvt) | |
4475 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 4476 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
4477 | else |
4478 | { | |
4479 | /* This is the last step of the conversion sequence. Store the | |
4480 | vectors in SLP_NODE or in vector info of the scalar statement | |
4481 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
4482 | if (slp_node) | |
9771b263 | 4483 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 4484 | else |
c689ce1e RB |
4485 | { |
4486 | if (!*prev_stmt_info) | |
4487 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
4488 | else | |
4489 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
4a00c761 | 4490 | |
c689ce1e RB |
4491 | *prev_stmt_info = vinfo_for_stmt (new_stmt); |
4492 | } | |
4a00c761 JJ |
4493 | } |
4494 | } | |
4495 | ||
4496 | /* For multi-step demotion operations we first generate demotion operations | |
4497 | from the source type to the intermediate types, and then combine the | |
4498 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
4499 | type. */ | |
4500 | if (multi_step_cvt) | |
4501 | { | |
4502 | /* At each level of recursion we have half of the operands we had at the | |
4503 | previous level. */ | |
9771b263 | 4504 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
4505 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
4506 | stmt, vec_dsts, gsi, slp_node, | |
4507 | VEC_PACK_TRUNC_EXPR, | |
4508 | prev_stmt_info); | |
4509 | } | |
4510 | ||
9771b263 | 4511 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4512 | } |
4513 | ||
4514 | ||
4515 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
4516 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
4517 | the resulting vectors and call the function recursively. */ | |
4518 | ||
4519 | static void | |
9771b263 DN |
4520 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
4521 | vec<tree> *vec_oprnds1, | |
355fe088 | 4522 | gimple *stmt, tree vec_dest, |
4a00c761 JJ |
4523 | gimple_stmt_iterator *gsi, |
4524 | enum tree_code code1, | |
4525 | enum tree_code code2, tree decl1, | |
4526 | tree decl2, int op_type) | |
4527 | { | |
4528 | int i; | |
4529 | tree vop0, vop1, new_tmp1, new_tmp2; | |
355fe088 | 4530 | gimple *new_stmt1, *new_stmt2; |
6e1aa848 | 4531 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 4532 | |
9771b263 DN |
4533 | vec_tmp.create (vec_oprnds0->length () * 2); |
4534 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
4535 | { |
4536 | if (op_type == binary_op) | |
9771b263 | 4537 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
4538 | else |
4539 | vop1 = NULL_TREE; | |
4540 | ||
4541 | /* Generate the two halves of promotion operation. */ | |
4542 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
4543 | op_type, vec_dest, gsi, stmt); | |
4544 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
4545 | op_type, vec_dest, gsi, stmt); | |
4546 | if (is_gimple_call (new_stmt1)) | |
4547 | { | |
4548 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
4549 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
4550 | } | |
4551 | else | |
4552 | { | |
4553 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
4554 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
4555 | } | |
4556 | ||
4557 | /* Store the results for the next step. */ | |
9771b263 DN |
4558 | vec_tmp.quick_push (new_tmp1); |
4559 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
4560 | } |
4561 | ||
689eaba3 | 4562 | vec_oprnds0->release (); |
4a00c761 JJ |
4563 | *vec_oprnds0 = vec_tmp; |
4564 | } | |
4565 | ||
4566 | ||
b8698a0f L |
4567 | /* Check if STMT performs a conversion operation, that can be vectorized. |
4568 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 4569 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
4570 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
4571 | ||
4572 | static bool | |
355fe088 | 4573 | vectorizable_conversion (gimple *stmt, gimple_stmt_iterator *gsi, |
68435eb2 RB |
4574 | gimple **vec_stmt, slp_tree slp_node, |
4575 | stmt_vector_for_cost *cost_vec) | |
ebfd146a IR |
4576 | { |
4577 | tree vec_dest; | |
4578 | tree scalar_dest; | |
4a00c761 | 4579 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
4580 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
4581 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4582 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4583 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 4584 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
4585 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
4586 | tree new_temp; | |
355fe088 | 4587 | gimple *def_stmt; |
ebfd146a | 4588 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 4589 | int ndts = 2; |
355fe088 | 4590 | gimple *new_stmt = NULL; |
ebfd146a | 4591 | stmt_vec_info prev_stmt_info; |
062d5ccc RS |
4592 | poly_uint64 nunits_in; |
4593 | poly_uint64 nunits_out; | |
ebfd146a | 4594 | tree vectype_out, vectype_in; |
4a00c761 JJ |
4595 | int ncopies, i, j; |
4596 | tree lhs_type, rhs_type; | |
ebfd146a | 4597 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
4598 | vec<tree> vec_oprnds0 = vNULL; |
4599 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 4600 | tree vop0; |
4a00c761 | 4601 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4602 | vec_info *vinfo = stmt_info->vinfo; |
4a00c761 | 4603 | int multi_step_cvt = 0; |
6e1aa848 | 4604 | vec<tree> interm_types = vNULL; |
4a00c761 JJ |
4605 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
4606 | int op_type; | |
4a00c761 | 4607 | unsigned short fltsz; |
ebfd146a IR |
4608 | |
4609 | /* Is STMT a vectorizable conversion? */ | |
4610 | ||
4a00c761 | 4611 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4612 | return false; |
4613 | ||
66c16fd9 RB |
4614 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4615 | && ! vec_stmt) | |
ebfd146a IR |
4616 | return false; |
4617 | ||
4618 | if (!is_gimple_assign (stmt)) | |
4619 | return false; | |
4620 | ||
4621 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4622 | return false; | |
4623 | ||
4624 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
4625 | if (!CONVERT_EXPR_CODE_P (code) |
4626 | && code != FIX_TRUNC_EXPR | |
4627 | && code != FLOAT_EXPR | |
4628 | && code != WIDEN_MULT_EXPR | |
4629 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
4630 | return false; |
4631 | ||
4a00c761 JJ |
4632 | op_type = TREE_CODE_LENGTH (code); |
4633 | ||
ebfd146a | 4634 | /* Check types of lhs and rhs. */ |
b690cc0f | 4635 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 4636 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
4637 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
4638 | ||
ebfd146a IR |
4639 | op0 = gimple_assign_rhs1 (stmt); |
4640 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
4641 | |
4642 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4643 | && !((INTEGRAL_TYPE_P (lhs_type) | |
4644 | && INTEGRAL_TYPE_P (rhs_type)) | |
4645 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
4646 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
4647 | return false; | |
4648 | ||
e6f5c25d IE |
4649 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4650 | && ((INTEGRAL_TYPE_P (lhs_type) | |
2be65d9e | 4651 | && !type_has_mode_precision_p (lhs_type)) |
e6f5c25d | 4652 | || (INTEGRAL_TYPE_P (rhs_type) |
2be65d9e | 4653 | && !type_has_mode_precision_p (rhs_type)))) |
4a00c761 | 4654 | { |
73fbfcad | 4655 | if (dump_enabled_p ()) |
78c60e3d | 4656 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
4657 | "type conversion to/from bit-precision unsupported." |
4658 | "\n"); | |
4a00c761 JJ |
4659 | return false; |
4660 | } | |
4661 | ||
b690cc0f | 4662 | /* Check the operands of the operation. */ |
81c40241 | 4663 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype_in)) |
b690cc0f | 4664 | { |
73fbfcad | 4665 | if (dump_enabled_p ()) |
78c60e3d | 4666 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4667 | "use not simple.\n"); |
b690cc0f RG |
4668 | return false; |
4669 | } | |
4a00c761 JJ |
4670 | if (op_type == binary_op) |
4671 | { | |
4672 | bool ok; | |
4673 | ||
4674 | op1 = gimple_assign_rhs2 (stmt); | |
4675 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
4676 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
4677 | OP1. */ | |
4678 | if (CONSTANT_CLASS_P (op0)) | |
81c40241 | 4679 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &vectype_in); |
4a00c761 | 4680 | else |
81c40241 | 4681 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1]); |
4a00c761 JJ |
4682 | |
4683 | if (!ok) | |
4684 | { | |
73fbfcad | 4685 | if (dump_enabled_p ()) |
78c60e3d | 4686 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4687 | "use not simple.\n"); |
4a00c761 JJ |
4688 | return false; |
4689 | } | |
4690 | } | |
4691 | ||
b690cc0f RG |
4692 | /* If op0 is an external or constant defs use a vector type of |
4693 | the same size as the output vector type. */ | |
ebfd146a | 4694 | if (!vectype_in) |
b690cc0f | 4695 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
4696 | if (vec_stmt) |
4697 | gcc_assert (vectype_in); | |
4698 | if (!vectype_in) | |
4699 | { | |
73fbfcad | 4700 | if (dump_enabled_p ()) |
4a00c761 | 4701 | { |
78c60e3d SS |
4702 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4703 | "no vectype for scalar type "); | |
4704 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 4705 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 4706 | } |
7d8930a0 IR |
4707 | |
4708 | return false; | |
4709 | } | |
ebfd146a | 4710 | |
e6f5c25d IE |
4711 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4712 | && !VECTOR_BOOLEAN_TYPE_P (vectype_in)) | |
4713 | { | |
4714 | if (dump_enabled_p ()) | |
4715 | { | |
4716 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
4717 | "can't convert between boolean and non " | |
4718 | "boolean vectors"); | |
4719 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
4720 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
4721 | } | |
4722 | ||
4723 | return false; | |
4724 | } | |
4725 | ||
b690cc0f RG |
4726 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
4727 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
062d5ccc | 4728 | if (known_eq (nunits_out, nunits_in)) |
ebfd146a | 4729 | modifier = NONE; |
062d5ccc RS |
4730 | else if (multiple_p (nunits_out, nunits_in)) |
4731 | modifier = NARROW; | |
ebfd146a | 4732 | else |
062d5ccc RS |
4733 | { |
4734 | gcc_checking_assert (multiple_p (nunits_in, nunits_out)); | |
4735 | modifier = WIDEN; | |
4736 | } | |
ebfd146a | 4737 | |
ff802fa1 IR |
4738 | /* Multiple types in SLP are handled by creating the appropriate number of |
4739 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4740 | case of SLP. */ | |
fce57248 | 4741 | if (slp_node) |
ebfd146a | 4742 | ncopies = 1; |
4a00c761 | 4743 | else if (modifier == NARROW) |
e8f142e2 | 4744 | ncopies = vect_get_num_copies (loop_vinfo, vectype_out); |
4a00c761 | 4745 | else |
e8f142e2 | 4746 | ncopies = vect_get_num_copies (loop_vinfo, vectype_in); |
b8698a0f | 4747 | |
ebfd146a IR |
4748 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
4749 | needs to be generated. */ | |
4750 | gcc_assert (ncopies >= 1); | |
4751 | ||
16d22000 RS |
4752 | bool found_mode = false; |
4753 | scalar_mode lhs_mode = SCALAR_TYPE_MODE (lhs_type); | |
4754 | scalar_mode rhs_mode = SCALAR_TYPE_MODE (rhs_type); | |
4755 | opt_scalar_mode rhs_mode_iter; | |
b397965c | 4756 | |
ebfd146a | 4757 | /* Supportable by target? */ |
4a00c761 | 4758 | switch (modifier) |
ebfd146a | 4759 | { |
4a00c761 JJ |
4760 | case NONE: |
4761 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4762 | return false; | |
4763 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
4764 | &decl1, &code1)) | |
4765 | break; | |
4766 | /* FALLTHRU */ | |
4767 | unsupported: | |
73fbfcad | 4768 | if (dump_enabled_p ()) |
78c60e3d | 4769 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4770 | "conversion not supported by target.\n"); |
ebfd146a | 4771 | return false; |
ebfd146a | 4772 | |
4a00c761 JJ |
4773 | case WIDEN: |
4774 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
4775 | &code1, &code2, &multi_step_cvt, |
4776 | &interm_types)) | |
4a00c761 JJ |
4777 | { |
4778 | /* Binary widening operation can only be supported directly by the | |
4779 | architecture. */ | |
4780 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
4781 | break; | |
4782 | } | |
4783 | ||
4784 | if (code != FLOAT_EXPR | |
b397965c | 4785 | || GET_MODE_SIZE (lhs_mode) <= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4786 | goto unsupported; |
4787 | ||
b397965c | 4788 | fltsz = GET_MODE_SIZE (lhs_mode); |
16d22000 | 4789 | FOR_EACH_2XWIDER_MODE (rhs_mode_iter, rhs_mode) |
4a00c761 | 4790 | { |
16d22000 | 4791 | rhs_mode = rhs_mode_iter.require (); |
c94843d2 RS |
4792 | if (GET_MODE_SIZE (rhs_mode) > fltsz) |
4793 | break; | |
4794 | ||
4a00c761 JJ |
4795 | cvt_type |
4796 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4797 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4798 | if (cvt_type == NULL_TREE) | |
4799 | goto unsupported; | |
4800 | ||
4801 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4802 | { | |
4803 | if (!supportable_convert_operation (code, vectype_out, | |
4804 | cvt_type, &decl1, &codecvt1)) | |
4805 | goto unsupported; | |
4806 | } | |
4807 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
4808 | cvt_type, &codecvt1, |
4809 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
4810 | &interm_types)) |
4811 | continue; | |
4812 | else | |
4813 | gcc_assert (multi_step_cvt == 0); | |
4814 | ||
4815 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
4816 | vectype_in, &code1, &code2, |
4817 | &multi_step_cvt, &interm_types)) | |
16d22000 RS |
4818 | { |
4819 | found_mode = true; | |
4820 | break; | |
4821 | } | |
4a00c761 JJ |
4822 | } |
4823 | ||
16d22000 | 4824 | if (!found_mode) |
4a00c761 JJ |
4825 | goto unsupported; |
4826 | ||
4827 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4828 | codecvt2 = ERROR_MARK; | |
4829 | else | |
4830 | { | |
4831 | multi_step_cvt++; | |
9771b263 | 4832 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
4833 | cvt_type = NULL_TREE; |
4834 | } | |
4835 | break; | |
4836 | ||
4837 | case NARROW: | |
4838 | gcc_assert (op_type == unary_op); | |
4839 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
4840 | &code1, &multi_step_cvt, | |
4841 | &interm_types)) | |
4842 | break; | |
4843 | ||
4844 | if (code != FIX_TRUNC_EXPR | |
b397965c | 4845 | || GET_MODE_SIZE (lhs_mode) >= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4846 | goto unsupported; |
4847 | ||
4a00c761 JJ |
4848 | cvt_type |
4849 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4850 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4851 | if (cvt_type == NULL_TREE) | |
4852 | goto unsupported; | |
4853 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
4854 | &decl1, &codecvt1)) | |
4855 | goto unsupported; | |
4856 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
4857 | &code1, &multi_step_cvt, | |
4858 | &interm_types)) | |
4859 | break; | |
4860 | goto unsupported; | |
4861 | ||
4862 | default: | |
4863 | gcc_unreachable (); | |
ebfd146a IR |
4864 | } |
4865 | ||
4866 | if (!vec_stmt) /* transformation not required. */ | |
4867 | { | |
73fbfcad | 4868 | if (dump_enabled_p ()) |
78c60e3d | 4869 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4870 | "=== vectorizable_conversion ===\n"); |
4a00c761 | 4871 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
4872 | { |
4873 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
68435eb2 RB |
4874 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, |
4875 | cost_vec); | |
8bd37302 | 4876 | } |
4a00c761 JJ |
4877 | else if (modifier == NARROW) |
4878 | { | |
4879 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
68435eb2 RB |
4880 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt, |
4881 | cost_vec); | |
4a00c761 JJ |
4882 | } |
4883 | else | |
4884 | { | |
4885 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
68435eb2 RB |
4886 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt, |
4887 | cost_vec); | |
4a00c761 | 4888 | } |
9771b263 | 4889 | interm_types.release (); |
ebfd146a IR |
4890 | return true; |
4891 | } | |
4892 | ||
67b8dbac | 4893 | /* Transform. */ |
73fbfcad | 4894 | if (dump_enabled_p ()) |
78c60e3d | 4895 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4896 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 4897 | |
4a00c761 JJ |
4898 | if (op_type == binary_op) |
4899 | { | |
4900 | if (CONSTANT_CLASS_P (op0)) | |
4901 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
4902 | else if (CONSTANT_CLASS_P (op1)) | |
4903 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
4904 | } | |
4905 | ||
4906 | /* In case of multi-step conversion, we first generate conversion operations | |
4907 | to the intermediate types, and then from that types to the final one. | |
4908 | We create vector destinations for the intermediate type (TYPES) received | |
4909 | from supportable_*_operation, and store them in the correct order | |
4910 | for future use in vect_create_vectorized_*_stmts (). */ | |
8c681247 | 4911 | auto_vec<tree> vec_dsts (multi_step_cvt + 1); |
82294ec1 JJ |
4912 | vec_dest = vect_create_destination_var (scalar_dest, |
4913 | (cvt_type && modifier == WIDEN) | |
4914 | ? cvt_type : vectype_out); | |
9771b263 | 4915 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4916 | |
4917 | if (multi_step_cvt) | |
4918 | { | |
9771b263 DN |
4919 | for (i = interm_types.length () - 1; |
4920 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
4921 | { |
4922 | vec_dest = vect_create_destination_var (scalar_dest, | |
4923 | intermediate_type); | |
9771b263 | 4924 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4925 | } |
4926 | } | |
ebfd146a | 4927 | |
4a00c761 | 4928 | if (cvt_type) |
82294ec1 JJ |
4929 | vec_dest = vect_create_destination_var (scalar_dest, |
4930 | modifier == WIDEN | |
4931 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
4932 | |
4933 | if (!slp_node) | |
4934 | { | |
30862efc | 4935 | if (modifier == WIDEN) |
4a00c761 | 4936 | { |
c3284718 | 4937 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 4938 | if (op_type == binary_op) |
9771b263 | 4939 | vec_oprnds1.create (1); |
4a00c761 | 4940 | } |
30862efc | 4941 | else if (modifier == NARROW) |
9771b263 DN |
4942 | vec_oprnds0.create ( |
4943 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
4944 | } |
4945 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 4946 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 4947 | |
4a00c761 | 4948 | last_oprnd = op0; |
ebfd146a IR |
4949 | prev_stmt_info = NULL; |
4950 | switch (modifier) | |
4951 | { | |
4952 | case NONE: | |
4953 | for (j = 0; j < ncopies; j++) | |
4954 | { | |
ebfd146a | 4955 | if (j == 0) |
306b0c92 | 4956 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node); |
ebfd146a IR |
4957 | else |
4958 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
4959 | ||
9771b263 | 4960 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4961 | { |
4962 | /* Arguments are ready, create the new vector stmt. */ | |
4963 | if (code1 == CALL_EXPR) | |
4964 | { | |
4965 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4966 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4967 | gimple_call_set_lhs (new_stmt, new_temp); | |
4968 | } | |
4969 | else | |
4970 | { | |
4971 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
0d0e4a03 | 4972 | new_stmt = gimple_build_assign (vec_dest, code1, vop0); |
4a00c761 JJ |
4973 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4974 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4975 | } | |
4976 | ||
4977 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4978 | if (slp_node) | |
9771b263 | 4979 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
225ce44b RB |
4980 | else |
4981 | { | |
4982 | if (!prev_stmt_info) | |
4983 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4984 | else | |
4985 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4986 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4987 | } | |
4a00c761 | 4988 | } |
ebfd146a IR |
4989 | } |
4990 | break; | |
4991 | ||
4992 | case WIDEN: | |
4993 | /* In case the vectorization factor (VF) is bigger than the number | |
4994 | of elements that we can fit in a vectype (nunits), we have to | |
4995 | generate more than one vector stmt - i.e - we need to "unroll" | |
4996 | the vector stmt by a factor VF/nunits. */ | |
4997 | for (j = 0; j < ncopies; j++) | |
4998 | { | |
4a00c761 | 4999 | /* Handle uses. */ |
ebfd146a | 5000 | if (j == 0) |
4a00c761 JJ |
5001 | { |
5002 | if (slp_node) | |
5003 | { | |
5004 | if (code == WIDEN_LSHIFT_EXPR) | |
5005 | { | |
5006 | unsigned int k; | |
ebfd146a | 5007 | |
4a00c761 JJ |
5008 | vec_oprnd1 = op1; |
5009 | /* Store vec_oprnd1 for every vector stmt to be created | |
5010 | for SLP_NODE. We check during the analysis that all | |
5011 | the shift arguments are the same. */ | |
5012 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 5013 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
5014 | |
5015 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5016 | slp_node); |
4a00c761 JJ |
5017 | } |
5018 | else | |
5019 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
306b0c92 | 5020 | &vec_oprnds1, slp_node); |
4a00c761 JJ |
5021 | } |
5022 | else | |
5023 | { | |
81c40241 | 5024 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 5025 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
5026 | if (op_type == binary_op) |
5027 | { | |
5028 | if (code == WIDEN_LSHIFT_EXPR) | |
5029 | vec_oprnd1 = op1; | |
5030 | else | |
81c40241 | 5031 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 5032 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
5033 | } |
5034 | } | |
5035 | } | |
ebfd146a | 5036 | else |
4a00c761 JJ |
5037 | { |
5038 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
5039 | vec_oprnds0.truncate (0); |
5040 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
5041 | if (op_type == binary_op) |
5042 | { | |
5043 | if (code == WIDEN_LSHIFT_EXPR) | |
5044 | vec_oprnd1 = op1; | |
5045 | else | |
5046 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
5047 | vec_oprnd1); | |
9771b263 DN |
5048 | vec_oprnds1.truncate (0); |
5049 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
5050 | } |
5051 | } | |
ebfd146a | 5052 | |
4a00c761 JJ |
5053 | /* Arguments are ready. Create the new vector stmts. */ |
5054 | for (i = multi_step_cvt; i >= 0; i--) | |
5055 | { | |
9771b263 | 5056 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
5057 | enum tree_code c1 = code1, c2 = code2; |
5058 | if (i == 0 && codecvt2 != ERROR_MARK) | |
5059 | { | |
5060 | c1 = codecvt1; | |
5061 | c2 = codecvt2; | |
5062 | } | |
5063 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
5064 | &vec_oprnds1, | |
5065 | stmt, this_dest, gsi, | |
5066 | c1, c2, decl1, decl2, | |
5067 | op_type); | |
5068 | } | |
5069 | ||
9771b263 | 5070 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
5071 | { |
5072 | if (cvt_type) | |
5073 | { | |
5074 | if (codecvt1 == CALL_EXPR) | |
5075 | { | |
5076 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
5077 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5078 | gimple_call_set_lhs (new_stmt, new_temp); | |
5079 | } | |
5080 | else | |
5081 | { | |
5082 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 5083 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
5084 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
5085 | vop0); | |
4a00c761 JJ |
5086 | } |
5087 | ||
5088 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5089 | } | |
5090 | else | |
5091 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
5092 | ||
5093 | if (slp_node) | |
9771b263 | 5094 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 5095 | else |
c689ce1e RB |
5096 | { |
5097 | if (!prev_stmt_info) | |
5098 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
5099 | else | |
5100 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5101 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5102 | } | |
4a00c761 | 5103 | } |
ebfd146a | 5104 | } |
4a00c761 JJ |
5105 | |
5106 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
5107 | break; |
5108 | ||
5109 | case NARROW: | |
5110 | /* In case the vectorization factor (VF) is bigger than the number | |
5111 | of elements that we can fit in a vectype (nunits), we have to | |
5112 | generate more than one vector stmt - i.e - we need to "unroll" | |
5113 | the vector stmt by a factor VF/nunits. */ | |
5114 | for (j = 0; j < ncopies; j++) | |
5115 | { | |
5116 | /* Handle uses. */ | |
4a00c761 JJ |
5117 | if (slp_node) |
5118 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5119 | slp_node); |
ebfd146a IR |
5120 | else |
5121 | { | |
9771b263 | 5122 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
5123 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
5124 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
5125 | } |
5126 | ||
4a00c761 JJ |
5127 | /* Arguments are ready. Create the new vector stmts. */ |
5128 | if (cvt_type) | |
9771b263 | 5129 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
5130 | { |
5131 | if (codecvt1 == CALL_EXPR) | |
5132 | { | |
5133 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
5134 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5135 | gimple_call_set_lhs (new_stmt, new_temp); | |
5136 | } | |
5137 | else | |
5138 | { | |
5139 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 5140 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
5141 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
5142 | vop0); | |
4a00c761 | 5143 | } |
ebfd146a | 5144 | |
4a00c761 | 5145 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 5146 | vec_oprnds0[i] = new_temp; |
4a00c761 | 5147 | } |
ebfd146a | 5148 | |
4a00c761 JJ |
5149 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
5150 | stmt, vec_dsts, gsi, | |
5151 | slp_node, code1, | |
5152 | &prev_stmt_info); | |
ebfd146a IR |
5153 | } |
5154 | ||
5155 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 5156 | break; |
ebfd146a IR |
5157 | } |
5158 | ||
9771b263 DN |
5159 | vec_oprnds0.release (); |
5160 | vec_oprnds1.release (); | |
9771b263 | 5161 | interm_types.release (); |
ebfd146a IR |
5162 | |
5163 | return true; | |
5164 | } | |
ff802fa1 IR |
5165 | |
5166 | ||
ebfd146a IR |
5167 | /* Function vectorizable_assignment. |
5168 | ||
b8698a0f L |
5169 | Check if STMT performs an assignment (copy) that can be vectorized. |
5170 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
5171 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5172 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5173 | ||
5174 | static bool | |
355fe088 | 5175 | vectorizable_assignment (gimple *stmt, gimple_stmt_iterator *gsi, |
68435eb2 RB |
5176 | gimple **vec_stmt, slp_tree slp_node, |
5177 | stmt_vector_for_cost *cost_vec) | |
ebfd146a IR |
5178 | { |
5179 | tree vec_dest; | |
5180 | tree scalar_dest; | |
5181 | tree op; | |
5182 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a IR |
5183 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
5184 | tree new_temp; | |
355fe088 | 5185 | gimple *def_stmt; |
4fc5ebf1 JG |
5186 | enum vect_def_type dt[1] = {vect_unknown_def_type}; |
5187 | int ndts = 1; | |
ebfd146a | 5188 | int ncopies; |
f18b55bd | 5189 | int i, j; |
6e1aa848 | 5190 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 5191 | tree vop; |
a70d6342 | 5192 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5193 | vec_info *vinfo = stmt_info->vinfo; |
355fe088 | 5194 | gimple *new_stmt = NULL; |
f18b55bd | 5195 | stmt_vec_info prev_stmt_info = NULL; |
fde9c428 RG |
5196 | enum tree_code code; |
5197 | tree vectype_in; | |
ebfd146a | 5198 | |
a70d6342 | 5199 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5200 | return false; |
5201 | ||
66c16fd9 RB |
5202 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5203 | && ! vec_stmt) | |
ebfd146a IR |
5204 | return false; |
5205 | ||
5206 | /* Is vectorizable assignment? */ | |
5207 | if (!is_gimple_assign (stmt)) | |
5208 | return false; | |
5209 | ||
5210 | scalar_dest = gimple_assign_lhs (stmt); | |
5211 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
5212 | return false; | |
5213 | ||
fde9c428 | 5214 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 5215 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
5216 | || code == PAREN_EXPR |
5217 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
5218 | op = gimple_assign_rhs1 (stmt); |
5219 | else | |
5220 | return false; | |
5221 | ||
7b7ec6c5 RG |
5222 | if (code == VIEW_CONVERT_EXPR) |
5223 | op = TREE_OPERAND (op, 0); | |
5224 | ||
465c8c19 | 5225 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
928686b1 | 5226 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 JJ |
5227 | |
5228 | /* Multiple types in SLP are handled by creating the appropriate number of | |
5229 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5230 | case of SLP. */ | |
fce57248 | 5231 | if (slp_node) |
465c8c19 JJ |
5232 | ncopies = 1; |
5233 | else | |
e8f142e2 | 5234 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
5235 | |
5236 | gcc_assert (ncopies >= 1); | |
5237 | ||
81c40241 | 5238 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[0], &vectype_in)) |
ebfd146a | 5239 | { |
73fbfcad | 5240 | if (dump_enabled_p ()) |
78c60e3d | 5241 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5242 | "use not simple.\n"); |
ebfd146a IR |
5243 | return false; |
5244 | } | |
5245 | ||
fde9c428 RG |
5246 | /* We can handle NOP_EXPR conversions that do not change the number |
5247 | of elements or the vector size. */ | |
7b7ec6c5 RG |
5248 | if ((CONVERT_EXPR_CODE_P (code) |
5249 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 | 5250 | && (!vectype_in |
928686b1 | 5251 | || maybe_ne (TYPE_VECTOR_SUBPARTS (vectype_in), nunits) |
cf098191 RS |
5252 | || maybe_ne (GET_MODE_SIZE (TYPE_MODE (vectype)), |
5253 | GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
fde9c428 RG |
5254 | return false; |
5255 | ||
7b7b1813 RG |
5256 | /* We do not handle bit-precision changes. */ |
5257 | if ((CONVERT_EXPR_CODE_P (code) | |
5258 | || code == VIEW_CONVERT_EXPR) | |
5259 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
2be65d9e RS |
5260 | && (!type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
5261 | || !type_has_mode_precision_p (TREE_TYPE (op))) | |
7b7b1813 RG |
5262 | /* But a conversion that does not change the bit-pattern is ok. */ |
5263 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
5264 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2dab46d5 IE |
5265 | && TYPE_UNSIGNED (TREE_TYPE (op))) |
5266 | /* Conversion between boolean types of different sizes is | |
5267 | a simple assignment in case their vectypes are same | |
5268 | boolean vectors. */ | |
5269 | && (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
5270 | || !VECTOR_BOOLEAN_TYPE_P (vectype_in))) | |
7b7b1813 | 5271 | { |
73fbfcad | 5272 | if (dump_enabled_p ()) |
78c60e3d SS |
5273 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5274 | "type conversion to/from bit-precision " | |
e645e942 | 5275 | "unsupported.\n"); |
7b7b1813 RG |
5276 | return false; |
5277 | } | |
5278 | ||
ebfd146a IR |
5279 | if (!vec_stmt) /* transformation not required. */ |
5280 | { | |
5281 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 5282 | if (dump_enabled_p ()) |
78c60e3d | 5283 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5284 | "=== vectorizable_assignment ===\n"); |
68435eb2 | 5285 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, cost_vec); |
ebfd146a IR |
5286 | return true; |
5287 | } | |
5288 | ||
67b8dbac | 5289 | /* Transform. */ |
73fbfcad | 5290 | if (dump_enabled_p ()) |
e645e942 | 5291 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
5292 | |
5293 | /* Handle def. */ | |
5294 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5295 | ||
5296 | /* Handle use. */ | |
f18b55bd | 5297 | for (j = 0; j < ncopies; j++) |
ebfd146a | 5298 | { |
f18b55bd IR |
5299 | /* Handle uses. */ |
5300 | if (j == 0) | |
306b0c92 | 5301 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); |
f18b55bd IR |
5302 | else |
5303 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
5304 | ||
5305 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 5306 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 5307 | { |
7b7ec6c5 RG |
5308 | if (CONVERT_EXPR_CODE_P (code) |
5309 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 5310 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
5311 | new_stmt = gimple_build_assign (vec_dest, vop); |
5312 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5313 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5314 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5315 | if (slp_node) | |
9771b263 | 5316 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 5317 | } |
ebfd146a IR |
5318 | |
5319 | if (slp_node) | |
f18b55bd IR |
5320 | continue; |
5321 | ||
5322 | if (j == 0) | |
5323 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5324 | else | |
5325 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5326 | ||
5327 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5328 | } | |
b8698a0f | 5329 | |
9771b263 | 5330 | vec_oprnds.release (); |
ebfd146a IR |
5331 | return true; |
5332 | } | |
5333 | ||
9dc3f7de | 5334 | |
1107f3ae IR |
5335 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
5336 | either as shift by a scalar or by a vector. */ | |
5337 | ||
5338 | bool | |
5339 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
5340 | { | |
5341 | ||
ef4bddc2 | 5342 | machine_mode vec_mode; |
1107f3ae IR |
5343 | optab optab; |
5344 | int icode; | |
5345 | tree vectype; | |
5346 | ||
5347 | vectype = get_vectype_for_scalar_type (scalar_type); | |
5348 | if (!vectype) | |
5349 | return false; | |
5350 | ||
5351 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
5352 | if (!optab | |
5353 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
5354 | { | |
5355 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
5356 | if (!optab | |
5357 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
5358 | == CODE_FOR_nothing)) | |
5359 | return false; | |
5360 | } | |
5361 | ||
5362 | vec_mode = TYPE_MODE (vectype); | |
5363 | icode = (int) optab_handler (optab, vec_mode); | |
5364 | if (icode == CODE_FOR_nothing) | |
5365 | return false; | |
5366 | ||
5367 | return true; | |
5368 | } | |
5369 | ||
5370 | ||
9dc3f7de IR |
5371 | /* Function vectorizable_shift. |
5372 | ||
5373 | Check if STMT performs a shift operation that can be vectorized. | |
5374 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
5375 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
5376 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5377 | ||
5378 | static bool | |
355fe088 | 5379 | vectorizable_shift (gimple *stmt, gimple_stmt_iterator *gsi, |
68435eb2 RB |
5380 | gimple **vec_stmt, slp_tree slp_node, |
5381 | stmt_vector_for_cost *cost_vec) | |
9dc3f7de IR |
5382 | { |
5383 | tree vec_dest; | |
5384 | tree scalar_dest; | |
5385 | tree op0, op1 = NULL; | |
5386 | tree vec_oprnd1 = NULL_TREE; | |
5387 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5388 | tree vectype; | |
5389 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
5390 | enum tree_code code; | |
ef4bddc2 | 5391 | machine_mode vec_mode; |
9dc3f7de IR |
5392 | tree new_temp; |
5393 | optab optab; | |
5394 | int icode; | |
ef4bddc2 | 5395 | machine_mode optab_op2_mode; |
355fe088 | 5396 | gimple *def_stmt; |
9dc3f7de | 5397 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 5398 | int ndts = 2; |
355fe088 | 5399 | gimple *new_stmt = NULL; |
9dc3f7de | 5400 | stmt_vec_info prev_stmt_info; |
928686b1 RS |
5401 | poly_uint64 nunits_in; |
5402 | poly_uint64 nunits_out; | |
9dc3f7de | 5403 | tree vectype_out; |
cede2577 | 5404 | tree op1_vectype; |
9dc3f7de IR |
5405 | int ncopies; |
5406 | int j, i; | |
6e1aa848 DN |
5407 | vec<tree> vec_oprnds0 = vNULL; |
5408 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
5409 | tree vop0, vop1; |
5410 | unsigned int k; | |
49eab32e | 5411 | bool scalar_shift_arg = true; |
9dc3f7de | 5412 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5413 | vec_info *vinfo = stmt_info->vinfo; |
9dc3f7de IR |
5414 | |
5415 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
5416 | return false; | |
5417 | ||
66c16fd9 RB |
5418 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5419 | && ! vec_stmt) | |
9dc3f7de IR |
5420 | return false; |
5421 | ||
5422 | /* Is STMT a vectorizable binary/unary operation? */ | |
5423 | if (!is_gimple_assign (stmt)) | |
5424 | return false; | |
5425 | ||
5426 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
5427 | return false; | |
5428 | ||
5429 | code = gimple_assign_rhs_code (stmt); | |
5430 | ||
5431 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
5432 | || code == RROTATE_EXPR)) | |
5433 | return false; | |
5434 | ||
5435 | scalar_dest = gimple_assign_lhs (stmt); | |
5436 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2be65d9e | 5437 | if (!type_has_mode_precision_p (TREE_TYPE (scalar_dest))) |
7b7b1813 | 5438 | { |
73fbfcad | 5439 | if (dump_enabled_p ()) |
78c60e3d | 5440 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5441 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
5442 | return false; |
5443 | } | |
9dc3f7de IR |
5444 | |
5445 | op0 = gimple_assign_rhs1 (stmt); | |
81c40241 | 5446 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
9dc3f7de | 5447 | { |
73fbfcad | 5448 | if (dump_enabled_p ()) |
78c60e3d | 5449 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5450 | "use not simple.\n"); |
9dc3f7de IR |
5451 | return false; |
5452 | } | |
5453 | /* If op0 is an external or constant def use a vector type with | |
5454 | the same size as the output vector type. */ | |
5455 | if (!vectype) | |
5456 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
5457 | if (vec_stmt) | |
5458 | gcc_assert (vectype); | |
5459 | if (!vectype) | |
5460 | { | |
73fbfcad | 5461 | if (dump_enabled_p ()) |
78c60e3d | 5462 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5463 | "no vectype for scalar type\n"); |
9dc3f7de IR |
5464 | return false; |
5465 | } | |
5466 | ||
5467 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
5468 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
928686b1 | 5469 | if (maybe_ne (nunits_out, nunits_in)) |
9dc3f7de IR |
5470 | return false; |
5471 | ||
5472 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 5473 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &op1_vectype)) |
9dc3f7de | 5474 | { |
73fbfcad | 5475 | if (dump_enabled_p ()) |
78c60e3d | 5476 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5477 | "use not simple.\n"); |
9dc3f7de IR |
5478 | return false; |
5479 | } | |
5480 | ||
9dc3f7de IR |
5481 | /* Multiple types in SLP are handled by creating the appropriate number of |
5482 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5483 | case of SLP. */ | |
fce57248 | 5484 | if (slp_node) |
9dc3f7de IR |
5485 | ncopies = 1; |
5486 | else | |
e8f142e2 | 5487 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
9dc3f7de IR |
5488 | |
5489 | gcc_assert (ncopies >= 1); | |
5490 | ||
5491 | /* Determine whether the shift amount is a vector, or scalar. If the | |
5492 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
5493 | ||
dbfa87aa YR |
5494 | if ((dt[1] == vect_internal_def |
5495 | || dt[1] == vect_induction_def) | |
5496 | && !slp_node) | |
49eab32e JJ |
5497 | scalar_shift_arg = false; |
5498 | else if (dt[1] == vect_constant_def | |
5499 | || dt[1] == vect_external_def | |
5500 | || dt[1] == vect_internal_def) | |
5501 | { | |
5502 | /* In SLP, need to check whether the shift count is the same, | |
5503 | in loops if it is a constant or invariant, it is always | |
5504 | a scalar shift. */ | |
5505 | if (slp_node) | |
5506 | { | |
355fe088 TS |
5507 | vec<gimple *> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
5508 | gimple *slpstmt; | |
49eab32e | 5509 | |
9771b263 | 5510 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
5511 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
5512 | scalar_shift_arg = false; | |
5513 | } | |
60d393e8 RB |
5514 | |
5515 | /* If the shift amount is computed by a pattern stmt we cannot | |
5516 | use the scalar amount directly thus give up and use a vector | |
5517 | shift. */ | |
5518 | if (dt[1] == vect_internal_def) | |
5519 | { | |
5520 | gimple *def = SSA_NAME_DEF_STMT (op1); | |
5521 | if (is_pattern_stmt_p (vinfo_for_stmt (def))) | |
5522 | scalar_shift_arg = false; | |
5523 | } | |
49eab32e JJ |
5524 | } |
5525 | else | |
5526 | { | |
73fbfcad | 5527 | if (dump_enabled_p ()) |
78c60e3d | 5528 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5529 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
5530 | return false; |
5531 | } | |
5532 | ||
9dc3f7de | 5533 | /* Vector shifted by vector. */ |
49eab32e | 5534 | if (!scalar_shift_arg) |
9dc3f7de IR |
5535 | { |
5536 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 5537 | if (dump_enabled_p ()) |
78c60e3d | 5538 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5539 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 5540 | |
aa948027 JJ |
5541 | if (!op1_vectype) |
5542 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
5543 | if (op1_vectype == NULL_TREE | |
5544 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 5545 | { |
73fbfcad | 5546 | if (dump_enabled_p ()) |
78c60e3d SS |
5547 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5548 | "unusable type for last operand in" | |
e645e942 | 5549 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
5550 | return false; |
5551 | } | |
9dc3f7de IR |
5552 | } |
5553 | /* See if the machine has a vector shifted by scalar insn and if not | |
5554 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 5555 | else |
9dc3f7de IR |
5556 | { |
5557 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
5558 | if (optab | |
5559 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
5560 | { | |
73fbfcad | 5561 | if (dump_enabled_p ()) |
78c60e3d | 5562 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5563 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
5564 | } |
5565 | else | |
5566 | { | |
5567 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
5568 | if (optab | |
5569 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
5570 | != CODE_FOR_nothing)) | |
5571 | { | |
49eab32e JJ |
5572 | scalar_shift_arg = false; |
5573 | ||
73fbfcad | 5574 | if (dump_enabled_p ()) |
78c60e3d | 5575 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5576 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
5577 | |
5578 | /* Unlike the other binary operators, shifts/rotates have | |
5579 | the rhs being int, instead of the same type as the lhs, | |
5580 | so make sure the scalar is the right type if we are | |
aa948027 | 5581 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
5582 | if (dt[1] == vect_constant_def) |
5583 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
5584 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
5585 | TREE_TYPE (op1))) | |
5586 | { | |
5587 | if (slp_node | |
5588 | && TYPE_MODE (TREE_TYPE (vectype)) | |
5589 | != TYPE_MODE (TREE_TYPE (op1))) | |
5590 | { | |
73fbfcad | 5591 | if (dump_enabled_p ()) |
78c60e3d SS |
5592 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5593 | "unusable type for last operand in" | |
e645e942 | 5594 | " vector/vector shift/rotate.\n"); |
21c0a521 | 5595 | return false; |
aa948027 JJ |
5596 | } |
5597 | if (vec_stmt && !slp_node) | |
5598 | { | |
5599 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
5600 | op1 = vect_init_vector (stmt, op1, | |
5601 | TREE_TYPE (vectype), NULL); | |
5602 | } | |
5603 | } | |
9dc3f7de IR |
5604 | } |
5605 | } | |
5606 | } | |
9dc3f7de IR |
5607 | |
5608 | /* Supportable by target? */ | |
5609 | if (!optab) | |
5610 | { | |
73fbfcad | 5611 | if (dump_enabled_p ()) |
78c60e3d | 5612 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5613 | "no optab.\n"); |
9dc3f7de IR |
5614 | return false; |
5615 | } | |
5616 | vec_mode = TYPE_MODE (vectype); | |
5617 | icode = (int) optab_handler (optab, vec_mode); | |
5618 | if (icode == CODE_FOR_nothing) | |
5619 | { | |
73fbfcad | 5620 | if (dump_enabled_p ()) |
78c60e3d | 5621 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5622 | "op not supported by target.\n"); |
9dc3f7de | 5623 | /* Check only during analysis. */ |
cf098191 | 5624 | if (maybe_ne (GET_MODE_SIZE (vec_mode), UNITS_PER_WORD) |
ca09abcb RS |
5625 | || (!vec_stmt |
5626 | && !vect_worthwhile_without_simd_p (vinfo, code))) | |
9dc3f7de | 5627 | return false; |
73fbfcad | 5628 | if (dump_enabled_p ()) |
e645e942 TJ |
5629 | dump_printf_loc (MSG_NOTE, vect_location, |
5630 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
5631 | } |
5632 | ||
5633 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
ca09abcb RS |
5634 | if (!vec_stmt |
5635 | && !VECTOR_MODE_P (TYPE_MODE (vectype)) | |
5636 | && !vect_worthwhile_without_simd_p (vinfo, code)) | |
9dc3f7de | 5637 | { |
73fbfcad | 5638 | if (dump_enabled_p ()) |
78c60e3d | 5639 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5640 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
5641 | return false; |
5642 | } | |
5643 | ||
5644 | if (!vec_stmt) /* transformation not required. */ | |
5645 | { | |
5646 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 5647 | if (dump_enabled_p ()) |
e645e942 TJ |
5648 | dump_printf_loc (MSG_NOTE, vect_location, |
5649 | "=== vectorizable_shift ===\n"); | |
68435eb2 | 5650 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, cost_vec); |
9dc3f7de IR |
5651 | return true; |
5652 | } | |
5653 | ||
67b8dbac | 5654 | /* Transform. */ |
9dc3f7de | 5655 | |
73fbfcad | 5656 | if (dump_enabled_p ()) |
78c60e3d | 5657 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5658 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
5659 | |
5660 | /* Handle def. */ | |
5661 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5662 | ||
9dc3f7de IR |
5663 | prev_stmt_info = NULL; |
5664 | for (j = 0; j < ncopies; j++) | |
5665 | { | |
5666 | /* Handle uses. */ | |
5667 | if (j == 0) | |
5668 | { | |
5669 | if (scalar_shift_arg) | |
5670 | { | |
5671 | /* Vector shl and shr insn patterns can be defined with scalar | |
5672 | operand 2 (shift operand). In this case, use constant or loop | |
5673 | invariant op1 directly, without extending it to vector mode | |
5674 | first. */ | |
5675 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
5676 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
5677 | { | |
73fbfcad | 5678 | if (dump_enabled_p ()) |
78c60e3d | 5679 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5680 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 5681 | vec_oprnd1 = op1; |
8930f723 | 5682 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 5683 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5684 | if (slp_node) |
5685 | { | |
5686 | /* Store vec_oprnd1 for every vector stmt to be created | |
5687 | for SLP_NODE. We check during the analysis that all | |
5688 | the shift arguments are the same. | |
5689 | TODO: Allow different constants for different vector | |
5690 | stmts generated for an SLP instance. */ | |
5691 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 5692 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5693 | } |
5694 | } | |
5695 | } | |
5696 | ||
5697 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
5698 | (a special case for certain kind of vector shifts); otherwise, | |
5699 | operand 1 should be of a vector type (the usual case). */ | |
5700 | if (vec_oprnd1) | |
5701 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5702 | slp_node); |
9dc3f7de IR |
5703 | else |
5704 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
306b0c92 | 5705 | slp_node); |
9dc3f7de IR |
5706 | } |
5707 | else | |
5708 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5709 | ||
5710 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 5711 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 5712 | { |
9771b263 | 5713 | vop1 = vec_oprnds1[i]; |
0d0e4a03 | 5714 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
9dc3f7de IR |
5715 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5716 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5717 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5718 | if (slp_node) | |
9771b263 | 5719 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
5720 | } |
5721 | ||
5722 | if (slp_node) | |
5723 | continue; | |
5724 | ||
5725 | if (j == 0) | |
5726 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5727 | else | |
5728 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5729 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5730 | } | |
5731 | ||
9771b263 DN |
5732 | vec_oprnds0.release (); |
5733 | vec_oprnds1.release (); | |
9dc3f7de IR |
5734 | |
5735 | return true; | |
5736 | } | |
5737 | ||
5738 | ||
ebfd146a IR |
5739 | /* Function vectorizable_operation. |
5740 | ||
16949072 RG |
5741 | Check if STMT performs a binary, unary or ternary operation that can |
5742 | be vectorized. | |
b8698a0f | 5743 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
5744 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5745 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5746 | ||
5747 | static bool | |
355fe088 | 5748 | vectorizable_operation (gimple *stmt, gimple_stmt_iterator *gsi, |
68435eb2 RB |
5749 | gimple **vec_stmt, slp_tree slp_node, |
5750 | stmt_vector_for_cost *cost_vec) | |
ebfd146a | 5751 | { |
00f07b86 | 5752 | tree vec_dest; |
ebfd146a | 5753 | tree scalar_dest; |
16949072 | 5754 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 5755 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 5756 | tree vectype; |
ebfd146a | 5757 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
0eb952ea | 5758 | enum tree_code code, orig_code; |
ef4bddc2 | 5759 | machine_mode vec_mode; |
ebfd146a IR |
5760 | tree new_temp; |
5761 | int op_type; | |
00f07b86 | 5762 | optab optab; |
523ba738 | 5763 | bool target_support_p; |
355fe088 | 5764 | gimple *def_stmt; |
16949072 RG |
5765 | enum vect_def_type dt[3] |
5766 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 5767 | int ndts = 3; |
355fe088 | 5768 | gimple *new_stmt = NULL; |
ebfd146a | 5769 | stmt_vec_info prev_stmt_info; |
928686b1 RS |
5770 | poly_uint64 nunits_in; |
5771 | poly_uint64 nunits_out; | |
ebfd146a IR |
5772 | tree vectype_out; |
5773 | int ncopies; | |
5774 | int j, i; | |
6e1aa848 DN |
5775 | vec<tree> vec_oprnds0 = vNULL; |
5776 | vec<tree> vec_oprnds1 = vNULL; | |
5777 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 5778 | tree vop0, vop1, vop2; |
a70d6342 | 5779 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5780 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 5781 | |
a70d6342 | 5782 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5783 | return false; |
5784 | ||
66c16fd9 RB |
5785 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5786 | && ! vec_stmt) | |
ebfd146a IR |
5787 | return false; |
5788 | ||
5789 | /* Is STMT a vectorizable binary/unary operation? */ | |
5790 | if (!is_gimple_assign (stmt)) | |
5791 | return false; | |
5792 | ||
5793 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
5794 | return false; | |
5795 | ||
0eb952ea | 5796 | orig_code = code = gimple_assign_rhs_code (stmt); |
ebfd146a | 5797 | |
1af4ebf5 MG |
5798 | /* For pointer addition and subtraction, we should use the normal |
5799 | plus and minus for the vector operation. */ | |
ebfd146a IR |
5800 | if (code == POINTER_PLUS_EXPR) |
5801 | code = PLUS_EXPR; | |
1af4ebf5 MG |
5802 | if (code == POINTER_DIFF_EXPR) |
5803 | code = MINUS_EXPR; | |
ebfd146a IR |
5804 | |
5805 | /* Support only unary or binary operations. */ | |
5806 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 5807 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 5808 | { |
73fbfcad | 5809 | if (dump_enabled_p ()) |
78c60e3d | 5810 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5811 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 5812 | op_type); |
ebfd146a IR |
5813 | return false; |
5814 | } | |
5815 | ||
b690cc0f RG |
5816 | scalar_dest = gimple_assign_lhs (stmt); |
5817 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
5818 | ||
7b7b1813 RG |
5819 | /* Most operations cannot handle bit-precision types without extra |
5820 | truncations. */ | |
045c1278 | 5821 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
2be65d9e | 5822 | && !type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
7b7b1813 RG |
5823 | /* Exception are bitwise binary operations. */ |
5824 | && code != BIT_IOR_EXPR | |
5825 | && code != BIT_XOR_EXPR | |
5826 | && code != BIT_AND_EXPR) | |
5827 | { | |
73fbfcad | 5828 | if (dump_enabled_p ()) |
78c60e3d | 5829 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5830 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
5831 | return false; |
5832 | } | |
5833 | ||
ebfd146a | 5834 | op0 = gimple_assign_rhs1 (stmt); |
81c40241 | 5835 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
ebfd146a | 5836 | { |
73fbfcad | 5837 | if (dump_enabled_p ()) |
78c60e3d | 5838 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5839 | "use not simple.\n"); |
ebfd146a IR |
5840 | return false; |
5841 | } | |
b690cc0f RG |
5842 | /* If op0 is an external or constant def use a vector type with |
5843 | the same size as the output vector type. */ | |
5844 | if (!vectype) | |
b036c6c5 IE |
5845 | { |
5846 | /* For boolean type we cannot determine vectype by | |
5847 | invariant value (don't know whether it is a vector | |
5848 | of booleans or vector of integers). We use output | |
5849 | vectype because operations on boolean don't change | |
5850 | type. */ | |
2568d8a1 | 5851 | if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op0))) |
b036c6c5 | 5852 | { |
2568d8a1 | 5853 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (scalar_dest))) |
b036c6c5 IE |
5854 | { |
5855 | if (dump_enabled_p ()) | |
5856 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5857 | "not supported operation on bool value.\n"); | |
5858 | return false; | |
5859 | } | |
5860 | vectype = vectype_out; | |
5861 | } | |
5862 | else | |
5863 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
5864 | } | |
7d8930a0 IR |
5865 | if (vec_stmt) |
5866 | gcc_assert (vectype); | |
5867 | if (!vectype) | |
5868 | { | |
73fbfcad | 5869 | if (dump_enabled_p ()) |
7d8930a0 | 5870 | { |
78c60e3d SS |
5871 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5872 | "no vectype for scalar type "); | |
5873 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
5874 | TREE_TYPE (op0)); | |
e645e942 | 5875 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
5876 | } |
5877 | ||
5878 | return false; | |
5879 | } | |
b690cc0f RG |
5880 | |
5881 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
5882 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
928686b1 | 5883 | if (maybe_ne (nunits_out, nunits_in)) |
b690cc0f | 5884 | return false; |
ebfd146a | 5885 | |
16949072 | 5886 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
5887 | { |
5888 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 5889 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1])) |
ebfd146a | 5890 | { |
73fbfcad | 5891 | if (dump_enabled_p ()) |
78c60e3d | 5892 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5893 | "use not simple.\n"); |
ebfd146a IR |
5894 | return false; |
5895 | } | |
5896 | } | |
16949072 RG |
5897 | if (op_type == ternary_op) |
5898 | { | |
5899 | op2 = gimple_assign_rhs3 (stmt); | |
81c40241 | 5900 | if (!vect_is_simple_use (op2, vinfo, &def_stmt, &dt[2])) |
16949072 | 5901 | { |
73fbfcad | 5902 | if (dump_enabled_p ()) |
78c60e3d | 5903 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5904 | "use not simple.\n"); |
16949072 RG |
5905 | return false; |
5906 | } | |
5907 | } | |
ebfd146a | 5908 | |
b690cc0f | 5909 | /* Multiple types in SLP are handled by creating the appropriate number of |
ff802fa1 | 5910 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 5911 | case of SLP. */ |
fce57248 | 5912 | if (slp_node) |
b690cc0f RG |
5913 | ncopies = 1; |
5914 | else | |
e8f142e2 | 5915 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
b690cc0f RG |
5916 | |
5917 | gcc_assert (ncopies >= 1); | |
5918 | ||
9dc3f7de | 5919 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
5920 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
5921 | || code == RROTATE_EXPR) | |
9dc3f7de | 5922 | return false; |
ebfd146a | 5923 | |
ebfd146a | 5924 | /* Supportable by target? */ |
00f07b86 RH |
5925 | |
5926 | vec_mode = TYPE_MODE (vectype); | |
5927 | if (code == MULT_HIGHPART_EXPR) | |
523ba738 | 5928 | target_support_p = can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype)); |
00f07b86 RH |
5929 | else |
5930 | { | |
5931 | optab = optab_for_tree_code (code, vectype, optab_default); | |
5932 | if (!optab) | |
5deb57cb | 5933 | { |
73fbfcad | 5934 | if (dump_enabled_p ()) |
78c60e3d | 5935 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5936 | "no optab.\n"); |
00f07b86 | 5937 | return false; |
5deb57cb | 5938 | } |
523ba738 RS |
5939 | target_support_p = (optab_handler (optab, vec_mode) |
5940 | != CODE_FOR_nothing); | |
5deb57cb JJ |
5941 | } |
5942 | ||
523ba738 | 5943 | if (!target_support_p) |
ebfd146a | 5944 | { |
73fbfcad | 5945 | if (dump_enabled_p ()) |
78c60e3d | 5946 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5947 | "op not supported by target.\n"); |
ebfd146a | 5948 | /* Check only during analysis. */ |
cf098191 | 5949 | if (maybe_ne (GET_MODE_SIZE (vec_mode), UNITS_PER_WORD) |
ca09abcb | 5950 | || (!vec_stmt && !vect_worthwhile_without_simd_p (vinfo, code))) |
ebfd146a | 5951 | return false; |
73fbfcad | 5952 | if (dump_enabled_p ()) |
e645e942 TJ |
5953 | dump_printf_loc (MSG_NOTE, vect_location, |
5954 | "proceeding using word mode.\n"); | |
383d9c83 IR |
5955 | } |
5956 | ||
4a00c761 | 5957 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
5958 | if (!VECTOR_MODE_P (vec_mode) |
5959 | && !vec_stmt | |
ca09abcb | 5960 | && !vect_worthwhile_without_simd_p (vinfo, code)) |
7d8930a0 | 5961 | { |
73fbfcad | 5962 | if (dump_enabled_p ()) |
78c60e3d | 5963 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5964 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 5965 | return false; |
7d8930a0 | 5966 | } |
ebfd146a | 5967 | |
ebfd146a IR |
5968 | if (!vec_stmt) /* transformation not required. */ |
5969 | { | |
4a00c761 | 5970 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 5971 | if (dump_enabled_p ()) |
78c60e3d | 5972 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5973 | "=== vectorizable_operation ===\n"); |
68435eb2 | 5974 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, slp_node, cost_vec); |
ebfd146a IR |
5975 | return true; |
5976 | } | |
5977 | ||
67b8dbac | 5978 | /* Transform. */ |
ebfd146a | 5979 | |
73fbfcad | 5980 | if (dump_enabled_p ()) |
78c60e3d | 5981 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5982 | "transform binary/unary operation.\n"); |
383d9c83 | 5983 | |
ebfd146a | 5984 | /* Handle def. */ |
00f07b86 | 5985 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 5986 | |
0eb952ea JJ |
5987 | /* POINTER_DIFF_EXPR has pointer arguments which are vectorized as |
5988 | vectors with unsigned elements, but the result is signed. So, we | |
5989 | need to compute the MINUS_EXPR into vectype temporary and | |
5990 | VIEW_CONVERT_EXPR it into the final vectype_out result. */ | |
5991 | tree vec_cvt_dest = NULL_TREE; | |
5992 | if (orig_code == POINTER_DIFF_EXPR) | |
5993 | vec_cvt_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
5994 | ||
ebfd146a IR |
5995 | /* In case the vectorization factor (VF) is bigger than the number |
5996 | of elements that we can fit in a vectype (nunits), we have to generate | |
5997 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
5998 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
5999 | from one copy of the vector stmt to the next, in the field | |
6000 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
6001 | stages to find the correct vector defs to be used when vectorizing | |
6002 | stmts that use the defs of the current stmt. The example below | |
6003 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
6004 | we need to create 4 vectorized stmts): | |
6005 | ||
6006 | before vectorization: | |
6007 | RELATED_STMT VEC_STMT | |
6008 | S1: x = memref - - | |
6009 | S2: z = x + 1 - - | |
6010 | ||
6011 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
6012 | there): | |
6013 | RELATED_STMT VEC_STMT | |
6014 | VS1_0: vx0 = memref0 VS1_1 - | |
6015 | VS1_1: vx1 = memref1 VS1_2 - | |
6016 | VS1_2: vx2 = memref2 VS1_3 - | |
6017 | VS1_3: vx3 = memref3 - - | |
6018 | S1: x = load - VS1_0 | |
6019 | S2: z = x + 1 - - | |
6020 | ||
6021 | step2: vectorize stmt S2 (done here): | |
6022 | To vectorize stmt S2 we first need to find the relevant vector | |
6023 | def for the first operand 'x'. This is, as usual, obtained from | |
6024 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
6025 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
6026 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
6027 | the vector stmt VS2_0, and as usual, record it in the | |
6028 | STMT_VINFO_VEC_STMT of stmt S2. | |
6029 | When creating the second copy (VS2_1), we obtain the relevant vector | |
6030 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
6031 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
6032 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
6033 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
6034 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
6035 | chain of stmts and pointers: | |
6036 | RELATED_STMT VEC_STMT | |
6037 | VS1_0: vx0 = memref0 VS1_1 - | |
6038 | VS1_1: vx1 = memref1 VS1_2 - | |
6039 | VS1_2: vx2 = memref2 VS1_3 - | |
6040 | VS1_3: vx3 = memref3 - - | |
6041 | S1: x = load - VS1_0 | |
6042 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
6043 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
6044 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
6045 | VS2_3: vz3 = vx3 + v1 - - | |
6046 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
6047 | |
6048 | prev_stmt_info = NULL; | |
6049 | for (j = 0; j < ncopies; j++) | |
6050 | { | |
6051 | /* Handle uses. */ | |
6052 | if (j == 0) | |
4a00c761 | 6053 | { |
d6476f90 | 6054 | if (op_type == binary_op) |
4a00c761 | 6055 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, |
306b0c92 | 6056 | slp_node); |
d6476f90 RB |
6057 | else if (op_type == ternary_op) |
6058 | { | |
6059 | if (slp_node) | |
6060 | { | |
6061 | auto_vec<tree> ops(3); | |
6062 | ops.quick_push (op0); | |
6063 | ops.quick_push (op1); | |
6064 | ops.quick_push (op2); | |
6065 | auto_vec<vec<tree> > vec_defs(3); | |
6066 | vect_get_slp_defs (ops, slp_node, &vec_defs); | |
6067 | vec_oprnds0 = vec_defs[0]; | |
6068 | vec_oprnds1 = vec_defs[1]; | |
6069 | vec_oprnds2 = vec_defs[2]; | |
6070 | } | |
6071 | else | |
6072 | { | |
6073 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
6074 | NULL); | |
6075 | vect_get_vec_defs (op2, NULL_TREE, stmt, &vec_oprnds2, NULL, | |
6076 | NULL); | |
6077 | } | |
6078 | } | |
4a00c761 JJ |
6079 | else |
6080 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 6081 | slp_node); |
4a00c761 | 6082 | } |
ebfd146a | 6083 | else |
4a00c761 JJ |
6084 | { |
6085 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
6086 | if (op_type == ternary_op) | |
6087 | { | |
9771b263 DN |
6088 | tree vec_oprnd = vec_oprnds2.pop (); |
6089 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
6090 | vec_oprnd)); | |
4a00c761 JJ |
6091 | } |
6092 | } | |
6093 | ||
6094 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 6095 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 6096 | { |
4a00c761 | 6097 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 6098 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 6099 | vop2 = ((op_type == ternary_op) |
9771b263 | 6100 | ? vec_oprnds2[i] : NULL_TREE); |
0d0e4a03 | 6101 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1, vop2); |
4a00c761 JJ |
6102 | new_temp = make_ssa_name (vec_dest, new_stmt); |
6103 | gimple_assign_set_lhs (new_stmt, new_temp); | |
6104 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
0eb952ea JJ |
6105 | if (vec_cvt_dest) |
6106 | { | |
6107 | new_temp = build1 (VIEW_CONVERT_EXPR, vectype_out, new_temp); | |
6108 | new_stmt = gimple_build_assign (vec_cvt_dest, VIEW_CONVERT_EXPR, | |
6109 | new_temp); | |
6110 | new_temp = make_ssa_name (vec_cvt_dest, new_stmt); | |
6111 | gimple_assign_set_lhs (new_stmt, new_temp); | |
6112 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6113 | } | |
4a00c761 | 6114 | if (slp_node) |
9771b263 | 6115 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
6116 | } |
6117 | ||
4a00c761 JJ |
6118 | if (slp_node) |
6119 | continue; | |
6120 | ||
6121 | if (j == 0) | |
6122 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6123 | else | |
6124 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6125 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
6126 | } |
6127 | ||
9771b263 DN |
6128 | vec_oprnds0.release (); |
6129 | vec_oprnds1.release (); | |
6130 | vec_oprnds2.release (); | |
ebfd146a | 6131 | |
ebfd146a IR |
6132 | return true; |
6133 | } | |
6134 | ||
f702e7d4 | 6135 | /* A helper function to ensure data reference DR's base alignment. */ |
c716e67f XDL |
6136 | |
6137 | static void | |
f702e7d4 | 6138 | ensure_base_align (struct data_reference *dr) |
c716e67f XDL |
6139 | { |
6140 | if (!dr->aux) | |
6141 | return; | |
6142 | ||
52639a61 | 6143 | if (DR_VECT_AUX (dr)->base_misaligned) |
c716e67f | 6144 | { |
52639a61 | 6145 | tree base_decl = DR_VECT_AUX (dr)->base_decl; |
c716e67f | 6146 | |
f702e7d4 RS |
6147 | unsigned int align_base_to = DR_TARGET_ALIGNMENT (dr) * BITS_PER_UNIT; |
6148 | ||
428f0c67 | 6149 | if (decl_in_symtab_p (base_decl)) |
f702e7d4 | 6150 | symtab_node::get (base_decl)->increase_alignment (align_base_to); |
428f0c67 JH |
6151 | else |
6152 | { | |
f702e7d4 | 6153 | SET_DECL_ALIGN (base_decl, align_base_to); |
428f0c67 JH |
6154 | DECL_USER_ALIGN (base_decl) = 1; |
6155 | } | |
52639a61 | 6156 | DR_VECT_AUX (dr)->base_misaligned = false; |
c716e67f XDL |
6157 | } |
6158 | } | |
6159 | ||
ebfd146a | 6160 | |
44fc7854 BE |
6161 | /* Function get_group_alias_ptr_type. |
6162 | ||
6163 | Return the alias type for the group starting at FIRST_STMT. */ | |
6164 | ||
6165 | static tree | |
6166 | get_group_alias_ptr_type (gimple *first_stmt) | |
6167 | { | |
6168 | struct data_reference *first_dr, *next_dr; | |
6169 | gimple *next_stmt; | |
6170 | ||
6171 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
2c53b149 | 6172 | next_stmt = DR_GROUP_NEXT_ELEMENT (vinfo_for_stmt (first_stmt)); |
44fc7854 BE |
6173 | while (next_stmt) |
6174 | { | |
6175 | next_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (next_stmt)); | |
6176 | if (get_alias_set (DR_REF (first_dr)) | |
6177 | != get_alias_set (DR_REF (next_dr))) | |
6178 | { | |
6179 | if (dump_enabled_p ()) | |
6180 | dump_printf_loc (MSG_NOTE, vect_location, | |
6181 | "conflicting alias set types.\n"); | |
6182 | return ptr_type_node; | |
6183 | } | |
2c53b149 | 6184 | next_stmt = DR_GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
44fc7854 BE |
6185 | } |
6186 | return reference_alias_ptr_type (DR_REF (first_dr)); | |
6187 | } | |
6188 | ||
6189 | ||
ebfd146a IR |
6190 | /* Function vectorizable_store. |
6191 | ||
b8698a0f L |
6192 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
6193 | can be vectorized. | |
6194 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
6195 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
6196 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
6197 | ||
6198 | static bool | |
355fe088 | 6199 | vectorizable_store (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
68435eb2 | 6200 | slp_tree slp_node, stmt_vector_for_cost *cost_vec) |
ebfd146a | 6201 | { |
ebfd146a IR |
6202 | tree data_ref; |
6203 | tree op; | |
6204 | tree vec_oprnd = NULL_TREE; | |
6205 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
6206 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
272c6793 | 6207 | tree elem_type; |
ebfd146a | 6208 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 6209 | struct loop *loop = NULL; |
ef4bddc2 | 6210 | machine_mode vec_mode; |
ebfd146a IR |
6211 | tree dummy; |
6212 | enum dr_alignment_support alignment_support_scheme; | |
355fe088 | 6213 | gimple *def_stmt; |
929b4411 RS |
6214 | enum vect_def_type rhs_dt = vect_unknown_def_type; |
6215 | enum vect_def_type mask_dt = vect_unknown_def_type; | |
ebfd146a IR |
6216 | stmt_vec_info prev_stmt_info = NULL; |
6217 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 6218 | tree dataref_offset = NULL_TREE; |
355fe088 | 6219 | gimple *ptr_incr = NULL; |
ebfd146a IR |
6220 | int ncopies; |
6221 | int j; | |
2de001ee RS |
6222 | gimple *next_stmt, *first_stmt; |
6223 | bool grouped_store; | |
ebfd146a | 6224 | unsigned int group_size, i; |
6e1aa848 DN |
6225 | vec<tree> oprnds = vNULL; |
6226 | vec<tree> result_chain = vNULL; | |
ebfd146a | 6227 | bool inv_p; |
09dfa495 | 6228 | tree offset = NULL_TREE; |
6e1aa848 | 6229 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 6230 | bool slp = (slp_node != NULL); |
ebfd146a | 6231 | unsigned int vec_num; |
a70d6342 | 6232 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 6233 | vec_info *vinfo = stmt_info->vinfo; |
272c6793 | 6234 | tree aggr_type; |
134c85ca | 6235 | gather_scatter_info gs_info; |
355fe088 | 6236 | gimple *new_stmt; |
d9f21f6a | 6237 | poly_uint64 vf; |
2de001ee | 6238 | vec_load_store_type vls_type; |
44fc7854 | 6239 | tree ref_type; |
a70d6342 | 6240 | |
a70d6342 | 6241 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
6242 | return false; |
6243 | ||
66c16fd9 RB |
6244 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
6245 | && ! vec_stmt) | |
ebfd146a IR |
6246 | return false; |
6247 | ||
6248 | /* Is vectorizable store? */ | |
6249 | ||
c3a8f964 RS |
6250 | tree mask = NULL_TREE, mask_vectype = NULL_TREE; |
6251 | if (is_gimple_assign (stmt)) | |
6252 | { | |
6253 | tree scalar_dest = gimple_assign_lhs (stmt); | |
6254 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR | |
6255 | && is_pattern_stmt_p (stmt_info)) | |
6256 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
6257 | if (TREE_CODE (scalar_dest) != ARRAY_REF | |
6258 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF | |
6259 | && TREE_CODE (scalar_dest) != INDIRECT_REF | |
6260 | && TREE_CODE (scalar_dest) != COMPONENT_REF | |
6261 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
6262 | && TREE_CODE (scalar_dest) != REALPART_EXPR | |
6263 | && TREE_CODE (scalar_dest) != MEM_REF) | |
6264 | return false; | |
6265 | } | |
6266 | else | |
6267 | { | |
6268 | gcall *call = dyn_cast <gcall *> (stmt); | |
f307441a RS |
6269 | if (!call || !gimple_call_internal_p (call)) |
6270 | return false; | |
6271 | ||
6272 | internal_fn ifn = gimple_call_internal_fn (call); | |
6273 | if (!internal_store_fn_p (ifn)) | |
c3a8f964 | 6274 | return false; |
ebfd146a | 6275 | |
c3a8f964 RS |
6276 | if (slp_node != NULL) |
6277 | { | |
6278 | if (dump_enabled_p ()) | |
6279 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6280 | "SLP of masked stores not supported.\n"); | |
6281 | return false; | |
6282 | } | |
6283 | ||
f307441a RS |
6284 | int mask_index = internal_fn_mask_index (ifn); |
6285 | if (mask_index >= 0) | |
6286 | { | |
6287 | mask = gimple_call_arg (call, mask_index); | |
929b4411 RS |
6288 | if (!vect_check_load_store_mask (stmt, mask, &mask_dt, |
6289 | &mask_vectype)) | |
f307441a RS |
6290 | return false; |
6291 | } | |
c3a8f964 RS |
6292 | } |
6293 | ||
6294 | op = vect_get_store_rhs (stmt); | |
ebfd146a | 6295 | |
fce57248 RS |
6296 | /* Cannot have hybrid store SLP -- that would mean storing to the |
6297 | same location twice. */ | |
6298 | gcc_assert (slp == PURE_SLP_STMT (stmt_info)); | |
6299 | ||
f4d09712 | 6300 | tree vectype = STMT_VINFO_VECTYPE (stmt_info), rhs_vectype = NULL_TREE; |
4d694b27 | 6301 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 JJ |
6302 | |
6303 | if (loop_vinfo) | |
b17dc4d4 RB |
6304 | { |
6305 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
6306 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
6307 | } | |
6308 | else | |
6309 | vf = 1; | |
465c8c19 JJ |
6310 | |
6311 | /* Multiple types in SLP are handled by creating the appropriate number of | |
6312 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
6313 | case of SLP. */ | |
fce57248 | 6314 | if (slp) |
465c8c19 JJ |
6315 | ncopies = 1; |
6316 | else | |
e8f142e2 | 6317 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
6318 | |
6319 | gcc_assert (ncopies >= 1); | |
6320 | ||
6321 | /* FORNOW. This restriction should be relaxed. */ | |
6322 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) | |
6323 | { | |
6324 | if (dump_enabled_p ()) | |
6325 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6326 | "multiple types in nested loop.\n"); | |
6327 | return false; | |
6328 | } | |
6329 | ||
929b4411 | 6330 | if (!vect_check_store_rhs (stmt, op, &rhs_dt, &rhs_vectype, &vls_type)) |
f4d09712 KY |
6331 | return false; |
6332 | ||
272c6793 | 6333 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 6334 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 6335 | |
ebfd146a IR |
6336 | if (!STMT_VINFO_DATA_REF (stmt_info)) |
6337 | return false; | |
6338 | ||
2de001ee | 6339 | vect_memory_access_type memory_access_type; |
7e11fc7f | 6340 | if (!get_load_store_type (stmt, vectype, slp, mask, vls_type, ncopies, |
2de001ee RS |
6341 | &memory_access_type, &gs_info)) |
6342 | return false; | |
3bab6342 | 6343 | |
c3a8f964 RS |
6344 | if (mask) |
6345 | { | |
7e11fc7f RS |
6346 | if (memory_access_type == VMAT_CONTIGUOUS) |
6347 | { | |
6348 | if (!VECTOR_MODE_P (vec_mode) | |
6349 | || !can_vec_mask_load_store_p (vec_mode, | |
6350 | TYPE_MODE (mask_vectype), false)) | |
6351 | return false; | |
6352 | } | |
f307441a RS |
6353 | else if (memory_access_type != VMAT_LOAD_STORE_LANES |
6354 | && (memory_access_type != VMAT_GATHER_SCATTER || gs_info.decl)) | |
c3a8f964 RS |
6355 | { |
6356 | if (dump_enabled_p ()) | |
6357 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6358 | "unsupported access type for masked store.\n"); | |
6359 | return false; | |
6360 | } | |
c3a8f964 RS |
6361 | } |
6362 | else | |
6363 | { | |
6364 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
6365 | (e.g. - array initialization with 0). */ | |
6366 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) | |
6367 | return false; | |
6368 | } | |
6369 | ||
f307441a | 6370 | grouped_store = (STMT_VINFO_GROUPED_ACCESS (stmt_info) |
b5ec4de7 RS |
6371 | && memory_access_type != VMAT_GATHER_SCATTER |
6372 | && (slp || memory_access_type != VMAT_CONTIGUOUS)); | |
7cfb4d93 RS |
6373 | if (grouped_store) |
6374 | { | |
2c53b149 | 6375 | first_stmt = DR_GROUP_FIRST_ELEMENT (stmt_info); |
7cfb4d93 | 6376 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
2c53b149 | 6377 | group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
7cfb4d93 RS |
6378 | } |
6379 | else | |
6380 | { | |
6381 | first_stmt = stmt; | |
6382 | first_dr = dr; | |
6383 | group_size = vec_num = 1; | |
6384 | } | |
6385 | ||
ebfd146a IR |
6386 | if (!vec_stmt) /* transformation not required. */ |
6387 | { | |
2de001ee | 6388 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; |
7cfb4d93 RS |
6389 | |
6390 | if (loop_vinfo | |
6391 | && LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo)) | |
6392 | check_load_store_masking (loop_vinfo, vectype, vls_type, group_size, | |
bfaa08b7 | 6393 | memory_access_type, &gs_info); |
7cfb4d93 | 6394 | |
ebfd146a | 6395 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; |
68435eb2 RB |
6396 | vect_model_store_cost (stmt_info, ncopies, rhs_dt, memory_access_type, |
6397 | vls_type, slp_node, cost_vec); | |
ebfd146a IR |
6398 | return true; |
6399 | } | |
2de001ee | 6400 | gcc_assert (memory_access_type == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); |
ebfd146a | 6401 | |
67b8dbac | 6402 | /* Transform. */ |
ebfd146a | 6403 | |
f702e7d4 | 6404 | ensure_base_align (dr); |
c716e67f | 6405 | |
f307441a | 6406 | if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.decl) |
3bab6342 | 6407 | { |
c3a8f964 | 6408 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE, src; |
134c85ca | 6409 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
6410 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; |
6411 | tree ptr, mask, var, scale, perm_mask = NULL_TREE; | |
6412 | edge pe = loop_preheader_edge (loop); | |
6413 | gimple_seq seq; | |
6414 | basic_block new_bb; | |
6415 | enum { NARROW, NONE, WIDEN } modifier; | |
4d694b27 RS |
6416 | poly_uint64 scatter_off_nunits |
6417 | = TYPE_VECTOR_SUBPARTS (gs_info.offset_vectype); | |
3bab6342 | 6418 | |
4d694b27 | 6419 | if (known_eq (nunits, scatter_off_nunits)) |
3bab6342 | 6420 | modifier = NONE; |
4d694b27 | 6421 | else if (known_eq (nunits * 2, scatter_off_nunits)) |
3bab6342 | 6422 | { |
3bab6342 AT |
6423 | modifier = WIDEN; |
6424 | ||
4d694b27 RS |
6425 | /* Currently gathers and scatters are only supported for |
6426 | fixed-length vectors. */ | |
6427 | unsigned int count = scatter_off_nunits.to_constant (); | |
6428 | vec_perm_builder sel (count, count, 1); | |
6429 | for (i = 0; i < (unsigned int) count; ++i) | |
6430 | sel.quick_push (i | (count / 2)); | |
3bab6342 | 6431 | |
4d694b27 | 6432 | vec_perm_indices indices (sel, 1, count); |
e3342de4 RS |
6433 | perm_mask = vect_gen_perm_mask_checked (gs_info.offset_vectype, |
6434 | indices); | |
3bab6342 AT |
6435 | gcc_assert (perm_mask != NULL_TREE); |
6436 | } | |
4d694b27 | 6437 | else if (known_eq (nunits, scatter_off_nunits * 2)) |
3bab6342 | 6438 | { |
3bab6342 AT |
6439 | modifier = NARROW; |
6440 | ||
4d694b27 RS |
6441 | /* Currently gathers and scatters are only supported for |
6442 | fixed-length vectors. */ | |
6443 | unsigned int count = nunits.to_constant (); | |
6444 | vec_perm_builder sel (count, count, 1); | |
6445 | for (i = 0; i < (unsigned int) count; ++i) | |
6446 | sel.quick_push (i | (count / 2)); | |
3bab6342 | 6447 | |
4d694b27 | 6448 | vec_perm_indices indices (sel, 2, count); |
e3342de4 | 6449 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); |
3bab6342 AT |
6450 | gcc_assert (perm_mask != NULL_TREE); |
6451 | ncopies *= 2; | |
6452 | } | |
6453 | else | |
6454 | gcc_unreachable (); | |
6455 | ||
134c85ca | 6456 | rettype = TREE_TYPE (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
6457 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
6458 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6459 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6460 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6461 | scaletype = TREE_VALUE (arglist); | |
6462 | ||
6463 | gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE | |
6464 | && TREE_CODE (rettype) == VOID_TYPE); | |
6465 | ||
134c85ca | 6466 | ptr = fold_convert (ptrtype, gs_info.base); |
3bab6342 AT |
6467 | if (!is_gimple_min_invariant (ptr)) |
6468 | { | |
6469 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
6470 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
6471 | gcc_assert (!new_bb); | |
6472 | } | |
6473 | ||
6474 | /* Currently we support only unconditional scatter stores, | |
6475 | so mask should be all ones. */ | |
6476 | mask = build_int_cst (masktype, -1); | |
6477 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
6478 | ||
134c85ca | 6479 | scale = build_int_cst (scaletype, gs_info.scale); |
3bab6342 AT |
6480 | |
6481 | prev_stmt_info = NULL; | |
6482 | for (j = 0; j < ncopies; ++j) | |
6483 | { | |
6484 | if (j == 0) | |
6485 | { | |
6486 | src = vec_oprnd1 | |
c3a8f964 | 6487 | = vect_get_vec_def_for_operand (op, stmt); |
3bab6342 | 6488 | op = vec_oprnd0 |
134c85ca | 6489 | = vect_get_vec_def_for_operand (gs_info.offset, stmt); |
3bab6342 AT |
6490 | } |
6491 | else if (modifier != NONE && (j & 1)) | |
6492 | { | |
6493 | if (modifier == WIDEN) | |
6494 | { | |
6495 | src = vec_oprnd1 | |
929b4411 | 6496 | = vect_get_vec_def_for_stmt_copy (rhs_dt, vec_oprnd1); |
3bab6342 AT |
6497 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, perm_mask, |
6498 | stmt, gsi); | |
6499 | } | |
6500 | else if (modifier == NARROW) | |
6501 | { | |
6502 | src = permute_vec_elements (vec_oprnd1, vec_oprnd1, perm_mask, | |
6503 | stmt, gsi); | |
6504 | op = vec_oprnd0 | |
134c85ca RS |
6505 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
6506 | vec_oprnd0); | |
3bab6342 AT |
6507 | } |
6508 | else | |
6509 | gcc_unreachable (); | |
6510 | } | |
6511 | else | |
6512 | { | |
6513 | src = vec_oprnd1 | |
929b4411 | 6514 | = vect_get_vec_def_for_stmt_copy (rhs_dt, vec_oprnd1); |
3bab6342 | 6515 | op = vec_oprnd0 |
134c85ca RS |
6516 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
6517 | vec_oprnd0); | |
3bab6342 AT |
6518 | } |
6519 | ||
6520 | if (!useless_type_conversion_p (srctype, TREE_TYPE (src))) | |
6521 | { | |
928686b1 RS |
6522 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)), |
6523 | TYPE_VECTOR_SUBPARTS (srctype))); | |
0e22bb5a | 6524 | var = vect_get_new_ssa_name (srctype, vect_simple_var); |
3bab6342 AT |
6525 | src = build1 (VIEW_CONVERT_EXPR, srctype, src); |
6526 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, src); | |
6527 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6528 | src = var; | |
6529 | } | |
6530 | ||
6531 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
6532 | { | |
928686b1 RS |
6533 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)), |
6534 | TYPE_VECTOR_SUBPARTS (idxtype))); | |
0e22bb5a | 6535 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
3bab6342 AT |
6536 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
6537 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
6538 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6539 | op = var; | |
6540 | } | |
6541 | ||
6542 | new_stmt | |
134c85ca | 6543 | = gimple_build_call (gs_info.decl, 5, ptr, mask, op, src, scale); |
3bab6342 AT |
6544 | |
6545 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6546 | ||
6547 | if (prev_stmt_info == NULL) | |
6548 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6549 | else | |
6550 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6551 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6552 | } | |
6553 | return true; | |
6554 | } | |
6555 | ||
f307441a | 6556 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 6557 | { |
2c53b149 RB |
6558 | gimple *group_stmt = DR_GROUP_FIRST_ELEMENT (stmt_info); |
6559 | DR_GROUP_STORE_COUNT (vinfo_for_stmt (group_stmt))++; | |
f307441a | 6560 | } |
ebfd146a | 6561 | |
f307441a RS |
6562 | if (grouped_store) |
6563 | { | |
ebfd146a | 6564 | /* FORNOW */ |
a70d6342 | 6565 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
6566 | |
6567 | /* We vectorize all the stmts of the interleaving group when we | |
6568 | reach the last stmt in the group. */ | |
2c53b149 RB |
6569 | if (DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
6570 | < DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
6571 | && !slp) |
6572 | { | |
6573 | *vec_stmt = NULL; | |
6574 | return true; | |
6575 | } | |
6576 | ||
6577 | if (slp) | |
4b5caab7 | 6578 | { |
0d0293ac | 6579 | grouped_store = false; |
4b5caab7 IR |
6580 | /* VEC_NUM is the number of vect stmts to be created for this |
6581 | group. */ | |
6582 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 6583 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
2c53b149 | 6584 | gcc_assert (DR_GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_stmt)) == first_stmt); |
4b5caab7 | 6585 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
c3a8f964 | 6586 | op = vect_get_store_rhs (first_stmt); |
4b5caab7 | 6587 | } |
ebfd146a | 6588 | else |
4b5caab7 IR |
6589 | /* VEC_NUM is the number of vect stmts to be created for this |
6590 | group. */ | |
ebfd146a | 6591 | vec_num = group_size; |
44fc7854 BE |
6592 | |
6593 | ref_type = get_group_alias_ptr_type (first_stmt); | |
ebfd146a | 6594 | } |
b8698a0f | 6595 | else |
7cfb4d93 | 6596 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); |
b8698a0f | 6597 | |
73fbfcad | 6598 | if (dump_enabled_p ()) |
78c60e3d | 6599 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6600 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 6601 | |
2de001ee RS |
6602 | if (memory_access_type == VMAT_ELEMENTWISE |
6603 | || memory_access_type == VMAT_STRIDED_SLP) | |
f2e2a985 MM |
6604 | { |
6605 | gimple_stmt_iterator incr_gsi; | |
6606 | bool insert_after; | |
355fe088 | 6607 | gimple *incr; |
f2e2a985 MM |
6608 | tree offvar; |
6609 | tree ivstep; | |
6610 | tree running_off; | |
f2e2a985 MM |
6611 | tree stride_base, stride_step, alias_off; |
6612 | tree vec_oprnd; | |
f502d50e | 6613 | unsigned int g; |
4d694b27 RS |
6614 | /* Checked by get_load_store_type. */ |
6615 | unsigned int const_nunits = nunits.to_constant (); | |
f2e2a985 | 6616 | |
7cfb4d93 | 6617 | gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); |
f2e2a985 MM |
6618 | gcc_assert (!nested_in_vect_loop_p (loop, stmt)); |
6619 | ||
6620 | stride_base | |
6621 | = fold_build_pointer_plus | |
b210f45f | 6622 | (DR_BASE_ADDRESS (first_dr), |
f2e2a985 | 6623 | size_binop (PLUS_EXPR, |
b210f45f | 6624 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
44fc7854 | 6625 | convert_to_ptrofftype (DR_INIT (first_dr)))); |
b210f45f | 6626 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); |
f2e2a985 MM |
6627 | |
6628 | /* For a store with loop-invariant (but other than power-of-2) | |
6629 | stride (i.e. not a grouped access) like so: | |
6630 | ||
6631 | for (i = 0; i < n; i += stride) | |
6632 | array[i] = ...; | |
6633 | ||
6634 | we generate a new induction variable and new stores from | |
6635 | the components of the (vectorized) rhs: | |
6636 | ||
6637 | for (j = 0; ; j += VF*stride) | |
6638 | vectemp = ...; | |
6639 | tmp1 = vectemp[0]; | |
6640 | array[j] = tmp1; | |
6641 | tmp2 = vectemp[1]; | |
6642 | array[j + stride] = tmp2; | |
6643 | ... | |
6644 | */ | |
6645 | ||
4d694b27 | 6646 | unsigned nstores = const_nunits; |
b17dc4d4 | 6647 | unsigned lnel = 1; |
cee62fee | 6648 | tree ltype = elem_type; |
04199738 | 6649 | tree lvectype = vectype; |
cee62fee MM |
6650 | if (slp) |
6651 | { | |
4d694b27 RS |
6652 | if (group_size < const_nunits |
6653 | && const_nunits % group_size == 0) | |
b17dc4d4 | 6654 | { |
4d694b27 | 6655 | nstores = const_nunits / group_size; |
b17dc4d4 RB |
6656 | lnel = group_size; |
6657 | ltype = build_vector_type (elem_type, group_size); | |
04199738 RB |
6658 | lvectype = vectype; |
6659 | ||
6660 | /* First check if vec_extract optab doesn't support extraction | |
6661 | of vector elts directly. */ | |
b397965c | 6662 | scalar_mode elmode = SCALAR_TYPE_MODE (elem_type); |
9da15d40 RS |
6663 | machine_mode vmode; |
6664 | if (!mode_for_vector (elmode, group_size).exists (&vmode) | |
6665 | || !VECTOR_MODE_P (vmode) | |
414fef4e | 6666 | || !targetm.vector_mode_supported_p (vmode) |
04199738 RB |
6667 | || (convert_optab_handler (vec_extract_optab, |
6668 | TYPE_MODE (vectype), vmode) | |
6669 | == CODE_FOR_nothing)) | |
6670 | { | |
6671 | /* Try to avoid emitting an extract of vector elements | |
6672 | by performing the extracts using an integer type of the | |
6673 | same size, extracting from a vector of those and then | |
6674 | re-interpreting it as the original vector type if | |
6675 | supported. */ | |
6676 | unsigned lsize | |
6677 | = group_size * GET_MODE_BITSIZE (elmode); | |
fffbab82 | 6678 | elmode = int_mode_for_size (lsize, 0).require (); |
4d694b27 | 6679 | unsigned int lnunits = const_nunits / group_size; |
04199738 RB |
6680 | /* If we can't construct such a vector fall back to |
6681 | element extracts from the original vector type and | |
6682 | element size stores. */ | |
4d694b27 | 6683 | if (mode_for_vector (elmode, lnunits).exists (&vmode) |
9da15d40 | 6684 | && VECTOR_MODE_P (vmode) |
414fef4e | 6685 | && targetm.vector_mode_supported_p (vmode) |
04199738 RB |
6686 | && (convert_optab_handler (vec_extract_optab, |
6687 | vmode, elmode) | |
6688 | != CODE_FOR_nothing)) | |
6689 | { | |
4d694b27 | 6690 | nstores = lnunits; |
04199738 RB |
6691 | lnel = group_size; |
6692 | ltype = build_nonstandard_integer_type (lsize, 1); | |
6693 | lvectype = build_vector_type (ltype, nstores); | |
6694 | } | |
6695 | /* Else fall back to vector extraction anyway. | |
6696 | Fewer stores are more important than avoiding spilling | |
6697 | of the vector we extract from. Compared to the | |
6698 | construction case in vectorizable_load no store-forwarding | |
6699 | issue exists here for reasonable archs. */ | |
6700 | } | |
b17dc4d4 | 6701 | } |
4d694b27 RS |
6702 | else if (group_size >= const_nunits |
6703 | && group_size % const_nunits == 0) | |
b17dc4d4 RB |
6704 | { |
6705 | nstores = 1; | |
4d694b27 | 6706 | lnel = const_nunits; |
b17dc4d4 | 6707 | ltype = vectype; |
04199738 | 6708 | lvectype = vectype; |
b17dc4d4 | 6709 | } |
cee62fee MM |
6710 | ltype = build_aligned_type (ltype, TYPE_ALIGN (elem_type)); |
6711 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
6712 | } | |
6713 | ||
f2e2a985 MM |
6714 | ivstep = stride_step; |
6715 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
b17dc4d4 | 6716 | build_int_cst (TREE_TYPE (ivstep), vf)); |
f2e2a985 MM |
6717 | |
6718 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
6719 | ||
b210f45f RB |
6720 | stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base); |
6721 | ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep); | |
f2e2a985 MM |
6722 | create_iv (stride_base, ivstep, NULL, |
6723 | loop, &incr_gsi, insert_after, | |
6724 | &offvar, NULL); | |
6725 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 6726 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
f2e2a985 | 6727 | |
b210f45f | 6728 | stride_step = cse_and_gimplify_to_preheader (loop_vinfo, stride_step); |
f2e2a985 MM |
6729 | |
6730 | prev_stmt_info = NULL; | |
44fc7854 | 6731 | alias_off = build_int_cst (ref_type, 0); |
f502d50e MM |
6732 | next_stmt = first_stmt; |
6733 | for (g = 0; g < group_size; g++) | |
f2e2a985 | 6734 | { |
f502d50e MM |
6735 | running_off = offvar; |
6736 | if (g) | |
f2e2a985 | 6737 | { |
f502d50e MM |
6738 | tree size = TYPE_SIZE_UNIT (ltype); |
6739 | tree pos = fold_build2 (MULT_EXPR, sizetype, size_int (g), | |
f2e2a985 | 6740 | size); |
f502d50e | 6741 | tree newoff = copy_ssa_name (running_off, NULL); |
f2e2a985 | 6742 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
f502d50e | 6743 | running_off, pos); |
f2e2a985 | 6744 | vect_finish_stmt_generation (stmt, incr, gsi); |
f2e2a985 | 6745 | running_off = newoff; |
f502d50e | 6746 | } |
b17dc4d4 RB |
6747 | unsigned int group_el = 0; |
6748 | unsigned HOST_WIDE_INT | |
6749 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
f502d50e MM |
6750 | for (j = 0; j < ncopies; j++) |
6751 | { | |
c3a8f964 | 6752 | /* We've set op and dt above, from vect_get_store_rhs, |
f502d50e MM |
6753 | and first_stmt == stmt. */ |
6754 | if (j == 0) | |
6755 | { | |
6756 | if (slp) | |
6757 | { | |
6758 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, NULL, | |
306b0c92 | 6759 | slp_node); |
f502d50e MM |
6760 | vec_oprnd = vec_oprnds[0]; |
6761 | } | |
6762 | else | |
6763 | { | |
c3a8f964 | 6764 | op = vect_get_store_rhs (next_stmt); |
81c40241 | 6765 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
f502d50e MM |
6766 | } |
6767 | } | |
f2e2a985 | 6768 | else |
f502d50e MM |
6769 | { |
6770 | if (slp) | |
6771 | vec_oprnd = vec_oprnds[j]; | |
6772 | else | |
c079cbac | 6773 | { |
929b4411 RS |
6774 | vect_is_simple_use (op, vinfo, &def_stmt, &rhs_dt); |
6775 | vec_oprnd = vect_get_vec_def_for_stmt_copy (rhs_dt, | |
6776 | vec_oprnd); | |
c079cbac | 6777 | } |
f502d50e | 6778 | } |
04199738 RB |
6779 | /* Pun the vector to extract from if necessary. */ |
6780 | if (lvectype != vectype) | |
6781 | { | |
6782 | tree tem = make_ssa_name (lvectype); | |
6783 | gimple *pun | |
6784 | = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
6785 | lvectype, vec_oprnd)); | |
6786 | vect_finish_stmt_generation (stmt, pun, gsi); | |
6787 | vec_oprnd = tem; | |
6788 | } | |
f502d50e MM |
6789 | for (i = 0; i < nstores; i++) |
6790 | { | |
6791 | tree newref, newoff; | |
355fe088 | 6792 | gimple *incr, *assign; |
f502d50e MM |
6793 | tree size = TYPE_SIZE (ltype); |
6794 | /* Extract the i'th component. */ | |
6795 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, | |
6796 | bitsize_int (i), size); | |
6797 | tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd, | |
6798 | size, pos); | |
6799 | ||
6800 | elem = force_gimple_operand_gsi (gsi, elem, true, | |
6801 | NULL_TREE, true, | |
6802 | GSI_SAME_STMT); | |
6803 | ||
b17dc4d4 RB |
6804 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
6805 | group_el * elsz); | |
f502d50e | 6806 | newref = build2 (MEM_REF, ltype, |
b17dc4d4 | 6807 | running_off, this_off); |
19986382 | 6808 | vect_copy_ref_info (newref, DR_REF (first_dr)); |
f502d50e MM |
6809 | |
6810 | /* And store it to *running_off. */ | |
6811 | assign = gimple_build_assign (newref, elem); | |
6812 | vect_finish_stmt_generation (stmt, assign, gsi); | |
6813 | ||
b17dc4d4 RB |
6814 | group_el += lnel; |
6815 | if (! slp | |
6816 | || group_el == group_size) | |
6817 | { | |
6818 | newoff = copy_ssa_name (running_off, NULL); | |
6819 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
6820 | running_off, stride_step); | |
6821 | vect_finish_stmt_generation (stmt, incr, gsi); | |
f502d50e | 6822 | |
b17dc4d4 RB |
6823 | running_off = newoff; |
6824 | group_el = 0; | |
6825 | } | |
225ce44b RB |
6826 | if (g == group_size - 1 |
6827 | && !slp) | |
f502d50e MM |
6828 | { |
6829 | if (j == 0 && i == 0) | |
225ce44b RB |
6830 | STMT_VINFO_VEC_STMT (stmt_info) |
6831 | = *vec_stmt = assign; | |
f502d50e MM |
6832 | else |
6833 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = assign; | |
6834 | prev_stmt_info = vinfo_for_stmt (assign); | |
6835 | } | |
6836 | } | |
f2e2a985 | 6837 | } |
2c53b149 | 6838 | next_stmt = DR_GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
b17dc4d4 RB |
6839 | if (slp) |
6840 | break; | |
f2e2a985 | 6841 | } |
778dd3b6 RB |
6842 | |
6843 | vec_oprnds.release (); | |
f2e2a985 MM |
6844 | return true; |
6845 | } | |
6846 | ||
8c681247 | 6847 | auto_vec<tree> dr_chain (group_size); |
9771b263 | 6848 | oprnds.create (group_size); |
ebfd146a | 6849 | |
720f5239 | 6850 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 6851 | gcc_assert (alignment_support_scheme); |
70088b95 RS |
6852 | vec_loop_masks *loop_masks |
6853 | = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) | |
6854 | ? &LOOP_VINFO_MASKS (loop_vinfo) | |
6855 | : NULL); | |
272c6793 | 6856 | /* Targets with store-lane instructions must not require explicit |
c3a8f964 RS |
6857 | realignment. vect_supportable_dr_alignment always returns either |
6858 | dr_aligned or dr_unaligned_supported for masked operations. */ | |
7cfb4d93 RS |
6859 | gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES |
6860 | && !mask | |
70088b95 | 6861 | && !loop_masks) |
272c6793 RS |
6862 | || alignment_support_scheme == dr_aligned |
6863 | || alignment_support_scheme == dr_unaligned_supported); | |
6864 | ||
62da9e14 RS |
6865 | if (memory_access_type == VMAT_CONTIGUOUS_DOWN |
6866 | || memory_access_type == VMAT_CONTIGUOUS_REVERSE) | |
09dfa495 BM |
6867 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
6868 | ||
f307441a RS |
6869 | tree bump; |
6870 | tree vec_offset = NULL_TREE; | |
6871 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
6872 | { | |
6873 | aggr_type = NULL_TREE; | |
6874 | bump = NULL_TREE; | |
6875 | } | |
6876 | else if (memory_access_type == VMAT_GATHER_SCATTER) | |
6877 | { | |
6878 | aggr_type = elem_type; | |
6879 | vect_get_strided_load_store_ops (stmt, loop_vinfo, &gs_info, | |
6880 | &bump, &vec_offset); | |
6881 | } | |
272c6793 | 6882 | else |
f307441a RS |
6883 | { |
6884 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
6885 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
6886 | else | |
6887 | aggr_type = vectype; | |
6888 | bump = vect_get_data_ptr_increment (dr, aggr_type, memory_access_type); | |
6889 | } | |
ebfd146a | 6890 | |
c3a8f964 RS |
6891 | if (mask) |
6892 | LOOP_VINFO_HAS_MASK_STORE (loop_vinfo) = true; | |
6893 | ||
ebfd146a IR |
6894 | /* In case the vectorization factor (VF) is bigger than the number |
6895 | of elements that we can fit in a vectype (nunits), we have to generate | |
6896 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 6897 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
6898 | vect_get_vec_def_for_copy_stmt. */ |
6899 | ||
0d0293ac | 6900 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
6901 | |
6902 | S1: &base + 2 = x2 | |
6903 | S2: &base = x0 | |
6904 | S3: &base + 1 = x1 | |
6905 | S4: &base + 3 = x3 | |
6906 | ||
6907 | We create vectorized stores starting from base address (the access of the | |
6908 | first stmt in the chain (S2 in the above example), when the last store stmt | |
6909 | of the chain (S4) is reached: | |
6910 | ||
6911 | VS1: &base = vx2 | |
6912 | VS2: &base + vec_size*1 = vx0 | |
6913 | VS3: &base + vec_size*2 = vx1 | |
6914 | VS4: &base + vec_size*3 = vx3 | |
6915 | ||
6916 | Then permutation statements are generated: | |
6917 | ||
3fcc1b55 JJ |
6918 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
6919 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 6920 | ... |
b8698a0f | 6921 | |
ebfd146a IR |
6922 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
6923 | (the order of the data-refs in the output of vect_permute_store_chain | |
6924 | corresponds to the order of scalar stmts in the interleaving chain - see | |
6925 | the documentation of vect_permute_store_chain()). | |
6926 | ||
6927 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 6928 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 6929 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 6930 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
6931 | */ |
6932 | ||
6933 | prev_stmt_info = NULL; | |
c3a8f964 | 6934 | tree vec_mask = NULL_TREE; |
ebfd146a IR |
6935 | for (j = 0; j < ncopies; j++) |
6936 | { | |
ebfd146a IR |
6937 | |
6938 | if (j == 0) | |
6939 | { | |
6940 | if (slp) | |
6941 | { | |
6942 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c | 6943 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
306b0c92 | 6944 | NULL, slp_node); |
ebfd146a | 6945 | |
9771b263 | 6946 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
6947 | } |
6948 | else | |
6949 | { | |
b8698a0f L |
6950 | /* For interleaved stores we collect vectorized defs for all the |
6951 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
6952 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
6953 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
6954 | ||
2c53b149 | 6955 | If the store is not grouped, DR_GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 6956 | OPRNDS are of size 1. */ |
b8698a0f | 6957 | next_stmt = first_stmt; |
ebfd146a IR |
6958 | for (i = 0; i < group_size; i++) |
6959 | { | |
b8698a0f | 6960 | /* Since gaps are not supported for interleaved stores, |
2c53b149 | 6961 | DR_GROUP_SIZE is the exact number of stmts in the chain. |
b8698a0f | 6962 | Therefore, NEXT_STMT can't be NULL_TREE. In case that |
2c53b149 | 6963 | there is no interleaving, DR_GROUP_SIZE is 1, and only one |
ebfd146a | 6964 | iteration of the loop will be executed. */ |
c3a8f964 | 6965 | op = vect_get_store_rhs (next_stmt); |
81c40241 | 6966 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
9771b263 DN |
6967 | dr_chain.quick_push (vec_oprnd); |
6968 | oprnds.quick_push (vec_oprnd); | |
2c53b149 | 6969 | next_stmt = DR_GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a | 6970 | } |
c3a8f964 RS |
6971 | if (mask) |
6972 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, | |
6973 | mask_vectype); | |
ebfd146a IR |
6974 | } |
6975 | ||
6976 | /* We should have catched mismatched types earlier. */ | |
6977 | gcc_assert (useless_type_conversion_p (vectype, | |
6978 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
6979 | bool simd_lane_access_p |
6980 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
6981 | if (simd_lane_access_p | |
6982 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
6983 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
6984 | && integer_zerop (DR_OFFSET (first_dr)) | |
6985 | && integer_zerop (DR_INIT (first_dr)) | |
6986 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
44fc7854 | 6987 | get_alias_set (TREE_TYPE (ref_type)))) |
74bf76ed JJ |
6988 | { |
6989 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
44fc7854 | 6990 | dataref_offset = build_int_cst (ref_type, 0); |
8928eff3 | 6991 | inv_p = false; |
74bf76ed | 6992 | } |
f307441a RS |
6993 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
6994 | { | |
6995 | vect_get_gather_scatter_ops (loop, stmt, &gs_info, | |
6996 | &dataref_ptr, &vec_offset); | |
6997 | inv_p = false; | |
6998 | } | |
74bf76ed JJ |
6999 | else |
7000 | dataref_ptr | |
7001 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
7002 | simd_lane_access_p ? loop : NULL, | |
09dfa495 | 7003 | offset, &dummy, gsi, &ptr_incr, |
f307441a RS |
7004 | simd_lane_access_p, &inv_p, |
7005 | NULL_TREE, bump); | |
a70d6342 | 7006 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 7007 | } |
b8698a0f | 7008 | else |
ebfd146a | 7009 | { |
b8698a0f L |
7010 | /* For interleaved stores we created vectorized defs for all the |
7011 | defs stored in OPRNDS in the previous iteration (previous copy). | |
7012 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
7013 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
7014 | next copy. | |
2c53b149 | 7015 | If the store is not grouped, DR_GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
7016 | OPRNDS are of size 1. */ |
7017 | for (i = 0; i < group_size; i++) | |
7018 | { | |
9771b263 | 7019 | op = oprnds[i]; |
929b4411 RS |
7020 | vect_is_simple_use (op, vinfo, &def_stmt, &rhs_dt); |
7021 | vec_oprnd = vect_get_vec_def_for_stmt_copy (rhs_dt, op); | |
9771b263 DN |
7022 | dr_chain[i] = vec_oprnd; |
7023 | oprnds[i] = vec_oprnd; | |
ebfd146a | 7024 | } |
c3a8f964 | 7025 | if (mask) |
929b4411 | 7026 | vec_mask = vect_get_vec_def_for_stmt_copy (mask_dt, vec_mask); |
74bf76ed JJ |
7027 | if (dataref_offset) |
7028 | dataref_offset | |
f307441a RS |
7029 | = int_const_binop (PLUS_EXPR, dataref_offset, bump); |
7030 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
929b4411 RS |
7031 | vec_offset = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
7032 | vec_offset); | |
74bf76ed JJ |
7033 | else |
7034 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
f307441a | 7035 | bump); |
ebfd146a IR |
7036 | } |
7037 | ||
2de001ee | 7038 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
ebfd146a | 7039 | { |
272c6793 | 7040 | tree vec_array; |
267d3070 | 7041 | |
3ba4ff41 | 7042 | /* Get an array into which we can store the individual vectors. */ |
272c6793 | 7043 | vec_array = create_vector_array (vectype, vec_num); |
3ba4ff41 RS |
7044 | |
7045 | /* Invalidate the current contents of VEC_ARRAY. This should | |
7046 | become an RTL clobber too, which prevents the vector registers | |
7047 | from being upward-exposed. */ | |
7048 | vect_clobber_variable (stmt, gsi, vec_array); | |
7049 | ||
7050 | /* Store the individual vectors into the array. */ | |
272c6793 | 7051 | for (i = 0; i < vec_num; i++) |
c2d7ab2a | 7052 | { |
9771b263 | 7053 | vec_oprnd = dr_chain[i]; |
272c6793 | 7054 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 7055 | } |
b8698a0f | 7056 | |
7cfb4d93 | 7057 | tree final_mask = NULL; |
70088b95 RS |
7058 | if (loop_masks) |
7059 | final_mask = vect_get_loop_mask (gsi, loop_masks, ncopies, | |
7060 | vectype, j); | |
7cfb4d93 RS |
7061 | if (vec_mask) |
7062 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
7063 | vec_mask, gsi); | |
7064 | ||
7e11fc7f | 7065 | gcall *call; |
7cfb4d93 | 7066 | if (final_mask) |
7e11fc7f RS |
7067 | { |
7068 | /* Emit: | |
7069 | MASK_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK, | |
7070 | VEC_ARRAY). */ | |
7071 | unsigned int align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype)); | |
7072 | tree alias_ptr = build_int_cst (ref_type, align); | |
7073 | call = gimple_build_call_internal (IFN_MASK_STORE_LANES, 4, | |
7074 | dataref_ptr, alias_ptr, | |
7cfb4d93 | 7075 | final_mask, vec_array); |
7e11fc7f RS |
7076 | } |
7077 | else | |
7078 | { | |
7079 | /* Emit: | |
7080 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
7081 | data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type); | |
7082 | call = gimple_build_call_internal (IFN_STORE_LANES, 1, | |
7083 | vec_array); | |
7084 | gimple_call_set_lhs (call, data_ref); | |
7085 | } | |
a844293d RS |
7086 | gimple_call_set_nothrow (call, true); |
7087 | new_stmt = call; | |
267d3070 | 7088 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3ba4ff41 RS |
7089 | |
7090 | /* Record that VEC_ARRAY is now dead. */ | |
7091 | vect_clobber_variable (stmt, gsi, vec_array); | |
272c6793 RS |
7092 | } |
7093 | else | |
7094 | { | |
7095 | new_stmt = NULL; | |
0d0293ac | 7096 | if (grouped_store) |
272c6793 | 7097 | { |
b6b9227d JJ |
7098 | if (j == 0) |
7099 | result_chain.create (group_size); | |
272c6793 RS |
7100 | /* Permute. */ |
7101 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
7102 | &result_chain); | |
7103 | } | |
c2d7ab2a | 7104 | |
272c6793 RS |
7105 | next_stmt = first_stmt; |
7106 | for (i = 0; i < vec_num; i++) | |
7107 | { | |
644ffefd | 7108 | unsigned align, misalign; |
272c6793 | 7109 | |
7cfb4d93 | 7110 | tree final_mask = NULL_TREE; |
70088b95 RS |
7111 | if (loop_masks) |
7112 | final_mask = vect_get_loop_mask (gsi, loop_masks, | |
7113 | vec_num * ncopies, | |
7cfb4d93 RS |
7114 | vectype, vec_num * j + i); |
7115 | if (vec_mask) | |
7116 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
7117 | vec_mask, gsi); | |
7118 | ||
f307441a RS |
7119 | if (memory_access_type == VMAT_GATHER_SCATTER) |
7120 | { | |
7121 | tree scale = size_int (gs_info.scale); | |
7122 | gcall *call; | |
70088b95 | 7123 | if (loop_masks) |
f307441a RS |
7124 | call = gimple_build_call_internal |
7125 | (IFN_MASK_SCATTER_STORE, 5, dataref_ptr, vec_offset, | |
7126 | scale, vec_oprnd, final_mask); | |
7127 | else | |
7128 | call = gimple_build_call_internal | |
7129 | (IFN_SCATTER_STORE, 4, dataref_ptr, vec_offset, | |
7130 | scale, vec_oprnd); | |
7131 | gimple_call_set_nothrow (call, true); | |
7132 | new_stmt = call; | |
7133 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7134 | break; | |
7135 | } | |
7136 | ||
272c6793 RS |
7137 | if (i > 0) |
7138 | /* Bump the vector pointer. */ | |
7139 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
f307441a | 7140 | stmt, bump); |
272c6793 RS |
7141 | |
7142 | if (slp) | |
9771b263 | 7143 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
7144 | else if (grouped_store) |
7145 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 7146 | vect_permute_store_chain(). */ |
9771b263 | 7147 | vec_oprnd = result_chain[i]; |
272c6793 | 7148 | |
f702e7d4 | 7149 | align = DR_TARGET_ALIGNMENT (first_dr); |
272c6793 | 7150 | if (aligned_access_p (first_dr)) |
644ffefd | 7151 | misalign = 0; |
272c6793 RS |
7152 | else if (DR_MISALIGNMENT (first_dr) == -1) |
7153 | { | |
25f68d90 | 7154 | align = dr_alignment (vect_dr_behavior (first_dr)); |
52639a61 | 7155 | misalign = 0; |
272c6793 RS |
7156 | } |
7157 | else | |
c3a8f964 | 7158 | misalign = DR_MISALIGNMENT (first_dr); |
aed93b23 RB |
7159 | if (dataref_offset == NULL_TREE |
7160 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
7161 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
7162 | misalign); | |
c2d7ab2a | 7163 | |
62da9e14 | 7164 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
09dfa495 BM |
7165 | { |
7166 | tree perm_mask = perm_mask_for_reverse (vectype); | |
7167 | tree perm_dest | |
c3a8f964 | 7168 | = vect_create_destination_var (vect_get_store_rhs (stmt), |
09dfa495 | 7169 | vectype); |
b731b390 | 7170 | tree new_temp = make_ssa_name (perm_dest); |
09dfa495 BM |
7171 | |
7172 | /* Generate the permute statement. */ | |
355fe088 | 7173 | gimple *perm_stmt |
0d0e4a03 JJ |
7174 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
7175 | vec_oprnd, perm_mask); | |
09dfa495 BM |
7176 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
7177 | ||
7178 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); | |
7179 | vec_oprnd = new_temp; | |
7180 | } | |
7181 | ||
272c6793 | 7182 | /* Arguments are ready. Create the new vector stmt. */ |
7cfb4d93 | 7183 | if (final_mask) |
c3a8f964 RS |
7184 | { |
7185 | align = least_bit_hwi (misalign | align); | |
7186 | tree ptr = build_int_cst (ref_type, align); | |
7187 | gcall *call | |
7188 | = gimple_build_call_internal (IFN_MASK_STORE, 4, | |
7189 | dataref_ptr, ptr, | |
7cfb4d93 | 7190 | final_mask, vec_oprnd); |
c3a8f964 RS |
7191 | gimple_call_set_nothrow (call, true); |
7192 | new_stmt = call; | |
7193 | } | |
7194 | else | |
7195 | { | |
7196 | data_ref = fold_build2 (MEM_REF, vectype, | |
7197 | dataref_ptr, | |
7198 | dataref_offset | |
7199 | ? dataref_offset | |
7200 | : build_int_cst (ref_type, 0)); | |
7201 | if (aligned_access_p (first_dr)) | |
7202 | ; | |
7203 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
7204 | TREE_TYPE (data_ref) | |
7205 | = build_aligned_type (TREE_TYPE (data_ref), | |
7206 | align * BITS_PER_UNIT); | |
7207 | else | |
7208 | TREE_TYPE (data_ref) | |
7209 | = build_aligned_type (TREE_TYPE (data_ref), | |
7210 | TYPE_ALIGN (elem_type)); | |
19986382 | 7211 | vect_copy_ref_info (data_ref, DR_REF (first_dr)); |
c3a8f964 RS |
7212 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); |
7213 | } | |
272c6793 | 7214 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
7215 | |
7216 | if (slp) | |
7217 | continue; | |
7218 | ||
2c53b149 | 7219 | next_stmt = DR_GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
7220 | if (!next_stmt) |
7221 | break; | |
7222 | } | |
ebfd146a | 7223 | } |
1da0876c RS |
7224 | if (!slp) |
7225 | { | |
7226 | if (j == 0) | |
7227 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7228 | else | |
7229 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7230 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7231 | } | |
ebfd146a IR |
7232 | } |
7233 | ||
9771b263 DN |
7234 | oprnds.release (); |
7235 | result_chain.release (); | |
7236 | vec_oprnds.release (); | |
ebfd146a IR |
7237 | |
7238 | return true; | |
7239 | } | |
7240 | ||
557be5a8 AL |
7241 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
7242 | VECTOR_CST mask. No checks are made that the target platform supports the | |
7ac7e286 | 7243 | mask, so callers may wish to test can_vec_perm_const_p separately, or use |
557be5a8 | 7244 | vect_gen_perm_mask_checked. */ |
a1e53f3f | 7245 | |
3fcc1b55 | 7246 | tree |
4aae3cb3 | 7247 | vect_gen_perm_mask_any (tree vectype, const vec_perm_indices &sel) |
a1e53f3f | 7248 | { |
b00cb3bf | 7249 | tree mask_type; |
a1e53f3f | 7250 | |
0ecc2b7d RS |
7251 | poly_uint64 nunits = sel.length (); |
7252 | gcc_assert (known_eq (nunits, TYPE_VECTOR_SUBPARTS (vectype))); | |
b00cb3bf RS |
7253 | |
7254 | mask_type = build_vector_type (ssizetype, nunits); | |
736d0f28 | 7255 | return vec_perm_indices_to_tree (mask_type, sel); |
a1e53f3f L |
7256 | } |
7257 | ||
7ac7e286 | 7258 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_const_p, |
cf7aa6a3 | 7259 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ |
557be5a8 AL |
7260 | |
7261 | tree | |
4aae3cb3 | 7262 | vect_gen_perm_mask_checked (tree vectype, const vec_perm_indices &sel) |
557be5a8 | 7263 | { |
7ac7e286 | 7264 | gcc_assert (can_vec_perm_const_p (TYPE_MODE (vectype), sel)); |
557be5a8 AL |
7265 | return vect_gen_perm_mask_any (vectype, sel); |
7266 | } | |
7267 | ||
aec7ae7d JJ |
7268 | /* Given a vector variable X and Y, that was generated for the scalar |
7269 | STMT, generate instructions to permute the vector elements of X and Y | |
7270 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
7271 | permuted vector variable. */ | |
a1e53f3f L |
7272 | |
7273 | static tree | |
355fe088 | 7274 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple *stmt, |
aec7ae7d | 7275 | gimple_stmt_iterator *gsi) |
a1e53f3f L |
7276 | { |
7277 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 7278 | tree perm_dest, data_ref; |
355fe088 | 7279 | gimple *perm_stmt; |
a1e53f3f | 7280 | |
7ad429a4 RS |
7281 | tree scalar_dest = gimple_get_lhs (stmt); |
7282 | if (TREE_CODE (scalar_dest) == SSA_NAME) | |
7283 | perm_dest = vect_create_destination_var (scalar_dest, vectype); | |
7284 | else | |
7285 | perm_dest = vect_get_new_vect_var (vectype, vect_simple_var, NULL); | |
b731b390 | 7286 | data_ref = make_ssa_name (perm_dest); |
a1e53f3f L |
7287 | |
7288 | /* Generate the permute statement. */ | |
0d0e4a03 | 7289 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
a1e53f3f L |
7290 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
7291 | ||
7292 | return data_ref; | |
7293 | } | |
7294 | ||
6b916b36 RB |
7295 | /* Hoist the definitions of all SSA uses on STMT out of the loop LOOP, |
7296 | inserting them on the loops preheader edge. Returns true if we | |
7297 | were successful in doing so (and thus STMT can be moved then), | |
7298 | otherwise returns false. */ | |
7299 | ||
7300 | static bool | |
355fe088 | 7301 | hoist_defs_of_uses (gimple *stmt, struct loop *loop) |
6b916b36 RB |
7302 | { |
7303 | ssa_op_iter i; | |
7304 | tree op; | |
7305 | bool any = false; | |
7306 | ||
7307 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
7308 | { | |
355fe088 | 7309 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
7310 | if (!gimple_nop_p (def_stmt) |
7311 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
7312 | { | |
7313 | /* Make sure we don't need to recurse. While we could do | |
7314 | so in simple cases when there are more complex use webs | |
7315 | we don't have an easy way to preserve stmt order to fulfil | |
7316 | dependencies within them. */ | |
7317 | tree op2; | |
7318 | ssa_op_iter i2; | |
d1417442 JJ |
7319 | if (gimple_code (def_stmt) == GIMPLE_PHI) |
7320 | return false; | |
6b916b36 RB |
7321 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
7322 | { | |
355fe088 | 7323 | gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2); |
6b916b36 RB |
7324 | if (!gimple_nop_p (def_stmt2) |
7325 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2))) | |
7326 | return false; | |
7327 | } | |
7328 | any = true; | |
7329 | } | |
7330 | } | |
7331 | ||
7332 | if (!any) | |
7333 | return true; | |
7334 | ||
7335 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
7336 | { | |
355fe088 | 7337 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
7338 | if (!gimple_nop_p (def_stmt) |
7339 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
7340 | { | |
7341 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); | |
7342 | gsi_remove (&gsi, false); | |
7343 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); | |
7344 | } | |
7345 | } | |
7346 | ||
7347 | return true; | |
7348 | } | |
7349 | ||
ebfd146a IR |
7350 | /* vectorizable_load. |
7351 | ||
b8698a0f L |
7352 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
7353 | can be vectorized. | |
7354 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
7355 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
7356 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7357 | ||
7358 | static bool | |
355fe088 | 7359 | vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
68435eb2 RB |
7360 | slp_tree slp_node, slp_instance slp_node_instance, |
7361 | stmt_vector_for_cost *cost_vec) | |
ebfd146a IR |
7362 | { |
7363 | tree scalar_dest; | |
7364 | tree vec_dest = NULL; | |
7365 | tree data_ref = NULL; | |
7366 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 7367 | stmt_vec_info prev_stmt_info; |
ebfd146a | 7368 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 7369 | struct loop *loop = NULL; |
ebfd146a | 7370 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 7371 | bool nested_in_vect_loop = false; |
c716e67f | 7372 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
272c6793 | 7373 | tree elem_type; |
ebfd146a | 7374 | tree new_temp; |
ef4bddc2 | 7375 | machine_mode mode; |
355fe088 | 7376 | gimple *new_stmt = NULL; |
ebfd146a IR |
7377 | tree dummy; |
7378 | enum dr_alignment_support alignment_support_scheme; | |
7379 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 7380 | tree dataref_offset = NULL_TREE; |
355fe088 | 7381 | gimple *ptr_incr = NULL; |
ebfd146a | 7382 | int ncopies; |
4d694b27 RS |
7383 | int i, j; |
7384 | unsigned int group_size; | |
7385 | poly_uint64 group_gap_adj; | |
ebfd146a IR |
7386 | tree msq = NULL_TREE, lsq; |
7387 | tree offset = NULL_TREE; | |
356bbc4c | 7388 | tree byte_offset = NULL_TREE; |
ebfd146a | 7389 | tree realignment_token = NULL_TREE; |
538dd0b7 | 7390 | gphi *phi = NULL; |
6e1aa848 | 7391 | vec<tree> dr_chain = vNULL; |
0d0293ac | 7392 | bool grouped_load = false; |
355fe088 | 7393 | gimple *first_stmt; |
4f0a0218 | 7394 | gimple *first_stmt_for_drptr = NULL; |
ebfd146a IR |
7395 | bool inv_p; |
7396 | bool compute_in_loop = false; | |
7397 | struct loop *at_loop; | |
7398 | int vec_num; | |
7399 | bool slp = (slp_node != NULL); | |
7400 | bool slp_perm = false; | |
a70d6342 | 7401 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
d9f21f6a | 7402 | poly_uint64 vf; |
272c6793 | 7403 | tree aggr_type; |
134c85ca | 7404 | gather_scatter_info gs_info; |
310213d4 | 7405 | vec_info *vinfo = stmt_info->vinfo; |
44fc7854 | 7406 | tree ref_type; |
929b4411 | 7407 | enum vect_def_type mask_dt = vect_unknown_def_type; |
a70d6342 | 7408 | |
465c8c19 JJ |
7409 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
7410 | return false; | |
7411 | ||
66c16fd9 RB |
7412 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7413 | && ! vec_stmt) | |
465c8c19 JJ |
7414 | return false; |
7415 | ||
c3a8f964 RS |
7416 | tree mask = NULL_TREE, mask_vectype = NULL_TREE; |
7417 | if (is_gimple_assign (stmt)) | |
7418 | { | |
7419 | scalar_dest = gimple_assign_lhs (stmt); | |
7420 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
7421 | return false; | |
465c8c19 | 7422 | |
c3a8f964 RS |
7423 | tree_code code = gimple_assign_rhs_code (stmt); |
7424 | if (code != ARRAY_REF | |
7425 | && code != BIT_FIELD_REF | |
7426 | && code != INDIRECT_REF | |
7427 | && code != COMPONENT_REF | |
7428 | && code != IMAGPART_EXPR | |
7429 | && code != REALPART_EXPR | |
7430 | && code != MEM_REF | |
7431 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
7432 | return false; | |
7433 | } | |
7434 | else | |
7435 | { | |
7436 | gcall *call = dyn_cast <gcall *> (stmt); | |
bfaa08b7 RS |
7437 | if (!call || !gimple_call_internal_p (call)) |
7438 | return false; | |
7439 | ||
7440 | internal_fn ifn = gimple_call_internal_fn (call); | |
7441 | if (!internal_load_fn_p (ifn)) | |
c3a8f964 | 7442 | return false; |
465c8c19 | 7443 | |
c3a8f964 RS |
7444 | scalar_dest = gimple_call_lhs (call); |
7445 | if (!scalar_dest) | |
7446 | return false; | |
7447 | ||
7448 | if (slp_node != NULL) | |
7449 | { | |
7450 | if (dump_enabled_p ()) | |
7451 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7452 | "SLP of masked loads not supported.\n"); | |
7453 | return false; | |
7454 | } | |
7455 | ||
bfaa08b7 RS |
7456 | int mask_index = internal_fn_mask_index (ifn); |
7457 | if (mask_index >= 0) | |
7458 | { | |
7459 | mask = gimple_call_arg (call, mask_index); | |
929b4411 RS |
7460 | if (!vect_check_load_store_mask (stmt, mask, &mask_dt, |
7461 | &mask_vectype)) | |
bfaa08b7 RS |
7462 | return false; |
7463 | } | |
c3a8f964 | 7464 | } |
465c8c19 JJ |
7465 | |
7466 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
7467 | return false; | |
7468 | ||
7469 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
4d694b27 | 7470 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 | 7471 | |
a70d6342 IR |
7472 | if (loop_vinfo) |
7473 | { | |
7474 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
7475 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
7476 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
7477 | } | |
7478 | else | |
3533e503 | 7479 | vf = 1; |
ebfd146a IR |
7480 | |
7481 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 7482 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 7483 | case of SLP. */ |
fce57248 | 7484 | if (slp) |
ebfd146a IR |
7485 | ncopies = 1; |
7486 | else | |
e8f142e2 | 7487 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
ebfd146a IR |
7488 | |
7489 | gcc_assert (ncopies >= 1); | |
7490 | ||
7491 | /* FORNOW. This restriction should be relaxed. */ | |
7492 | if (nested_in_vect_loop && ncopies > 1) | |
7493 | { | |
73fbfcad | 7494 | if (dump_enabled_p ()) |
78c60e3d | 7495 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7496 | "multiple types in nested loop.\n"); |
ebfd146a IR |
7497 | return false; |
7498 | } | |
7499 | ||
f2556b68 RB |
7500 | /* Invalidate assumptions made by dependence analysis when vectorization |
7501 | on the unrolled body effectively re-orders stmts. */ | |
7502 | if (ncopies > 1 | |
7503 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
d9f21f6a RS |
7504 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
7505 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
f2556b68 RB |
7506 | { |
7507 | if (dump_enabled_p ()) | |
7508 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7509 | "cannot perform implicit CSE when unrolling " | |
7510 | "with negative dependence distance\n"); | |
7511 | return false; | |
7512 | } | |
7513 | ||
7b7b1813 | 7514 | elem_type = TREE_TYPE (vectype); |
947131ba | 7515 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
7516 | |
7517 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
7518 | (e.g. - data copies). */ | |
947131ba | 7519 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 7520 | { |
73fbfcad | 7521 | if (dump_enabled_p ()) |
78c60e3d | 7522 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7523 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
7524 | return false; |
7525 | } | |
7526 | ||
ebfd146a | 7527 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 7528 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 7529 | { |
0d0293ac | 7530 | grouped_load = true; |
ebfd146a | 7531 | /* FORNOW */ |
2de001ee RS |
7532 | gcc_assert (!nested_in_vect_loop); |
7533 | gcc_assert (!STMT_VINFO_GATHER_SCATTER_P (stmt_info)); | |
ebfd146a | 7534 | |
2c53b149 RB |
7535 | first_stmt = DR_GROUP_FIRST_ELEMENT (stmt_info); |
7536 | group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
d5f035ea | 7537 | |
b1af7da6 RB |
7538 | if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
7539 | slp_perm = true; | |
7540 | ||
f2556b68 RB |
7541 | /* Invalidate assumptions made by dependence analysis when vectorization |
7542 | on the unrolled body effectively re-orders stmts. */ | |
7543 | if (!PURE_SLP_STMT (stmt_info) | |
7544 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
d9f21f6a RS |
7545 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
7546 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
f2556b68 RB |
7547 | { |
7548 | if (dump_enabled_p ()) | |
7549 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7550 | "cannot perform implicit CSE when performing " | |
7551 | "group loads with negative dependence distance\n"); | |
7552 | return false; | |
7553 | } | |
96bb56b2 RB |
7554 | |
7555 | /* Similarly when the stmt is a load that is both part of a SLP | |
7556 | instance and a loop vectorized stmt via the same-dr mechanism | |
7557 | we have to give up. */ | |
2c53b149 | 7558 | if (DR_GROUP_SAME_DR_STMT (stmt_info) |
96bb56b2 RB |
7559 | && (STMT_SLP_TYPE (stmt_info) |
7560 | != STMT_SLP_TYPE (vinfo_for_stmt | |
2c53b149 | 7561 | (DR_GROUP_SAME_DR_STMT (stmt_info))))) |
96bb56b2 RB |
7562 | { |
7563 | if (dump_enabled_p ()) | |
7564 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7565 | "conflicting SLP types for CSEd load\n"); | |
7566 | return false; | |
7567 | } | |
ebfd146a | 7568 | } |
7cfb4d93 RS |
7569 | else |
7570 | group_size = 1; | |
ebfd146a | 7571 | |
2de001ee | 7572 | vect_memory_access_type memory_access_type; |
7e11fc7f | 7573 | if (!get_load_store_type (stmt, vectype, slp, mask, VLS_LOAD, ncopies, |
2de001ee RS |
7574 | &memory_access_type, &gs_info)) |
7575 | return false; | |
a1e53f3f | 7576 | |
c3a8f964 RS |
7577 | if (mask) |
7578 | { | |
7579 | if (memory_access_type == VMAT_CONTIGUOUS) | |
7580 | { | |
7e11fc7f RS |
7581 | machine_mode vec_mode = TYPE_MODE (vectype); |
7582 | if (!VECTOR_MODE_P (vec_mode) | |
7583 | || !can_vec_mask_load_store_p (vec_mode, | |
c3a8f964 RS |
7584 | TYPE_MODE (mask_vectype), true)) |
7585 | return false; | |
7586 | } | |
bfaa08b7 | 7587 | else if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.decl) |
c3a8f964 RS |
7588 | { |
7589 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); | |
7590 | tree masktype | |
7591 | = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (TREE_CHAIN (arglist)))); | |
7592 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
7593 | { | |
7594 | if (dump_enabled_p ()) | |
7595 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7596 | "masked gather with integer mask not" | |
7597 | " supported."); | |
7598 | return false; | |
7599 | } | |
7600 | } | |
bfaa08b7 RS |
7601 | else if (memory_access_type != VMAT_LOAD_STORE_LANES |
7602 | && memory_access_type != VMAT_GATHER_SCATTER) | |
c3a8f964 RS |
7603 | { |
7604 | if (dump_enabled_p ()) | |
7605 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7606 | "unsupported access type for masked load.\n"); | |
7607 | return false; | |
7608 | } | |
7609 | } | |
7610 | ||
ebfd146a IR |
7611 | if (!vec_stmt) /* transformation not required. */ |
7612 | { | |
2de001ee RS |
7613 | if (!slp) |
7614 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; | |
7cfb4d93 RS |
7615 | |
7616 | if (loop_vinfo | |
7617 | && LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo)) | |
7618 | check_load_store_masking (loop_vinfo, vectype, VLS_LOAD, group_size, | |
bfaa08b7 | 7619 | memory_access_type, &gs_info); |
7cfb4d93 | 7620 | |
ebfd146a | 7621 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; |
68435eb2 RB |
7622 | vect_model_load_cost (stmt_info, ncopies, memory_access_type, |
7623 | slp_node_instance, slp_node, cost_vec); | |
ebfd146a IR |
7624 | return true; |
7625 | } | |
7626 | ||
2de001ee RS |
7627 | if (!slp) |
7628 | gcc_assert (memory_access_type | |
7629 | == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); | |
7630 | ||
73fbfcad | 7631 | if (dump_enabled_p ()) |
78c60e3d | 7632 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 7633 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a | 7634 | |
67b8dbac | 7635 | /* Transform. */ |
ebfd146a | 7636 | |
f702e7d4 | 7637 | ensure_base_align (dr); |
c716e67f | 7638 | |
bfaa08b7 | 7639 | if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.decl) |
aec7ae7d | 7640 | { |
929b4411 RS |
7641 | vect_build_gather_load_calls (stmt, gsi, vec_stmt, &gs_info, mask, |
7642 | mask_dt); | |
aec7ae7d JJ |
7643 | return true; |
7644 | } | |
2de001ee RS |
7645 | |
7646 | if (memory_access_type == VMAT_ELEMENTWISE | |
7647 | || memory_access_type == VMAT_STRIDED_SLP) | |
7d75abc8 MM |
7648 | { |
7649 | gimple_stmt_iterator incr_gsi; | |
7650 | bool insert_after; | |
355fe088 | 7651 | gimple *incr; |
7d75abc8 | 7652 | tree offvar; |
7d75abc8 MM |
7653 | tree ivstep; |
7654 | tree running_off; | |
9771b263 | 7655 | vec<constructor_elt, va_gc> *v = NULL; |
14ac6aa2 | 7656 | tree stride_base, stride_step, alias_off; |
4d694b27 RS |
7657 | /* Checked by get_load_store_type. */ |
7658 | unsigned int const_nunits = nunits.to_constant (); | |
b210f45f | 7659 | unsigned HOST_WIDE_INT cst_offset = 0; |
14ac6aa2 | 7660 | |
7cfb4d93 | 7661 | gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); |
14ac6aa2 | 7662 | gcc_assert (!nested_in_vect_loop); |
7d75abc8 | 7663 | |
b210f45f | 7664 | if (grouped_load) |
44fc7854 | 7665 | { |
2c53b149 | 7666 | first_stmt = DR_GROUP_FIRST_ELEMENT (stmt_info); |
44fc7854 | 7667 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
44fc7854 | 7668 | } |
ab313a8c | 7669 | else |
44fc7854 BE |
7670 | { |
7671 | first_stmt = stmt; | |
7672 | first_dr = dr; | |
b210f45f RB |
7673 | } |
7674 | if (slp && grouped_load) | |
7675 | { | |
2c53b149 | 7676 | group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
b210f45f RB |
7677 | ref_type = get_group_alias_ptr_type (first_stmt); |
7678 | } | |
7679 | else | |
7680 | { | |
7681 | if (grouped_load) | |
7682 | cst_offset | |
7683 | = (tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))) | |
7684 | * vect_get_place_in_interleaving_chain (stmt, first_stmt)); | |
44fc7854 | 7685 | group_size = 1; |
b210f45f | 7686 | ref_type = reference_alias_ptr_type (DR_REF (dr)); |
44fc7854 | 7687 | } |
ab313a8c | 7688 | |
14ac6aa2 RB |
7689 | stride_base |
7690 | = fold_build_pointer_plus | |
ab313a8c | 7691 | (DR_BASE_ADDRESS (first_dr), |
14ac6aa2 | 7692 | size_binop (PLUS_EXPR, |
ab313a8c RB |
7693 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
7694 | convert_to_ptrofftype (DR_INIT (first_dr)))); | |
7695 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); | |
7d75abc8 MM |
7696 | |
7697 | /* For a load with loop-invariant (but other than power-of-2) | |
7698 | stride (i.e. not a grouped access) like so: | |
7699 | ||
7700 | for (i = 0; i < n; i += stride) | |
7701 | ... = array[i]; | |
7702 | ||
7703 | we generate a new induction variable and new accesses to | |
7704 | form a new vector (or vectors, depending on ncopies): | |
7705 | ||
7706 | for (j = 0; ; j += VF*stride) | |
7707 | tmp1 = array[j]; | |
7708 | tmp2 = array[j + stride]; | |
7709 | ... | |
7710 | vectemp = {tmp1, tmp2, ...} | |
7711 | */ | |
7712 | ||
ab313a8c RB |
7713 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step, |
7714 | build_int_cst (TREE_TYPE (stride_step), vf)); | |
7d75abc8 MM |
7715 | |
7716 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
7717 | ||
b210f45f RB |
7718 | stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base); |
7719 | ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep); | |
7720 | create_iv (stride_base, ivstep, NULL, | |
7d75abc8 MM |
7721 | loop, &incr_gsi, insert_after, |
7722 | &offvar, NULL); | |
7723 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 7724 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
7d75abc8 | 7725 | |
b210f45f | 7726 | stride_step = cse_and_gimplify_to_preheader (loop_vinfo, stride_step); |
7d75abc8 MM |
7727 | |
7728 | prev_stmt_info = NULL; | |
7729 | running_off = offvar; | |
44fc7854 | 7730 | alias_off = build_int_cst (ref_type, 0); |
4d694b27 | 7731 | int nloads = const_nunits; |
e09b4c37 | 7732 | int lnel = 1; |
7b5fc413 | 7733 | tree ltype = TREE_TYPE (vectype); |
ea60dd34 | 7734 | tree lvectype = vectype; |
b266b968 | 7735 | auto_vec<tree> dr_chain; |
2de001ee | 7736 | if (memory_access_type == VMAT_STRIDED_SLP) |
7b5fc413 | 7737 | { |
4d694b27 | 7738 | if (group_size < const_nunits) |
e09b4c37 | 7739 | { |
ff03930a JJ |
7740 | /* First check if vec_init optab supports construction from |
7741 | vector elts directly. */ | |
b397965c | 7742 | scalar_mode elmode = SCALAR_TYPE_MODE (TREE_TYPE (vectype)); |
9da15d40 RS |
7743 | machine_mode vmode; |
7744 | if (mode_for_vector (elmode, group_size).exists (&vmode) | |
7745 | && VECTOR_MODE_P (vmode) | |
414fef4e | 7746 | && targetm.vector_mode_supported_p (vmode) |
ff03930a JJ |
7747 | && (convert_optab_handler (vec_init_optab, |
7748 | TYPE_MODE (vectype), vmode) | |
7749 | != CODE_FOR_nothing)) | |
ea60dd34 | 7750 | { |
4d694b27 | 7751 | nloads = const_nunits / group_size; |
ea60dd34 | 7752 | lnel = group_size; |
ff03930a JJ |
7753 | ltype = build_vector_type (TREE_TYPE (vectype), group_size); |
7754 | } | |
7755 | else | |
7756 | { | |
7757 | /* Otherwise avoid emitting a constructor of vector elements | |
7758 | by performing the loads using an integer type of the same | |
7759 | size, constructing a vector of those and then | |
7760 | re-interpreting it as the original vector type. | |
7761 | This avoids a huge runtime penalty due to the general | |
7762 | inability to perform store forwarding from smaller stores | |
7763 | to a larger load. */ | |
7764 | unsigned lsize | |
7765 | = group_size * TYPE_PRECISION (TREE_TYPE (vectype)); | |
fffbab82 | 7766 | elmode = int_mode_for_size (lsize, 0).require (); |
4d694b27 | 7767 | unsigned int lnunits = const_nunits / group_size; |
ff03930a JJ |
7768 | /* If we can't construct such a vector fall back to |
7769 | element loads of the original vector type. */ | |
4d694b27 | 7770 | if (mode_for_vector (elmode, lnunits).exists (&vmode) |
9da15d40 | 7771 | && VECTOR_MODE_P (vmode) |
414fef4e | 7772 | && targetm.vector_mode_supported_p (vmode) |
ff03930a JJ |
7773 | && (convert_optab_handler (vec_init_optab, vmode, elmode) |
7774 | != CODE_FOR_nothing)) | |
7775 | { | |
4d694b27 | 7776 | nloads = lnunits; |
ff03930a JJ |
7777 | lnel = group_size; |
7778 | ltype = build_nonstandard_integer_type (lsize, 1); | |
7779 | lvectype = build_vector_type (ltype, nloads); | |
7780 | } | |
ea60dd34 | 7781 | } |
e09b4c37 | 7782 | } |
2de001ee | 7783 | else |
e09b4c37 | 7784 | { |
ea60dd34 | 7785 | nloads = 1; |
4d694b27 | 7786 | lnel = const_nunits; |
e09b4c37 | 7787 | ltype = vectype; |
e09b4c37 | 7788 | } |
2de001ee RS |
7789 | ltype = build_aligned_type (ltype, TYPE_ALIGN (TREE_TYPE (vectype))); |
7790 | } | |
bb4e4747 BC |
7791 | /* Load vector(1) scalar_type if it's 1 element-wise vectype. */ |
7792 | else if (nloads == 1) | |
7793 | ltype = vectype; | |
7794 | ||
2de001ee RS |
7795 | if (slp) |
7796 | { | |
66c16fd9 RB |
7797 | /* For SLP permutation support we need to load the whole group, |
7798 | not only the number of vector stmts the permutation result | |
7799 | fits in. */ | |
b266b968 | 7800 | if (slp_perm) |
66c16fd9 | 7801 | { |
d9f21f6a RS |
7802 | /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for |
7803 | variable VF. */ | |
7804 | unsigned int const_vf = vf.to_constant (); | |
4d694b27 | 7805 | ncopies = CEIL (group_size * const_vf, const_nunits); |
66c16fd9 RB |
7806 | dr_chain.create (ncopies); |
7807 | } | |
7808 | else | |
7809 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
7b5fc413 | 7810 | } |
4d694b27 | 7811 | unsigned int group_el = 0; |
e09b4c37 RB |
7812 | unsigned HOST_WIDE_INT |
7813 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
7d75abc8 MM |
7814 | for (j = 0; j < ncopies; j++) |
7815 | { | |
7b5fc413 | 7816 | if (nloads > 1) |
e09b4c37 RB |
7817 | vec_alloc (v, nloads); |
7818 | for (i = 0; i < nloads; i++) | |
7b5fc413 | 7819 | { |
e09b4c37 | 7820 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
b210f45f | 7821 | group_el * elsz + cst_offset); |
19986382 RB |
7822 | tree data_ref = build2 (MEM_REF, ltype, running_off, this_off); |
7823 | vect_copy_ref_info (data_ref, DR_REF (first_dr)); | |
7824 | new_stmt = gimple_build_assign (make_ssa_name (ltype), data_ref); | |
e09b4c37 RB |
7825 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7826 | if (nloads > 1) | |
7827 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, | |
7828 | gimple_assign_lhs (new_stmt)); | |
7829 | ||
7830 | group_el += lnel; | |
7831 | if (! slp | |
7832 | || group_el == group_size) | |
7b5fc413 | 7833 | { |
e09b4c37 RB |
7834 | tree newoff = copy_ssa_name (running_off); |
7835 | gimple *incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
7836 | running_off, stride_step); | |
7b5fc413 RB |
7837 | vect_finish_stmt_generation (stmt, incr, gsi); |
7838 | ||
7839 | running_off = newoff; | |
e09b4c37 | 7840 | group_el = 0; |
7b5fc413 | 7841 | } |
7b5fc413 | 7842 | } |
e09b4c37 | 7843 | if (nloads > 1) |
7d75abc8 | 7844 | { |
ea60dd34 RB |
7845 | tree vec_inv = build_constructor (lvectype, v); |
7846 | new_temp = vect_init_vector (stmt, vec_inv, lvectype, gsi); | |
e09b4c37 | 7847 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
ea60dd34 RB |
7848 | if (lvectype != vectype) |
7849 | { | |
7850 | new_stmt = gimple_build_assign (make_ssa_name (vectype), | |
7851 | VIEW_CONVERT_EXPR, | |
7852 | build1 (VIEW_CONVERT_EXPR, | |
7853 | vectype, new_temp)); | |
7854 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7855 | } | |
7d75abc8 MM |
7856 | } |
7857 | ||
7b5fc413 | 7858 | if (slp) |
b266b968 | 7859 | { |
b266b968 RB |
7860 | if (slp_perm) |
7861 | dr_chain.quick_push (gimple_assign_lhs (new_stmt)); | |
66c16fd9 RB |
7862 | else |
7863 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
b266b968 | 7864 | } |
7d75abc8 | 7865 | else |
225ce44b RB |
7866 | { |
7867 | if (j == 0) | |
7868 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7869 | else | |
7870 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7871 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7872 | } | |
7d75abc8 | 7873 | } |
b266b968 | 7874 | if (slp_perm) |
29afecdf RB |
7875 | { |
7876 | unsigned n_perms; | |
7877 | vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, | |
7878 | slp_node_instance, false, &n_perms); | |
7879 | } | |
7d75abc8 MM |
7880 | return true; |
7881 | } | |
aec7ae7d | 7882 | |
b5ec4de7 RS |
7883 | if (memory_access_type == VMAT_GATHER_SCATTER |
7884 | || (!slp && memory_access_type == VMAT_CONTIGUOUS)) | |
ab2fc782 RS |
7885 | grouped_load = false; |
7886 | ||
0d0293ac | 7887 | if (grouped_load) |
ebfd146a | 7888 | { |
2c53b149 RB |
7889 | first_stmt = DR_GROUP_FIRST_ELEMENT (stmt_info); |
7890 | group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
4f0a0218 | 7891 | /* For SLP vectorization we directly vectorize a subchain |
52eab378 RB |
7892 | without permutation. */ |
7893 | if (slp && ! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
4f0a0218 RB |
7894 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
7895 | /* For BB vectorization always use the first stmt to base | |
7896 | the data ref pointer on. */ | |
7897 | if (bb_vinfo) | |
7898 | first_stmt_for_drptr = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 7899 | |
ebfd146a | 7900 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
7901 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
7902 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
7903 | ??? But we can only do so if there is exactly one | |
7904 | as we have no way to get at the rest. Leave the CSE | |
7905 | opportunity alone. | |
7906 | ??? With the group load eventually participating | |
7907 | in multiple different permutations (having multiple | |
7908 | slp nodes which refer to the same group) the CSE | |
7909 | is even wrong code. See PR56270. */ | |
7910 | && !slp) | |
ebfd146a IR |
7911 | { |
7912 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
7913 | return true; | |
7914 | } | |
7915 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
9b999e8c | 7916 | group_gap_adj = 0; |
ebfd146a IR |
7917 | |
7918 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
7919 | if (slp) | |
7920 | { | |
0d0293ac | 7921 | grouped_load = false; |
91ff1504 RB |
7922 | /* For SLP permutation support we need to load the whole group, |
7923 | not only the number of vector stmts the permutation result | |
7924 | fits in. */ | |
7925 | if (slp_perm) | |
b267968e | 7926 | { |
d9f21f6a RS |
7927 | /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for |
7928 | variable VF. */ | |
7929 | unsigned int const_vf = vf.to_constant (); | |
4d694b27 RS |
7930 | unsigned int const_nunits = nunits.to_constant (); |
7931 | vec_num = CEIL (group_size * const_vf, const_nunits); | |
b267968e RB |
7932 | group_gap_adj = vf * group_size - nunits * vec_num; |
7933 | } | |
91ff1504 | 7934 | else |
b267968e RB |
7935 | { |
7936 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
796bd467 RB |
7937 | group_gap_adj |
7938 | = group_size - SLP_INSTANCE_GROUP_SIZE (slp_node_instance); | |
b267968e | 7939 | } |
a70d6342 | 7940 | } |
ebfd146a | 7941 | else |
9b999e8c | 7942 | vec_num = group_size; |
44fc7854 BE |
7943 | |
7944 | ref_type = get_group_alias_ptr_type (first_stmt); | |
ebfd146a IR |
7945 | } |
7946 | else | |
7947 | { | |
7948 | first_stmt = stmt; | |
7949 | first_dr = dr; | |
7950 | group_size = vec_num = 1; | |
9b999e8c | 7951 | group_gap_adj = 0; |
44fc7854 | 7952 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); |
ebfd146a IR |
7953 | } |
7954 | ||
720f5239 | 7955 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 7956 | gcc_assert (alignment_support_scheme); |
70088b95 RS |
7957 | vec_loop_masks *loop_masks |
7958 | = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) | |
7959 | ? &LOOP_VINFO_MASKS (loop_vinfo) | |
7960 | : NULL); | |
7cfb4d93 RS |
7961 | /* Targets with store-lane instructions must not require explicit |
7962 | realignment. vect_supportable_dr_alignment always returns either | |
7963 | dr_aligned or dr_unaligned_supported for masked operations. */ | |
7964 | gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES | |
7965 | && !mask | |
70088b95 | 7966 | && !loop_masks) |
272c6793 RS |
7967 | || alignment_support_scheme == dr_aligned |
7968 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
7969 | |
7970 | /* In case the vectorization factor (VF) is bigger than the number | |
7971 | of elements that we can fit in a vectype (nunits), we have to generate | |
7972 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 7973 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 7974 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 7975 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 7976 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
7977 | stmts that use the defs of the current stmt. The example below |
7978 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
7979 | need to create 4 vectorized stmts): | |
ebfd146a IR |
7980 | |
7981 | before vectorization: | |
7982 | RELATED_STMT VEC_STMT | |
7983 | S1: x = memref - - | |
7984 | S2: z = x + 1 - - | |
7985 | ||
7986 | step 1: vectorize stmt S1: | |
7987 | We first create the vector stmt VS1_0, and, as usual, record a | |
7988 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
7989 | Next, we create the vector stmt VS1_1, and record a pointer to | |
7990 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 7991 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
7992 | stmts and pointers: |
7993 | RELATED_STMT VEC_STMT | |
7994 | VS1_0: vx0 = memref0 VS1_1 - | |
7995 | VS1_1: vx1 = memref1 VS1_2 - | |
7996 | VS1_2: vx2 = memref2 VS1_3 - | |
7997 | VS1_3: vx3 = memref3 - - | |
7998 | S1: x = load - VS1_0 | |
7999 | S2: z = x + 1 - - | |
8000 | ||
b8698a0f L |
8001 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
8002 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
8003 | stmt S2. */ |
8004 | ||
0d0293ac | 8005 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
8006 | |
8007 | S1: x2 = &base + 2 | |
8008 | S2: x0 = &base | |
8009 | S3: x1 = &base + 1 | |
8010 | S4: x3 = &base + 3 | |
8011 | ||
b8698a0f | 8012 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
8013 | starting from the access of the first stmt of the chain: |
8014 | ||
8015 | VS1: vx0 = &base | |
8016 | VS2: vx1 = &base + vec_size*1 | |
8017 | VS3: vx3 = &base + vec_size*2 | |
8018 | VS4: vx4 = &base + vec_size*3 | |
8019 | ||
8020 | Then permutation statements are generated: | |
8021 | ||
e2c83630 RH |
8022 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
8023 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
8024 | ... |
8025 | ||
8026 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
8027 | (the order of the data-refs in the output of vect_permute_load_chain | |
8028 | corresponds to the order of scalar stmts in the interleaving chain - see | |
8029 | the documentation of vect_permute_load_chain()). | |
8030 | The generation of permutation stmts and recording them in | |
0d0293ac | 8031 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 8032 | |
b8698a0f | 8033 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
8034 | permutation stmts above are created for every copy. The result vector |
8035 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
8036 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
8037 | |
8038 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
8039 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
8040 | we generate the following code: | |
8041 | p = initial_addr; | |
8042 | indx = 0; | |
8043 | loop { | |
8044 | p = p + indx * vectype_size; | |
8045 | vec_dest = *(p); | |
8046 | indx = indx + 1; | |
8047 | } | |
8048 | ||
8049 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 8050 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
8051 | then generate the following code, in which the data in each iteration is |
8052 | obtained by two vector loads, one from the previous iteration, and one | |
8053 | from the current iteration: | |
8054 | p1 = initial_addr; | |
8055 | msq_init = *(floor(p1)) | |
8056 | p2 = initial_addr + VS - 1; | |
8057 | realignment_token = call target_builtin; | |
8058 | indx = 0; | |
8059 | loop { | |
8060 | p2 = p2 + indx * vectype_size | |
8061 | lsq = *(floor(p2)) | |
8062 | vec_dest = realign_load (msq, lsq, realignment_token) | |
8063 | indx = indx + 1; | |
8064 | msq = lsq; | |
8065 | } */ | |
8066 | ||
8067 | /* If the misalignment remains the same throughout the execution of the | |
8068 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 8069 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
8070 | This can only occur when vectorizing memory accesses in the inner-loop |
8071 | nested within an outer-loop that is being vectorized. */ | |
8072 | ||
d1e4b493 | 8073 | if (nested_in_vect_loop |
cf098191 RS |
8074 | && !multiple_p (DR_STEP_ALIGNMENT (dr), |
8075 | GET_MODE_SIZE (TYPE_MODE (vectype)))) | |
ebfd146a IR |
8076 | { |
8077 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
8078 | compute_in_loop = true; | |
8079 | } | |
8080 | ||
8081 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
8082 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 8083 | && !compute_in_loop) |
ebfd146a IR |
8084 | { |
8085 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
8086 | alignment_support_scheme, NULL_TREE, | |
8087 | &at_loop); | |
8088 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
8089 | { | |
538dd0b7 | 8090 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (msq)); |
356bbc4c JJ |
8091 | byte_offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
8092 | size_one_node); | |
ebfd146a IR |
8093 | } |
8094 | } | |
8095 | else | |
8096 | at_loop = loop; | |
8097 | ||
62da9e14 | 8098 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
a1e53f3f L |
8099 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
8100 | ||
ab2fc782 RS |
8101 | tree bump; |
8102 | tree vec_offset = NULL_TREE; | |
8103 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
8104 | { | |
8105 | aggr_type = NULL_TREE; | |
8106 | bump = NULL_TREE; | |
8107 | } | |
8108 | else if (memory_access_type == VMAT_GATHER_SCATTER) | |
8109 | { | |
8110 | aggr_type = elem_type; | |
8111 | vect_get_strided_load_store_ops (stmt, loop_vinfo, &gs_info, | |
8112 | &bump, &vec_offset); | |
8113 | } | |
272c6793 | 8114 | else |
ab2fc782 RS |
8115 | { |
8116 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
8117 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
8118 | else | |
8119 | aggr_type = vectype; | |
8120 | bump = vect_get_data_ptr_increment (dr, aggr_type, memory_access_type); | |
8121 | } | |
272c6793 | 8122 | |
c3a8f964 | 8123 | tree vec_mask = NULL_TREE; |
ebfd146a | 8124 | prev_stmt_info = NULL; |
4d694b27 | 8125 | poly_uint64 group_elt = 0; |
ebfd146a | 8126 | for (j = 0; j < ncopies; j++) |
b8698a0f | 8127 | { |
272c6793 | 8128 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 8129 | if (j == 0) |
74bf76ed JJ |
8130 | { |
8131 | bool simd_lane_access_p | |
8132 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
8133 | if (simd_lane_access_p | |
8134 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
8135 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
8136 | && integer_zerop (DR_OFFSET (first_dr)) | |
8137 | && integer_zerop (DR_INIT (first_dr)) | |
8138 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
44fc7854 | 8139 | get_alias_set (TREE_TYPE (ref_type))) |
74bf76ed JJ |
8140 | && (alignment_support_scheme == dr_aligned |
8141 | || alignment_support_scheme == dr_unaligned_supported)) | |
8142 | { | |
8143 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
44fc7854 | 8144 | dataref_offset = build_int_cst (ref_type, 0); |
8928eff3 | 8145 | inv_p = false; |
74bf76ed | 8146 | } |
4f0a0218 RB |
8147 | else if (first_stmt_for_drptr |
8148 | && first_stmt != first_stmt_for_drptr) | |
8149 | { | |
8150 | dataref_ptr | |
8151 | = vect_create_data_ref_ptr (first_stmt_for_drptr, aggr_type, | |
8152 | at_loop, offset, &dummy, gsi, | |
8153 | &ptr_incr, simd_lane_access_p, | |
ab2fc782 | 8154 | &inv_p, byte_offset, bump); |
4f0a0218 RB |
8155 | /* Adjust the pointer by the difference to first_stmt. */ |
8156 | data_reference_p ptrdr | |
8157 | = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt_for_drptr)); | |
8158 | tree diff = fold_convert (sizetype, | |
8159 | size_binop (MINUS_EXPR, | |
8160 | DR_INIT (first_dr), | |
8161 | DR_INIT (ptrdr))); | |
8162 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
8163 | stmt, diff); | |
8164 | } | |
bfaa08b7 RS |
8165 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
8166 | { | |
8167 | vect_get_gather_scatter_ops (loop, stmt, &gs_info, | |
8168 | &dataref_ptr, &vec_offset); | |
8169 | inv_p = false; | |
8170 | } | |
74bf76ed JJ |
8171 | else |
8172 | dataref_ptr | |
8173 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
8174 | offset, &dummy, gsi, &ptr_incr, | |
356bbc4c | 8175 | simd_lane_access_p, &inv_p, |
ab2fc782 | 8176 | byte_offset, bump); |
c3a8f964 RS |
8177 | if (mask) |
8178 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, | |
8179 | mask_vectype); | |
74bf76ed | 8180 | } |
ebfd146a | 8181 | else |
c3a8f964 RS |
8182 | { |
8183 | if (dataref_offset) | |
8184 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
ab2fc782 | 8185 | bump); |
bfaa08b7 | 8186 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
929b4411 RS |
8187 | vec_offset = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
8188 | vec_offset); | |
c3a8f964 | 8189 | else |
ab2fc782 RS |
8190 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
8191 | stmt, bump); | |
c3a8f964 | 8192 | if (mask) |
929b4411 | 8193 | vec_mask = vect_get_vec_def_for_stmt_copy (mask_dt, vec_mask); |
c3a8f964 | 8194 | } |
ebfd146a | 8195 | |
0d0293ac | 8196 | if (grouped_load || slp_perm) |
9771b263 | 8197 | dr_chain.create (vec_num); |
5ce1ee7f | 8198 | |
2de001ee | 8199 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
ebfd146a | 8200 | { |
272c6793 RS |
8201 | tree vec_array; |
8202 | ||
8203 | vec_array = create_vector_array (vectype, vec_num); | |
8204 | ||
7cfb4d93 | 8205 | tree final_mask = NULL_TREE; |
70088b95 RS |
8206 | if (loop_masks) |
8207 | final_mask = vect_get_loop_mask (gsi, loop_masks, ncopies, | |
8208 | vectype, j); | |
7cfb4d93 RS |
8209 | if (vec_mask) |
8210 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
8211 | vec_mask, gsi); | |
8212 | ||
7e11fc7f | 8213 | gcall *call; |
7cfb4d93 | 8214 | if (final_mask) |
7e11fc7f RS |
8215 | { |
8216 | /* Emit: | |
8217 | VEC_ARRAY = MASK_LOAD_LANES (DATAREF_PTR, ALIAS_PTR, | |
8218 | VEC_MASK). */ | |
8219 | unsigned int align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype)); | |
8220 | tree alias_ptr = build_int_cst (ref_type, align); | |
8221 | call = gimple_build_call_internal (IFN_MASK_LOAD_LANES, 3, | |
8222 | dataref_ptr, alias_ptr, | |
7cfb4d93 | 8223 | final_mask); |
7e11fc7f RS |
8224 | } |
8225 | else | |
8226 | { | |
8227 | /* Emit: | |
8228 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
8229 | data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type); | |
8230 | call = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
8231 | } | |
a844293d RS |
8232 | gimple_call_set_lhs (call, vec_array); |
8233 | gimple_call_set_nothrow (call, true); | |
8234 | new_stmt = call; | |
272c6793 | 8235 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a | 8236 | |
272c6793 RS |
8237 | /* Extract each vector into an SSA_NAME. */ |
8238 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 8239 | { |
272c6793 RS |
8240 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
8241 | vec_array, i); | |
9771b263 | 8242 | dr_chain.quick_push (new_temp); |
272c6793 RS |
8243 | } |
8244 | ||
8245 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 8246 | vect_record_grouped_load_vectors (stmt, dr_chain); |
3ba4ff41 RS |
8247 | |
8248 | /* Record that VEC_ARRAY is now dead. */ | |
8249 | vect_clobber_variable (stmt, gsi, vec_array); | |
272c6793 RS |
8250 | } |
8251 | else | |
8252 | { | |
8253 | for (i = 0; i < vec_num; i++) | |
8254 | { | |
7cfb4d93 | 8255 | tree final_mask = NULL_TREE; |
70088b95 | 8256 | if (loop_masks |
7cfb4d93 | 8257 | && memory_access_type != VMAT_INVARIANT) |
70088b95 RS |
8258 | final_mask = vect_get_loop_mask (gsi, loop_masks, |
8259 | vec_num * ncopies, | |
7cfb4d93 RS |
8260 | vectype, vec_num * j + i); |
8261 | if (vec_mask) | |
8262 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
8263 | vec_mask, gsi); | |
8264 | ||
272c6793 RS |
8265 | if (i > 0) |
8266 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
ab2fc782 | 8267 | stmt, bump); |
272c6793 RS |
8268 | |
8269 | /* 2. Create the vector-load in the loop. */ | |
8270 | switch (alignment_support_scheme) | |
8271 | { | |
8272 | case dr_aligned: | |
8273 | case dr_unaligned_supported: | |
be1ac4ec | 8274 | { |
644ffefd MJ |
8275 | unsigned int align, misalign; |
8276 | ||
bfaa08b7 RS |
8277 | if (memory_access_type == VMAT_GATHER_SCATTER) |
8278 | { | |
8279 | tree scale = size_int (gs_info.scale); | |
8280 | gcall *call; | |
70088b95 | 8281 | if (loop_masks) |
bfaa08b7 RS |
8282 | call = gimple_build_call_internal |
8283 | (IFN_MASK_GATHER_LOAD, 4, dataref_ptr, | |
8284 | vec_offset, scale, final_mask); | |
8285 | else | |
8286 | call = gimple_build_call_internal | |
8287 | (IFN_GATHER_LOAD, 3, dataref_ptr, | |
8288 | vec_offset, scale); | |
8289 | gimple_call_set_nothrow (call, true); | |
8290 | new_stmt = call; | |
8291 | data_ref = NULL_TREE; | |
8292 | break; | |
8293 | } | |
8294 | ||
f702e7d4 | 8295 | align = DR_TARGET_ALIGNMENT (dr); |
272c6793 RS |
8296 | if (alignment_support_scheme == dr_aligned) |
8297 | { | |
8298 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 8299 | misalign = 0; |
272c6793 RS |
8300 | } |
8301 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
8302 | { | |
25f68d90 | 8303 | align = dr_alignment (vect_dr_behavior (first_dr)); |
52639a61 | 8304 | misalign = 0; |
272c6793 RS |
8305 | } |
8306 | else | |
c3a8f964 | 8307 | misalign = DR_MISALIGNMENT (first_dr); |
aed93b23 RB |
8308 | if (dataref_offset == NULL_TREE |
8309 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
8310 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), |
8311 | align, misalign); | |
c3a8f964 | 8312 | |
7cfb4d93 | 8313 | if (final_mask) |
c3a8f964 RS |
8314 | { |
8315 | align = least_bit_hwi (misalign | align); | |
8316 | tree ptr = build_int_cst (ref_type, align); | |
8317 | gcall *call | |
8318 | = gimple_build_call_internal (IFN_MASK_LOAD, 3, | |
8319 | dataref_ptr, ptr, | |
7cfb4d93 | 8320 | final_mask); |
c3a8f964 RS |
8321 | gimple_call_set_nothrow (call, true); |
8322 | new_stmt = call; | |
8323 | data_ref = NULL_TREE; | |
8324 | } | |
8325 | else | |
8326 | { | |
8327 | data_ref | |
8328 | = fold_build2 (MEM_REF, vectype, dataref_ptr, | |
8329 | dataref_offset | |
8330 | ? dataref_offset | |
8331 | : build_int_cst (ref_type, 0)); | |
8332 | if (alignment_support_scheme == dr_aligned) | |
8333 | ; | |
8334 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
8335 | TREE_TYPE (data_ref) | |
8336 | = build_aligned_type (TREE_TYPE (data_ref), | |
8337 | align * BITS_PER_UNIT); | |
8338 | else | |
8339 | TREE_TYPE (data_ref) | |
8340 | = build_aligned_type (TREE_TYPE (data_ref), | |
8341 | TYPE_ALIGN (elem_type)); | |
8342 | } | |
272c6793 | 8343 | break; |
be1ac4ec | 8344 | } |
272c6793 | 8345 | case dr_explicit_realign: |
267d3070 | 8346 | { |
272c6793 | 8347 | tree ptr, bump; |
272c6793 | 8348 | |
d88981fc | 8349 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
272c6793 RS |
8350 | |
8351 | if (compute_in_loop) | |
8352 | msq = vect_setup_realignment (first_stmt, gsi, | |
8353 | &realignment_token, | |
8354 | dr_explicit_realign, | |
8355 | dataref_ptr, NULL); | |
8356 | ||
aed93b23 RB |
8357 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
8358 | ptr = copy_ssa_name (dataref_ptr); | |
8359 | else | |
8360 | ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
f702e7d4 | 8361 | unsigned int align = DR_TARGET_ALIGNMENT (first_dr); |
0d0e4a03 JJ |
8362 | new_stmt = gimple_build_assign |
8363 | (ptr, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
8364 | build_int_cst |
8365 | (TREE_TYPE (dataref_ptr), | |
f702e7d4 | 8366 | -(HOST_WIDE_INT) align)); |
272c6793 RS |
8367 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
8368 | data_ref | |
8369 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 8370 | build_int_cst (ref_type, 0)); |
19986382 | 8371 | vect_copy_ref_info (data_ref, DR_REF (first_dr)); |
272c6793 RS |
8372 | vec_dest = vect_create_destination_var (scalar_dest, |
8373 | vectype); | |
8374 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
8375 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
8376 | gimple_assign_set_lhs (new_stmt, new_temp); | |
8377 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
8378 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
8379 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8380 | msq = new_temp; | |
8381 | ||
d88981fc | 8382 | bump = size_binop (MULT_EXPR, vs, |
7b7b1813 | 8383 | TYPE_SIZE_UNIT (elem_type)); |
d88981fc | 8384 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
272c6793 | 8385 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
0d0e4a03 JJ |
8386 | new_stmt = gimple_build_assign |
8387 | (NULL_TREE, BIT_AND_EXPR, ptr, | |
272c6793 | 8388 | build_int_cst |
f702e7d4 | 8389 | (TREE_TYPE (ptr), -(HOST_WIDE_INT) align)); |
aed93b23 | 8390 | ptr = copy_ssa_name (ptr, new_stmt); |
272c6793 RS |
8391 | gimple_assign_set_lhs (new_stmt, ptr); |
8392 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8393 | data_ref | |
8394 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 8395 | build_int_cst (ref_type, 0)); |
272c6793 | 8396 | break; |
267d3070 | 8397 | } |
272c6793 | 8398 | case dr_explicit_realign_optimized: |
f702e7d4 RS |
8399 | { |
8400 | if (TREE_CODE (dataref_ptr) == SSA_NAME) | |
8401 | new_temp = copy_ssa_name (dataref_ptr); | |
8402 | else | |
8403 | new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
8404 | unsigned int align = DR_TARGET_ALIGNMENT (first_dr); | |
8405 | new_stmt = gimple_build_assign | |
8406 | (new_temp, BIT_AND_EXPR, dataref_ptr, | |
8407 | build_int_cst (TREE_TYPE (dataref_ptr), | |
8408 | -(HOST_WIDE_INT) align)); | |
8409 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8410 | data_ref | |
8411 | = build2 (MEM_REF, vectype, new_temp, | |
8412 | build_int_cst (ref_type, 0)); | |
8413 | break; | |
8414 | } | |
272c6793 RS |
8415 | default: |
8416 | gcc_unreachable (); | |
8417 | } | |
ebfd146a | 8418 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
c3a8f964 RS |
8419 | /* DATA_REF is null if we've already built the statement. */ |
8420 | if (data_ref) | |
19986382 RB |
8421 | { |
8422 | vect_copy_ref_info (data_ref, DR_REF (first_dr)); | |
8423 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
8424 | } | |
ebfd146a | 8425 | new_temp = make_ssa_name (vec_dest, new_stmt); |
c3a8f964 | 8426 | gimple_set_lhs (new_stmt, new_temp); |
ebfd146a IR |
8427 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
8428 | ||
272c6793 RS |
8429 | /* 3. Handle explicit realignment if necessary/supported. |
8430 | Create in loop: | |
8431 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
8432 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
8433 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 8434 | { |
272c6793 RS |
8435 | lsq = gimple_assign_lhs (new_stmt); |
8436 | if (!realignment_token) | |
8437 | realignment_token = dataref_ptr; | |
8438 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
0d0e4a03 JJ |
8439 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, |
8440 | msq, lsq, realignment_token); | |
272c6793 RS |
8441 | new_temp = make_ssa_name (vec_dest, new_stmt); |
8442 | gimple_assign_set_lhs (new_stmt, new_temp); | |
8443 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8444 | ||
8445 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
8446 | { | |
8447 | gcc_assert (phi); | |
8448 | if (i == vec_num - 1 && j == ncopies - 1) | |
8449 | add_phi_arg (phi, lsq, | |
8450 | loop_latch_edge (containing_loop), | |
9e227d60 | 8451 | UNKNOWN_LOCATION); |
272c6793 RS |
8452 | msq = lsq; |
8453 | } | |
ebfd146a | 8454 | } |
ebfd146a | 8455 | |
59fd17e3 RB |
8456 | /* 4. Handle invariant-load. */ |
8457 | if (inv_p && !bb_vinfo) | |
8458 | { | |
59fd17e3 | 8459 | gcc_assert (!grouped_load); |
d1417442 JJ |
8460 | /* If we have versioned for aliasing or the loop doesn't |
8461 | have any data dependencies that would preclude this, | |
8462 | then we are sure this is a loop invariant load and | |
8463 | thus we can insert it on the preheader edge. */ | |
8464 | if (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) | |
8465 | && !nested_in_vect_loop | |
6b916b36 | 8466 | && hoist_defs_of_uses (stmt, loop)) |
a0e35eb0 RB |
8467 | { |
8468 | if (dump_enabled_p ()) | |
8469 | { | |
8470 | dump_printf_loc (MSG_NOTE, vect_location, | |
8471 | "hoisting out of the vectorized " | |
8472 | "loop: "); | |
8473 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
a0e35eb0 | 8474 | } |
b731b390 | 8475 | tree tem = copy_ssa_name (scalar_dest); |
a0e35eb0 RB |
8476 | gsi_insert_on_edge_immediate |
8477 | (loop_preheader_edge (loop), | |
8478 | gimple_build_assign (tem, | |
8479 | unshare_expr | |
8480 | (gimple_assign_rhs1 (stmt)))); | |
8481 | new_temp = vect_init_vector (stmt, tem, vectype, NULL); | |
34cd48e5 RB |
8482 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
8483 | set_vinfo_for_stmt (new_stmt, | |
8484 | new_stmt_vec_info (new_stmt, vinfo)); | |
a0e35eb0 RB |
8485 | } |
8486 | else | |
8487 | { | |
8488 | gimple_stmt_iterator gsi2 = *gsi; | |
8489 | gsi_next (&gsi2); | |
8490 | new_temp = vect_init_vector (stmt, scalar_dest, | |
8491 | vectype, &gsi2); | |
34cd48e5 | 8492 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
a0e35eb0 | 8493 | } |
59fd17e3 RB |
8494 | } |
8495 | ||
62da9e14 | 8496 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
272c6793 | 8497 | { |
aec7ae7d JJ |
8498 | tree perm_mask = perm_mask_for_reverse (vectype); |
8499 | new_temp = permute_vec_elements (new_temp, new_temp, | |
8500 | perm_mask, stmt, gsi); | |
ebfd146a IR |
8501 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
8502 | } | |
267d3070 | 8503 | |
272c6793 | 8504 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
8505 | vect_transform_grouped_load (). */ |
8506 | if (grouped_load || slp_perm) | |
9771b263 | 8507 | dr_chain.quick_push (new_temp); |
267d3070 | 8508 | |
272c6793 RS |
8509 | /* Store vector loads in the corresponding SLP_NODE. */ |
8510 | if (slp && !slp_perm) | |
9771b263 | 8511 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
b267968e RB |
8512 | |
8513 | /* With SLP permutation we load the gaps as well, without | |
8514 | we need to skip the gaps after we manage to fully load | |
2c53b149 | 8515 | all elements. group_gap_adj is DR_GROUP_SIZE here. */ |
b267968e | 8516 | group_elt += nunits; |
d9f21f6a RS |
8517 | if (maybe_ne (group_gap_adj, 0U) |
8518 | && !slp_perm | |
8519 | && known_eq (group_elt, group_size - group_gap_adj)) | |
b267968e | 8520 | { |
d9f21f6a RS |
8521 | poly_wide_int bump_val |
8522 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) | |
8523 | * group_gap_adj); | |
8e6cdc90 | 8524 | tree bump = wide_int_to_tree (sizetype, bump_val); |
b267968e RB |
8525 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
8526 | stmt, bump); | |
8527 | group_elt = 0; | |
8528 | } | |
272c6793 | 8529 | } |
9b999e8c RB |
8530 | /* Bump the vector pointer to account for a gap or for excess |
8531 | elements loaded for a permuted SLP load. */ | |
d9f21f6a | 8532 | if (maybe_ne (group_gap_adj, 0U) && slp_perm) |
a64b9c26 | 8533 | { |
d9f21f6a RS |
8534 | poly_wide_int bump_val |
8535 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) | |
8536 | * group_gap_adj); | |
8e6cdc90 | 8537 | tree bump = wide_int_to_tree (sizetype, bump_val); |
a64b9c26 RB |
8538 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
8539 | stmt, bump); | |
8540 | } | |
ebfd146a IR |
8541 | } |
8542 | ||
8543 | if (slp && !slp_perm) | |
8544 | continue; | |
8545 | ||
8546 | if (slp_perm) | |
8547 | { | |
29afecdf | 8548 | unsigned n_perms; |
01d8bf07 | 8549 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
29afecdf RB |
8550 | slp_node_instance, false, |
8551 | &n_perms)) | |
ebfd146a | 8552 | { |
9771b263 | 8553 | dr_chain.release (); |
ebfd146a IR |
8554 | return false; |
8555 | } | |
8556 | } | |
8557 | else | |
8558 | { | |
0d0293ac | 8559 | if (grouped_load) |
ebfd146a | 8560 | { |
2de001ee | 8561 | if (memory_access_type != VMAT_LOAD_STORE_LANES) |
0d0293ac | 8562 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 8563 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
8564 | } |
8565 | else | |
8566 | { | |
8567 | if (j == 0) | |
8568 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
8569 | else | |
8570 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
8571 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
8572 | } | |
8573 | } | |
9771b263 | 8574 | dr_chain.release (); |
ebfd146a IR |
8575 | } |
8576 | ||
ebfd146a IR |
8577 | return true; |
8578 | } | |
8579 | ||
8580 | /* Function vect_is_simple_cond. | |
b8698a0f | 8581 | |
ebfd146a IR |
8582 | Input: |
8583 | LOOP - the loop that is being vectorized. | |
8584 | COND - Condition that is checked for simple use. | |
8585 | ||
e9e1d143 RG |
8586 | Output: |
8587 | *COMP_VECTYPE - the vector type for the comparison. | |
4fc5ebf1 | 8588 | *DTS - The def types for the arguments of the comparison |
e9e1d143 | 8589 | |
ebfd146a IR |
8590 | Returns whether a COND can be vectorized. Checks whether |
8591 | condition operands are supportable using vec_is_simple_use. */ | |
8592 | ||
87aab9b2 | 8593 | static bool |
4fc5ebf1 | 8594 | vect_is_simple_cond (tree cond, vec_info *vinfo, |
8da4c8d8 RB |
8595 | tree *comp_vectype, enum vect_def_type *dts, |
8596 | tree vectype) | |
ebfd146a IR |
8597 | { |
8598 | tree lhs, rhs; | |
e9e1d143 | 8599 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a | 8600 | |
a414c77f IE |
8601 | /* Mask case. */ |
8602 | if (TREE_CODE (cond) == SSA_NAME | |
2568d8a1 | 8603 | && VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (cond))) |
a414c77f IE |
8604 | { |
8605 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (cond); | |
8606 | if (!vect_is_simple_use (cond, vinfo, &lhs_def_stmt, | |
4fc5ebf1 | 8607 | &dts[0], comp_vectype) |
a414c77f IE |
8608 | || !*comp_vectype |
8609 | || !VECTOR_BOOLEAN_TYPE_P (*comp_vectype)) | |
8610 | return false; | |
8611 | return true; | |
8612 | } | |
8613 | ||
ebfd146a IR |
8614 | if (!COMPARISON_CLASS_P (cond)) |
8615 | return false; | |
8616 | ||
8617 | lhs = TREE_OPERAND (cond, 0); | |
8618 | rhs = TREE_OPERAND (cond, 1); | |
8619 | ||
8620 | if (TREE_CODE (lhs) == SSA_NAME) | |
8621 | { | |
355fe088 | 8622 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); |
4fc5ebf1 | 8623 | if (!vect_is_simple_use (lhs, vinfo, &lhs_def_stmt, &dts[0], &vectype1)) |
ebfd146a IR |
8624 | return false; |
8625 | } | |
4fc5ebf1 JG |
8626 | else if (TREE_CODE (lhs) == INTEGER_CST || TREE_CODE (lhs) == REAL_CST |
8627 | || TREE_CODE (lhs) == FIXED_CST) | |
8628 | dts[0] = vect_constant_def; | |
8629 | else | |
ebfd146a IR |
8630 | return false; |
8631 | ||
8632 | if (TREE_CODE (rhs) == SSA_NAME) | |
8633 | { | |
355fe088 | 8634 | gimple *rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); |
4fc5ebf1 | 8635 | if (!vect_is_simple_use (rhs, vinfo, &rhs_def_stmt, &dts[1], &vectype2)) |
ebfd146a IR |
8636 | return false; |
8637 | } | |
4fc5ebf1 JG |
8638 | else if (TREE_CODE (rhs) == INTEGER_CST || TREE_CODE (rhs) == REAL_CST |
8639 | || TREE_CODE (rhs) == FIXED_CST) | |
8640 | dts[1] = vect_constant_def; | |
8641 | else | |
ebfd146a IR |
8642 | return false; |
8643 | ||
28b33016 | 8644 | if (vectype1 && vectype2 |
928686b1 RS |
8645 | && maybe_ne (TYPE_VECTOR_SUBPARTS (vectype1), |
8646 | TYPE_VECTOR_SUBPARTS (vectype2))) | |
28b33016 IE |
8647 | return false; |
8648 | ||
e9e1d143 | 8649 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
8da4c8d8 | 8650 | /* Invariant comparison. */ |
4515e413 | 8651 | if (! *comp_vectype && vectype) |
8da4c8d8 RB |
8652 | { |
8653 | tree scalar_type = TREE_TYPE (lhs); | |
8654 | /* If we can widen the comparison to match vectype do so. */ | |
8655 | if (INTEGRAL_TYPE_P (scalar_type) | |
8656 | && tree_int_cst_lt (TYPE_SIZE (scalar_type), | |
8657 | TYPE_SIZE (TREE_TYPE (vectype)))) | |
8658 | scalar_type = build_nonstandard_integer_type | |
8659 | (tree_to_uhwi (TYPE_SIZE (TREE_TYPE (vectype))), | |
8660 | TYPE_UNSIGNED (scalar_type)); | |
8661 | *comp_vectype = get_vectype_for_scalar_type (scalar_type); | |
8662 | } | |
8663 | ||
ebfd146a IR |
8664 | return true; |
8665 | } | |
8666 | ||
8667 | /* vectorizable_condition. | |
8668 | ||
b8698a0f L |
8669 | Check if STMT is conditional modify expression that can be vectorized. |
8670 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
8671 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
8672 | at GSI. |
8673 | ||
8674 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
8675 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
0ad23163 | 8676 | else clause if it is 2). |
ebfd146a IR |
8677 | |
8678 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
8679 | ||
4bbe8262 | 8680 | bool |
355fe088 TS |
8681 | vectorizable_condition (gimple *stmt, gimple_stmt_iterator *gsi, |
8682 | gimple **vec_stmt, tree reduc_def, int reduc_index, | |
68435eb2 | 8683 | slp_tree slp_node, stmt_vector_for_cost *cost_vec) |
ebfd146a IR |
8684 | { |
8685 | tree scalar_dest = NULL_TREE; | |
8686 | tree vec_dest = NULL_TREE; | |
01216d27 JJ |
8687 | tree cond_expr, cond_expr0 = NULL_TREE, cond_expr1 = NULL_TREE; |
8688 | tree then_clause, else_clause; | |
ebfd146a | 8689 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
df11cc78 | 8690 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
8691 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
8692 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
5958f9e2 | 8693 | tree vec_compare; |
ebfd146a IR |
8694 | tree new_temp; |
8695 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4fc5ebf1 JG |
8696 | enum vect_def_type dts[4] |
8697 | = {vect_unknown_def_type, vect_unknown_def_type, | |
8698 | vect_unknown_def_type, vect_unknown_def_type}; | |
8699 | int ndts = 4; | |
f7e531cf | 8700 | int ncopies; |
01216d27 | 8701 | enum tree_code code, cond_code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
a855b1b1 | 8702 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
8703 | int i, j; |
8704 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
8705 | vec<tree> vec_oprnds0 = vNULL; |
8706 | vec<tree> vec_oprnds1 = vNULL; | |
8707 | vec<tree> vec_oprnds2 = vNULL; | |
8708 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 8709 | tree vec_cmp_type; |
a414c77f | 8710 | bool masked = false; |
b8698a0f | 8711 | |
f7e531cf IR |
8712 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
8713 | return false; | |
8714 | ||
bb6c2b68 RS |
8715 | vect_reduction_type reduction_type |
8716 | = STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info); | |
8717 | if (reduction_type == TREE_CODE_REDUCTION) | |
af29617a AH |
8718 | { |
8719 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
8720 | return false; | |
ebfd146a | 8721 | |
af29617a AH |
8722 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
8723 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
8724 | && reduc_def)) | |
8725 | return false; | |
ebfd146a | 8726 | |
af29617a AH |
8727 | /* FORNOW: not yet supported. */ |
8728 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
8729 | { | |
8730 | if (dump_enabled_p ()) | |
8731 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
8732 | "value used after loop.\n"); | |
8733 | return false; | |
8734 | } | |
ebfd146a IR |
8735 | } |
8736 | ||
8737 | /* Is vectorizable conditional operation? */ | |
8738 | if (!is_gimple_assign (stmt)) | |
8739 | return false; | |
8740 | ||
8741 | code = gimple_assign_rhs_code (stmt); | |
8742 | ||
8743 | if (code != COND_EXPR) | |
8744 | return false; | |
8745 | ||
465c8c19 | 8746 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
2947d3b2 | 8747 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
465c8c19 | 8748 | |
fce57248 | 8749 | if (slp_node) |
465c8c19 JJ |
8750 | ncopies = 1; |
8751 | else | |
e8f142e2 | 8752 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
8753 | |
8754 | gcc_assert (ncopies >= 1); | |
8755 | if (reduc_index && ncopies > 1) | |
8756 | return false; /* FORNOW */ | |
8757 | ||
4e71066d RG |
8758 | cond_expr = gimple_assign_rhs1 (stmt); |
8759 | then_clause = gimple_assign_rhs2 (stmt); | |
8760 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 8761 | |
4fc5ebf1 | 8762 | if (!vect_is_simple_cond (cond_expr, stmt_info->vinfo, |
4515e413 | 8763 | &comp_vectype, &dts[0], slp_node ? NULL : vectype) |
e9e1d143 | 8764 | || !comp_vectype) |
ebfd146a IR |
8765 | return false; |
8766 | ||
81c40241 | 8767 | gimple *def_stmt; |
4fc5ebf1 | 8768 | if (!vect_is_simple_use (then_clause, stmt_info->vinfo, &def_stmt, &dts[2], |
2947d3b2 IE |
8769 | &vectype1)) |
8770 | return false; | |
4fc5ebf1 | 8771 | if (!vect_is_simple_use (else_clause, stmt_info->vinfo, &def_stmt, &dts[3], |
2947d3b2 | 8772 | &vectype2)) |
ebfd146a | 8773 | return false; |
2947d3b2 IE |
8774 | |
8775 | if (vectype1 && !useless_type_conversion_p (vectype, vectype1)) | |
8776 | return false; | |
8777 | ||
8778 | if (vectype2 && !useless_type_conversion_p (vectype, vectype2)) | |
ebfd146a IR |
8779 | return false; |
8780 | ||
28b33016 IE |
8781 | masked = !COMPARISON_CLASS_P (cond_expr); |
8782 | vec_cmp_type = build_same_sized_truth_vector_type (comp_vectype); | |
8783 | ||
74946978 MP |
8784 | if (vec_cmp_type == NULL_TREE) |
8785 | return false; | |
784fb9b3 | 8786 | |
01216d27 JJ |
8787 | cond_code = TREE_CODE (cond_expr); |
8788 | if (!masked) | |
8789 | { | |
8790 | cond_expr0 = TREE_OPERAND (cond_expr, 0); | |
8791 | cond_expr1 = TREE_OPERAND (cond_expr, 1); | |
8792 | } | |
8793 | ||
8794 | if (!masked && VECTOR_BOOLEAN_TYPE_P (comp_vectype)) | |
8795 | { | |
8796 | /* Boolean values may have another representation in vectors | |
8797 | and therefore we prefer bit operations over comparison for | |
8798 | them (which also works for scalar masks). We store opcodes | |
8799 | to use in bitop1 and bitop2. Statement is vectorized as | |
8800 | BITOP2 (rhs1 BITOP1 rhs2) or rhs1 BITOP2 (BITOP1 rhs2) | |
8801 | depending on bitop1 and bitop2 arity. */ | |
8802 | switch (cond_code) | |
8803 | { | |
8804 | case GT_EXPR: | |
8805 | bitop1 = BIT_NOT_EXPR; | |
8806 | bitop2 = BIT_AND_EXPR; | |
8807 | break; | |
8808 | case GE_EXPR: | |
8809 | bitop1 = BIT_NOT_EXPR; | |
8810 | bitop2 = BIT_IOR_EXPR; | |
8811 | break; | |
8812 | case LT_EXPR: | |
8813 | bitop1 = BIT_NOT_EXPR; | |
8814 | bitop2 = BIT_AND_EXPR; | |
8815 | std::swap (cond_expr0, cond_expr1); | |
8816 | break; | |
8817 | case LE_EXPR: | |
8818 | bitop1 = BIT_NOT_EXPR; | |
8819 | bitop2 = BIT_IOR_EXPR; | |
8820 | std::swap (cond_expr0, cond_expr1); | |
8821 | break; | |
8822 | case NE_EXPR: | |
8823 | bitop1 = BIT_XOR_EXPR; | |
8824 | break; | |
8825 | case EQ_EXPR: | |
8826 | bitop1 = BIT_XOR_EXPR; | |
8827 | bitop2 = BIT_NOT_EXPR; | |
8828 | break; | |
8829 | default: | |
8830 | return false; | |
8831 | } | |
8832 | cond_code = SSA_NAME; | |
8833 | } | |
8834 | ||
b8698a0f | 8835 | if (!vec_stmt) |
ebfd146a | 8836 | { |
01216d27 JJ |
8837 | if (bitop1 != NOP_EXPR) |
8838 | { | |
8839 | machine_mode mode = TYPE_MODE (comp_vectype); | |
8840 | optab optab; | |
8841 | ||
8842 | optab = optab_for_tree_code (bitop1, comp_vectype, optab_default); | |
8843 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8844 | return false; | |
8845 | ||
8846 | if (bitop2 != NOP_EXPR) | |
8847 | { | |
8848 | optab = optab_for_tree_code (bitop2, comp_vectype, | |
8849 | optab_default); | |
8850 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8851 | return false; | |
8852 | } | |
8853 | } | |
4fc5ebf1 JG |
8854 | if (expand_vec_cond_expr_p (vectype, comp_vectype, |
8855 | cond_code)) | |
8856 | { | |
68435eb2 RB |
8857 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; |
8858 | vect_model_simple_cost (stmt_info, ncopies, dts, ndts, slp_node, | |
8859 | cost_vec); | |
4fc5ebf1 JG |
8860 | return true; |
8861 | } | |
8862 | return false; | |
ebfd146a IR |
8863 | } |
8864 | ||
f7e531cf IR |
8865 | /* Transform. */ |
8866 | ||
8867 | if (!slp_node) | |
8868 | { | |
9771b263 DN |
8869 | vec_oprnds0.create (1); |
8870 | vec_oprnds1.create (1); | |
8871 | vec_oprnds2.create (1); | |
8872 | vec_oprnds3.create (1); | |
f7e531cf | 8873 | } |
ebfd146a IR |
8874 | |
8875 | /* Handle def. */ | |
8876 | scalar_dest = gimple_assign_lhs (stmt); | |
bb6c2b68 RS |
8877 | if (reduction_type != EXTRACT_LAST_REDUCTION) |
8878 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
ebfd146a IR |
8879 | |
8880 | /* Handle cond expr. */ | |
a855b1b1 MM |
8881 | for (j = 0; j < ncopies; j++) |
8882 | { | |
bb6c2b68 | 8883 | gimple *new_stmt = NULL; |
a855b1b1 MM |
8884 | if (j == 0) |
8885 | { | |
f7e531cf IR |
8886 | if (slp_node) |
8887 | { | |
00f96dc9 TS |
8888 | auto_vec<tree, 4> ops; |
8889 | auto_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 8890 | |
a414c77f | 8891 | if (masked) |
01216d27 | 8892 | ops.safe_push (cond_expr); |
a414c77f IE |
8893 | else |
8894 | { | |
01216d27 JJ |
8895 | ops.safe_push (cond_expr0); |
8896 | ops.safe_push (cond_expr1); | |
a414c77f | 8897 | } |
9771b263 DN |
8898 | ops.safe_push (then_clause); |
8899 | ops.safe_push (else_clause); | |
306b0c92 | 8900 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
37b5ec8f JJ |
8901 | vec_oprnds3 = vec_defs.pop (); |
8902 | vec_oprnds2 = vec_defs.pop (); | |
a414c77f IE |
8903 | if (!masked) |
8904 | vec_oprnds1 = vec_defs.pop (); | |
37b5ec8f | 8905 | vec_oprnds0 = vec_defs.pop (); |
f7e531cf IR |
8906 | } |
8907 | else | |
8908 | { | |
355fe088 | 8909 | gimple *gtemp; |
a414c77f IE |
8910 | if (masked) |
8911 | { | |
8912 | vec_cond_lhs | |
8913 | = vect_get_vec_def_for_operand (cond_expr, stmt, | |
8914 | comp_vectype); | |
8915 | vect_is_simple_use (cond_expr, stmt_info->vinfo, | |
8916 | >emp, &dts[0]); | |
8917 | } | |
8918 | else | |
8919 | { | |
01216d27 JJ |
8920 | vec_cond_lhs |
8921 | = vect_get_vec_def_for_operand (cond_expr0, | |
8922 | stmt, comp_vectype); | |
8923 | vect_is_simple_use (cond_expr0, loop_vinfo, >emp, &dts[0]); | |
8924 | ||
8925 | vec_cond_rhs | |
8926 | = vect_get_vec_def_for_operand (cond_expr1, | |
8927 | stmt, comp_vectype); | |
8928 | vect_is_simple_use (cond_expr1, loop_vinfo, >emp, &dts[1]); | |
a414c77f | 8929 | } |
f7e531cf IR |
8930 | if (reduc_index == 1) |
8931 | vec_then_clause = reduc_def; | |
8932 | else | |
8933 | { | |
8934 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
81c40241 RB |
8935 | stmt); |
8936 | vect_is_simple_use (then_clause, loop_vinfo, | |
8937 | >emp, &dts[2]); | |
f7e531cf IR |
8938 | } |
8939 | if (reduc_index == 2) | |
8940 | vec_else_clause = reduc_def; | |
8941 | else | |
8942 | { | |
8943 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
81c40241 RB |
8944 | stmt); |
8945 | vect_is_simple_use (else_clause, loop_vinfo, >emp, &dts[3]); | |
f7e531cf | 8946 | } |
a855b1b1 MM |
8947 | } |
8948 | } | |
8949 | else | |
8950 | { | |
a414c77f IE |
8951 | vec_cond_lhs |
8952 | = vect_get_vec_def_for_stmt_copy (dts[0], | |
8953 | vec_oprnds0.pop ()); | |
8954 | if (!masked) | |
8955 | vec_cond_rhs | |
8956 | = vect_get_vec_def_for_stmt_copy (dts[1], | |
8957 | vec_oprnds1.pop ()); | |
8958 | ||
a855b1b1 | 8959 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 8960 | vec_oprnds2.pop ()); |
a855b1b1 | 8961 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 8962 | vec_oprnds3.pop ()); |
f7e531cf IR |
8963 | } |
8964 | ||
8965 | if (!slp_node) | |
8966 | { | |
9771b263 | 8967 | vec_oprnds0.quick_push (vec_cond_lhs); |
a414c77f IE |
8968 | if (!masked) |
8969 | vec_oprnds1.quick_push (vec_cond_rhs); | |
9771b263 DN |
8970 | vec_oprnds2.quick_push (vec_then_clause); |
8971 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
8972 | } |
8973 | ||
9dc3f7de | 8974 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 8975 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 8976 | { |
9771b263 DN |
8977 | vec_then_clause = vec_oprnds2[i]; |
8978 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 8979 | |
a414c77f IE |
8980 | if (masked) |
8981 | vec_compare = vec_cond_lhs; | |
8982 | else | |
8983 | { | |
8984 | vec_cond_rhs = vec_oprnds1[i]; | |
01216d27 JJ |
8985 | if (bitop1 == NOP_EXPR) |
8986 | vec_compare = build2 (cond_code, vec_cmp_type, | |
8987 | vec_cond_lhs, vec_cond_rhs); | |
8988 | else | |
8989 | { | |
8990 | new_temp = make_ssa_name (vec_cmp_type); | |
8991 | if (bitop1 == BIT_NOT_EXPR) | |
8992 | new_stmt = gimple_build_assign (new_temp, bitop1, | |
8993 | vec_cond_rhs); | |
8994 | else | |
8995 | new_stmt | |
8996 | = gimple_build_assign (new_temp, bitop1, vec_cond_lhs, | |
8997 | vec_cond_rhs); | |
8998 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8999 | if (bitop2 == NOP_EXPR) | |
9000 | vec_compare = new_temp; | |
9001 | else if (bitop2 == BIT_NOT_EXPR) | |
9002 | { | |
9003 | /* Instead of doing ~x ? y : z do x ? z : y. */ | |
9004 | vec_compare = new_temp; | |
9005 | std::swap (vec_then_clause, vec_else_clause); | |
9006 | } | |
9007 | else | |
9008 | { | |
9009 | vec_compare = make_ssa_name (vec_cmp_type); | |
9010 | new_stmt | |
9011 | = gimple_build_assign (vec_compare, bitop2, | |
9012 | vec_cond_lhs, new_temp); | |
9013 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9014 | } | |
9015 | } | |
a414c77f | 9016 | } |
bb6c2b68 RS |
9017 | if (reduction_type == EXTRACT_LAST_REDUCTION) |
9018 | { | |
9019 | if (!is_gimple_val (vec_compare)) | |
9020 | { | |
9021 | tree vec_compare_name = make_ssa_name (vec_cmp_type); | |
9022 | new_stmt = gimple_build_assign (vec_compare_name, | |
9023 | vec_compare); | |
9024 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9025 | vec_compare = vec_compare_name; | |
9026 | } | |
9027 | gcc_assert (reduc_index == 2); | |
9028 | new_stmt = gimple_build_call_internal | |
9029 | (IFN_FOLD_EXTRACT_LAST, 3, else_clause, vec_compare, | |
9030 | vec_then_clause); | |
9031 | gimple_call_set_lhs (new_stmt, scalar_dest); | |
9032 | SSA_NAME_DEF_STMT (scalar_dest) = new_stmt; | |
9033 | if (stmt == gsi_stmt (*gsi)) | |
9034 | vect_finish_replace_stmt (stmt, new_stmt); | |
9035 | else | |
9036 | { | |
9037 | /* In this case we're moving the definition to later in the | |
9038 | block. That doesn't matter because the only uses of the | |
9039 | lhs are in phi statements. */ | |
9040 | gimple_stmt_iterator old_gsi = gsi_for_stmt (stmt); | |
9041 | gsi_remove (&old_gsi, true); | |
9042 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9043 | } | |
9044 | } | |
9045 | else | |
9046 | { | |
9047 | new_temp = make_ssa_name (vec_dest); | |
9048 | new_stmt = gimple_build_assign (new_temp, VEC_COND_EXPR, | |
9049 | vec_compare, vec_then_clause, | |
9050 | vec_else_clause); | |
9051 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9052 | } | |
f7e531cf | 9053 | if (slp_node) |
9771b263 | 9054 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
9055 | } |
9056 | ||
9057 | if (slp_node) | |
9058 | continue; | |
9059 | ||
9060 | if (j == 0) | |
9061 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
9062 | else | |
9063 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
9064 | ||
9065 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 9066 | } |
b8698a0f | 9067 | |
9771b263 DN |
9068 | vec_oprnds0.release (); |
9069 | vec_oprnds1.release (); | |
9070 | vec_oprnds2.release (); | |
9071 | vec_oprnds3.release (); | |
f7e531cf | 9072 | |
ebfd146a IR |
9073 | return true; |
9074 | } | |
9075 | ||
42fd8198 IE |
9076 | /* vectorizable_comparison. |
9077 | ||
9078 | Check if STMT is comparison expression that can be vectorized. | |
9079 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
9080 | comparison, put it in VEC_STMT, and insert it at GSI. | |
9081 | ||
9082 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
9083 | ||
fce57248 | 9084 | static bool |
42fd8198 IE |
9085 | vectorizable_comparison (gimple *stmt, gimple_stmt_iterator *gsi, |
9086 | gimple **vec_stmt, tree reduc_def, | |
68435eb2 | 9087 | slp_tree slp_node, stmt_vector_for_cost *cost_vec) |
42fd8198 IE |
9088 | { |
9089 | tree lhs, rhs1, rhs2; | |
9090 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
9091 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; | |
9092 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
9093 | tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE; | |
9094 | tree new_temp; | |
9095 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
9096 | enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 9097 | int ndts = 2; |
928686b1 | 9098 | poly_uint64 nunits; |
42fd8198 | 9099 | int ncopies; |
49e76ff1 | 9100 | enum tree_code code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
42fd8198 IE |
9101 | stmt_vec_info prev_stmt_info = NULL; |
9102 | int i, j; | |
9103 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
9104 | vec<tree> vec_oprnds0 = vNULL; | |
9105 | vec<tree> vec_oprnds1 = vNULL; | |
9106 | gimple *def_stmt; | |
9107 | tree mask_type; | |
9108 | tree mask; | |
9109 | ||
c245362b IE |
9110 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
9111 | return false; | |
9112 | ||
30480bcd | 9113 | if (!vectype || !VECTOR_BOOLEAN_TYPE_P (vectype)) |
42fd8198 IE |
9114 | return false; |
9115 | ||
9116 | mask_type = vectype; | |
9117 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
9118 | ||
fce57248 | 9119 | if (slp_node) |
42fd8198 IE |
9120 | ncopies = 1; |
9121 | else | |
e8f142e2 | 9122 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
42fd8198 IE |
9123 | |
9124 | gcc_assert (ncopies >= 1); | |
42fd8198 IE |
9125 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
9126 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
9127 | && reduc_def)) | |
9128 | return false; | |
9129 | ||
9130 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
9131 | { | |
9132 | if (dump_enabled_p ()) | |
9133 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
9134 | "value used after loop.\n"); | |
9135 | return false; | |
9136 | } | |
9137 | ||
9138 | if (!is_gimple_assign (stmt)) | |
9139 | return false; | |
9140 | ||
9141 | code = gimple_assign_rhs_code (stmt); | |
9142 | ||
9143 | if (TREE_CODE_CLASS (code) != tcc_comparison) | |
9144 | return false; | |
9145 | ||
9146 | rhs1 = gimple_assign_rhs1 (stmt); | |
9147 | rhs2 = gimple_assign_rhs2 (stmt); | |
9148 | ||
9149 | if (!vect_is_simple_use (rhs1, stmt_info->vinfo, &def_stmt, | |
9150 | &dts[0], &vectype1)) | |
9151 | return false; | |
9152 | ||
9153 | if (!vect_is_simple_use (rhs2, stmt_info->vinfo, &def_stmt, | |
9154 | &dts[1], &vectype2)) | |
9155 | return false; | |
9156 | ||
9157 | if (vectype1 && vectype2 | |
928686b1 RS |
9158 | && maybe_ne (TYPE_VECTOR_SUBPARTS (vectype1), |
9159 | TYPE_VECTOR_SUBPARTS (vectype2))) | |
42fd8198 IE |
9160 | return false; |
9161 | ||
9162 | vectype = vectype1 ? vectype1 : vectype2; | |
9163 | ||
9164 | /* Invariant comparison. */ | |
9165 | if (!vectype) | |
9166 | { | |
69a9a66f | 9167 | vectype = get_vectype_for_scalar_type (TREE_TYPE (rhs1)); |
928686b1 | 9168 | if (maybe_ne (TYPE_VECTOR_SUBPARTS (vectype), nunits)) |
42fd8198 IE |
9169 | return false; |
9170 | } | |
928686b1 | 9171 | else if (maybe_ne (nunits, TYPE_VECTOR_SUBPARTS (vectype))) |
42fd8198 IE |
9172 | return false; |
9173 | ||
49e76ff1 IE |
9174 | /* Can't compare mask and non-mask types. */ |
9175 | if (vectype1 && vectype2 | |
9176 | && (VECTOR_BOOLEAN_TYPE_P (vectype1) ^ VECTOR_BOOLEAN_TYPE_P (vectype2))) | |
9177 | return false; | |
9178 | ||
9179 | /* Boolean values may have another representation in vectors | |
9180 | and therefore we prefer bit operations over comparison for | |
9181 | them (which also works for scalar masks). We store opcodes | |
9182 | to use in bitop1 and bitop2. Statement is vectorized as | |
9183 | BITOP2 (rhs1 BITOP1 rhs2) or | |
9184 | rhs1 BITOP2 (BITOP1 rhs2) | |
9185 | depending on bitop1 and bitop2 arity. */ | |
9186 | if (VECTOR_BOOLEAN_TYPE_P (vectype)) | |
9187 | { | |
9188 | if (code == GT_EXPR) | |
9189 | { | |
9190 | bitop1 = BIT_NOT_EXPR; | |
9191 | bitop2 = BIT_AND_EXPR; | |
9192 | } | |
9193 | else if (code == GE_EXPR) | |
9194 | { | |
9195 | bitop1 = BIT_NOT_EXPR; | |
9196 | bitop2 = BIT_IOR_EXPR; | |
9197 | } | |
9198 | else if (code == LT_EXPR) | |
9199 | { | |
9200 | bitop1 = BIT_NOT_EXPR; | |
9201 | bitop2 = BIT_AND_EXPR; | |
9202 | std::swap (rhs1, rhs2); | |
264d951a | 9203 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
9204 | } |
9205 | else if (code == LE_EXPR) | |
9206 | { | |
9207 | bitop1 = BIT_NOT_EXPR; | |
9208 | bitop2 = BIT_IOR_EXPR; | |
9209 | std::swap (rhs1, rhs2); | |
264d951a | 9210 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
9211 | } |
9212 | else | |
9213 | { | |
9214 | bitop1 = BIT_XOR_EXPR; | |
9215 | if (code == EQ_EXPR) | |
9216 | bitop2 = BIT_NOT_EXPR; | |
9217 | } | |
9218 | } | |
9219 | ||
42fd8198 IE |
9220 | if (!vec_stmt) |
9221 | { | |
49e76ff1 | 9222 | if (bitop1 == NOP_EXPR) |
68435eb2 RB |
9223 | { |
9224 | if (!expand_vec_cmp_expr_p (vectype, mask_type, code)) | |
9225 | return false; | |
9226 | } | |
49e76ff1 IE |
9227 | else |
9228 | { | |
9229 | machine_mode mode = TYPE_MODE (vectype); | |
9230 | optab optab; | |
9231 | ||
9232 | optab = optab_for_tree_code (bitop1, vectype, optab_default); | |
9233 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
9234 | return false; | |
9235 | ||
9236 | if (bitop2 != NOP_EXPR) | |
9237 | { | |
9238 | optab = optab_for_tree_code (bitop2, vectype, optab_default); | |
9239 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
9240 | return false; | |
9241 | } | |
49e76ff1 | 9242 | } |
68435eb2 RB |
9243 | |
9244 | STMT_VINFO_TYPE (stmt_info) = comparison_vec_info_type; | |
9245 | vect_model_simple_cost (stmt_info, ncopies * (1 + (bitop2 != NOP_EXPR)), | |
9246 | dts, ndts, slp_node, cost_vec); | |
9247 | return true; | |
42fd8198 IE |
9248 | } |
9249 | ||
9250 | /* Transform. */ | |
9251 | if (!slp_node) | |
9252 | { | |
9253 | vec_oprnds0.create (1); | |
9254 | vec_oprnds1.create (1); | |
9255 | } | |
9256 | ||
9257 | /* Handle def. */ | |
9258 | lhs = gimple_assign_lhs (stmt); | |
9259 | mask = vect_create_destination_var (lhs, mask_type); | |
9260 | ||
9261 | /* Handle cmp expr. */ | |
9262 | for (j = 0; j < ncopies; j++) | |
9263 | { | |
9264 | gassign *new_stmt = NULL; | |
9265 | if (j == 0) | |
9266 | { | |
9267 | if (slp_node) | |
9268 | { | |
9269 | auto_vec<tree, 2> ops; | |
9270 | auto_vec<vec<tree>, 2> vec_defs; | |
9271 | ||
9272 | ops.safe_push (rhs1); | |
9273 | ops.safe_push (rhs2); | |
306b0c92 | 9274 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
42fd8198 IE |
9275 | vec_oprnds1 = vec_defs.pop (); |
9276 | vec_oprnds0 = vec_defs.pop (); | |
9277 | } | |
9278 | else | |
9279 | { | |
e4af0bc4 IE |
9280 | vec_rhs1 = vect_get_vec_def_for_operand (rhs1, stmt, vectype); |
9281 | vec_rhs2 = vect_get_vec_def_for_operand (rhs2, stmt, vectype); | |
42fd8198 IE |
9282 | } |
9283 | } | |
9284 | else | |
9285 | { | |
9286 | vec_rhs1 = vect_get_vec_def_for_stmt_copy (dts[0], | |
9287 | vec_oprnds0.pop ()); | |
9288 | vec_rhs2 = vect_get_vec_def_for_stmt_copy (dts[1], | |
9289 | vec_oprnds1.pop ()); | |
9290 | } | |
9291 | ||
9292 | if (!slp_node) | |
9293 | { | |
9294 | vec_oprnds0.quick_push (vec_rhs1); | |
9295 | vec_oprnds1.quick_push (vec_rhs2); | |
9296 | } | |
9297 | ||
9298 | /* Arguments are ready. Create the new vector stmt. */ | |
9299 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1) | |
9300 | { | |
9301 | vec_rhs2 = vec_oprnds1[i]; | |
9302 | ||
9303 | new_temp = make_ssa_name (mask); | |
49e76ff1 IE |
9304 | if (bitop1 == NOP_EXPR) |
9305 | { | |
9306 | new_stmt = gimple_build_assign (new_temp, code, | |
9307 | vec_rhs1, vec_rhs2); | |
9308 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9309 | } | |
9310 | else | |
9311 | { | |
9312 | if (bitop1 == BIT_NOT_EXPR) | |
9313 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs2); | |
9314 | else | |
9315 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs1, | |
9316 | vec_rhs2); | |
9317 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9318 | if (bitop2 != NOP_EXPR) | |
9319 | { | |
9320 | tree res = make_ssa_name (mask); | |
9321 | if (bitop2 == BIT_NOT_EXPR) | |
9322 | new_stmt = gimple_build_assign (res, bitop2, new_temp); | |
9323 | else | |
9324 | new_stmt = gimple_build_assign (res, bitop2, vec_rhs1, | |
9325 | new_temp); | |
9326 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9327 | } | |
9328 | } | |
42fd8198 IE |
9329 | if (slp_node) |
9330 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
9331 | } | |
9332 | ||
9333 | if (slp_node) | |
9334 | continue; | |
9335 | ||
9336 | if (j == 0) | |
9337 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
9338 | else | |
9339 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
9340 | ||
9341 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
9342 | } | |
9343 | ||
9344 | vec_oprnds0.release (); | |
9345 | vec_oprnds1.release (); | |
9346 | ||
9347 | return true; | |
9348 | } | |
ebfd146a | 9349 | |
68a0f2ff RS |
9350 | /* If SLP_NODE is nonnull, return true if vectorizable_live_operation |
9351 | can handle all live statements in the node. Otherwise return true | |
9352 | if STMT is not live or if vectorizable_live_operation can handle it. | |
9353 | GSI and VEC_STMT are as for vectorizable_live_operation. */ | |
9354 | ||
9355 | static bool | |
9356 | can_vectorize_live_stmts (gimple *stmt, gimple_stmt_iterator *gsi, | |
68435eb2 RB |
9357 | slp_tree slp_node, gimple **vec_stmt, |
9358 | stmt_vector_for_cost *cost_vec) | |
68a0f2ff RS |
9359 | { |
9360 | if (slp_node) | |
9361 | { | |
9362 | gimple *slp_stmt; | |
9363 | unsigned int i; | |
9364 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (slp_node), i, slp_stmt) | |
9365 | { | |
9366 | stmt_vec_info slp_stmt_info = vinfo_for_stmt (slp_stmt); | |
9367 | if (STMT_VINFO_LIVE_P (slp_stmt_info) | |
9368 | && !vectorizable_live_operation (slp_stmt, gsi, slp_node, i, | |
68435eb2 | 9369 | vec_stmt, cost_vec)) |
68a0f2ff RS |
9370 | return false; |
9371 | } | |
9372 | } | |
9373 | else if (STMT_VINFO_LIVE_P (vinfo_for_stmt (stmt)) | |
68435eb2 RB |
9374 | && !vectorizable_live_operation (stmt, gsi, slp_node, -1, vec_stmt, |
9375 | cost_vec)) | |
68a0f2ff RS |
9376 | return false; |
9377 | ||
9378 | return true; | |
9379 | } | |
9380 | ||
8644a673 | 9381 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
9382 | |
9383 | bool | |
891ad31c | 9384 | vect_analyze_stmt (gimple *stmt, bool *need_to_vectorize, slp_tree node, |
68435eb2 | 9385 | slp_instance node_instance, stmt_vector_for_cost *cost_vec) |
ebfd146a | 9386 | { |
8644a673 | 9387 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 9388 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 9389 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 9390 | bool ok; |
355fe088 | 9391 | gimple *pattern_stmt; |
363477c0 | 9392 | gimple_seq pattern_def_seq; |
ebfd146a | 9393 | |
73fbfcad | 9394 | if (dump_enabled_p ()) |
ebfd146a | 9395 | { |
78c60e3d SS |
9396 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
9397 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
8644a673 | 9398 | } |
ebfd146a | 9399 | |
1825a1f3 | 9400 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 9401 | { |
73fbfcad | 9402 | if (dump_enabled_p ()) |
78c60e3d | 9403 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9404 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
9405 | |
9406 | return false; | |
9407 | } | |
b8698a0f L |
9408 | |
9409 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
9410 | to include: |
9411 | - the COND_EXPR which is the loop exit condition | |
9412 | - any LABEL_EXPRs in the loop | |
b8698a0f | 9413 | - computations that are used only for array indexing or loop control. |
8644a673 | 9414 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 9415 | instance, therefore, all the statements are relevant. |
ebfd146a | 9416 | |
d092494c | 9417 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 9418 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
9419 | statements. In basic blocks we are called from some SLP instance |
9420 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
9421 | already will be part of SLP instance. */ | |
83197f37 IR |
9422 | |
9423 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 9424 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 9425 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 9426 | { |
9d5e7640 | 9427 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 9428 | && pattern_stmt |
9d5e7640 IR |
9429 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
9430 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
9431 | { | |
83197f37 | 9432 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
9433 | stmt = pattern_stmt; |
9434 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 9435 | if (dump_enabled_p ()) |
9d5e7640 | 9436 | { |
78c60e3d SS |
9437 | dump_printf_loc (MSG_NOTE, vect_location, |
9438 | "==> examining pattern statement: "); | |
9439 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
9d5e7640 IR |
9440 | } |
9441 | } | |
9442 | else | |
9443 | { | |
73fbfcad | 9444 | if (dump_enabled_p ()) |
e645e942 | 9445 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 9446 | |
9d5e7640 IR |
9447 | return true; |
9448 | } | |
8644a673 | 9449 | } |
83197f37 | 9450 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 9451 | && node == NULL |
83197f37 IR |
9452 | && pattern_stmt |
9453 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
9454 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
9455 | { | |
9456 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 9457 | if (dump_enabled_p ()) |
83197f37 | 9458 | { |
78c60e3d SS |
9459 | dump_printf_loc (MSG_NOTE, vect_location, |
9460 | "==> examining pattern statement: "); | |
9461 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
83197f37 IR |
9462 | } |
9463 | ||
891ad31c | 9464 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node, |
68435eb2 | 9465 | node_instance, cost_vec)) |
83197f37 IR |
9466 | return false; |
9467 | } | |
ebfd146a | 9468 | |
1107f3ae | 9469 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 9470 | && node == NULL |
363477c0 | 9471 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 9472 | { |
363477c0 | 9473 | gimple_stmt_iterator si; |
1107f3ae | 9474 | |
363477c0 JJ |
9475 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
9476 | { | |
355fe088 | 9477 | gimple *pattern_def_stmt = gsi_stmt (si); |
363477c0 JJ |
9478 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) |
9479 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
9480 | { | |
9481 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 9482 | if (dump_enabled_p ()) |
363477c0 | 9483 | { |
78c60e3d SS |
9484 | dump_printf_loc (MSG_NOTE, vect_location, |
9485 | "==> examining pattern def statement: "); | |
9486 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
363477c0 | 9487 | } |
1107f3ae | 9488 | |
363477c0 | 9489 | if (!vect_analyze_stmt (pattern_def_stmt, |
68435eb2 RB |
9490 | need_to_vectorize, node, node_instance, |
9491 | cost_vec)) | |
363477c0 JJ |
9492 | return false; |
9493 | } | |
9494 | } | |
9495 | } | |
1107f3ae | 9496 | |
8644a673 IR |
9497 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
9498 | { | |
9499 | case vect_internal_def: | |
9500 | break; | |
ebfd146a | 9501 | |
8644a673 | 9502 | case vect_reduction_def: |
7c5222ff | 9503 | case vect_nested_cycle: |
14a61437 RB |
9504 | gcc_assert (!bb_vinfo |
9505 | && (relevance == vect_used_in_outer | |
9506 | || relevance == vect_used_in_outer_by_reduction | |
9507 | || relevance == vect_used_by_reduction | |
b28ead45 AH |
9508 | || relevance == vect_unused_in_scope |
9509 | || relevance == vect_used_only_live)); | |
8644a673 IR |
9510 | break; |
9511 | ||
9512 | case vect_induction_def: | |
e7baeb39 RB |
9513 | gcc_assert (!bb_vinfo); |
9514 | break; | |
9515 | ||
8644a673 IR |
9516 | case vect_constant_def: |
9517 | case vect_external_def: | |
9518 | case vect_unknown_def_type: | |
9519 | default: | |
9520 | gcc_unreachable (); | |
9521 | } | |
ebfd146a | 9522 | |
8644a673 | 9523 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 9524 | { |
8644a673 | 9525 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
0136f8f0 AH |
9526 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
9527 | || (is_gimple_call (stmt) | |
9528 | && gimple_call_lhs (stmt) == NULL_TREE)); | |
8644a673 | 9529 | *need_to_vectorize = true; |
ebfd146a IR |
9530 | } |
9531 | ||
b1af7da6 RB |
9532 | if (PURE_SLP_STMT (stmt_info) && !node) |
9533 | { | |
9534 | dump_printf_loc (MSG_NOTE, vect_location, | |
9535 | "handled only by SLP analysis\n"); | |
9536 | return true; | |
9537 | } | |
9538 | ||
9539 | ok = true; | |
9540 | if (!bb_vinfo | |
9541 | && (STMT_VINFO_RELEVANT_P (stmt_info) | |
9542 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
68435eb2 RB |
9543 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node, cost_vec) |
9544 | || vectorizable_conversion (stmt, NULL, NULL, node, cost_vec) | |
9545 | || vectorizable_shift (stmt, NULL, NULL, node, cost_vec) | |
9546 | || vectorizable_operation (stmt, NULL, NULL, node, cost_vec) | |
9547 | || vectorizable_assignment (stmt, NULL, NULL, node, cost_vec) | |
9548 | || vectorizable_load (stmt, NULL, NULL, node, node_instance, cost_vec) | |
9549 | || vectorizable_call (stmt, NULL, NULL, node, cost_vec) | |
9550 | || vectorizable_store (stmt, NULL, NULL, node, cost_vec) | |
9551 | || vectorizable_reduction (stmt, NULL, NULL, node, node_instance, | |
9552 | cost_vec) | |
9553 | || vectorizable_induction (stmt, NULL, NULL, node, cost_vec) | |
9554 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node, cost_vec) | |
9555 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node, cost_vec)); | |
b1af7da6 RB |
9556 | else |
9557 | { | |
9558 | if (bb_vinfo) | |
68435eb2 RB |
9559 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node, cost_vec) |
9560 | || vectorizable_conversion (stmt, NULL, NULL, node, cost_vec) | |
9561 | || vectorizable_shift (stmt, NULL, NULL, node, cost_vec) | |
9562 | || vectorizable_operation (stmt, NULL, NULL, node, cost_vec) | |
9563 | || vectorizable_assignment (stmt, NULL, NULL, node, cost_vec) | |
9564 | || vectorizable_load (stmt, NULL, NULL, node, node_instance, | |
9565 | cost_vec) | |
9566 | || vectorizable_call (stmt, NULL, NULL, node, cost_vec) | |
9567 | || vectorizable_store (stmt, NULL, NULL, node, cost_vec) | |
9568 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node, | |
9569 | cost_vec) | |
9570 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node, | |
9571 | cost_vec)); | |
b1af7da6 | 9572 | } |
8644a673 IR |
9573 | |
9574 | if (!ok) | |
ebfd146a | 9575 | { |
73fbfcad | 9576 | if (dump_enabled_p ()) |
8644a673 | 9577 | { |
78c60e3d SS |
9578 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9579 | "not vectorized: relevant stmt not "); | |
9580 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
9581 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 9582 | } |
b8698a0f | 9583 | |
ebfd146a IR |
9584 | return false; |
9585 | } | |
9586 | ||
8644a673 IR |
9587 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
9588 | need extra handling, except for vectorizable reductions. */ | |
68435eb2 RB |
9589 | if (!bb_vinfo |
9590 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type | |
9591 | && !can_vectorize_live_stmts (stmt, NULL, node, NULL, cost_vec)) | |
ebfd146a | 9592 | { |
73fbfcad | 9593 | if (dump_enabled_p ()) |
8644a673 | 9594 | { |
78c60e3d | 9595 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
68a0f2ff | 9596 | "not vectorized: live stmt not supported: "); |
78c60e3d | 9597 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); |
8644a673 | 9598 | } |
b8698a0f | 9599 | |
8644a673 | 9600 | return false; |
ebfd146a IR |
9601 | } |
9602 | ||
ebfd146a IR |
9603 | return true; |
9604 | } | |
9605 | ||
9606 | ||
9607 | /* Function vect_transform_stmt. | |
9608 | ||
9609 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
9610 | ||
9611 | bool | |
355fe088 | 9612 | vect_transform_stmt (gimple *stmt, gimple_stmt_iterator *gsi, |
0d0293ac | 9613 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
9614 | slp_instance slp_node_instance) |
9615 | { | |
9616 | bool is_store = false; | |
355fe088 | 9617 | gimple *vec_stmt = NULL; |
ebfd146a | 9618 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
ebfd146a | 9619 | bool done; |
ebfd146a | 9620 | |
fce57248 | 9621 | gcc_assert (slp_node || !PURE_SLP_STMT (stmt_info)); |
355fe088 | 9622 | gimple *old_vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
225ce44b | 9623 | |
e57d9a82 RB |
9624 | bool nested_p = (STMT_VINFO_LOOP_VINFO (stmt_info) |
9625 | && nested_in_vect_loop_p | |
9626 | (LOOP_VINFO_LOOP (STMT_VINFO_LOOP_VINFO (stmt_info)), | |
9627 | stmt)); | |
9628 | ||
ebfd146a IR |
9629 | switch (STMT_VINFO_TYPE (stmt_info)) |
9630 | { | |
9631 | case type_demotion_vec_info_type: | |
ebfd146a | 9632 | case type_promotion_vec_info_type: |
ebfd146a | 9633 | case type_conversion_vec_info_type: |
68435eb2 | 9634 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node, NULL); |
ebfd146a IR |
9635 | gcc_assert (done); |
9636 | break; | |
9637 | ||
9638 | case induc_vec_info_type: | |
68435eb2 | 9639 | done = vectorizable_induction (stmt, gsi, &vec_stmt, slp_node, NULL); |
ebfd146a IR |
9640 | gcc_assert (done); |
9641 | break; | |
9642 | ||
9dc3f7de | 9643 | case shift_vec_info_type: |
68435eb2 | 9644 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node, NULL); |
9dc3f7de IR |
9645 | gcc_assert (done); |
9646 | break; | |
9647 | ||
ebfd146a | 9648 | case op_vec_info_type: |
68435eb2 | 9649 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node, NULL); |
ebfd146a IR |
9650 | gcc_assert (done); |
9651 | break; | |
9652 | ||
9653 | case assignment_vec_info_type: | |
68435eb2 | 9654 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node, NULL); |
ebfd146a IR |
9655 | gcc_assert (done); |
9656 | break; | |
9657 | ||
9658 | case load_vec_info_type: | |
b8698a0f | 9659 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
68435eb2 | 9660 | slp_node_instance, NULL); |
ebfd146a IR |
9661 | gcc_assert (done); |
9662 | break; | |
9663 | ||
9664 | case store_vec_info_type: | |
68435eb2 | 9665 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node, NULL); |
ebfd146a | 9666 | gcc_assert (done); |
0d0293ac | 9667 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
9668 | { |
9669 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 9670 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
9671 | one are skipped, and there vec_stmt_info shouldn't be freed |
9672 | meanwhile. */ | |
0d0293ac | 9673 | *grouped_store = true; |
f307441a | 9674 | stmt_vec_info group_info |
2c53b149 RB |
9675 | = vinfo_for_stmt (DR_GROUP_FIRST_ELEMENT (stmt_info)); |
9676 | if (DR_GROUP_STORE_COUNT (group_info) == DR_GROUP_SIZE (group_info)) | |
ebfd146a | 9677 | is_store = true; |
f307441a | 9678 | } |
ebfd146a IR |
9679 | else |
9680 | is_store = true; | |
9681 | break; | |
9682 | ||
9683 | case condition_vec_info_type: | |
68435eb2 | 9684 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node, NULL); |
ebfd146a IR |
9685 | gcc_assert (done); |
9686 | break; | |
9687 | ||
42fd8198 | 9688 | case comparison_vec_info_type: |
68435eb2 | 9689 | done = vectorizable_comparison (stmt, gsi, &vec_stmt, NULL, slp_node, NULL); |
42fd8198 IE |
9690 | gcc_assert (done); |
9691 | break; | |
9692 | ||
ebfd146a | 9693 | case call_vec_info_type: |
68435eb2 | 9694 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node, NULL); |
039d9ea1 | 9695 | stmt = gsi_stmt (*gsi); |
ebfd146a IR |
9696 | break; |
9697 | ||
0136f8f0 | 9698 | case call_simd_clone_vec_info_type: |
68435eb2 | 9699 | done = vectorizable_simd_clone_call (stmt, gsi, &vec_stmt, slp_node, NULL); |
0136f8f0 AH |
9700 | stmt = gsi_stmt (*gsi); |
9701 | break; | |
9702 | ||
ebfd146a | 9703 | case reduc_vec_info_type: |
891ad31c | 9704 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node, |
68435eb2 | 9705 | slp_node_instance, NULL); |
ebfd146a IR |
9706 | gcc_assert (done); |
9707 | break; | |
9708 | ||
9709 | default: | |
9710 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
9711 | { | |
73fbfcad | 9712 | if (dump_enabled_p ()) |
78c60e3d | 9713 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9714 | "stmt not supported.\n"); |
ebfd146a IR |
9715 | gcc_unreachable (); |
9716 | } | |
9717 | } | |
9718 | ||
225ce44b RB |
9719 | /* Verify SLP vectorization doesn't mess with STMT_VINFO_VEC_STMT. |
9720 | This would break hybrid SLP vectorization. */ | |
9721 | if (slp_node) | |
d90f8440 RB |
9722 | gcc_assert (!vec_stmt |
9723 | && STMT_VINFO_VEC_STMT (stmt_info) == old_vec_stmt); | |
225ce44b | 9724 | |
ebfd146a IR |
9725 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that |
9726 | is being vectorized, but outside the immediately enclosing loop. */ | |
9727 | if (vec_stmt | |
e57d9a82 | 9728 | && nested_p |
ebfd146a IR |
9729 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
9730 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 9731 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 9732 | vect_used_in_outer_by_reduction)) |
ebfd146a | 9733 | { |
a70d6342 IR |
9734 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
9735 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
9736 | imm_use_iterator imm_iter; |
9737 | use_operand_p use_p; | |
9738 | tree scalar_dest; | |
355fe088 | 9739 | gimple *exit_phi; |
ebfd146a | 9740 | |
73fbfcad | 9741 | if (dump_enabled_p ()) |
78c60e3d | 9742 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 9743 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
9744 | |
9745 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
9746 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
9747 | STMT). */ | |
9748 | if (gimple_code (stmt) == GIMPLE_PHI) | |
9749 | scalar_dest = PHI_RESULT (stmt); | |
9750 | else | |
9751 | scalar_dest = gimple_assign_lhs (stmt); | |
9752 | ||
9753 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
9754 | { | |
9755 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
9756 | { | |
9757 | exit_phi = USE_STMT (use_p); | |
9758 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
9759 | } | |
9760 | } | |
9761 | } | |
9762 | ||
9763 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
9764 | being vectorized. */ | |
68a0f2ff | 9765 | if (STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) |
ebfd146a | 9766 | { |
68435eb2 | 9767 | done = can_vectorize_live_stmts (stmt, gsi, slp_node, &vec_stmt, NULL); |
ebfd146a IR |
9768 | gcc_assert (done); |
9769 | } | |
9770 | ||
9771 | if (vec_stmt) | |
83197f37 | 9772 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 9773 | |
b8698a0f | 9774 | return is_store; |
ebfd146a IR |
9775 | } |
9776 | ||
9777 | ||
b8698a0f | 9778 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
9779 | stmt_vec_info. */ |
9780 | ||
9781 | void | |
355fe088 | 9782 | vect_remove_stores (gimple *first_stmt) |
ebfd146a | 9783 | { |
355fe088 TS |
9784 | gimple *next = first_stmt; |
9785 | gimple *tmp; | |
ebfd146a IR |
9786 | gimple_stmt_iterator next_si; |
9787 | ||
9788 | while (next) | |
9789 | { | |
78048b1c JJ |
9790 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
9791 | ||
2c53b149 | 9792 | tmp = DR_GROUP_NEXT_ELEMENT (stmt_info); |
78048b1c JJ |
9793 | if (is_pattern_stmt_p (stmt_info)) |
9794 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
9795 | /* Free the attached stmt_vec_info and remove the stmt. */ |
9796 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 9797 | unlink_stmt_vdef (next); |
ebfd146a | 9798 | gsi_remove (&next_si, true); |
3d3f2249 | 9799 | release_defs (next); |
ebfd146a IR |
9800 | free_stmt_vec_info (next); |
9801 | next = tmp; | |
9802 | } | |
9803 | } | |
9804 | ||
9805 | ||
9806 | /* Function new_stmt_vec_info. | |
9807 | ||
9808 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
9809 | ||
9810 | stmt_vec_info | |
310213d4 | 9811 | new_stmt_vec_info (gimple *stmt, vec_info *vinfo) |
ebfd146a IR |
9812 | { |
9813 | stmt_vec_info res; | |
9814 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
9815 | ||
9816 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
9817 | STMT_VINFO_STMT (res) = stmt; | |
310213d4 | 9818 | res->vinfo = vinfo; |
8644a673 | 9819 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
9820 | STMT_VINFO_LIVE_P (res) = false; |
9821 | STMT_VINFO_VECTYPE (res) = NULL; | |
9822 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 9823 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
9824 | STMT_VINFO_IN_PATTERN_P (res) = false; |
9825 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 9826 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a | 9827 | STMT_VINFO_DATA_REF (res) = NULL; |
af29617a | 9828 | STMT_VINFO_VEC_REDUCTION_TYPE (res) = TREE_CODE_REDUCTION; |
7e16ce79 | 9829 | STMT_VINFO_VEC_CONST_COND_REDUC_CODE (res) = ERROR_MARK; |
ebfd146a | 9830 | |
ebfd146a IR |
9831 | if (gimple_code (stmt) == GIMPLE_PHI |
9832 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
9833 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
9834 | else | |
8644a673 IR |
9835 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
9836 | ||
9771b263 | 9837 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 9838 | STMT_SLP_TYPE (res) = loop_vect; |
78810bd3 RB |
9839 | STMT_VINFO_NUM_SLP_USES (res) = 0; |
9840 | ||
2c53b149 RB |
9841 | res->first_element = NULL; /* GROUP_FIRST_ELEMENT */ |
9842 | res->next_element = NULL; /* GROUP_NEXT_ELEMENT */ | |
9843 | res->size = 0; /* GROUP_SIZE */ | |
9844 | res->store_count = 0; /* GROUP_STORE_COUNT */ | |
9845 | res->gap = 0; /* GROUP_GAP */ | |
9846 | res->same_dr_stmt = NULL; /* GROUP_SAME_DR_STMT */ | |
ebfd146a IR |
9847 | |
9848 | return res; | |
9849 | } | |
9850 | ||
9851 | ||
f8c0baaf | 9852 | /* Set the current stmt_vec_info vector to V. */ |
ebfd146a IR |
9853 | |
9854 | void | |
f8c0baaf | 9855 | set_stmt_vec_info_vec (vec<stmt_vec_info> *v) |
ebfd146a | 9856 | { |
f8c0baaf | 9857 | stmt_vec_info_vec = v; |
ebfd146a IR |
9858 | } |
9859 | ||
f8c0baaf | 9860 | /* Free the stmt_vec_info entries in V and release V. */ |
ebfd146a IR |
9861 | |
9862 | void | |
f8c0baaf | 9863 | free_stmt_vec_infos (vec<stmt_vec_info> *v) |
ebfd146a | 9864 | { |
93675444 | 9865 | unsigned int i; |
3161455c | 9866 | stmt_vec_info info; |
f8c0baaf | 9867 | FOR_EACH_VEC_ELT (*v, i, info) |
93675444 | 9868 | if (info != NULL) |
3161455c | 9869 | free_stmt_vec_info (STMT_VINFO_STMT (info)); |
f8c0baaf RB |
9870 | if (v == stmt_vec_info_vec) |
9871 | stmt_vec_info_vec = NULL; | |
9872 | v->release (); | |
ebfd146a IR |
9873 | } |
9874 | ||
9875 | ||
9876 | /* Free stmt vectorization related info. */ | |
9877 | ||
9878 | void | |
355fe088 | 9879 | free_stmt_vec_info (gimple *stmt) |
ebfd146a IR |
9880 | { |
9881 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
9882 | ||
9883 | if (!stmt_info) | |
9884 | return; | |
9885 | ||
78048b1c JJ |
9886 | /* Check if this statement has a related "pattern stmt" |
9887 | (introduced by the vectorizer during the pattern recognition | |
9888 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
9889 | too. */ | |
9890 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
9891 | { | |
9892 | stmt_vec_info patt_info | |
9893 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
9894 | if (patt_info) | |
9895 | { | |
363477c0 | 9896 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
355fe088 | 9897 | gimple *patt_stmt = STMT_VINFO_STMT (patt_info); |
f0281fde RB |
9898 | gimple_set_bb (patt_stmt, NULL); |
9899 | tree lhs = gimple_get_lhs (patt_stmt); | |
e6f5c25d | 9900 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde | 9901 | release_ssa_name (lhs); |
363477c0 JJ |
9902 | if (seq) |
9903 | { | |
9904 | gimple_stmt_iterator si; | |
9905 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
f0281fde | 9906 | { |
355fe088 | 9907 | gimple *seq_stmt = gsi_stmt (si); |
f0281fde | 9908 | gimple_set_bb (seq_stmt, NULL); |
7532abf2 | 9909 | lhs = gimple_get_lhs (seq_stmt); |
e6f5c25d | 9910 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde RB |
9911 | release_ssa_name (lhs); |
9912 | free_stmt_vec_info (seq_stmt); | |
9913 | } | |
363477c0 | 9914 | } |
f0281fde | 9915 | free_stmt_vec_info (patt_stmt); |
78048b1c JJ |
9916 | } |
9917 | } | |
9918 | ||
9771b263 | 9919 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
6c9e85fb | 9920 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).release (); |
ebfd146a IR |
9921 | set_vinfo_for_stmt (stmt, NULL); |
9922 | free (stmt_info); | |
9923 | } | |
9924 | ||
9925 | ||
bb67d9c7 | 9926 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 9927 | |
bb67d9c7 | 9928 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
9929 | by the target. */ |
9930 | ||
c803b2a9 | 9931 | tree |
86e36728 | 9932 | get_vectype_for_scalar_type_and_size (tree scalar_type, poly_uint64 size) |
ebfd146a | 9933 | { |
c7d97b28 | 9934 | tree orig_scalar_type = scalar_type; |
3bd8f481 | 9935 | scalar_mode inner_mode; |
ef4bddc2 | 9936 | machine_mode simd_mode; |
86e36728 | 9937 | poly_uint64 nunits; |
ebfd146a IR |
9938 | tree vectype; |
9939 | ||
3bd8f481 RS |
9940 | if (!is_int_mode (TYPE_MODE (scalar_type), &inner_mode) |
9941 | && !is_float_mode (TYPE_MODE (scalar_type), &inner_mode)) | |
ebfd146a IR |
9942 | return NULL_TREE; |
9943 | ||
3bd8f481 | 9944 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
48f2e373 | 9945 | |
7b7b1813 RG |
9946 | /* For vector types of elements whose mode precision doesn't |
9947 | match their types precision we use a element type of mode | |
9948 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
9949 | they support the proper result truncation/extension. |
9950 | We also make sure to build vector types with INTEGER_TYPE | |
9951 | component type only. */ | |
6d7971b8 | 9952 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
9953 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
9954 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
9955 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
9956 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 9957 | |
ccbf5bb4 RG |
9958 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
9959 | When the component mode passes the above test simply use a type | |
9960 | corresponding to that mode. The theory is that any use that | |
9961 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 9962 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 9963 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
9964 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
9965 | ||
9966 | /* We can't build a vector type of elements with alignment bigger than | |
9967 | their size. */ | |
dfc2e2ac | 9968 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
9969 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
9970 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 9971 | |
dfc2e2ac RB |
9972 | /* If we felt back to using the mode fail if there was |
9973 | no scalar type for it. */ | |
9974 | if (scalar_type == NULL_TREE) | |
9975 | return NULL_TREE; | |
9976 | ||
bb67d9c7 RG |
9977 | /* If no size was supplied use the mode the target prefers. Otherwise |
9978 | lookup a vector mode of the specified size. */ | |
86e36728 | 9979 | if (known_eq (size, 0U)) |
bb67d9c7 | 9980 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); |
86e36728 RS |
9981 | else if (!multiple_p (size, nbytes, &nunits) |
9982 | || !mode_for_vector (inner_mode, nunits).exists (&simd_mode)) | |
9da15d40 | 9983 | return NULL_TREE; |
4c8fd8ac | 9984 | /* NOTE: nunits == 1 is allowed to support single element vector types. */ |
86e36728 | 9985 | if (!multiple_p (GET_MODE_SIZE (simd_mode), nbytes, &nunits)) |
cc4b5170 | 9986 | return NULL_TREE; |
ebfd146a IR |
9987 | |
9988 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
9989 | |
9990 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
9991 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 9992 | return NULL_TREE; |
ebfd146a | 9993 | |
c7d97b28 RB |
9994 | /* Re-attach the address-space qualifier if we canonicalized the scalar |
9995 | type. */ | |
9996 | if (TYPE_ADDR_SPACE (orig_scalar_type) != TYPE_ADDR_SPACE (vectype)) | |
9997 | return build_qualified_type | |
9998 | (vectype, KEEP_QUAL_ADDR_SPACE (TYPE_QUALS (orig_scalar_type))); | |
9999 | ||
ebfd146a IR |
10000 | return vectype; |
10001 | } | |
10002 | ||
86e36728 | 10003 | poly_uint64 current_vector_size; |
bb67d9c7 RG |
10004 | |
10005 | /* Function get_vectype_for_scalar_type. | |
10006 | ||
10007 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
10008 | by the target. */ | |
10009 | ||
10010 | tree | |
10011 | get_vectype_for_scalar_type (tree scalar_type) | |
10012 | { | |
10013 | tree vectype; | |
10014 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
10015 | current_vector_size); | |
10016 | if (vectype | |
86e36728 | 10017 | && known_eq (current_vector_size, 0U)) |
bb67d9c7 RG |
10018 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); |
10019 | return vectype; | |
10020 | } | |
10021 | ||
42fd8198 IE |
10022 | /* Function get_mask_type_for_scalar_type. |
10023 | ||
10024 | Returns the mask type corresponding to a result of comparison | |
10025 | of vectors of specified SCALAR_TYPE as supported by target. */ | |
10026 | ||
10027 | tree | |
10028 | get_mask_type_for_scalar_type (tree scalar_type) | |
10029 | { | |
10030 | tree vectype = get_vectype_for_scalar_type (scalar_type); | |
10031 | ||
10032 | if (!vectype) | |
10033 | return NULL; | |
10034 | ||
10035 | return build_truth_vector_type (TYPE_VECTOR_SUBPARTS (vectype), | |
10036 | current_vector_size); | |
10037 | } | |
10038 | ||
b690cc0f RG |
10039 | /* Function get_same_sized_vectype |
10040 | ||
10041 | Returns a vector type corresponding to SCALAR_TYPE of size | |
10042 | VECTOR_TYPE if supported by the target. */ | |
10043 | ||
10044 | tree | |
bb67d9c7 | 10045 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 10046 | { |
2568d8a1 | 10047 | if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type)) |
9f47c7e5 IE |
10048 | return build_same_sized_truth_vector_type (vector_type); |
10049 | ||
bb67d9c7 RG |
10050 | return get_vectype_for_scalar_type_and_size |
10051 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
10052 | } |
10053 | ||
ebfd146a IR |
10054 | /* Function vect_is_simple_use. |
10055 | ||
10056 | Input: | |
81c40241 RB |
10057 | VINFO - the vect info of the loop or basic block that is being vectorized. |
10058 | OPERAND - operand in the loop or bb. | |
10059 | Output: | |
10060 | DEF_STMT - the defining stmt in case OPERAND is an SSA_NAME. | |
10061 | DT - the type of definition | |
ebfd146a IR |
10062 | |
10063 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 10064 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 10065 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 10066 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
10067 | is the case in reduction/induction computations). |
10068 | For basic blocks, supportable operands are constants and bb invariants. | |
10069 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
10070 | |
10071 | bool | |
81c40241 RB |
10072 | vect_is_simple_use (tree operand, vec_info *vinfo, |
10073 | gimple **def_stmt, enum vect_def_type *dt) | |
b8698a0f | 10074 | { |
ebfd146a | 10075 | *def_stmt = NULL; |
3fc356dc | 10076 | *dt = vect_unknown_def_type; |
b8698a0f | 10077 | |
73fbfcad | 10078 | if (dump_enabled_p ()) |
ebfd146a | 10079 | { |
78c60e3d SS |
10080 | dump_printf_loc (MSG_NOTE, vect_location, |
10081 | "vect_is_simple_use: operand "); | |
10082 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
e645e942 | 10083 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 10084 | } |
b8698a0f | 10085 | |
b758f602 | 10086 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
10087 | { |
10088 | *dt = vect_constant_def; | |
10089 | return true; | |
10090 | } | |
b8698a0f | 10091 | |
ebfd146a IR |
10092 | if (is_gimple_min_invariant (operand)) |
10093 | { | |
8644a673 | 10094 | *dt = vect_external_def; |
ebfd146a IR |
10095 | return true; |
10096 | } | |
10097 | ||
ebfd146a IR |
10098 | if (TREE_CODE (operand) != SSA_NAME) |
10099 | { | |
73fbfcad | 10100 | if (dump_enabled_p ()) |
af29617a AH |
10101 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
10102 | "not ssa-name.\n"); | |
ebfd146a IR |
10103 | return false; |
10104 | } | |
b8698a0f | 10105 | |
3fc356dc | 10106 | if (SSA_NAME_IS_DEFAULT_DEF (operand)) |
ebfd146a | 10107 | { |
3fc356dc RB |
10108 | *dt = vect_external_def; |
10109 | return true; | |
ebfd146a IR |
10110 | } |
10111 | ||
3fc356dc | 10112 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
73fbfcad | 10113 | if (dump_enabled_p ()) |
ebfd146a | 10114 | { |
78c60e3d SS |
10115 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
10116 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
ebfd146a IR |
10117 | } |
10118 | ||
61d371eb | 10119 | if (! vect_stmt_in_region_p (vinfo, *def_stmt)) |
8644a673 | 10120 | *dt = vect_external_def; |
ebfd146a IR |
10121 | else |
10122 | { | |
3fc356dc | 10123 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (*def_stmt); |
603cca93 | 10124 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
ebfd146a IR |
10125 | } |
10126 | ||
2e8ab70c RB |
10127 | if (dump_enabled_p ()) |
10128 | { | |
10129 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: "); | |
10130 | switch (*dt) | |
10131 | { | |
10132 | case vect_uninitialized_def: | |
10133 | dump_printf (MSG_NOTE, "uninitialized\n"); | |
10134 | break; | |
10135 | case vect_constant_def: | |
10136 | dump_printf (MSG_NOTE, "constant\n"); | |
10137 | break; | |
10138 | case vect_external_def: | |
10139 | dump_printf (MSG_NOTE, "external\n"); | |
10140 | break; | |
10141 | case vect_internal_def: | |
10142 | dump_printf (MSG_NOTE, "internal\n"); | |
10143 | break; | |
10144 | case vect_induction_def: | |
10145 | dump_printf (MSG_NOTE, "induction\n"); | |
10146 | break; | |
10147 | case vect_reduction_def: | |
10148 | dump_printf (MSG_NOTE, "reduction\n"); | |
10149 | break; | |
10150 | case vect_double_reduction_def: | |
10151 | dump_printf (MSG_NOTE, "double reduction\n"); | |
10152 | break; | |
10153 | case vect_nested_cycle: | |
10154 | dump_printf (MSG_NOTE, "nested cycle\n"); | |
10155 | break; | |
10156 | case vect_unknown_def_type: | |
10157 | dump_printf (MSG_NOTE, "unknown\n"); | |
10158 | break; | |
10159 | } | |
10160 | } | |
10161 | ||
81c40241 | 10162 | if (*dt == vect_unknown_def_type) |
ebfd146a | 10163 | { |
73fbfcad | 10164 | if (dump_enabled_p ()) |
78c60e3d | 10165 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 10166 | "Unsupported pattern.\n"); |
ebfd146a IR |
10167 | return false; |
10168 | } | |
10169 | ||
ebfd146a IR |
10170 | switch (gimple_code (*def_stmt)) |
10171 | { | |
10172 | case GIMPLE_PHI: | |
ebfd146a | 10173 | case GIMPLE_ASSIGN: |
ebfd146a | 10174 | case GIMPLE_CALL: |
81c40241 | 10175 | break; |
ebfd146a | 10176 | default: |
73fbfcad | 10177 | if (dump_enabled_p ()) |
78c60e3d | 10178 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 10179 | "unsupported defining stmt:\n"); |
ebfd146a IR |
10180 | return false; |
10181 | } | |
10182 | ||
10183 | return true; | |
10184 | } | |
10185 | ||
81c40241 | 10186 | /* Function vect_is_simple_use. |
b690cc0f | 10187 | |
81c40241 | 10188 | Same as vect_is_simple_use but also determines the vector operand |
b690cc0f RG |
10189 | type of OPERAND and stores it to *VECTYPE. If the definition of |
10190 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
10191 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
10192 | is responsible to compute the best suited vector type for the | |
10193 | scalar operand. */ | |
10194 | ||
10195 | bool | |
81c40241 RB |
10196 | vect_is_simple_use (tree operand, vec_info *vinfo, |
10197 | gimple **def_stmt, enum vect_def_type *dt, tree *vectype) | |
b690cc0f | 10198 | { |
81c40241 | 10199 | if (!vect_is_simple_use (operand, vinfo, def_stmt, dt)) |
b690cc0f RG |
10200 | return false; |
10201 | ||
10202 | /* Now get a vector type if the def is internal, otherwise supply | |
10203 | NULL_TREE and leave it up to the caller to figure out a proper | |
10204 | type for the use stmt. */ | |
10205 | if (*dt == vect_internal_def | |
10206 | || *dt == vect_induction_def | |
10207 | || *dt == vect_reduction_def | |
10208 | || *dt == vect_double_reduction_def | |
10209 | || *dt == vect_nested_cycle) | |
10210 | { | |
10211 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
10212 | |
10213 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
10214 | && !STMT_VINFO_RELEVANT (stmt_info) | |
10215 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 10216 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 10217 | |
b690cc0f RG |
10218 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
10219 | gcc_assert (*vectype != NULL_TREE); | |
10220 | } | |
10221 | else if (*dt == vect_uninitialized_def | |
10222 | || *dt == vect_constant_def | |
10223 | || *dt == vect_external_def) | |
10224 | *vectype = NULL_TREE; | |
10225 | else | |
10226 | gcc_unreachable (); | |
10227 | ||
10228 | return true; | |
10229 | } | |
10230 | ||
ebfd146a IR |
10231 | |
10232 | /* Function supportable_widening_operation | |
10233 | ||
b8698a0f L |
10234 | Check whether an operation represented by the code CODE is a |
10235 | widening operation that is supported by the target platform in | |
b690cc0f RG |
10236 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
10237 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 10238 | |
1bda738b JJ |
10239 | Widening operations we currently support are NOP (CONVERT), FLOAT, |
10240 | FIX_TRUNC and WIDEN_MULT. This function checks if these operations | |
10241 | are supported by the target platform either directly (via vector | |
10242 | tree-codes), or via target builtins. | |
ebfd146a IR |
10243 | |
10244 | Output: | |
b8698a0f L |
10245 | - CODE1 and CODE2 are codes of vector operations to be used when |
10246 | vectorizing the operation, if available. | |
ebfd146a IR |
10247 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
10248 | case of multi-step conversion (like char->short->int - in that case | |
10249 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
10250 | - INTERM_TYPES contains the intermediate type required to perform the |
10251 | widening operation (short in the above example). */ | |
ebfd146a IR |
10252 | |
10253 | bool | |
355fe088 | 10254 | supportable_widening_operation (enum tree_code code, gimple *stmt, |
b690cc0f | 10255 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
10256 | enum tree_code *code1, enum tree_code *code2, |
10257 | int *multi_step_cvt, | |
9771b263 | 10258 | vec<tree> *interm_types) |
ebfd146a IR |
10259 | { |
10260 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
10261 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 10262 | struct loop *vect_loop = NULL; |
ef4bddc2 | 10263 | machine_mode vec_mode; |
81f40b79 | 10264 | enum insn_code icode1, icode2; |
ebfd146a | 10265 | optab optab1, optab2; |
b690cc0f RG |
10266 | tree vectype = vectype_in; |
10267 | tree wide_vectype = vectype_out; | |
ebfd146a | 10268 | enum tree_code c1, c2; |
4a00c761 JJ |
10269 | int i; |
10270 | tree prev_type, intermediate_type; | |
ef4bddc2 | 10271 | machine_mode intermediate_mode, prev_mode; |
4a00c761 | 10272 | optab optab3, optab4; |
ebfd146a | 10273 | |
4a00c761 | 10274 | *multi_step_cvt = 0; |
4ef69dfc IR |
10275 | if (loop_info) |
10276 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
10277 | ||
ebfd146a IR |
10278 | switch (code) |
10279 | { | |
10280 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
10281 | /* The result of a vectorized widening operation usually requires |
10282 | two vectors (because the widened results do not fit into one vector). | |
10283 | The generated vector results would normally be expected to be | |
10284 | generated in the same order as in the original scalar computation, | |
10285 | i.e. if 8 results are generated in each vector iteration, they are | |
10286 | to be organized as follows: | |
10287 | vect1: [res1,res2,res3,res4], | |
10288 | vect2: [res5,res6,res7,res8]. | |
10289 | ||
10290 | However, in the special case that the result of the widening | |
10291 | operation is used in a reduction computation only, the order doesn't | |
10292 | matter (because when vectorizing a reduction we change the order of | |
10293 | the computation). Some targets can take advantage of this and | |
10294 | generate more efficient code. For example, targets like Altivec, | |
10295 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
10296 | generate the following vectors: | |
10297 | vect1: [res1,res3,res5,res7], | |
10298 | vect2: [res2,res4,res6,res8]. | |
10299 | ||
10300 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
10301 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
10302 | iterations in parallel). We therefore don't allow to change the | |
10303 | order of the computation in the inner-loop during outer-loop | |
10304 | vectorization. */ | |
10305 | /* TODO: Another case in which order doesn't *really* matter is when we | |
10306 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
10307 | Normally, pack_trunc performs an even/odd permute, whereas the | |
10308 | repack from an even/odd expansion would be an interleave, which | |
10309 | would be significantly simpler for e.g. AVX2. */ | |
10310 | /* In any case, in order to avoid duplicating the code below, recurse | |
10311 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
10312 | are properly set up for the caller. If we fail, we'll continue with | |
10313 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
10314 | if (vect_loop | |
10315 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
10316 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
10317 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
10318 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
10319 | code1, code2, multi_step_cvt, |
10320 | interm_types)) | |
ebc047a2 CH |
10321 | { |
10322 | /* Elements in a vector with vect_used_by_reduction property cannot | |
10323 | be reordered if the use chain with this property does not have the | |
10324 | same operation. One such an example is s += a * b, where elements | |
10325 | in a and b cannot be reordered. Here we check if the vector defined | |
10326 | by STMT is only directly used in the reduction statement. */ | |
10327 | tree lhs = gimple_assign_lhs (stmt); | |
10328 | use_operand_p dummy; | |
355fe088 | 10329 | gimple *use_stmt; |
ebc047a2 CH |
10330 | stmt_vec_info use_stmt_info = NULL; |
10331 | if (single_imm_use (lhs, &dummy, &use_stmt) | |
10332 | && (use_stmt_info = vinfo_for_stmt (use_stmt)) | |
10333 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) | |
10334 | return true; | |
10335 | } | |
4a00c761 JJ |
10336 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
10337 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
10338 | break; |
10339 | ||
81c40241 RB |
10340 | case DOT_PROD_EXPR: |
10341 | c1 = DOT_PROD_EXPR; | |
10342 | c2 = DOT_PROD_EXPR; | |
10343 | break; | |
10344 | ||
10345 | case SAD_EXPR: | |
10346 | c1 = SAD_EXPR; | |
10347 | c2 = SAD_EXPR; | |
10348 | break; | |
10349 | ||
6ae6116f RH |
10350 | case VEC_WIDEN_MULT_EVEN_EXPR: |
10351 | /* Support the recursion induced just above. */ | |
10352 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
10353 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
10354 | break; | |
10355 | ||
36ba4aae | 10356 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
10357 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
10358 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
10359 | break; |
10360 | ||
ebfd146a | 10361 | CASE_CONVERT: |
4a00c761 JJ |
10362 | c1 = VEC_UNPACK_LO_EXPR; |
10363 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
10364 | break; |
10365 | ||
10366 | case FLOAT_EXPR: | |
4a00c761 JJ |
10367 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
10368 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
10369 | break; |
10370 | ||
10371 | case FIX_TRUNC_EXPR: | |
1bda738b JJ |
10372 | c1 = VEC_UNPACK_FIX_TRUNC_LO_EXPR; |
10373 | c2 = VEC_UNPACK_FIX_TRUNC_HI_EXPR; | |
10374 | break; | |
ebfd146a IR |
10375 | |
10376 | default: | |
10377 | gcc_unreachable (); | |
10378 | } | |
10379 | ||
6ae6116f | 10380 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
6b4db501 | 10381 | std::swap (c1, c2); |
4a00c761 | 10382 | |
ebfd146a IR |
10383 | if (code == FIX_TRUNC_EXPR) |
10384 | { | |
10385 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
10386 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
10387 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
10388 | } |
10389 | else | |
10390 | { | |
10391 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
10392 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
10393 | } | |
10394 | ||
10395 | if (!optab1 || !optab2) | |
10396 | return false; | |
10397 | ||
10398 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
10399 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
10400 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
10401 | return false; |
10402 | ||
4a00c761 JJ |
10403 | *code1 = c1; |
10404 | *code2 = c2; | |
10405 | ||
10406 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
10407 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff IE |
10408 | /* For scalar masks we may have different boolean |
10409 | vector types having the same QImode. Thus we | |
10410 | add additional check for elements number. */ | |
10411 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
928686b1 RS |
10412 | || known_eq (TYPE_VECTOR_SUBPARTS (vectype), |
10413 | TYPE_VECTOR_SUBPARTS (wide_vectype) * 2)); | |
4a00c761 | 10414 | |
b8698a0f | 10415 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 10416 | types. */ |
ebfd146a | 10417 | |
4a00c761 JJ |
10418 | prev_type = vectype; |
10419 | prev_mode = vec_mode; | |
b8698a0f | 10420 | |
4a00c761 JJ |
10421 | if (!CONVERT_EXPR_CODE_P (code)) |
10422 | return false; | |
b8698a0f | 10423 | |
4a00c761 JJ |
10424 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
10425 | intermediate steps in promotion sequence. We try | |
10426 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
10427 | not. */ | |
9771b263 | 10428 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
10429 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
10430 | { | |
10431 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
10432 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
10433 | { | |
7cfb4d93 | 10434 | intermediate_type = vect_halve_mask_nunits (prev_type); |
3ae0661a IE |
10435 | if (intermediate_mode != TYPE_MODE (intermediate_type)) |
10436 | return false; | |
10437 | } | |
10438 | else | |
10439 | intermediate_type | |
10440 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
10441 | TYPE_UNSIGNED (prev_type)); | |
10442 | ||
4a00c761 JJ |
10443 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); |
10444 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
10445 | ||
10446 | if (!optab3 || !optab4 | |
10447 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
10448 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
10449 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
10450 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
10451 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
10452 | == CODE_FOR_nothing) | |
10453 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
10454 | == CODE_FOR_nothing)) | |
10455 | break; | |
ebfd146a | 10456 | |
9771b263 | 10457 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
10458 | (*multi_step_cvt)++; |
10459 | ||
10460 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
10461 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff | 10462 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
928686b1 RS |
10463 | || known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type), |
10464 | TYPE_VECTOR_SUBPARTS (wide_vectype) * 2)); | |
4a00c761 JJ |
10465 | |
10466 | prev_type = intermediate_type; | |
10467 | prev_mode = intermediate_mode; | |
ebfd146a IR |
10468 | } |
10469 | ||
9771b263 | 10470 | interm_types->release (); |
4a00c761 | 10471 | return false; |
ebfd146a IR |
10472 | } |
10473 | ||
10474 | ||
10475 | /* Function supportable_narrowing_operation | |
10476 | ||
b8698a0f L |
10477 | Check whether an operation represented by the code CODE is a |
10478 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
10479 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
10480 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 10481 | |
1bda738b JJ |
10482 | Narrowing operations we currently support are NOP (CONVERT), FIX_TRUNC |
10483 | and FLOAT. This function checks if these operations are supported by | |
ebfd146a IR |
10484 | the target platform directly via vector tree-codes. |
10485 | ||
10486 | Output: | |
b8698a0f L |
10487 | - CODE1 is the code of a vector operation to be used when |
10488 | vectorizing the operation, if available. | |
ebfd146a IR |
10489 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
10490 | case of multi-step conversion (like int->short->char - in that case | |
10491 | MULTI_STEP_CVT will be 1). | |
10492 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 10493 | narrowing operation (short in the above example). */ |
ebfd146a IR |
10494 | |
10495 | bool | |
10496 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 10497 | tree vectype_out, tree vectype_in, |
ebfd146a | 10498 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 10499 | vec<tree> *interm_types) |
ebfd146a | 10500 | { |
ef4bddc2 | 10501 | machine_mode vec_mode; |
ebfd146a IR |
10502 | enum insn_code icode1; |
10503 | optab optab1, interm_optab; | |
b690cc0f RG |
10504 | tree vectype = vectype_in; |
10505 | tree narrow_vectype = vectype_out; | |
ebfd146a | 10506 | enum tree_code c1; |
3ae0661a | 10507 | tree intermediate_type, prev_type; |
ef4bddc2 | 10508 | machine_mode intermediate_mode, prev_mode; |
ebfd146a | 10509 | int i; |
4a00c761 | 10510 | bool uns; |
ebfd146a | 10511 | |
4a00c761 | 10512 | *multi_step_cvt = 0; |
ebfd146a IR |
10513 | switch (code) |
10514 | { | |
10515 | CASE_CONVERT: | |
10516 | c1 = VEC_PACK_TRUNC_EXPR; | |
10517 | break; | |
10518 | ||
10519 | case FIX_TRUNC_EXPR: | |
10520 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
10521 | break; | |
10522 | ||
10523 | case FLOAT_EXPR: | |
1bda738b JJ |
10524 | c1 = VEC_PACK_FLOAT_EXPR; |
10525 | break; | |
ebfd146a IR |
10526 | |
10527 | default: | |
10528 | gcc_unreachable (); | |
10529 | } | |
10530 | ||
10531 | if (code == FIX_TRUNC_EXPR) | |
10532 | /* The signedness is determined from output operand. */ | |
b690cc0f | 10533 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
10534 | else |
10535 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
10536 | ||
10537 | if (!optab1) | |
10538 | return false; | |
10539 | ||
10540 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 10541 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
10542 | return false; |
10543 | ||
4a00c761 JJ |
10544 | *code1 = c1; |
10545 | ||
10546 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff IE |
10547 | /* For scalar masks we may have different boolean |
10548 | vector types having the same QImode. Thus we | |
10549 | add additional check for elements number. */ | |
10550 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
928686b1 RS |
10551 | || known_eq (TYPE_VECTOR_SUBPARTS (vectype) * 2, |
10552 | TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 | 10553 | |
1bda738b JJ |
10554 | if (code == FLOAT_EXPR) |
10555 | return false; | |
10556 | ||
ebfd146a IR |
10557 | /* Check if it's a multi-step conversion that can be done using intermediate |
10558 | types. */ | |
4a00c761 | 10559 | prev_mode = vec_mode; |
3ae0661a | 10560 | prev_type = vectype; |
4a00c761 JJ |
10561 | if (code == FIX_TRUNC_EXPR) |
10562 | uns = TYPE_UNSIGNED (vectype_out); | |
10563 | else | |
10564 | uns = TYPE_UNSIGNED (vectype); | |
10565 | ||
10566 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
10567 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
10568 | costly than signed. */ | |
10569 | if (code == FIX_TRUNC_EXPR && uns) | |
10570 | { | |
10571 | enum insn_code icode2; | |
10572 | ||
10573 | intermediate_type | |
10574 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
10575 | interm_optab | |
10576 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 10577 | if (interm_optab != unknown_optab |
4a00c761 JJ |
10578 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
10579 | && insn_data[icode1].operand[0].mode | |
10580 | == insn_data[icode2].operand[0].mode) | |
10581 | { | |
10582 | uns = false; | |
10583 | optab1 = interm_optab; | |
10584 | icode1 = icode2; | |
10585 | } | |
10586 | } | |
ebfd146a | 10587 | |
4a00c761 JJ |
10588 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
10589 | intermediate steps in promotion sequence. We try | |
10590 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 10591 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
10592 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
10593 | { | |
10594 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
10595 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
10596 | { | |
7cfb4d93 | 10597 | intermediate_type = vect_double_mask_nunits (prev_type); |
3ae0661a | 10598 | if (intermediate_mode != TYPE_MODE (intermediate_type)) |
7cfb4d93 | 10599 | return false; |
3ae0661a IE |
10600 | } |
10601 | else | |
10602 | intermediate_type | |
10603 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
4a00c761 JJ |
10604 | interm_optab |
10605 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
10606 | optab_default); | |
10607 | if (!interm_optab | |
10608 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
10609 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
10610 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
10611 | == CODE_FOR_nothing)) | |
10612 | break; | |
10613 | ||
9771b263 | 10614 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
10615 | (*multi_step_cvt)++; |
10616 | ||
10617 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff | 10618 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
928686b1 RS |
10619 | || known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type) * 2, |
10620 | TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 JJ |
10621 | |
10622 | prev_mode = intermediate_mode; | |
3ae0661a | 10623 | prev_type = intermediate_type; |
4a00c761 | 10624 | optab1 = interm_optab; |
ebfd146a IR |
10625 | } |
10626 | ||
9771b263 | 10627 | interm_types->release (); |
4a00c761 | 10628 | return false; |
ebfd146a | 10629 | } |
7cfb4d93 RS |
10630 | |
10631 | /* Generate and return a statement that sets vector mask MASK such that | |
10632 | MASK[I] is true iff J + START_INDEX < END_INDEX for all J <= I. */ | |
10633 | ||
10634 | gcall * | |
10635 | vect_gen_while (tree mask, tree start_index, tree end_index) | |
10636 | { | |
10637 | tree cmp_type = TREE_TYPE (start_index); | |
10638 | tree mask_type = TREE_TYPE (mask); | |
10639 | gcc_checking_assert (direct_internal_fn_supported_p (IFN_WHILE_ULT, | |
10640 | cmp_type, mask_type, | |
10641 | OPTIMIZE_FOR_SPEED)); | |
10642 | gcall *call = gimple_build_call_internal (IFN_WHILE_ULT, 3, | |
10643 | start_index, end_index, | |
10644 | build_zero_cst (mask_type)); | |
10645 | gimple_call_set_lhs (call, mask); | |
10646 | return call; | |
10647 | } | |
535e7c11 RS |
10648 | |
10649 | /* Generate a vector mask of type MASK_TYPE for which index I is false iff | |
10650 | J + START_INDEX < END_INDEX for all J <= I. Add the statements to SEQ. */ | |
10651 | ||
10652 | tree | |
10653 | vect_gen_while_not (gimple_seq *seq, tree mask_type, tree start_index, | |
10654 | tree end_index) | |
10655 | { | |
10656 | tree tmp = make_ssa_name (mask_type); | |
10657 | gcall *call = vect_gen_while (tmp, start_index, end_index); | |
10658 | gimple_seq_add_stmt (seq, call); | |
10659 | return gimple_build (seq, BIT_NOT_EXPR, mask_type, tmp); | |
10660 | } | |
1f3cb663 RS |
10661 | |
10662 | /* Try to compute the vector types required to vectorize STMT_INFO, | |
10663 | returning true on success and false if vectorization isn't possible. | |
10664 | ||
10665 | On success: | |
10666 | ||
10667 | - Set *STMT_VECTYPE_OUT to: | |
10668 | - NULL_TREE if the statement doesn't need to be vectorized; | |
10669 | - boolean_type_node if the statement is a boolean operation whose | |
10670 | vector type can only be determined once all the other vector types | |
10671 | are known; and | |
10672 | - the equivalent of STMT_VINFO_VECTYPE otherwise. | |
10673 | ||
10674 | - Set *NUNITS_VECTYPE_OUT to the vector type that contains the maximum | |
10675 | number of units needed to vectorize STMT_INFO, or NULL_TREE if the | |
10676 | statement does not help to determine the overall number of units. */ | |
10677 | ||
10678 | bool | |
10679 | vect_get_vector_types_for_stmt (stmt_vec_info stmt_info, | |
10680 | tree *stmt_vectype_out, | |
10681 | tree *nunits_vectype_out) | |
10682 | { | |
10683 | gimple *stmt = stmt_info->stmt; | |
10684 | ||
10685 | *stmt_vectype_out = NULL_TREE; | |
10686 | *nunits_vectype_out = NULL_TREE; | |
10687 | ||
10688 | if (gimple_get_lhs (stmt) == NULL_TREE | |
10689 | /* MASK_STORE has no lhs, but is ok. */ | |
10690 | && !gimple_call_internal_p (stmt, IFN_MASK_STORE)) | |
10691 | { | |
10692 | if (is_a <gcall *> (stmt)) | |
10693 | { | |
10694 | /* Ignore calls with no lhs. These must be calls to | |
10695 | #pragma omp simd functions, and what vectorization factor | |
10696 | it really needs can't be determined until | |
10697 | vectorizable_simd_clone_call. */ | |
10698 | if (dump_enabled_p ()) | |
10699 | dump_printf_loc (MSG_NOTE, vect_location, | |
10700 | "defer to SIMD clone analysis.\n"); | |
10701 | return true; | |
10702 | } | |
10703 | ||
10704 | if (dump_enabled_p ()) | |
10705 | { | |
10706 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10707 | "not vectorized: irregular stmt."); | |
10708 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
10709 | } | |
10710 | return false; | |
10711 | } | |
10712 | ||
10713 | if (VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))) | |
10714 | { | |
10715 | if (dump_enabled_p ()) | |
10716 | { | |
10717 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10718 | "not vectorized: vector stmt in loop:"); | |
10719 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
10720 | } | |
10721 | return false; | |
10722 | } | |
10723 | ||
10724 | tree vectype; | |
10725 | tree scalar_type = NULL_TREE; | |
10726 | if (STMT_VINFO_VECTYPE (stmt_info)) | |
10727 | *stmt_vectype_out = vectype = STMT_VINFO_VECTYPE (stmt_info); | |
10728 | else | |
10729 | { | |
10730 | gcc_assert (!STMT_VINFO_DATA_REF (stmt_info)); | |
10731 | if (gimple_call_internal_p (stmt, IFN_MASK_STORE)) | |
10732 | scalar_type = TREE_TYPE (gimple_call_arg (stmt, 3)); | |
10733 | else | |
10734 | scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); | |
10735 | ||
10736 | /* Pure bool ops don't participate in number-of-units computation. | |
10737 | For comparisons use the types being compared. */ | |
10738 | if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type) | |
10739 | && is_gimple_assign (stmt) | |
10740 | && gimple_assign_rhs_code (stmt) != COND_EXPR) | |
10741 | { | |
10742 | *stmt_vectype_out = boolean_type_node; | |
10743 | ||
10744 | tree rhs1 = gimple_assign_rhs1 (stmt); | |
10745 | if (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)) == tcc_comparison | |
10746 | && !VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (rhs1))) | |
10747 | scalar_type = TREE_TYPE (rhs1); | |
10748 | else | |
10749 | { | |
10750 | if (dump_enabled_p ()) | |
10751 | dump_printf_loc (MSG_NOTE, vect_location, | |
10752 | "pure bool operation.\n"); | |
10753 | return true; | |
10754 | } | |
10755 | } | |
10756 | ||
10757 | if (dump_enabled_p ()) | |
10758 | { | |
10759 | dump_printf_loc (MSG_NOTE, vect_location, | |
10760 | "get vectype for scalar type: "); | |
10761 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
10762 | dump_printf (MSG_NOTE, "\n"); | |
10763 | } | |
10764 | vectype = get_vectype_for_scalar_type (scalar_type); | |
10765 | if (!vectype) | |
10766 | { | |
10767 | if (dump_enabled_p ()) | |
10768 | { | |
10769 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10770 | "not vectorized: unsupported data-type "); | |
10771 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
10772 | scalar_type); | |
10773 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
10774 | } | |
10775 | return false; | |
10776 | } | |
10777 | ||
10778 | if (!*stmt_vectype_out) | |
10779 | *stmt_vectype_out = vectype; | |
10780 | ||
10781 | if (dump_enabled_p ()) | |
10782 | { | |
10783 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); | |
10784 | dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); | |
10785 | dump_printf (MSG_NOTE, "\n"); | |
10786 | } | |
10787 | } | |
10788 | ||
10789 | /* Don't try to compute scalar types if the stmt produces a boolean | |
10790 | vector; use the existing vector type instead. */ | |
10791 | tree nunits_vectype; | |
10792 | if (VECTOR_BOOLEAN_TYPE_P (vectype)) | |
10793 | nunits_vectype = vectype; | |
10794 | else | |
10795 | { | |
10796 | /* The number of units is set according to the smallest scalar | |
10797 | type (or the largest vector size, but we only support one | |
10798 | vector size per vectorization). */ | |
10799 | if (*stmt_vectype_out != boolean_type_node) | |
10800 | { | |
10801 | HOST_WIDE_INT dummy; | |
10802 | scalar_type = vect_get_smallest_scalar_type (stmt, &dummy, &dummy); | |
10803 | } | |
10804 | if (dump_enabled_p ()) | |
10805 | { | |
10806 | dump_printf_loc (MSG_NOTE, vect_location, | |
10807 | "get vectype for scalar type: "); | |
10808 | dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); | |
10809 | dump_printf (MSG_NOTE, "\n"); | |
10810 | } | |
10811 | nunits_vectype = get_vectype_for_scalar_type (scalar_type); | |
10812 | } | |
10813 | if (!nunits_vectype) | |
10814 | { | |
10815 | if (dump_enabled_p ()) | |
10816 | { | |
10817 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10818 | "not vectorized: unsupported data-type "); | |
10819 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, scalar_type); | |
10820 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
10821 | } | |
10822 | return false; | |
10823 | } | |
10824 | ||
10825 | if (maybe_ne (GET_MODE_SIZE (TYPE_MODE (vectype)), | |
10826 | GET_MODE_SIZE (TYPE_MODE (nunits_vectype)))) | |
10827 | { | |
10828 | if (dump_enabled_p ()) | |
10829 | { | |
10830 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10831 | "not vectorized: different sized vector " | |
10832 | "types in statement, "); | |
10833 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, vectype); | |
10834 | dump_printf (MSG_MISSED_OPTIMIZATION, " and "); | |
10835 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, nunits_vectype); | |
10836 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
10837 | } | |
10838 | return false; | |
10839 | } | |
10840 | ||
10841 | if (dump_enabled_p ()) | |
10842 | { | |
10843 | dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); | |
10844 | dump_generic_expr (MSG_NOTE, TDF_SLIM, nunits_vectype); | |
10845 | dump_printf (MSG_NOTE, "\n"); | |
10846 | ||
10847 | dump_printf_loc (MSG_NOTE, vect_location, "nunits = "); | |
10848 | dump_dec (MSG_NOTE, TYPE_VECTOR_SUBPARTS (nunits_vectype)); | |
10849 | dump_printf (MSG_NOTE, "\n"); | |
10850 | } | |
10851 | ||
10852 | *nunits_vectype_out = nunits_vectype; | |
10853 | return true; | |
10854 | } | |
10855 | ||
10856 | /* Try to determine the correct vector type for STMT_INFO, which is a | |
10857 | statement that produces a scalar boolean result. Return the vector | |
10858 | type on success, otherwise return NULL_TREE. */ | |
10859 | ||
10860 | tree | |
10861 | vect_get_mask_type_for_stmt (stmt_vec_info stmt_info) | |
10862 | { | |
10863 | gimple *stmt = stmt_info->stmt; | |
10864 | tree mask_type = NULL; | |
10865 | tree vectype, scalar_type; | |
10866 | ||
10867 | if (is_gimple_assign (stmt) | |
10868 | && TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)) == tcc_comparison | |
10869 | && !VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (gimple_assign_rhs1 (stmt)))) | |
10870 | { | |
10871 | scalar_type = TREE_TYPE (gimple_assign_rhs1 (stmt)); | |
10872 | mask_type = get_mask_type_for_scalar_type (scalar_type); | |
10873 | ||
10874 | if (!mask_type) | |
10875 | { | |
10876 | if (dump_enabled_p ()) | |
10877 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10878 | "not vectorized: unsupported mask\n"); | |
10879 | return NULL_TREE; | |
10880 | } | |
10881 | } | |
10882 | else | |
10883 | { | |
10884 | tree rhs; | |
10885 | ssa_op_iter iter; | |
10886 | gimple *def_stmt; | |
10887 | enum vect_def_type dt; | |
10888 | ||
10889 | FOR_EACH_SSA_TREE_OPERAND (rhs, stmt, iter, SSA_OP_USE) | |
10890 | { | |
10891 | if (!vect_is_simple_use (rhs, stmt_info->vinfo, | |
10892 | &def_stmt, &dt, &vectype)) | |
10893 | { | |
10894 | if (dump_enabled_p ()) | |
10895 | { | |
10896 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10897 | "not vectorized: can't compute mask type " | |
10898 | "for statement, "); | |
10899 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, | |
10900 | 0); | |
10901 | } | |
10902 | return NULL_TREE; | |
10903 | } | |
10904 | ||
10905 | /* No vectype probably means external definition. | |
10906 | Allow it in case there is another operand which | |
10907 | allows to determine mask type. */ | |
10908 | if (!vectype) | |
10909 | continue; | |
10910 | ||
10911 | if (!mask_type) | |
10912 | mask_type = vectype; | |
10913 | else if (maybe_ne (TYPE_VECTOR_SUBPARTS (mask_type), | |
10914 | TYPE_VECTOR_SUBPARTS (vectype))) | |
10915 | { | |
10916 | if (dump_enabled_p ()) | |
10917 | { | |
10918 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10919 | "not vectorized: different sized masks " | |
10920 | "types in statement, "); | |
10921 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
10922 | mask_type); | |
10923 | dump_printf (MSG_MISSED_OPTIMIZATION, " and "); | |
10924 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
10925 | vectype); | |
10926 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
10927 | } | |
10928 | return NULL_TREE; | |
10929 | } | |
10930 | else if (VECTOR_BOOLEAN_TYPE_P (mask_type) | |
10931 | != VECTOR_BOOLEAN_TYPE_P (vectype)) | |
10932 | { | |
10933 | if (dump_enabled_p ()) | |
10934 | { | |
10935 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10936 | "not vectorized: mixed mask and " | |
10937 | "nonmask vector types in statement, "); | |
10938 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
10939 | mask_type); | |
10940 | dump_printf (MSG_MISSED_OPTIMIZATION, " and "); | |
10941 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
10942 | vectype); | |
10943 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
10944 | } | |
10945 | return NULL_TREE; | |
10946 | } | |
10947 | } | |
10948 | ||
10949 | /* We may compare boolean value loaded as vector of integers. | |
10950 | Fix mask_type in such case. */ | |
10951 | if (mask_type | |
10952 | && !VECTOR_BOOLEAN_TYPE_P (mask_type) | |
10953 | && gimple_code (stmt) == GIMPLE_ASSIGN | |
10954 | && TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)) == tcc_comparison) | |
10955 | mask_type = build_same_sized_truth_vector_type (mask_type); | |
10956 | } | |
10957 | ||
10958 | /* No mask_type should mean loop invariant predicate. | |
10959 | This is probably a subject for optimization in if-conversion. */ | |
10960 | if (!mask_type && dump_enabled_p ()) | |
10961 | { | |
10962 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
10963 | "not vectorized: can't compute mask type " | |
10964 | "for statement, "); | |
10965 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
10966 | } | |
10967 | return mask_type; | |
10968 | } |