]>
Commit | Line | Data |
---|---|---|
ebfd146a | 1 | /* Statement Analysis and Transformation for Vectorization |
85ec4feb | 2 | Copyright (C) 2003-2018 Free Software Foundation, Inc. |
b8698a0f | 3 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
4 | and Ira Rosen <irar@il.ibm.com> |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 3, or (at your option) any later | |
11 | version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
c7131fb2 | 25 | #include "backend.h" |
957060b5 AM |
26 | #include "target.h" |
27 | #include "rtl.h" | |
ebfd146a | 28 | #include "tree.h" |
c7131fb2 | 29 | #include "gimple.h" |
c7131fb2 | 30 | #include "ssa.h" |
957060b5 AM |
31 | #include "optabs-tree.h" |
32 | #include "insn-config.h" | |
33 | #include "recog.h" /* FIXME: for insn_data */ | |
34 | #include "cgraph.h" | |
957060b5 | 35 | #include "dumpfile.h" |
c7131fb2 | 36 | #include "alias.h" |
40e23961 | 37 | #include "fold-const.h" |
d8a2d370 | 38 | #include "stor-layout.h" |
2fb9a547 | 39 | #include "tree-eh.h" |
45b0be94 | 40 | #include "gimplify.h" |
5be5c238 | 41 | #include "gimple-iterator.h" |
18f429e2 | 42 | #include "gimplify-me.h" |
442b4905 | 43 | #include "tree-cfg.h" |
e28030cf | 44 | #include "tree-ssa-loop-manip.h" |
ebfd146a | 45 | #include "cfgloop.h" |
0136f8f0 AH |
46 | #include "tree-ssa-loop.h" |
47 | #include "tree-scalar-evolution.h" | |
ebfd146a | 48 | #include "tree-vectorizer.h" |
9b2b7279 | 49 | #include "builtins.h" |
70439f0d | 50 | #include "internal-fn.h" |
5ebaa477 | 51 | #include "tree-vector-builder.h" |
f151c9e1 | 52 | #include "vec-perm-indices.h" |
7cfb4d93 RS |
53 | #include "tree-ssa-loop-niter.h" |
54 | #include "gimple-fold.h" | |
ebfd146a | 55 | |
7ee2468b SB |
56 | /* For lang_hooks.types.type_for_mode. */ |
57 | #include "langhooks.h" | |
ebfd146a | 58 | |
c3e7ee41 BS |
59 | /* Return the vectorized type for the given statement. */ |
60 | ||
61 | tree | |
62 | stmt_vectype (struct _stmt_vec_info *stmt_info) | |
63 | { | |
64 | return STMT_VINFO_VECTYPE (stmt_info); | |
65 | } | |
66 | ||
67 | /* Return TRUE iff the given statement is in an inner loop relative to | |
68 | the loop being vectorized. */ | |
69 | bool | |
70 | stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info) | |
71 | { | |
355fe088 | 72 | gimple *stmt = STMT_VINFO_STMT (stmt_info); |
c3e7ee41 BS |
73 | basic_block bb = gimple_bb (stmt); |
74 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
75 | struct loop* loop; | |
76 | ||
77 | if (!loop_vinfo) | |
78 | return false; | |
79 | ||
80 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
81 | ||
82 | return (bb->loop_father == loop->inner); | |
83 | } | |
84 | ||
85 | /* Record the cost of a statement, either by directly informing the | |
86 | target model or by saving it in a vector for later processing. | |
87 | Return a preliminary estimate of the statement's cost. */ | |
88 | ||
89 | unsigned | |
92345349 | 90 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
c3e7ee41 | 91 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
92345349 | 92 | int misalign, enum vect_cost_model_location where) |
c3e7ee41 | 93 | { |
cc9fe6bb JH |
94 | if ((kind == vector_load || kind == unaligned_load) |
95 | && STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
96 | kind = vector_gather_load; | |
97 | if ((kind == vector_store || kind == unaligned_store) | |
98 | && STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
99 | kind = vector_scatter_store; | |
92345349 | 100 | if (body_cost_vec) |
c3e7ee41 | 101 | { |
92345349 | 102 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
ddf56386 RB |
103 | stmt_info_for_cost si = { count, kind, |
104 | stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL, | |
105 | misalign }; | |
106 | body_cost_vec->safe_push (si); | |
c3e7ee41 | 107 | return (unsigned) |
92345349 | 108 | (builtin_vectorization_cost (kind, vectype, misalign) * count); |
c3e7ee41 BS |
109 | } |
110 | else | |
310213d4 RB |
111 | return add_stmt_cost (stmt_info->vinfo->target_cost_data, |
112 | count, kind, stmt_info, misalign, where); | |
c3e7ee41 BS |
113 | } |
114 | ||
272c6793 RS |
115 | /* Return a variable of type ELEM_TYPE[NELEMS]. */ |
116 | ||
117 | static tree | |
118 | create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems) | |
119 | { | |
120 | return create_tmp_var (build_array_type_nelts (elem_type, nelems), | |
121 | "vect_array"); | |
122 | } | |
123 | ||
124 | /* ARRAY is an array of vectors created by create_vector_array. | |
125 | Return an SSA_NAME for the vector in index N. The reference | |
126 | is part of the vectorization of STMT and the vector is associated | |
127 | with scalar destination SCALAR_DEST. */ | |
128 | ||
129 | static tree | |
355fe088 | 130 | read_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree scalar_dest, |
272c6793 RS |
131 | tree array, unsigned HOST_WIDE_INT n) |
132 | { | |
133 | tree vect_type, vect, vect_name, array_ref; | |
355fe088 | 134 | gimple *new_stmt; |
272c6793 RS |
135 | |
136 | gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE); | |
137 | vect_type = TREE_TYPE (TREE_TYPE (array)); | |
138 | vect = vect_create_destination_var (scalar_dest, vect_type); | |
139 | array_ref = build4 (ARRAY_REF, vect_type, array, | |
140 | build_int_cst (size_type_node, n), | |
141 | NULL_TREE, NULL_TREE); | |
142 | ||
143 | new_stmt = gimple_build_assign (vect, array_ref); | |
144 | vect_name = make_ssa_name (vect, new_stmt); | |
145 | gimple_assign_set_lhs (new_stmt, vect_name); | |
146 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
147 | |
148 | return vect_name; | |
149 | } | |
150 | ||
151 | /* ARRAY is an array of vectors created by create_vector_array. | |
152 | Emit code to store SSA_NAME VECT in index N of the array. | |
153 | The store is part of the vectorization of STMT. */ | |
154 | ||
155 | static void | |
355fe088 | 156 | write_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree vect, |
272c6793 RS |
157 | tree array, unsigned HOST_WIDE_INT n) |
158 | { | |
159 | tree array_ref; | |
355fe088 | 160 | gimple *new_stmt; |
272c6793 RS |
161 | |
162 | array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array, | |
163 | build_int_cst (size_type_node, n), | |
164 | NULL_TREE, NULL_TREE); | |
165 | ||
166 | new_stmt = gimple_build_assign (array_ref, vect); | |
167 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
272c6793 RS |
168 | } |
169 | ||
170 | /* PTR is a pointer to an array of type TYPE. Return a representation | |
171 | of *PTR. The memory reference replaces those in FIRST_DR | |
172 | (and its group). */ | |
173 | ||
174 | static tree | |
44fc7854 | 175 | create_array_ref (tree type, tree ptr, tree alias_ptr_type) |
272c6793 | 176 | { |
44fc7854 | 177 | tree mem_ref; |
272c6793 | 178 | |
272c6793 RS |
179 | mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0)); |
180 | /* Arrays have the same alignment as their type. */ | |
644ffefd | 181 | set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0); |
272c6793 RS |
182 | return mem_ref; |
183 | } | |
184 | ||
3ba4ff41 RS |
185 | /* Add a clobber of variable VAR to the vectorization of STMT. |
186 | Emit the clobber before *GSI. */ | |
187 | ||
188 | static void | |
189 | vect_clobber_variable (gimple *stmt, gimple_stmt_iterator *gsi, tree var) | |
190 | { | |
191 | tree clobber = build_clobber (TREE_TYPE (var)); | |
192 | gimple *new_stmt = gimple_build_assign (var, clobber); | |
193 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
194 | } | |
195 | ||
ebfd146a IR |
196 | /* Utility functions used by vect_mark_stmts_to_be_vectorized. */ |
197 | ||
198 | /* Function vect_mark_relevant. | |
199 | ||
200 | Mark STMT as "relevant for vectorization" and add it to WORKLIST. */ | |
201 | ||
202 | static void | |
355fe088 | 203 | vect_mark_relevant (vec<gimple *> *worklist, gimple *stmt, |
97ecdb46 | 204 | enum vect_relevant relevant, bool live_p) |
ebfd146a IR |
205 | { |
206 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
207 | enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
208 | bool save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
355fe088 | 209 | gimple *pattern_stmt; |
ebfd146a | 210 | |
73fbfcad | 211 | if (dump_enabled_p ()) |
66c16fd9 RB |
212 | { |
213 | dump_printf_loc (MSG_NOTE, vect_location, | |
214 | "mark relevant %d, live %d: ", relevant, live_p); | |
215 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
216 | } | |
ebfd146a | 217 | |
83197f37 IR |
218 | /* If this stmt is an original stmt in a pattern, we might need to mark its |
219 | related pattern stmt instead of the original stmt. However, such stmts | |
220 | may have their own uses that are not in any pattern, in such cases the | |
221 | stmt itself should be marked. */ | |
ebfd146a IR |
222 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
223 | { | |
97ecdb46 JJ |
224 | /* This is the last stmt in a sequence that was detected as a |
225 | pattern that can potentially be vectorized. Don't mark the stmt | |
226 | as relevant/live because it's not going to be vectorized. | |
227 | Instead mark the pattern-stmt that replaces it. */ | |
83197f37 | 228 | |
97ecdb46 JJ |
229 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); |
230 | ||
231 | if (dump_enabled_p ()) | |
232 | dump_printf_loc (MSG_NOTE, vect_location, | |
233 | "last stmt in pattern. don't mark" | |
234 | " relevant/live.\n"); | |
235 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
236 | gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt); | |
237 | save_relevant = STMT_VINFO_RELEVANT (stmt_info); | |
238 | save_live_p = STMT_VINFO_LIVE_P (stmt_info); | |
239 | stmt = pattern_stmt; | |
ebfd146a IR |
240 | } |
241 | ||
242 | STMT_VINFO_LIVE_P (stmt_info) |= live_p; | |
243 | if (relevant > STMT_VINFO_RELEVANT (stmt_info)) | |
244 | STMT_VINFO_RELEVANT (stmt_info) = relevant; | |
245 | ||
246 | if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant | |
247 | && STMT_VINFO_LIVE_P (stmt_info) == save_live_p) | |
248 | { | |
73fbfcad | 249 | if (dump_enabled_p ()) |
78c60e3d | 250 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 251 | "already marked relevant/live.\n"); |
ebfd146a IR |
252 | return; |
253 | } | |
254 | ||
9771b263 | 255 | worklist->safe_push (stmt); |
ebfd146a IR |
256 | } |
257 | ||
258 | ||
b28ead45 AH |
259 | /* Function is_simple_and_all_uses_invariant |
260 | ||
261 | Return true if STMT is simple and all uses of it are invariant. */ | |
262 | ||
263 | bool | |
264 | is_simple_and_all_uses_invariant (gimple *stmt, loop_vec_info loop_vinfo) | |
265 | { | |
266 | tree op; | |
267 | gimple *def_stmt; | |
268 | ssa_op_iter iter; | |
269 | ||
270 | if (!is_gimple_assign (stmt)) | |
271 | return false; | |
272 | ||
273 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) | |
274 | { | |
275 | enum vect_def_type dt = vect_uninitialized_def; | |
276 | ||
277 | if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt)) | |
278 | { | |
279 | if (dump_enabled_p ()) | |
280 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
281 | "use not simple.\n"); | |
282 | return false; | |
283 | } | |
284 | ||
285 | if (dt != vect_external_def && dt != vect_constant_def) | |
286 | return false; | |
287 | } | |
288 | return true; | |
289 | } | |
290 | ||
ebfd146a IR |
291 | /* Function vect_stmt_relevant_p. |
292 | ||
293 | Return true if STMT in loop that is represented by LOOP_VINFO is | |
294 | "relevant for vectorization". | |
295 | ||
296 | A stmt is considered "relevant for vectorization" if: | |
297 | - it has uses outside the loop. | |
298 | - it has vdefs (it alters memory). | |
299 | - control stmts in the loop (except for the exit condition). | |
300 | ||
301 | CHECKME: what other side effects would the vectorizer allow? */ | |
302 | ||
303 | static bool | |
355fe088 | 304 | vect_stmt_relevant_p (gimple *stmt, loop_vec_info loop_vinfo, |
ebfd146a IR |
305 | enum vect_relevant *relevant, bool *live_p) |
306 | { | |
307 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
308 | ssa_op_iter op_iter; | |
309 | imm_use_iterator imm_iter; | |
310 | use_operand_p use_p; | |
311 | def_operand_p def_p; | |
312 | ||
8644a673 | 313 | *relevant = vect_unused_in_scope; |
ebfd146a IR |
314 | *live_p = false; |
315 | ||
316 | /* cond stmt other than loop exit cond. */ | |
b8698a0f L |
317 | if (is_ctrl_stmt (stmt) |
318 | && STMT_VINFO_TYPE (vinfo_for_stmt (stmt)) | |
319 | != loop_exit_ctrl_vec_info_type) | |
8644a673 | 320 | *relevant = vect_used_in_scope; |
ebfd146a IR |
321 | |
322 | /* changing memory. */ | |
323 | if (gimple_code (stmt) != GIMPLE_PHI) | |
ac6aeab4 RB |
324 | if (gimple_vdef (stmt) |
325 | && !gimple_clobber_p (stmt)) | |
ebfd146a | 326 | { |
73fbfcad | 327 | if (dump_enabled_p ()) |
78c60e3d | 328 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 329 | "vec_stmt_relevant_p: stmt has vdefs.\n"); |
8644a673 | 330 | *relevant = vect_used_in_scope; |
ebfd146a IR |
331 | } |
332 | ||
333 | /* uses outside the loop. */ | |
334 | FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) | |
335 | { | |
336 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p)) | |
337 | { | |
338 | basic_block bb = gimple_bb (USE_STMT (use_p)); | |
339 | if (!flow_bb_inside_loop_p (loop, bb)) | |
340 | { | |
73fbfcad | 341 | if (dump_enabled_p ()) |
78c60e3d | 342 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 343 | "vec_stmt_relevant_p: used out of loop.\n"); |
ebfd146a | 344 | |
3157b0c2 AO |
345 | if (is_gimple_debug (USE_STMT (use_p))) |
346 | continue; | |
347 | ||
ebfd146a IR |
348 | /* We expect all such uses to be in the loop exit phis |
349 | (because of loop closed form) */ | |
350 | gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI); | |
351 | gcc_assert (bb == single_exit (loop)->dest); | |
352 | ||
353 | *live_p = true; | |
354 | } | |
355 | } | |
356 | } | |
357 | ||
3a2edf4c AH |
358 | if (*live_p && *relevant == vect_unused_in_scope |
359 | && !is_simple_and_all_uses_invariant (stmt, loop_vinfo)) | |
b28ead45 AH |
360 | { |
361 | if (dump_enabled_p ()) | |
362 | dump_printf_loc (MSG_NOTE, vect_location, | |
363 | "vec_stmt_relevant_p: stmt live but not relevant.\n"); | |
364 | *relevant = vect_used_only_live; | |
365 | } | |
366 | ||
ebfd146a IR |
367 | return (*live_p || *relevant); |
368 | } | |
369 | ||
370 | ||
b8698a0f | 371 | /* Function exist_non_indexing_operands_for_use_p |
ebfd146a | 372 | |
ff802fa1 | 373 | USE is one of the uses attached to STMT. Check if USE is |
ebfd146a IR |
374 | used in STMT for anything other than indexing an array. */ |
375 | ||
376 | static bool | |
355fe088 | 377 | exist_non_indexing_operands_for_use_p (tree use, gimple *stmt) |
ebfd146a IR |
378 | { |
379 | tree operand; | |
380 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
59a05b0c | 381 | |
ff802fa1 | 382 | /* USE corresponds to some operand in STMT. If there is no data |
ebfd146a IR |
383 | reference in STMT, then any operand that corresponds to USE |
384 | is not indexing an array. */ | |
385 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
386 | return true; | |
59a05b0c | 387 | |
ebfd146a IR |
388 | /* STMT has a data_ref. FORNOW this means that its of one of |
389 | the following forms: | |
390 | -1- ARRAY_REF = var | |
391 | -2- var = ARRAY_REF | |
392 | (This should have been verified in analyze_data_refs). | |
393 | ||
394 | 'var' in the second case corresponds to a def, not a use, | |
b8698a0f | 395 | so USE cannot correspond to any operands that are not used |
ebfd146a IR |
396 | for array indexing. |
397 | ||
398 | Therefore, all we need to check is if STMT falls into the | |
399 | first case, and whether var corresponds to USE. */ | |
ebfd146a IR |
400 | |
401 | if (!gimple_assign_copy_p (stmt)) | |
5ce9450f JJ |
402 | { |
403 | if (is_gimple_call (stmt) | |
404 | && gimple_call_internal_p (stmt)) | |
bfaa08b7 RS |
405 | { |
406 | internal_fn ifn = gimple_call_internal_fn (stmt); | |
407 | int mask_index = internal_fn_mask_index (ifn); | |
408 | if (mask_index >= 0 | |
409 | && use == gimple_call_arg (stmt, mask_index)) | |
410 | return true; | |
f307441a RS |
411 | int stored_value_index = internal_fn_stored_value_index (ifn); |
412 | if (stored_value_index >= 0 | |
413 | && use == gimple_call_arg (stmt, stored_value_index)) | |
414 | return true; | |
bfaa08b7 RS |
415 | if (internal_gather_scatter_fn_p (ifn) |
416 | && use == gimple_call_arg (stmt, 1)) | |
417 | return true; | |
bfaa08b7 | 418 | } |
5ce9450f JJ |
419 | return false; |
420 | } | |
421 | ||
59a05b0c EB |
422 | if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME) |
423 | return false; | |
ebfd146a | 424 | operand = gimple_assign_rhs1 (stmt); |
ebfd146a IR |
425 | if (TREE_CODE (operand) != SSA_NAME) |
426 | return false; | |
427 | ||
428 | if (operand == use) | |
429 | return true; | |
430 | ||
431 | return false; | |
432 | } | |
433 | ||
434 | ||
b8698a0f | 435 | /* |
ebfd146a IR |
436 | Function process_use. |
437 | ||
438 | Inputs: | |
439 | - a USE in STMT in a loop represented by LOOP_VINFO | |
b28ead45 | 440 | - RELEVANT - enum value to be set in the STMT_VINFO of the stmt |
ff802fa1 | 441 | that defined USE. This is done by calling mark_relevant and passing it |
ebfd146a | 442 | the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant). |
aec7ae7d JJ |
443 | - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't |
444 | be performed. | |
ebfd146a IR |
445 | |
446 | Outputs: | |
447 | Generally, LIVE_P and RELEVANT are used to define the liveness and | |
448 | relevance info of the DEF_STMT of this USE: | |
449 | STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p | |
450 | STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant | |
451 | Exceptions: | |
452 | - case 1: If USE is used only for address computations (e.g. array indexing), | |
b8698a0f | 453 | which does not need to be directly vectorized, then the liveness/relevance |
ebfd146a | 454 | of the respective DEF_STMT is left unchanged. |
b8698a0f L |
455 | - case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we |
456 | skip DEF_STMT cause it had already been processed. | |
ebfd146a IR |
457 | - case 3: If DEF_STMT and STMT are in different nests, then "relevant" will |
458 | be modified accordingly. | |
459 | ||
460 | Return true if everything is as expected. Return false otherwise. */ | |
461 | ||
462 | static bool | |
b28ead45 | 463 | process_use (gimple *stmt, tree use, loop_vec_info loop_vinfo, |
355fe088 | 464 | enum vect_relevant relevant, vec<gimple *> *worklist, |
aec7ae7d | 465 | bool force) |
ebfd146a IR |
466 | { |
467 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
468 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
469 | stmt_vec_info dstmt_vinfo; | |
470 | basic_block bb, def_bb; | |
355fe088 | 471 | gimple *def_stmt; |
ebfd146a IR |
472 | enum vect_def_type dt; |
473 | ||
b8698a0f | 474 | /* case 1: we are only interested in uses that need to be vectorized. Uses |
ebfd146a | 475 | that are used for address computation are not considered relevant. */ |
aec7ae7d | 476 | if (!force && !exist_non_indexing_operands_for_use_p (use, stmt)) |
ebfd146a IR |
477 | return true; |
478 | ||
81c40241 | 479 | if (!vect_is_simple_use (use, loop_vinfo, &def_stmt, &dt)) |
b8698a0f | 480 | { |
73fbfcad | 481 | if (dump_enabled_p ()) |
78c60e3d | 482 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 483 | "not vectorized: unsupported use in stmt.\n"); |
ebfd146a IR |
484 | return false; |
485 | } | |
486 | ||
487 | if (!def_stmt || gimple_nop_p (def_stmt)) | |
488 | return true; | |
489 | ||
490 | def_bb = gimple_bb (def_stmt); | |
491 | if (!flow_bb_inside_loop_p (loop, def_bb)) | |
492 | { | |
73fbfcad | 493 | if (dump_enabled_p ()) |
e645e942 | 494 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n"); |
ebfd146a IR |
495 | return true; |
496 | } | |
497 | ||
b8698a0f L |
498 | /* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT). |
499 | DEF_STMT must have already been processed, because this should be the | |
500 | only way that STMT, which is a reduction-phi, was put in the worklist, | |
501 | as there should be no other uses for DEF_STMT in the loop. So we just | |
ebfd146a IR |
502 | check that everything is as expected, and we are done. */ |
503 | dstmt_vinfo = vinfo_for_stmt (def_stmt); | |
504 | bb = gimple_bb (stmt); | |
505 | if (gimple_code (stmt) == GIMPLE_PHI | |
506 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def | |
507 | && gimple_code (def_stmt) != GIMPLE_PHI | |
508 | && STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def | |
509 | && bb->loop_father == def_bb->loop_father) | |
510 | { | |
73fbfcad | 511 | if (dump_enabled_p ()) |
78c60e3d | 512 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 513 | "reduc-stmt defining reduc-phi in the same nest.\n"); |
ebfd146a IR |
514 | if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo)) |
515 | dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo)); | |
516 | gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction); | |
b8698a0f | 517 | gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo) |
8644a673 | 518 | || STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope); |
ebfd146a IR |
519 | return true; |
520 | } | |
521 | ||
522 | /* case 3a: outer-loop stmt defining an inner-loop stmt: | |
523 | outer-loop-header-bb: | |
524 | d = def_stmt | |
525 | inner-loop: | |
526 | stmt # use (d) | |
527 | outer-loop-tail-bb: | |
528 | ... */ | |
529 | if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father)) | |
530 | { | |
73fbfcad | 531 | if (dump_enabled_p ()) |
78c60e3d | 532 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 533 | "outer-loop def-stmt defining inner-loop stmt.\n"); |
7c5222ff | 534 | |
ebfd146a IR |
535 | switch (relevant) |
536 | { | |
8644a673 | 537 | case vect_unused_in_scope: |
7c5222ff IR |
538 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ? |
539 | vect_used_in_scope : vect_unused_in_scope; | |
ebfd146a | 540 | break; |
7c5222ff | 541 | |
ebfd146a | 542 | case vect_used_in_outer_by_reduction: |
7c5222ff | 543 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
ebfd146a IR |
544 | relevant = vect_used_by_reduction; |
545 | break; | |
7c5222ff | 546 | |
ebfd146a | 547 | case vect_used_in_outer: |
7c5222ff | 548 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def); |
8644a673 | 549 | relevant = vect_used_in_scope; |
ebfd146a | 550 | break; |
7c5222ff | 551 | |
8644a673 | 552 | case vect_used_in_scope: |
ebfd146a IR |
553 | break; |
554 | ||
555 | default: | |
556 | gcc_unreachable (); | |
b8698a0f | 557 | } |
ebfd146a IR |
558 | } |
559 | ||
560 | /* case 3b: inner-loop stmt defining an outer-loop stmt: | |
561 | outer-loop-header-bb: | |
562 | ... | |
563 | inner-loop: | |
564 | d = def_stmt | |
06066f92 | 565 | outer-loop-tail-bb (or outer-loop-exit-bb in double reduction): |
ebfd146a IR |
566 | stmt # use (d) */ |
567 | else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father)) | |
568 | { | |
73fbfcad | 569 | if (dump_enabled_p ()) |
78c60e3d | 570 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 571 | "inner-loop def-stmt defining outer-loop stmt.\n"); |
7c5222ff | 572 | |
ebfd146a IR |
573 | switch (relevant) |
574 | { | |
8644a673 | 575 | case vect_unused_in_scope: |
b8698a0f | 576 | relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def |
06066f92 | 577 | || STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ? |
a70d6342 | 578 | vect_used_in_outer_by_reduction : vect_unused_in_scope; |
ebfd146a IR |
579 | break; |
580 | ||
ebfd146a | 581 | case vect_used_by_reduction: |
b28ead45 | 582 | case vect_used_only_live: |
ebfd146a IR |
583 | relevant = vect_used_in_outer_by_reduction; |
584 | break; | |
585 | ||
8644a673 | 586 | case vect_used_in_scope: |
ebfd146a IR |
587 | relevant = vect_used_in_outer; |
588 | break; | |
589 | ||
590 | default: | |
591 | gcc_unreachable (); | |
592 | } | |
593 | } | |
643a9684 RB |
594 | /* We are also not interested in uses on loop PHI backedges that are |
595 | inductions. Otherwise we'll needlessly vectorize the IV increment | |
e294f495 RB |
596 | and cause hybrid SLP for SLP inductions. Unless the PHI is live |
597 | of course. */ | |
643a9684 RB |
598 | else if (gimple_code (stmt) == GIMPLE_PHI |
599 | && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_induction_def | |
e294f495 | 600 | && ! STMT_VINFO_LIVE_P (stmt_vinfo) |
643a9684 RB |
601 | && (PHI_ARG_DEF_FROM_EDGE (stmt, loop_latch_edge (bb->loop_father)) |
602 | == use)) | |
603 | { | |
604 | if (dump_enabled_p ()) | |
605 | dump_printf_loc (MSG_NOTE, vect_location, | |
606 | "induction value on backedge.\n"); | |
607 | return true; | |
608 | } | |
609 | ||
ebfd146a | 610 | |
b28ead45 | 611 | vect_mark_relevant (worklist, def_stmt, relevant, false); |
ebfd146a IR |
612 | return true; |
613 | } | |
614 | ||
615 | ||
616 | /* Function vect_mark_stmts_to_be_vectorized. | |
617 | ||
618 | Not all stmts in the loop need to be vectorized. For example: | |
619 | ||
620 | for i... | |
621 | for j... | |
622 | 1. T0 = i + j | |
623 | 2. T1 = a[T0] | |
624 | ||
625 | 3. j = j + 1 | |
626 | ||
627 | Stmt 1 and 3 do not need to be vectorized, because loop control and | |
628 | addressing of vectorized data-refs are handled differently. | |
629 | ||
630 | This pass detects such stmts. */ | |
631 | ||
632 | bool | |
633 | vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo) | |
634 | { | |
ebfd146a IR |
635 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
636 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
637 | unsigned int nbbs = loop->num_nodes; | |
638 | gimple_stmt_iterator si; | |
355fe088 | 639 | gimple *stmt; |
ebfd146a IR |
640 | unsigned int i; |
641 | stmt_vec_info stmt_vinfo; | |
642 | basic_block bb; | |
355fe088 | 643 | gimple *phi; |
ebfd146a | 644 | bool live_p; |
b28ead45 | 645 | enum vect_relevant relevant; |
ebfd146a | 646 | |
73fbfcad | 647 | if (dump_enabled_p ()) |
78c60e3d | 648 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 649 | "=== vect_mark_stmts_to_be_vectorized ===\n"); |
ebfd146a | 650 | |
355fe088 | 651 | auto_vec<gimple *, 64> worklist; |
ebfd146a IR |
652 | |
653 | /* 1. Init worklist. */ | |
654 | for (i = 0; i < nbbs; i++) | |
655 | { | |
656 | bb = bbs[i]; | |
657 | for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) | |
b8698a0f | 658 | { |
ebfd146a | 659 | phi = gsi_stmt (si); |
73fbfcad | 660 | if (dump_enabled_p ()) |
ebfd146a | 661 | { |
78c60e3d SS |
662 | dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? "); |
663 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
664 | } |
665 | ||
666 | if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 667 | vect_mark_relevant (&worklist, phi, relevant, live_p); |
ebfd146a IR |
668 | } |
669 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
670 | { | |
671 | stmt = gsi_stmt (si); | |
73fbfcad | 672 | if (dump_enabled_p ()) |
ebfd146a | 673 | { |
78c60e3d SS |
674 | dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? "); |
675 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
b8698a0f | 676 | } |
ebfd146a IR |
677 | |
678 | if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p)) | |
97ecdb46 | 679 | vect_mark_relevant (&worklist, stmt, relevant, live_p); |
ebfd146a IR |
680 | } |
681 | } | |
682 | ||
683 | /* 2. Process_worklist */ | |
9771b263 | 684 | while (worklist.length () > 0) |
ebfd146a IR |
685 | { |
686 | use_operand_p use_p; | |
687 | ssa_op_iter iter; | |
688 | ||
9771b263 | 689 | stmt = worklist.pop (); |
73fbfcad | 690 | if (dump_enabled_p ()) |
ebfd146a | 691 | { |
78c60e3d SS |
692 | dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: "); |
693 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
ebfd146a IR |
694 | } |
695 | ||
b8698a0f | 696 | /* Examine the USEs of STMT. For each USE, mark the stmt that defines it |
b28ead45 AH |
697 | (DEF_STMT) as relevant/irrelevant according to the relevance property |
698 | of STMT. */ | |
ebfd146a IR |
699 | stmt_vinfo = vinfo_for_stmt (stmt); |
700 | relevant = STMT_VINFO_RELEVANT (stmt_vinfo); | |
ebfd146a | 701 | |
b28ead45 AH |
702 | /* Generally, the relevance property of STMT (in STMT_VINFO_RELEVANT) is |
703 | propagated as is to the DEF_STMTs of its USEs. | |
ebfd146a IR |
704 | |
705 | One exception is when STMT has been identified as defining a reduction | |
b28ead45 | 706 | variable; in this case we set the relevance to vect_used_by_reduction. |
ebfd146a | 707 | This is because we distinguish between two kinds of relevant stmts - |
b8698a0f | 708 | those that are used by a reduction computation, and those that are |
ff802fa1 | 709 | (also) used by a regular computation. This allows us later on to |
b8698a0f | 710 | identify stmts that are used solely by a reduction, and therefore the |
7c5222ff | 711 | order of the results that they produce does not have to be kept. */ |
ebfd146a | 712 | |
b28ead45 | 713 | switch (STMT_VINFO_DEF_TYPE (stmt_vinfo)) |
ebfd146a | 714 | { |
06066f92 | 715 | case vect_reduction_def: |
b28ead45 AH |
716 | gcc_assert (relevant != vect_unused_in_scope); |
717 | if (relevant != vect_unused_in_scope | |
718 | && relevant != vect_used_in_scope | |
719 | && relevant != vect_used_by_reduction | |
720 | && relevant != vect_used_only_live) | |
06066f92 | 721 | { |
b28ead45 AH |
722 | if (dump_enabled_p ()) |
723 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
724 | "unsupported use of reduction.\n"); | |
725 | return false; | |
06066f92 | 726 | } |
06066f92 | 727 | break; |
b8698a0f | 728 | |
06066f92 | 729 | case vect_nested_cycle: |
b28ead45 AH |
730 | if (relevant != vect_unused_in_scope |
731 | && relevant != vect_used_in_outer_by_reduction | |
732 | && relevant != vect_used_in_outer) | |
06066f92 | 733 | { |
73fbfcad | 734 | if (dump_enabled_p ()) |
78c60e3d | 735 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 736 | "unsupported use of nested cycle.\n"); |
7c5222ff | 737 | |
06066f92 IR |
738 | return false; |
739 | } | |
b8698a0f L |
740 | break; |
741 | ||
06066f92 | 742 | case vect_double_reduction_def: |
b28ead45 AH |
743 | if (relevant != vect_unused_in_scope |
744 | && relevant != vect_used_by_reduction | |
745 | && relevant != vect_used_only_live) | |
06066f92 | 746 | { |
73fbfcad | 747 | if (dump_enabled_p ()) |
78c60e3d | 748 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 749 | "unsupported use of double reduction.\n"); |
7c5222ff | 750 | |
7c5222ff | 751 | return false; |
06066f92 | 752 | } |
b8698a0f | 753 | break; |
7c5222ff | 754 | |
06066f92 IR |
755 | default: |
756 | break; | |
7c5222ff | 757 | } |
b8698a0f | 758 | |
aec7ae7d | 759 | if (is_pattern_stmt_p (stmt_vinfo)) |
9d5e7640 IR |
760 | { |
761 | /* Pattern statements are not inserted into the code, so | |
762 | FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we | |
763 | have to scan the RHS or function arguments instead. */ | |
764 | if (is_gimple_assign (stmt)) | |
765 | { | |
69d2aade JJ |
766 | enum tree_code rhs_code = gimple_assign_rhs_code (stmt); |
767 | tree op = gimple_assign_rhs1 (stmt); | |
768 | ||
769 | i = 1; | |
770 | if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op)) | |
771 | { | |
772 | if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo, | |
b28ead45 | 773 | relevant, &worklist, false) |
69d2aade | 774 | || !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo, |
b28ead45 | 775 | relevant, &worklist, false)) |
566d377a | 776 | return false; |
69d2aade JJ |
777 | i = 2; |
778 | } | |
779 | for (; i < gimple_num_ops (stmt); i++) | |
9d5e7640 | 780 | { |
69d2aade | 781 | op = gimple_op (stmt, i); |
afbe6325 | 782 | if (TREE_CODE (op) == SSA_NAME |
b28ead45 | 783 | && !process_use (stmt, op, loop_vinfo, relevant, |
afbe6325 | 784 | &worklist, false)) |
07687835 | 785 | return false; |
9d5e7640 IR |
786 | } |
787 | } | |
788 | else if (is_gimple_call (stmt)) | |
789 | { | |
790 | for (i = 0; i < gimple_call_num_args (stmt); i++) | |
791 | { | |
792 | tree arg = gimple_call_arg (stmt, i); | |
b28ead45 | 793 | if (!process_use (stmt, arg, loop_vinfo, relevant, |
aec7ae7d | 794 | &worklist, false)) |
07687835 | 795 | return false; |
9d5e7640 IR |
796 | } |
797 | } | |
798 | } | |
799 | else | |
800 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) | |
801 | { | |
802 | tree op = USE_FROM_PTR (use_p); | |
b28ead45 | 803 | if (!process_use (stmt, op, loop_vinfo, relevant, |
aec7ae7d | 804 | &worklist, false)) |
07687835 | 805 | return false; |
9d5e7640 | 806 | } |
aec7ae7d | 807 | |
3bab6342 | 808 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo)) |
aec7ae7d | 809 | { |
134c85ca RS |
810 | gather_scatter_info gs_info; |
811 | if (!vect_check_gather_scatter (stmt, loop_vinfo, &gs_info)) | |
812 | gcc_unreachable (); | |
813 | if (!process_use (stmt, gs_info.offset, loop_vinfo, relevant, | |
814 | &worklist, true)) | |
566d377a | 815 | return false; |
aec7ae7d | 816 | } |
ebfd146a IR |
817 | } /* while worklist */ |
818 | ||
ebfd146a IR |
819 | return true; |
820 | } | |
821 | ||
822 | ||
b8698a0f | 823 | /* Function vect_model_simple_cost. |
ebfd146a | 824 | |
b8698a0f | 825 | Models cost for simple operations, i.e. those that only emit ncopies of a |
ebfd146a IR |
826 | single op. Right now, this does not account for multiple insns that could |
827 | be generated for the single vector op. We will handle that shortly. */ | |
828 | ||
829 | void | |
b8698a0f | 830 | vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, |
92345349 | 831 | enum vect_def_type *dt, |
4fc5ebf1 | 832 | int ndts, |
92345349 BS |
833 | stmt_vector_for_cost *prologue_cost_vec, |
834 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a IR |
835 | { |
836 | int i; | |
92345349 | 837 | int inside_cost = 0, prologue_cost = 0; |
ebfd146a IR |
838 | |
839 | /* The SLP costs were already calculated during SLP tree build. */ | |
78604de0 | 840 | gcc_assert (!PURE_SLP_STMT (stmt_info)); |
ebfd146a | 841 | |
4fc5ebf1 JG |
842 | /* Cost the "broadcast" of a scalar operand in to a vector operand. |
843 | Use scalar_to_vec to cost the broadcast, as elsewhere in the vector | |
844 | cost model. */ | |
845 | for (i = 0; i < ndts; i++) | |
92345349 | 846 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
4fc5ebf1 | 847 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
92345349 | 848 | stmt_info, 0, vect_prologue); |
c3e7ee41 BS |
849 | |
850 | /* Pass the inside-of-loop statements to the target-specific cost model. */ | |
92345349 BS |
851 | inside_cost = record_stmt_cost (body_cost_vec, ncopies, vector_stmt, |
852 | stmt_info, 0, vect_body); | |
c3e7ee41 | 853 | |
73fbfcad | 854 | if (dump_enabled_p ()) |
78c60e3d SS |
855 | dump_printf_loc (MSG_NOTE, vect_location, |
856 | "vect_model_simple_cost: inside_cost = %d, " | |
e645e942 | 857 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
858 | } |
859 | ||
860 | ||
8bd37302 BS |
861 | /* Model cost for type demotion and promotion operations. PWR is normally |
862 | zero for single-step promotions and demotions. It will be one if | |
863 | two-step promotion/demotion is required, and so on. Each additional | |
864 | step doubles the number of instructions required. */ | |
865 | ||
866 | static void | |
867 | vect_model_promotion_demotion_cost (stmt_vec_info stmt_info, | |
868 | enum vect_def_type *dt, int pwr) | |
869 | { | |
870 | int i, tmp; | |
92345349 | 871 | int inside_cost = 0, prologue_cost = 0; |
c3e7ee41 BS |
872 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
873 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
874 | void *target_cost_data; | |
8bd37302 BS |
875 | |
876 | /* The SLP costs were already calculated during SLP tree build. */ | |
78604de0 | 877 | gcc_assert (!PURE_SLP_STMT (stmt_info)); |
8bd37302 | 878 | |
c3e7ee41 BS |
879 | if (loop_vinfo) |
880 | target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); | |
881 | else | |
882 | target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo); | |
883 | ||
8bd37302 BS |
884 | for (i = 0; i < pwr + 1; i++) |
885 | { | |
886 | tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ? | |
887 | (i + 1) : i; | |
c3e7ee41 | 888 | inside_cost += add_stmt_cost (target_cost_data, vect_pow2 (tmp), |
92345349 BS |
889 | vec_promote_demote, stmt_info, 0, |
890 | vect_body); | |
8bd37302 BS |
891 | } |
892 | ||
893 | /* FORNOW: Assuming maximum 2 args per stmts. */ | |
894 | for (i = 0; i < 2; i++) | |
92345349 BS |
895 | if (dt[i] == vect_constant_def || dt[i] == vect_external_def) |
896 | prologue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt, | |
897 | stmt_info, 0, vect_prologue); | |
8bd37302 | 898 | |
73fbfcad | 899 | if (dump_enabled_p ()) |
78c60e3d SS |
900 | dump_printf_loc (MSG_NOTE, vect_location, |
901 | "vect_model_promotion_demotion_cost: inside_cost = %d, " | |
e645e942 | 902 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
8bd37302 BS |
903 | } |
904 | ||
ebfd146a IR |
905 | /* Function vect_model_store_cost |
906 | ||
0d0293ac MM |
907 | Models cost for stores. In the case of grouped accesses, one access |
908 | has the overhead of the grouped access attributed to it. */ | |
ebfd146a IR |
909 | |
910 | void | |
b8698a0f | 911 | vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, |
2de001ee | 912 | vect_memory_access_type memory_access_type, |
9ce4345a | 913 | vec_load_store_type vls_type, slp_tree slp_node, |
92345349 BS |
914 | stmt_vector_for_cost *prologue_cost_vec, |
915 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a | 916 | { |
92345349 | 917 | unsigned int inside_cost = 0, prologue_cost = 0; |
892a981f RS |
918 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
919 | gimple *first_stmt = STMT_VINFO_STMT (stmt_info); | |
920 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); | |
ebfd146a | 921 | |
9ce4345a | 922 | if (vls_type == VLS_STORE_INVARIANT) |
92345349 BS |
923 | prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec, |
924 | stmt_info, 0, vect_prologue); | |
ebfd146a | 925 | |
892a981f RS |
926 | /* Grouped stores update all elements in the group at once, |
927 | so we want the DR for the first statement. */ | |
928 | if (!slp_node && grouped_access_p) | |
720f5239 | 929 | { |
892a981f RS |
930 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
931 | dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
720f5239 | 932 | } |
ebfd146a | 933 | |
892a981f RS |
934 | /* True if we should include any once-per-group costs as well as |
935 | the cost of the statement itself. For SLP we only get called | |
936 | once per group anyhow. */ | |
937 | bool first_stmt_p = (first_stmt == STMT_VINFO_STMT (stmt_info)); | |
938 | ||
272c6793 | 939 | /* We assume that the cost of a single store-lanes instruction is |
0d0293ac | 940 | equivalent to the cost of GROUP_SIZE separate stores. If a grouped |
272c6793 | 941 | access is instead being provided by a permute-and-store operation, |
2de001ee RS |
942 | include the cost of the permutes. */ |
943 | if (first_stmt_p | |
944 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 945 | { |
e1377713 ES |
946 | /* Uses a high and low interleave or shuffle operations for each |
947 | needed permute. */ | |
892a981f | 948 | int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
e1377713 | 949 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
92345349 BS |
950 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, |
951 | stmt_info, 0, vect_body); | |
ebfd146a | 952 | |
73fbfcad | 953 | if (dump_enabled_p ()) |
78c60e3d | 954 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 955 | "vect_model_store_cost: strided group_size = %d .\n", |
78c60e3d | 956 | group_size); |
ebfd146a IR |
957 | } |
958 | ||
cee62fee | 959 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
ebfd146a | 960 | /* Costs of the stores. */ |
067bc855 RB |
961 | if (memory_access_type == VMAT_ELEMENTWISE |
962 | || memory_access_type == VMAT_GATHER_SCATTER) | |
c5126ce8 RS |
963 | { |
964 | /* N scalar stores plus extracting the elements. */ | |
965 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); | |
966 | inside_cost += record_stmt_cost (body_cost_vec, | |
967 | ncopies * assumed_nunits, | |
968 | scalar_store, stmt_info, 0, vect_body); | |
969 | } | |
f2e2a985 | 970 | else |
892a981f | 971 | vect_get_store_cost (dr, ncopies, &inside_cost, body_cost_vec); |
ebfd146a | 972 | |
2de001ee RS |
973 | if (memory_access_type == VMAT_ELEMENTWISE |
974 | || memory_access_type == VMAT_STRIDED_SLP) | |
c5126ce8 RS |
975 | { |
976 | /* N scalar stores plus extracting the elements. */ | |
977 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); | |
978 | inside_cost += record_stmt_cost (body_cost_vec, | |
979 | ncopies * assumed_nunits, | |
980 | vec_to_scalar, stmt_info, 0, vect_body); | |
981 | } | |
cee62fee | 982 | |
73fbfcad | 983 | if (dump_enabled_p ()) |
78c60e3d SS |
984 | dump_printf_loc (MSG_NOTE, vect_location, |
985 | "vect_model_store_cost: inside_cost = %d, " | |
e645e942 | 986 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
ebfd146a IR |
987 | } |
988 | ||
989 | ||
720f5239 IR |
990 | /* Calculate cost of DR's memory access. */ |
991 | void | |
992 | vect_get_store_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 993 | unsigned int *inside_cost, |
92345349 | 994 | stmt_vector_for_cost *body_cost_vec) |
720f5239 IR |
995 | { |
996 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 997 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 998 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
999 | |
1000 | switch (alignment_support_scheme) | |
1001 | { | |
1002 | case dr_aligned: | |
1003 | { | |
92345349 BS |
1004 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
1005 | vector_store, stmt_info, 0, | |
1006 | vect_body); | |
720f5239 | 1007 | |
73fbfcad | 1008 | if (dump_enabled_p ()) |
78c60e3d | 1009 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1010 | "vect_model_store_cost: aligned.\n"); |
720f5239 IR |
1011 | break; |
1012 | } | |
1013 | ||
1014 | case dr_unaligned_supported: | |
1015 | { | |
720f5239 | 1016 | /* Here, we assign an additional cost for the unaligned store. */ |
92345349 | 1017 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1018 | unaligned_store, stmt_info, |
92345349 | 1019 | DR_MISALIGNMENT (dr), vect_body); |
73fbfcad | 1020 | if (dump_enabled_p ()) |
78c60e3d SS |
1021 | dump_printf_loc (MSG_NOTE, vect_location, |
1022 | "vect_model_store_cost: unaligned supported by " | |
e645e942 | 1023 | "hardware.\n"); |
720f5239 IR |
1024 | break; |
1025 | } | |
1026 | ||
38eec4c6 UW |
1027 | case dr_unaligned_unsupported: |
1028 | { | |
1029 | *inside_cost = VECT_MAX_COST; | |
1030 | ||
73fbfcad | 1031 | if (dump_enabled_p ()) |
78c60e3d | 1032 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1033 | "vect_model_store_cost: unsupported access.\n"); |
38eec4c6 UW |
1034 | break; |
1035 | } | |
1036 | ||
720f5239 IR |
1037 | default: |
1038 | gcc_unreachable (); | |
1039 | } | |
1040 | } | |
1041 | ||
1042 | ||
ebfd146a IR |
1043 | /* Function vect_model_load_cost |
1044 | ||
892a981f RS |
1045 | Models cost for loads. In the case of grouped accesses, one access has |
1046 | the overhead of the grouped access attributed to it. Since unaligned | |
b8698a0f | 1047 | accesses are supported for loads, we also account for the costs of the |
ebfd146a IR |
1048 | access scheme chosen. */ |
1049 | ||
1050 | void | |
92345349 | 1051 | vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, |
2de001ee RS |
1052 | vect_memory_access_type memory_access_type, |
1053 | slp_tree slp_node, | |
92345349 BS |
1054 | stmt_vector_for_cost *prologue_cost_vec, |
1055 | stmt_vector_for_cost *body_cost_vec) | |
ebfd146a | 1056 | { |
892a981f RS |
1057 | gimple *first_stmt = STMT_VINFO_STMT (stmt_info); |
1058 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
92345349 | 1059 | unsigned int inside_cost = 0, prologue_cost = 0; |
892a981f | 1060 | bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info); |
ebfd146a | 1061 | |
892a981f RS |
1062 | /* Grouped loads read all elements in the group at once, |
1063 | so we want the DR for the first statement. */ | |
1064 | if (!slp_node && grouped_access_p) | |
ebfd146a | 1065 | { |
892a981f RS |
1066 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
1067 | dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
ebfd146a IR |
1068 | } |
1069 | ||
892a981f RS |
1070 | /* True if we should include any once-per-group costs as well as |
1071 | the cost of the statement itself. For SLP we only get called | |
1072 | once per group anyhow. */ | |
1073 | bool first_stmt_p = (first_stmt == STMT_VINFO_STMT (stmt_info)); | |
1074 | ||
272c6793 | 1075 | /* We assume that the cost of a single load-lanes instruction is |
0d0293ac | 1076 | equivalent to the cost of GROUP_SIZE separate loads. If a grouped |
272c6793 | 1077 | access is instead being provided by a load-and-permute operation, |
2de001ee RS |
1078 | include the cost of the permutes. */ |
1079 | if (first_stmt_p | |
1080 | && memory_access_type == VMAT_CONTIGUOUS_PERMUTE) | |
ebfd146a | 1081 | { |
2c23db6d ES |
1082 | /* Uses an even and odd extract operations or shuffle operations |
1083 | for each needed permute. */ | |
892a981f | 1084 | int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
2c23db6d ES |
1085 | int nstmts = ncopies * ceil_log2 (group_size) * group_size; |
1086 | inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm, | |
1087 | stmt_info, 0, vect_body); | |
ebfd146a | 1088 | |
73fbfcad | 1089 | if (dump_enabled_p ()) |
e645e942 TJ |
1090 | dump_printf_loc (MSG_NOTE, vect_location, |
1091 | "vect_model_load_cost: strided group_size = %d .\n", | |
78c60e3d | 1092 | group_size); |
ebfd146a IR |
1093 | } |
1094 | ||
1095 | /* The loads themselves. */ | |
067bc855 RB |
1096 | if (memory_access_type == VMAT_ELEMENTWISE |
1097 | || memory_access_type == VMAT_GATHER_SCATTER) | |
a82960aa | 1098 | { |
a21892ad BS |
1099 | /* N scalar loads plus gathering them into a vector. */ |
1100 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
c5126ce8 | 1101 | unsigned int assumed_nunits = vect_nunits_for_cost (vectype); |
92345349 | 1102 | inside_cost += record_stmt_cost (body_cost_vec, |
c5126ce8 | 1103 | ncopies * assumed_nunits, |
92345349 | 1104 | scalar_load, stmt_info, 0, vect_body); |
a82960aa RG |
1105 | } |
1106 | else | |
892a981f | 1107 | vect_get_load_cost (dr, ncopies, first_stmt_p, |
92345349 BS |
1108 | &inside_cost, &prologue_cost, |
1109 | prologue_cost_vec, body_cost_vec, true); | |
2de001ee RS |
1110 | if (memory_access_type == VMAT_ELEMENTWISE |
1111 | || memory_access_type == VMAT_STRIDED_SLP) | |
892a981f RS |
1112 | inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct, |
1113 | stmt_info, 0, vect_body); | |
720f5239 | 1114 | |
73fbfcad | 1115 | if (dump_enabled_p ()) |
78c60e3d SS |
1116 | dump_printf_loc (MSG_NOTE, vect_location, |
1117 | "vect_model_load_cost: inside_cost = %d, " | |
e645e942 | 1118 | "prologue_cost = %d .\n", inside_cost, prologue_cost); |
720f5239 IR |
1119 | } |
1120 | ||
1121 | ||
1122 | /* Calculate cost of DR's memory access. */ | |
1123 | void | |
1124 | vect_get_load_cost (struct data_reference *dr, int ncopies, | |
c3e7ee41 | 1125 | bool add_realign_cost, unsigned int *inside_cost, |
92345349 BS |
1126 | unsigned int *prologue_cost, |
1127 | stmt_vector_for_cost *prologue_cost_vec, | |
1128 | stmt_vector_for_cost *body_cost_vec, | |
1129 | bool record_prologue_costs) | |
720f5239 IR |
1130 | { |
1131 | int alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
355fe088 | 1132 | gimple *stmt = DR_STMT (dr); |
c3e7ee41 | 1133 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
720f5239 IR |
1134 | |
1135 | switch (alignment_support_scheme) | |
ebfd146a IR |
1136 | { |
1137 | case dr_aligned: | |
1138 | { | |
92345349 BS |
1139 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1140 | stmt_info, 0, vect_body); | |
ebfd146a | 1141 | |
73fbfcad | 1142 | if (dump_enabled_p ()) |
78c60e3d | 1143 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1144 | "vect_model_load_cost: aligned.\n"); |
ebfd146a IR |
1145 | |
1146 | break; | |
1147 | } | |
1148 | case dr_unaligned_supported: | |
1149 | { | |
720f5239 | 1150 | /* Here, we assign an additional cost for the unaligned load. */ |
92345349 | 1151 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, |
c3e7ee41 | 1152 | unaligned_load, stmt_info, |
92345349 | 1153 | DR_MISALIGNMENT (dr), vect_body); |
c3e7ee41 | 1154 | |
73fbfcad | 1155 | if (dump_enabled_p ()) |
78c60e3d SS |
1156 | dump_printf_loc (MSG_NOTE, vect_location, |
1157 | "vect_model_load_cost: unaligned supported by " | |
e645e942 | 1158 | "hardware.\n"); |
ebfd146a IR |
1159 | |
1160 | break; | |
1161 | } | |
1162 | case dr_explicit_realign: | |
1163 | { | |
92345349 BS |
1164 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2, |
1165 | vector_load, stmt_info, 0, vect_body); | |
1166 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, | |
1167 | vec_perm, stmt_info, 0, vect_body); | |
ebfd146a IR |
1168 | |
1169 | /* FIXME: If the misalignment remains fixed across the iterations of | |
1170 | the containing loop, the following cost should be added to the | |
92345349 | 1171 | prologue costs. */ |
ebfd146a | 1172 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1173 | *inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt, |
1174 | stmt_info, 0, vect_body); | |
ebfd146a | 1175 | |
73fbfcad | 1176 | if (dump_enabled_p ()) |
e645e942 TJ |
1177 | dump_printf_loc (MSG_NOTE, vect_location, |
1178 | "vect_model_load_cost: explicit realign\n"); | |
8bd37302 | 1179 | |
ebfd146a IR |
1180 | break; |
1181 | } | |
1182 | case dr_explicit_realign_optimized: | |
1183 | { | |
73fbfcad | 1184 | if (dump_enabled_p ()) |
e645e942 | 1185 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1186 | "vect_model_load_cost: unaligned software " |
e645e942 | 1187 | "pipelined.\n"); |
ebfd146a IR |
1188 | |
1189 | /* Unaligned software pipeline has a load of an address, an initial | |
ff802fa1 | 1190 | load, and possibly a mask operation to "prime" the loop. However, |
0d0293ac | 1191 | if this is an access in a group of loads, which provide grouped |
ebfd146a | 1192 | access, then the above cost should only be considered for one |
ff802fa1 | 1193 | access in the group. Inside the loop, there is a load op |
ebfd146a IR |
1194 | and a realignment op. */ |
1195 | ||
92345349 | 1196 | if (add_realign_cost && record_prologue_costs) |
ebfd146a | 1197 | { |
92345349 BS |
1198 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 2, |
1199 | vector_stmt, stmt_info, | |
1200 | 0, vect_prologue); | |
ebfd146a | 1201 | if (targetm.vectorize.builtin_mask_for_load) |
92345349 BS |
1202 | *prologue_cost += record_stmt_cost (prologue_cost_vec, 1, |
1203 | vector_stmt, stmt_info, | |
1204 | 0, vect_prologue); | |
ebfd146a IR |
1205 | } |
1206 | ||
92345349 BS |
1207 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load, |
1208 | stmt_info, 0, vect_body); | |
1209 | *inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm, | |
1210 | stmt_info, 0, vect_body); | |
8bd37302 | 1211 | |
73fbfcad | 1212 | if (dump_enabled_p ()) |
78c60e3d | 1213 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 TJ |
1214 | "vect_model_load_cost: explicit realign optimized" |
1215 | "\n"); | |
8bd37302 | 1216 | |
ebfd146a IR |
1217 | break; |
1218 | } | |
1219 | ||
38eec4c6 UW |
1220 | case dr_unaligned_unsupported: |
1221 | { | |
1222 | *inside_cost = VECT_MAX_COST; | |
1223 | ||
73fbfcad | 1224 | if (dump_enabled_p ()) |
78c60e3d | 1225 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1226 | "vect_model_load_cost: unsupported access.\n"); |
38eec4c6 UW |
1227 | break; |
1228 | } | |
1229 | ||
ebfd146a IR |
1230 | default: |
1231 | gcc_unreachable (); | |
1232 | } | |
ebfd146a IR |
1233 | } |
1234 | ||
418b7df3 RG |
1235 | /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in |
1236 | the loop preheader for the vectorized stmt STMT. */ | |
ebfd146a | 1237 | |
418b7df3 | 1238 | static void |
355fe088 | 1239 | vect_init_vector_1 (gimple *stmt, gimple *new_stmt, gimple_stmt_iterator *gsi) |
ebfd146a | 1240 | { |
ebfd146a | 1241 | if (gsi) |
418b7df3 | 1242 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a IR |
1243 | else |
1244 | { | |
418b7df3 | 1245 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); |
ebfd146a | 1246 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
b8698a0f | 1247 | |
a70d6342 IR |
1248 | if (loop_vinfo) |
1249 | { | |
1250 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
418b7df3 RG |
1251 | basic_block new_bb; |
1252 | edge pe; | |
a70d6342 IR |
1253 | |
1254 | if (nested_in_vect_loop_p (loop, stmt)) | |
1255 | loop = loop->inner; | |
b8698a0f | 1256 | |
a70d6342 | 1257 | pe = loop_preheader_edge (loop); |
418b7df3 | 1258 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); |
a70d6342 IR |
1259 | gcc_assert (!new_bb); |
1260 | } | |
1261 | else | |
1262 | { | |
1263 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo); | |
1264 | basic_block bb; | |
1265 | gimple_stmt_iterator gsi_bb_start; | |
1266 | ||
1267 | gcc_assert (bb_vinfo); | |
1268 | bb = BB_VINFO_BB (bb_vinfo); | |
12aaf609 | 1269 | gsi_bb_start = gsi_after_labels (bb); |
418b7df3 | 1270 | gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT); |
a70d6342 | 1271 | } |
ebfd146a IR |
1272 | } |
1273 | ||
73fbfcad | 1274 | if (dump_enabled_p ()) |
ebfd146a | 1275 | { |
78c60e3d SS |
1276 | dump_printf_loc (MSG_NOTE, vect_location, |
1277 | "created new init_stmt: "); | |
1278 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); | |
ebfd146a | 1279 | } |
418b7df3 RG |
1280 | } |
1281 | ||
1282 | /* Function vect_init_vector. | |
ebfd146a | 1283 | |
5467ee52 RG |
1284 | Insert a new stmt (INIT_STMT) that initializes a new variable of type |
1285 | TYPE with the value VAL. If TYPE is a vector type and VAL does not have | |
1286 | vector type a vector with all elements equal to VAL is created first. | |
1287 | Place the initialization at BSI if it is not NULL. Otherwise, place the | |
1288 | initialization at the loop preheader. | |
418b7df3 RG |
1289 | Return the DEF of INIT_STMT. |
1290 | It will be used in the vectorization of STMT. */ | |
1291 | ||
1292 | tree | |
355fe088 | 1293 | vect_init_vector (gimple *stmt, tree val, tree type, gimple_stmt_iterator *gsi) |
418b7df3 | 1294 | { |
355fe088 | 1295 | gimple *init_stmt; |
418b7df3 RG |
1296 | tree new_temp; |
1297 | ||
e412ece4 RB |
1298 | /* We abuse this function to push sth to a SSA name with initial 'val'. */ |
1299 | if (! useless_type_conversion_p (type, TREE_TYPE (val))) | |
418b7df3 | 1300 | { |
e412ece4 RB |
1301 | gcc_assert (TREE_CODE (type) == VECTOR_TYPE); |
1302 | if (! types_compatible_p (TREE_TYPE (type), TREE_TYPE (val))) | |
418b7df3 | 1303 | { |
5a308cf1 IE |
1304 | /* Scalar boolean value should be transformed into |
1305 | all zeros or all ones value before building a vector. */ | |
1306 | if (VECTOR_BOOLEAN_TYPE_P (type)) | |
1307 | { | |
b3d51f23 IE |
1308 | tree true_val = build_all_ones_cst (TREE_TYPE (type)); |
1309 | tree false_val = build_zero_cst (TREE_TYPE (type)); | |
5a308cf1 IE |
1310 | |
1311 | if (CONSTANT_CLASS_P (val)) | |
1312 | val = integer_zerop (val) ? false_val : true_val; | |
1313 | else | |
1314 | { | |
1315 | new_temp = make_ssa_name (TREE_TYPE (type)); | |
1316 | init_stmt = gimple_build_assign (new_temp, COND_EXPR, | |
1317 | val, true_val, false_val); | |
1318 | vect_init_vector_1 (stmt, init_stmt, gsi); | |
1319 | val = new_temp; | |
1320 | } | |
1321 | } | |
1322 | else if (CONSTANT_CLASS_P (val)) | |
42fd8198 | 1323 | val = fold_convert (TREE_TYPE (type), val); |
418b7df3 RG |
1324 | else |
1325 | { | |
b731b390 | 1326 | new_temp = make_ssa_name (TREE_TYPE (type)); |
e412ece4 RB |
1327 | if (! INTEGRAL_TYPE_P (TREE_TYPE (val))) |
1328 | init_stmt = gimple_build_assign (new_temp, | |
1329 | fold_build1 (VIEW_CONVERT_EXPR, | |
1330 | TREE_TYPE (type), | |
1331 | val)); | |
1332 | else | |
1333 | init_stmt = gimple_build_assign (new_temp, NOP_EXPR, val); | |
418b7df3 | 1334 | vect_init_vector_1 (stmt, init_stmt, gsi); |
5467ee52 | 1335 | val = new_temp; |
418b7df3 RG |
1336 | } |
1337 | } | |
5467ee52 | 1338 | val = build_vector_from_val (type, val); |
418b7df3 RG |
1339 | } |
1340 | ||
0e22bb5a RB |
1341 | new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_"); |
1342 | init_stmt = gimple_build_assign (new_temp, val); | |
418b7df3 | 1343 | vect_init_vector_1 (stmt, init_stmt, gsi); |
0e22bb5a | 1344 | return new_temp; |
ebfd146a IR |
1345 | } |
1346 | ||
c83a894c | 1347 | /* Function vect_get_vec_def_for_operand_1. |
a70d6342 | 1348 | |
c83a894c AH |
1349 | For a defining stmt DEF_STMT of a scalar stmt, return a vector def with type |
1350 | DT that will be used in the vectorized stmt. */ | |
ebfd146a IR |
1351 | |
1352 | tree | |
c83a894c | 1353 | vect_get_vec_def_for_operand_1 (gimple *def_stmt, enum vect_def_type dt) |
ebfd146a IR |
1354 | { |
1355 | tree vec_oprnd; | |
355fe088 | 1356 | gimple *vec_stmt; |
ebfd146a | 1357 | stmt_vec_info def_stmt_info = NULL; |
ebfd146a IR |
1358 | |
1359 | switch (dt) | |
1360 | { | |
81c40241 | 1361 | /* operand is a constant or a loop invariant. */ |
ebfd146a | 1362 | case vect_constant_def: |
81c40241 | 1363 | case vect_external_def: |
c83a894c AH |
1364 | /* Code should use vect_get_vec_def_for_operand. */ |
1365 | gcc_unreachable (); | |
ebfd146a | 1366 | |
81c40241 | 1367 | /* operand is defined inside the loop. */ |
8644a673 | 1368 | case vect_internal_def: |
ebfd146a | 1369 | { |
ebfd146a IR |
1370 | /* Get the def from the vectorized stmt. */ |
1371 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
83197f37 | 1372 | |
ebfd146a | 1373 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); |
83197f37 IR |
1374 | /* Get vectorized pattern statement. */ |
1375 | if (!vec_stmt | |
1376 | && STMT_VINFO_IN_PATTERN_P (def_stmt_info) | |
1377 | && !STMT_VINFO_RELEVANT (def_stmt_info)) | |
1378 | vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt ( | |
1379 | STMT_VINFO_RELATED_STMT (def_stmt_info))); | |
ebfd146a IR |
1380 | gcc_assert (vec_stmt); |
1381 | if (gimple_code (vec_stmt) == GIMPLE_PHI) | |
1382 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1383 | else if (is_gimple_call (vec_stmt)) | |
1384 | vec_oprnd = gimple_call_lhs (vec_stmt); | |
1385 | else | |
1386 | vec_oprnd = gimple_assign_lhs (vec_stmt); | |
1387 | return vec_oprnd; | |
1388 | } | |
1389 | ||
c78e3652 | 1390 | /* operand is defined by a loop header phi. */ |
ebfd146a | 1391 | case vect_reduction_def: |
06066f92 | 1392 | case vect_double_reduction_def: |
7c5222ff | 1393 | case vect_nested_cycle: |
ebfd146a IR |
1394 | case vect_induction_def: |
1395 | { | |
1396 | gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI); | |
1397 | ||
1398 | /* Get the def from the vectorized stmt. */ | |
1399 | def_stmt_info = vinfo_for_stmt (def_stmt); | |
1400 | vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); | |
6dbbece6 RG |
1401 | if (gimple_code (vec_stmt) == GIMPLE_PHI) |
1402 | vec_oprnd = PHI_RESULT (vec_stmt); | |
1403 | else | |
1404 | vec_oprnd = gimple_get_lhs (vec_stmt); | |
ebfd146a IR |
1405 | return vec_oprnd; |
1406 | } | |
1407 | ||
1408 | default: | |
1409 | gcc_unreachable (); | |
1410 | } | |
1411 | } | |
1412 | ||
1413 | ||
c83a894c AH |
1414 | /* Function vect_get_vec_def_for_operand. |
1415 | ||
1416 | OP is an operand in STMT. This function returns a (vector) def that will be | |
1417 | used in the vectorized stmt for STMT. | |
1418 | ||
1419 | In the case that OP is an SSA_NAME which is defined in the loop, then | |
1420 | STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. | |
1421 | ||
1422 | In case OP is an invariant or constant, a new stmt that creates a vector def | |
1423 | needs to be introduced. VECTYPE may be used to specify a required type for | |
1424 | vector invariant. */ | |
1425 | ||
1426 | tree | |
1427 | vect_get_vec_def_for_operand (tree op, gimple *stmt, tree vectype) | |
1428 | { | |
1429 | gimple *def_stmt; | |
1430 | enum vect_def_type dt; | |
1431 | bool is_simple_use; | |
1432 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); | |
1433 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
1434 | ||
1435 | if (dump_enabled_p ()) | |
1436 | { | |
1437 | dump_printf_loc (MSG_NOTE, vect_location, | |
1438 | "vect_get_vec_def_for_operand: "); | |
1439 | dump_generic_expr (MSG_NOTE, TDF_SLIM, op); | |
1440 | dump_printf (MSG_NOTE, "\n"); | |
1441 | } | |
1442 | ||
1443 | is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt); | |
1444 | gcc_assert (is_simple_use); | |
1445 | if (def_stmt && dump_enabled_p ()) | |
1446 | { | |
1447 | dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = "); | |
1448 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0); | |
1449 | } | |
1450 | ||
1451 | if (dt == vect_constant_def || dt == vect_external_def) | |
1452 | { | |
1453 | tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo); | |
1454 | tree vector_type; | |
1455 | ||
1456 | if (vectype) | |
1457 | vector_type = vectype; | |
2568d8a1 | 1458 | else if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op)) |
c83a894c AH |
1459 | && VECTOR_BOOLEAN_TYPE_P (stmt_vectype)) |
1460 | vector_type = build_same_sized_truth_vector_type (stmt_vectype); | |
1461 | else | |
1462 | vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); | |
1463 | ||
1464 | gcc_assert (vector_type); | |
1465 | return vect_init_vector (stmt, op, vector_type, NULL); | |
1466 | } | |
1467 | else | |
1468 | return vect_get_vec_def_for_operand_1 (def_stmt, dt); | |
1469 | } | |
1470 | ||
1471 | ||
ebfd146a IR |
1472 | /* Function vect_get_vec_def_for_stmt_copy |
1473 | ||
ff802fa1 | 1474 | Return a vector-def for an operand. This function is used when the |
b8698a0f L |
1475 | vectorized stmt to be created (by the caller to this function) is a "copy" |
1476 | created in case the vectorized result cannot fit in one vector, and several | |
ff802fa1 | 1477 | copies of the vector-stmt are required. In this case the vector-def is |
ebfd146a | 1478 | retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field |
b8698a0f | 1479 | of the stmt that defines VEC_OPRND. |
ebfd146a IR |
1480 | DT is the type of the vector def VEC_OPRND. |
1481 | ||
1482 | Context: | |
1483 | In case the vectorization factor (VF) is bigger than the number | |
1484 | of elements that can fit in a vectype (nunits), we have to generate | |
ff802fa1 | 1485 | more than one vector stmt to vectorize the scalar stmt. This situation |
b8698a0f | 1486 | arises when there are multiple data-types operated upon in the loop; the |
ebfd146a IR |
1487 | smallest data-type determines the VF, and as a result, when vectorizing |
1488 | stmts operating on wider types we need to create 'VF/nunits' "copies" of the | |
1489 | vector stmt (each computing a vector of 'nunits' results, and together | |
b8698a0f | 1490 | computing 'VF' results in each iteration). This function is called when |
ebfd146a IR |
1491 | vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in |
1492 | which VF=16 and nunits=4, so the number of copies required is 4): | |
1493 | ||
1494 | scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT | |
b8698a0f | 1495 | |
ebfd146a IR |
1496 | S1: x = load VS1.0: vx.0 = memref0 VS1.1 |
1497 | VS1.1: vx.1 = memref1 VS1.2 | |
1498 | VS1.2: vx.2 = memref2 VS1.3 | |
b8698a0f | 1499 | VS1.3: vx.3 = memref3 |
ebfd146a IR |
1500 | |
1501 | S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 | |
1502 | VSnew.1: vz1 = vx.1 + ... VSnew.2 | |
1503 | VSnew.2: vz2 = vx.2 + ... VSnew.3 | |
1504 | VSnew.3: vz3 = vx.3 + ... | |
1505 | ||
1506 | The vectorization of S1 is explained in vectorizable_load. | |
1507 | The vectorization of S2: | |
b8698a0f L |
1508 | To create the first vector-stmt out of the 4 copies - VSnew.0 - |
1509 | the function 'vect_get_vec_def_for_operand' is called to | |
ff802fa1 | 1510 | get the relevant vector-def for each operand of S2. For operand x it |
ebfd146a IR |
1511 | returns the vector-def 'vx.0'. |
1512 | ||
b8698a0f L |
1513 | To create the remaining copies of the vector-stmt (VSnew.j), this |
1514 | function is called to get the relevant vector-def for each operand. It is | |
1515 | obtained from the respective VS1.j stmt, which is recorded in the | |
ebfd146a IR |
1516 | STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. |
1517 | ||
b8698a0f L |
1518 | For example, to obtain the vector-def 'vx.1' in order to create the |
1519 | vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. | |
1520 | Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the | |
ebfd146a IR |
1521 | STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', |
1522 | and return its def ('vx.1'). | |
1523 | Overall, to create the above sequence this function will be called 3 times: | |
1524 | vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); | |
1525 | vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); | |
1526 | vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ | |
1527 | ||
1528 | tree | |
1529 | vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) | |
1530 | { | |
355fe088 | 1531 | gimple *vec_stmt_for_operand; |
ebfd146a IR |
1532 | stmt_vec_info def_stmt_info; |
1533 | ||
1534 | /* Do nothing; can reuse same def. */ | |
8644a673 | 1535 | if (dt == vect_external_def || dt == vect_constant_def ) |
ebfd146a IR |
1536 | return vec_oprnd; |
1537 | ||
1538 | vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); | |
1539 | def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); | |
1540 | gcc_assert (def_stmt_info); | |
1541 | vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); | |
1542 | gcc_assert (vec_stmt_for_operand); | |
ebfd146a IR |
1543 | if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI) |
1544 | vec_oprnd = PHI_RESULT (vec_stmt_for_operand); | |
1545 | else | |
1546 | vec_oprnd = gimple_get_lhs (vec_stmt_for_operand); | |
1547 | return vec_oprnd; | |
1548 | } | |
1549 | ||
1550 | ||
1551 | /* Get vectorized definitions for the operands to create a copy of an original | |
ff802fa1 | 1552 | stmt. See vect_get_vec_def_for_stmt_copy () for details. */ |
ebfd146a | 1553 | |
c78e3652 | 1554 | void |
b8698a0f | 1555 | vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt, |
9771b263 DN |
1556 | vec<tree> *vec_oprnds0, |
1557 | vec<tree> *vec_oprnds1) | |
ebfd146a | 1558 | { |
9771b263 | 1559 | tree vec_oprnd = vec_oprnds0->pop (); |
ebfd146a IR |
1560 | |
1561 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd); | |
9771b263 | 1562 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a | 1563 | |
9771b263 | 1564 | if (vec_oprnds1 && vec_oprnds1->length ()) |
ebfd146a | 1565 | { |
9771b263 | 1566 | vec_oprnd = vec_oprnds1->pop (); |
ebfd146a | 1567 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd); |
9771b263 | 1568 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1569 | } |
1570 | } | |
1571 | ||
1572 | ||
c78e3652 | 1573 | /* Get vectorized definitions for OP0 and OP1. */ |
ebfd146a | 1574 | |
c78e3652 | 1575 | void |
355fe088 | 1576 | vect_get_vec_defs (tree op0, tree op1, gimple *stmt, |
9771b263 DN |
1577 | vec<tree> *vec_oprnds0, |
1578 | vec<tree> *vec_oprnds1, | |
306b0c92 | 1579 | slp_tree slp_node) |
ebfd146a IR |
1580 | { |
1581 | if (slp_node) | |
d092494c IR |
1582 | { |
1583 | int nops = (op1 == NULL_TREE) ? 1 : 2; | |
ef062b13 TS |
1584 | auto_vec<tree> ops (nops); |
1585 | auto_vec<vec<tree> > vec_defs (nops); | |
d092494c | 1586 | |
9771b263 | 1587 | ops.quick_push (op0); |
d092494c | 1588 | if (op1) |
9771b263 | 1589 | ops.quick_push (op1); |
d092494c | 1590 | |
306b0c92 | 1591 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
d092494c | 1592 | |
37b5ec8f | 1593 | *vec_oprnds0 = vec_defs[0]; |
d092494c | 1594 | if (op1) |
37b5ec8f | 1595 | *vec_oprnds1 = vec_defs[1]; |
d092494c | 1596 | } |
ebfd146a IR |
1597 | else |
1598 | { | |
1599 | tree vec_oprnd; | |
1600 | ||
9771b263 | 1601 | vec_oprnds0->create (1); |
81c40241 | 1602 | vec_oprnd = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 1603 | vec_oprnds0->quick_push (vec_oprnd); |
ebfd146a IR |
1604 | |
1605 | if (op1) | |
1606 | { | |
9771b263 | 1607 | vec_oprnds1->create (1); |
81c40241 | 1608 | vec_oprnd = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 1609 | vec_oprnds1->quick_push (vec_oprnd); |
ebfd146a IR |
1610 | } |
1611 | } | |
1612 | } | |
1613 | ||
bb6c2b68 RS |
1614 | /* Helper function called by vect_finish_replace_stmt and |
1615 | vect_finish_stmt_generation. Set the location of the new | |
1616 | statement and create a stmt_vec_info for it. */ | |
1617 | ||
1618 | static void | |
1619 | vect_finish_stmt_generation_1 (gimple *stmt, gimple *vec_stmt) | |
1620 | { | |
1621 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1622 | vec_info *vinfo = stmt_info->vinfo; | |
1623 | ||
1624 | set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, vinfo)); | |
1625 | ||
1626 | if (dump_enabled_p ()) | |
1627 | { | |
1628 | dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: "); | |
1629 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0); | |
1630 | } | |
1631 | ||
1632 | gimple_set_location (vec_stmt, gimple_location (stmt)); | |
1633 | ||
1634 | /* While EH edges will generally prevent vectorization, stmt might | |
1635 | e.g. be in a must-not-throw region. Ensure newly created stmts | |
1636 | that could throw are part of the same region. */ | |
1637 | int lp_nr = lookup_stmt_eh_lp (stmt); | |
1638 | if (lp_nr != 0 && stmt_could_throw_p (vec_stmt)) | |
1639 | add_stmt_to_eh_lp (vec_stmt, lp_nr); | |
1640 | } | |
1641 | ||
1642 | /* Replace the scalar statement STMT with a new vector statement VEC_STMT, | |
1643 | which sets the same scalar result as STMT did. */ | |
1644 | ||
1645 | void | |
1646 | vect_finish_replace_stmt (gimple *stmt, gimple *vec_stmt) | |
1647 | { | |
1648 | gcc_assert (gimple_get_lhs (stmt) == gimple_get_lhs (vec_stmt)); | |
1649 | ||
1650 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); | |
1651 | gsi_replace (&gsi, vec_stmt, false); | |
1652 | ||
1653 | vect_finish_stmt_generation_1 (stmt, vec_stmt); | |
1654 | } | |
ebfd146a IR |
1655 | |
1656 | /* Function vect_finish_stmt_generation. | |
1657 | ||
1658 | Insert a new stmt. */ | |
1659 | ||
1660 | void | |
355fe088 | 1661 | vect_finish_stmt_generation (gimple *stmt, gimple *vec_stmt, |
ebfd146a IR |
1662 | gimple_stmt_iterator *gsi) |
1663 | { | |
ebfd146a IR |
1664 | gcc_assert (gimple_code (stmt) != GIMPLE_LABEL); |
1665 | ||
54e8e2c3 RG |
1666 | if (!gsi_end_p (*gsi) |
1667 | && gimple_has_mem_ops (vec_stmt)) | |
1668 | { | |
355fe088 | 1669 | gimple *at_stmt = gsi_stmt (*gsi); |
54e8e2c3 RG |
1670 | tree vuse = gimple_vuse (at_stmt); |
1671 | if (vuse && TREE_CODE (vuse) == SSA_NAME) | |
1672 | { | |
1673 | tree vdef = gimple_vdef (at_stmt); | |
1674 | gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt)); | |
1675 | /* If we have an SSA vuse and insert a store, update virtual | |
1676 | SSA form to avoid triggering the renamer. Do so only | |
1677 | if we can easily see all uses - which is what almost always | |
1678 | happens with the way vectorized stmts are inserted. */ | |
1679 | if ((vdef && TREE_CODE (vdef) == SSA_NAME) | |
1680 | && ((is_gimple_assign (vec_stmt) | |
1681 | && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) | |
1682 | || (is_gimple_call (vec_stmt) | |
1683 | && !(gimple_call_flags (vec_stmt) | |
1684 | & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) | |
1685 | { | |
1686 | tree new_vdef = copy_ssa_name (vuse, vec_stmt); | |
1687 | gimple_set_vdef (vec_stmt, new_vdef); | |
1688 | SET_USE (gimple_vuse_op (at_stmt), new_vdef); | |
1689 | } | |
1690 | } | |
1691 | } | |
ebfd146a | 1692 | gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); |
bb6c2b68 | 1693 | vect_finish_stmt_generation_1 (stmt, vec_stmt); |
ebfd146a IR |
1694 | } |
1695 | ||
70439f0d RS |
1696 | /* We want to vectorize a call to combined function CFN with function |
1697 | decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN | |
1698 | as the types of all inputs. Check whether this is possible using | |
1699 | an internal function, returning its code if so or IFN_LAST if not. */ | |
ebfd146a | 1700 | |
70439f0d RS |
1701 | static internal_fn |
1702 | vectorizable_internal_function (combined_fn cfn, tree fndecl, | |
1703 | tree vectype_out, tree vectype_in) | |
ebfd146a | 1704 | { |
70439f0d RS |
1705 | internal_fn ifn; |
1706 | if (internal_fn_p (cfn)) | |
1707 | ifn = as_internal_fn (cfn); | |
1708 | else | |
1709 | ifn = associated_internal_fn (fndecl); | |
1710 | if (ifn != IFN_LAST && direct_internal_fn_p (ifn)) | |
1711 | { | |
1712 | const direct_internal_fn_info &info = direct_internal_fn (ifn); | |
1713 | if (info.vectorizable) | |
1714 | { | |
1715 | tree type0 = (info.type0 < 0 ? vectype_out : vectype_in); | |
1716 | tree type1 = (info.type1 < 0 ? vectype_out : vectype_in); | |
d95ab70a RS |
1717 | if (direct_internal_fn_supported_p (ifn, tree_pair (type0, type1), |
1718 | OPTIMIZE_FOR_SPEED)) | |
70439f0d RS |
1719 | return ifn; |
1720 | } | |
1721 | } | |
1722 | return IFN_LAST; | |
ebfd146a IR |
1723 | } |
1724 | ||
5ce9450f | 1725 | |
355fe088 | 1726 | static tree permute_vec_elements (tree, tree, tree, gimple *, |
5ce9450f JJ |
1727 | gimple_stmt_iterator *); |
1728 | ||
7cfb4d93 RS |
1729 | /* Check whether a load or store statement in the loop described by |
1730 | LOOP_VINFO is possible in a fully-masked loop. This is testing | |
1731 | whether the vectorizer pass has the appropriate support, as well as | |
1732 | whether the target does. | |
1733 | ||
1734 | VLS_TYPE says whether the statement is a load or store and VECTYPE | |
1735 | is the type of the vector being loaded or stored. MEMORY_ACCESS_TYPE | |
1736 | says how the load or store is going to be implemented and GROUP_SIZE | |
1737 | is the number of load or store statements in the containing group. | |
bfaa08b7 RS |
1738 | If the access is a gather load or scatter store, GS_INFO describes |
1739 | its arguments. | |
7cfb4d93 RS |
1740 | |
1741 | Clear LOOP_VINFO_CAN_FULLY_MASK_P if a fully-masked loop is not | |
1742 | supported, otherwise record the required mask types. */ | |
1743 | ||
1744 | static void | |
1745 | check_load_store_masking (loop_vec_info loop_vinfo, tree vectype, | |
1746 | vec_load_store_type vls_type, int group_size, | |
bfaa08b7 RS |
1747 | vect_memory_access_type memory_access_type, |
1748 | gather_scatter_info *gs_info) | |
7cfb4d93 RS |
1749 | { |
1750 | /* Invariant loads need no special support. */ | |
1751 | if (memory_access_type == VMAT_INVARIANT) | |
1752 | return; | |
1753 | ||
1754 | vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); | |
1755 | machine_mode vecmode = TYPE_MODE (vectype); | |
1756 | bool is_load = (vls_type == VLS_LOAD); | |
1757 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
1758 | { | |
1759 | if (is_load | |
1760 | ? !vect_load_lanes_supported (vectype, group_size, true) | |
1761 | : !vect_store_lanes_supported (vectype, group_size, true)) | |
1762 | { | |
1763 | if (dump_enabled_p ()) | |
1764 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1765 | "can't use a fully-masked loop because the" | |
1766 | " target doesn't have an appropriate masked" | |
1767 | " load/store-lanes instruction.\n"); | |
1768 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1769 | return; | |
1770 | } | |
1771 | unsigned int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
1772 | vect_record_loop_mask (loop_vinfo, masks, ncopies, vectype); | |
1773 | return; | |
1774 | } | |
1775 | ||
bfaa08b7 RS |
1776 | if (memory_access_type == VMAT_GATHER_SCATTER) |
1777 | { | |
f307441a RS |
1778 | internal_fn ifn = (is_load |
1779 | ? IFN_MASK_GATHER_LOAD | |
1780 | : IFN_MASK_SCATTER_STORE); | |
bfaa08b7 | 1781 | tree offset_type = TREE_TYPE (gs_info->offset); |
f307441a | 1782 | if (!internal_gather_scatter_fn_supported_p (ifn, vectype, |
bfaa08b7 RS |
1783 | gs_info->memory_type, |
1784 | TYPE_SIGN (offset_type), | |
1785 | gs_info->scale)) | |
1786 | { | |
1787 | if (dump_enabled_p ()) | |
1788 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1789 | "can't use a fully-masked loop because the" | |
1790 | " target doesn't have an appropriate masked" | |
f307441a | 1791 | " gather load or scatter store instruction.\n"); |
bfaa08b7 RS |
1792 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; |
1793 | return; | |
1794 | } | |
1795 | unsigned int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
1796 | vect_record_loop_mask (loop_vinfo, masks, ncopies, vectype); | |
1797 | return; | |
1798 | } | |
1799 | ||
7cfb4d93 RS |
1800 | if (memory_access_type != VMAT_CONTIGUOUS |
1801 | && memory_access_type != VMAT_CONTIGUOUS_PERMUTE) | |
1802 | { | |
1803 | /* Element X of the data must come from iteration i * VF + X of the | |
1804 | scalar loop. We need more work to support other mappings. */ | |
1805 | if (dump_enabled_p ()) | |
1806 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1807 | "can't use a fully-masked loop because an access" | |
1808 | " isn't contiguous.\n"); | |
1809 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1810 | return; | |
1811 | } | |
1812 | ||
1813 | machine_mode mask_mode; | |
1814 | if (!(targetm.vectorize.get_mask_mode | |
1815 | (GET_MODE_NUNITS (vecmode), | |
1816 | GET_MODE_SIZE (vecmode)).exists (&mask_mode)) | |
1817 | || !can_vec_mask_load_store_p (vecmode, mask_mode, is_load)) | |
1818 | { | |
1819 | if (dump_enabled_p ()) | |
1820 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1821 | "can't use a fully-masked loop because the target" | |
1822 | " doesn't have the appropriate masked load or" | |
1823 | " store.\n"); | |
1824 | LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; | |
1825 | return; | |
1826 | } | |
1827 | /* We might load more scalars than we need for permuting SLP loads. | |
1828 | We checked in get_group_load_store_type that the extra elements | |
1829 | don't leak into a new vector. */ | |
1830 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
1831 | poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
1832 | unsigned int nvectors; | |
1833 | if (can_div_away_from_zero_p (group_size * vf, nunits, &nvectors)) | |
1834 | vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype); | |
1835 | else | |
1836 | gcc_unreachable (); | |
1837 | } | |
1838 | ||
1839 | /* Return the mask input to a masked load or store. VEC_MASK is the vectorized | |
1840 | form of the scalar mask condition and LOOP_MASK, if nonnull, is the mask | |
1841 | that needs to be applied to all loads and stores in a vectorized loop. | |
1842 | Return VEC_MASK if LOOP_MASK is null, otherwise return VEC_MASK & LOOP_MASK. | |
1843 | ||
1844 | MASK_TYPE is the type of both masks. If new statements are needed, | |
1845 | insert them before GSI. */ | |
1846 | ||
1847 | static tree | |
1848 | prepare_load_store_mask (tree mask_type, tree loop_mask, tree vec_mask, | |
1849 | gimple_stmt_iterator *gsi) | |
1850 | { | |
1851 | gcc_assert (useless_type_conversion_p (mask_type, TREE_TYPE (vec_mask))); | |
1852 | if (!loop_mask) | |
1853 | return vec_mask; | |
1854 | ||
1855 | gcc_assert (TREE_TYPE (loop_mask) == mask_type); | |
1856 | tree and_res = make_temp_ssa_name (mask_type, NULL, "vec_mask_and"); | |
1857 | gimple *and_stmt = gimple_build_assign (and_res, BIT_AND_EXPR, | |
1858 | vec_mask, loop_mask); | |
1859 | gsi_insert_before (gsi, and_stmt, GSI_SAME_STMT); | |
1860 | return and_res; | |
1861 | } | |
1862 | ||
429ef523 RS |
1863 | /* Determine whether we can use a gather load or scatter store to vectorize |
1864 | strided load or store STMT by truncating the current offset to a smaller | |
1865 | width. We need to be able to construct an offset vector: | |
1866 | ||
1867 | { 0, X, X*2, X*3, ... } | |
1868 | ||
1869 | without loss of precision, where X is STMT's DR_STEP. | |
1870 | ||
1871 | Return true if this is possible, describing the gather load or scatter | |
1872 | store in GS_INFO. MASKED_P is true if the load or store is conditional. */ | |
1873 | ||
1874 | static bool | |
1875 | vect_truncate_gather_scatter_offset (gimple *stmt, loop_vec_info loop_vinfo, | |
1876 | bool masked_p, | |
1877 | gather_scatter_info *gs_info) | |
1878 | { | |
1879 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1880 | data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
1881 | tree step = DR_STEP (dr); | |
1882 | if (TREE_CODE (step) != INTEGER_CST) | |
1883 | { | |
1884 | /* ??? Perhaps we could use range information here? */ | |
1885 | if (dump_enabled_p ()) | |
1886 | dump_printf_loc (MSG_NOTE, vect_location, | |
1887 | "cannot truncate variable step.\n"); | |
1888 | return false; | |
1889 | } | |
1890 | ||
1891 | /* Get the number of bits in an element. */ | |
1892 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
1893 | scalar_mode element_mode = SCALAR_TYPE_MODE (TREE_TYPE (vectype)); | |
1894 | unsigned int element_bits = GET_MODE_BITSIZE (element_mode); | |
1895 | ||
1896 | /* Set COUNT to the upper limit on the number of elements - 1. | |
1897 | Start with the maximum vectorization factor. */ | |
1898 | unsigned HOST_WIDE_INT count = vect_max_vf (loop_vinfo) - 1; | |
1899 | ||
1900 | /* Try lowering COUNT to the number of scalar latch iterations. */ | |
1901 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1902 | widest_int max_iters; | |
1903 | if (max_loop_iterations (loop, &max_iters) | |
1904 | && max_iters < count) | |
1905 | count = max_iters.to_shwi (); | |
1906 | ||
1907 | /* Try scales of 1 and the element size. */ | |
1908 | int scales[] = { 1, vect_get_scalar_dr_size (dr) }; | |
1909 | bool overflow_p = false; | |
1910 | for (int i = 0; i < 2; ++i) | |
1911 | { | |
1912 | int scale = scales[i]; | |
1913 | widest_int factor; | |
1914 | if (!wi::multiple_of_p (wi::to_widest (step), scale, SIGNED, &factor)) | |
1915 | continue; | |
1916 | ||
1917 | /* See whether we can calculate (COUNT - 1) * STEP / SCALE | |
1918 | in OFFSET_BITS bits. */ | |
1919 | widest_int range = wi::mul (count, factor, SIGNED, &overflow_p); | |
1920 | if (overflow_p) | |
1921 | continue; | |
1922 | signop sign = range >= 0 ? UNSIGNED : SIGNED; | |
1923 | if (wi::min_precision (range, sign) > element_bits) | |
1924 | { | |
1925 | overflow_p = true; | |
1926 | continue; | |
1927 | } | |
1928 | ||
1929 | /* See whether the target supports the operation. */ | |
1930 | tree memory_type = TREE_TYPE (DR_REF (dr)); | |
1931 | if (!vect_gather_scatter_fn_p (DR_IS_READ (dr), masked_p, vectype, | |
1932 | memory_type, element_bits, sign, scale, | |
1933 | &gs_info->ifn, &gs_info->element_type)) | |
1934 | continue; | |
1935 | ||
1936 | tree offset_type = build_nonstandard_integer_type (element_bits, | |
1937 | sign == UNSIGNED); | |
1938 | ||
1939 | gs_info->decl = NULL_TREE; | |
1940 | /* Logically the sum of DR_BASE_ADDRESS, DR_INIT and DR_OFFSET, | |
1941 | but we don't need to store that here. */ | |
1942 | gs_info->base = NULL_TREE; | |
1943 | gs_info->offset = fold_convert (offset_type, step); | |
929b4411 | 1944 | gs_info->offset_dt = vect_constant_def; |
429ef523 RS |
1945 | gs_info->offset_vectype = NULL_TREE; |
1946 | gs_info->scale = scale; | |
1947 | gs_info->memory_type = memory_type; | |
1948 | return true; | |
1949 | } | |
1950 | ||
1951 | if (overflow_p && dump_enabled_p ()) | |
1952 | dump_printf_loc (MSG_NOTE, vect_location, | |
1953 | "truncating gather/scatter offset to %d bits" | |
1954 | " might change its value.\n", element_bits); | |
1955 | ||
1956 | return false; | |
1957 | } | |
1958 | ||
ab2fc782 RS |
1959 | /* Return true if we can use gather/scatter internal functions to |
1960 | vectorize STMT, which is a grouped or strided load or store. | |
429ef523 RS |
1961 | MASKED_P is true if load or store is conditional. When returning |
1962 | true, fill in GS_INFO with the information required to perform the | |
1963 | operation. */ | |
ab2fc782 RS |
1964 | |
1965 | static bool | |
1966 | vect_use_strided_gather_scatters_p (gimple *stmt, loop_vec_info loop_vinfo, | |
429ef523 | 1967 | bool masked_p, |
ab2fc782 RS |
1968 | gather_scatter_info *gs_info) |
1969 | { | |
1970 | if (!vect_check_gather_scatter (stmt, loop_vinfo, gs_info) | |
1971 | || gs_info->decl) | |
429ef523 RS |
1972 | return vect_truncate_gather_scatter_offset (stmt, loop_vinfo, |
1973 | masked_p, gs_info); | |
ab2fc782 RS |
1974 | |
1975 | scalar_mode element_mode = SCALAR_TYPE_MODE (gs_info->element_type); | |
1976 | unsigned int element_bits = GET_MODE_BITSIZE (element_mode); | |
1977 | tree offset_type = TREE_TYPE (gs_info->offset); | |
1978 | unsigned int offset_bits = TYPE_PRECISION (offset_type); | |
1979 | ||
1980 | /* Enforced by vect_check_gather_scatter. */ | |
1981 | gcc_assert (element_bits >= offset_bits); | |
1982 | ||
1983 | /* If the elements are wider than the offset, convert the offset to the | |
1984 | same width, without changing its sign. */ | |
1985 | if (element_bits > offset_bits) | |
1986 | { | |
1987 | bool unsigned_p = TYPE_UNSIGNED (offset_type); | |
1988 | offset_type = build_nonstandard_integer_type (element_bits, unsigned_p); | |
1989 | gs_info->offset = fold_convert (offset_type, gs_info->offset); | |
1990 | } | |
1991 | ||
1992 | if (dump_enabled_p ()) | |
1993 | dump_printf_loc (MSG_NOTE, vect_location, | |
1994 | "using gather/scatter for strided/grouped access," | |
1995 | " scale = %d\n", gs_info->scale); | |
1996 | ||
1997 | return true; | |
1998 | } | |
1999 | ||
62da9e14 RS |
2000 | /* STMT is a non-strided load or store, meaning that it accesses |
2001 | elements with a known constant step. Return -1 if that step | |
2002 | is negative, 0 if it is zero, and 1 if it is greater than zero. */ | |
2003 | ||
2004 | static int | |
2005 | compare_step_with_zero (gimple *stmt) | |
2006 | { | |
2007 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3f5e8a76 RS |
2008 | data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); |
2009 | return tree_int_cst_compare (vect_dr_behavior (dr)->step, | |
2010 | size_zero_node); | |
62da9e14 RS |
2011 | } |
2012 | ||
2013 | /* If the target supports a permute mask that reverses the elements in | |
2014 | a vector of type VECTYPE, return that mask, otherwise return null. */ | |
2015 | ||
2016 | static tree | |
2017 | perm_mask_for_reverse (tree vectype) | |
2018 | { | |
928686b1 | 2019 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
62da9e14 | 2020 | |
d980067b RS |
2021 | /* The encoding has a single stepped pattern. */ |
2022 | vec_perm_builder sel (nunits, 1, 3); | |
928686b1 | 2023 | for (int i = 0; i < 3; ++i) |
908a1a16 | 2024 | sel.quick_push (nunits - 1 - i); |
62da9e14 | 2025 | |
e3342de4 RS |
2026 | vec_perm_indices indices (sel, 1, nunits); |
2027 | if (!can_vec_perm_const_p (TYPE_MODE (vectype), indices)) | |
62da9e14 | 2028 | return NULL_TREE; |
e3342de4 | 2029 | return vect_gen_perm_mask_checked (vectype, indices); |
62da9e14 | 2030 | } |
5ce9450f | 2031 | |
c3a8f964 RS |
2032 | /* STMT is either a masked or unconditional store. Return the value |
2033 | being stored. */ | |
2034 | ||
f307441a | 2035 | tree |
c3a8f964 RS |
2036 | vect_get_store_rhs (gimple *stmt) |
2037 | { | |
2038 | if (gassign *assign = dyn_cast <gassign *> (stmt)) | |
2039 | { | |
2040 | gcc_assert (gimple_assign_single_p (assign)); | |
2041 | return gimple_assign_rhs1 (assign); | |
2042 | } | |
2043 | if (gcall *call = dyn_cast <gcall *> (stmt)) | |
2044 | { | |
2045 | internal_fn ifn = gimple_call_internal_fn (call); | |
f307441a RS |
2046 | int index = internal_fn_stored_value_index (ifn); |
2047 | gcc_assert (index >= 0); | |
2048 | return gimple_call_arg (stmt, index); | |
c3a8f964 RS |
2049 | } |
2050 | gcc_unreachable (); | |
2051 | } | |
2052 | ||
2de001ee RS |
2053 | /* A subroutine of get_load_store_type, with a subset of the same |
2054 | arguments. Handle the case where STMT is part of a grouped load | |
2055 | or store. | |
2056 | ||
2057 | For stores, the statements in the group are all consecutive | |
2058 | and there is no gap at the end. For loads, the statements in the | |
2059 | group might not be consecutive; there can be gaps between statements | |
2060 | as well as at the end. */ | |
2061 | ||
2062 | static bool | |
2063 | get_group_load_store_type (gimple *stmt, tree vectype, bool slp, | |
7e11fc7f | 2064 | bool masked_p, vec_load_store_type vls_type, |
429ef523 RS |
2065 | vect_memory_access_type *memory_access_type, |
2066 | gather_scatter_info *gs_info) | |
2de001ee RS |
2067 | { |
2068 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2069 | vec_info *vinfo = stmt_info->vinfo; | |
2070 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2071 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; | |
2072 | gimple *first_stmt = GROUP_FIRST_ELEMENT (stmt_info); | |
f702e7d4 | 2073 | data_reference *first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
2de001ee RS |
2074 | unsigned int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
2075 | bool single_element_p = (stmt == first_stmt | |
2076 | && !GROUP_NEXT_ELEMENT (stmt_info)); | |
2077 | unsigned HOST_WIDE_INT gap = GROUP_GAP (vinfo_for_stmt (first_stmt)); | |
928686b1 | 2078 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2de001ee RS |
2079 | |
2080 | /* True if the vectorized statements would access beyond the last | |
2081 | statement in the group. */ | |
2082 | bool overrun_p = false; | |
2083 | ||
2084 | /* True if we can cope with such overrun by peeling for gaps, so that | |
2085 | there is at least one final scalar iteration after the vector loop. */ | |
7e11fc7f RS |
2086 | bool can_overrun_p = (!masked_p |
2087 | && vls_type == VLS_LOAD | |
2088 | && loop_vinfo | |
2089 | && !loop->inner); | |
2de001ee RS |
2090 | |
2091 | /* There can only be a gap at the end of the group if the stride is | |
2092 | known at compile time. */ | |
2093 | gcc_assert (!STMT_VINFO_STRIDED_P (stmt_info) || gap == 0); | |
2094 | ||
2095 | /* Stores can't yet have gaps. */ | |
2096 | gcc_assert (slp || vls_type == VLS_LOAD || gap == 0); | |
2097 | ||
2098 | if (slp) | |
2099 | { | |
2100 | if (STMT_VINFO_STRIDED_P (stmt_info)) | |
2101 | { | |
2102 | /* Try to use consecutive accesses of GROUP_SIZE elements, | |
2103 | separated by the stride, until we have a complete vector. | |
2104 | Fall back to scalar accesses if that isn't possible. */ | |
928686b1 | 2105 | if (multiple_p (nunits, group_size)) |
2de001ee RS |
2106 | *memory_access_type = VMAT_STRIDED_SLP; |
2107 | else | |
2108 | *memory_access_type = VMAT_ELEMENTWISE; | |
2109 | } | |
2110 | else | |
2111 | { | |
2112 | overrun_p = loop_vinfo && gap != 0; | |
2113 | if (overrun_p && vls_type != VLS_LOAD) | |
2114 | { | |
2115 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2116 | "Grouped store with gaps requires" | |
2117 | " non-consecutive accesses\n"); | |
2118 | return false; | |
2119 | } | |
f702e7d4 RS |
2120 | /* An overrun is fine if the trailing elements are smaller |
2121 | than the alignment boundary B. Every vector access will | |
2122 | be a multiple of B and so we are guaranteed to access a | |
2123 | non-gap element in the same B-sized block. */ | |
f9ef2c76 | 2124 | if (overrun_p |
f702e7d4 RS |
2125 | && gap < (vect_known_alignment_in_bytes (first_dr) |
2126 | / vect_get_scalar_dr_size (first_dr))) | |
f9ef2c76 | 2127 | overrun_p = false; |
2de001ee RS |
2128 | if (overrun_p && !can_overrun_p) |
2129 | { | |
2130 | if (dump_enabled_p ()) | |
2131 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2132 | "Peeling for outer loop is not supported\n"); | |
2133 | return false; | |
2134 | } | |
2135 | *memory_access_type = VMAT_CONTIGUOUS; | |
2136 | } | |
2137 | } | |
2138 | else | |
2139 | { | |
2140 | /* We can always handle this case using elementwise accesses, | |
2141 | but see if something more efficient is available. */ | |
2142 | *memory_access_type = VMAT_ELEMENTWISE; | |
2143 | ||
2144 | /* If there is a gap at the end of the group then these optimizations | |
2145 | would access excess elements in the last iteration. */ | |
2146 | bool would_overrun_p = (gap != 0); | |
f702e7d4 RS |
2147 | /* An overrun is fine if the trailing elements are smaller than the |
2148 | alignment boundary B. Every vector access will be a multiple of B | |
2149 | and so we are guaranteed to access a non-gap element in the | |
2150 | same B-sized block. */ | |
f9ef2c76 | 2151 | if (would_overrun_p |
7e11fc7f | 2152 | && !masked_p |
f702e7d4 RS |
2153 | && gap < (vect_known_alignment_in_bytes (first_dr) |
2154 | / vect_get_scalar_dr_size (first_dr))) | |
f9ef2c76 | 2155 | would_overrun_p = false; |
f702e7d4 | 2156 | |
2de001ee | 2157 | if (!STMT_VINFO_STRIDED_P (stmt_info) |
62da9e14 RS |
2158 | && (can_overrun_p || !would_overrun_p) |
2159 | && compare_step_with_zero (stmt) > 0) | |
2de001ee | 2160 | { |
6737facb RS |
2161 | /* First cope with the degenerate case of a single-element |
2162 | vector. */ | |
2163 | if (known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U)) | |
2164 | *memory_access_type = VMAT_CONTIGUOUS; | |
2165 | ||
2166 | /* Otherwise try using LOAD/STORE_LANES. */ | |
2167 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2168 | && (vls_type == VLS_LOAD | |
7e11fc7f RS |
2169 | ? vect_load_lanes_supported (vectype, group_size, masked_p) |
2170 | : vect_store_lanes_supported (vectype, group_size, | |
2171 | masked_p))) | |
2de001ee RS |
2172 | { |
2173 | *memory_access_type = VMAT_LOAD_STORE_LANES; | |
2174 | overrun_p = would_overrun_p; | |
2175 | } | |
2176 | ||
2177 | /* If that fails, try using permuting loads. */ | |
2178 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2179 | && (vls_type == VLS_LOAD | |
2180 | ? vect_grouped_load_supported (vectype, single_element_p, | |
2181 | group_size) | |
2182 | : vect_grouped_store_supported (vectype, group_size))) | |
2183 | { | |
2184 | *memory_access_type = VMAT_CONTIGUOUS_PERMUTE; | |
2185 | overrun_p = would_overrun_p; | |
2186 | } | |
2187 | } | |
429ef523 RS |
2188 | |
2189 | /* As a last resort, trying using a gather load or scatter store. | |
2190 | ||
2191 | ??? Although the code can handle all group sizes correctly, | |
2192 | it probably isn't a win to use separate strided accesses based | |
2193 | on nearby locations. Or, even if it's a win over scalar code, | |
2194 | it might not be a win over vectorizing at a lower VF, if that | |
2195 | allows us to use contiguous accesses. */ | |
2196 | if (*memory_access_type == VMAT_ELEMENTWISE | |
2197 | && single_element_p | |
2198 | && loop_vinfo | |
2199 | && vect_use_strided_gather_scatters_p (stmt, loop_vinfo, | |
2200 | masked_p, gs_info)) | |
2201 | *memory_access_type = VMAT_GATHER_SCATTER; | |
2de001ee RS |
2202 | } |
2203 | ||
2204 | if (vls_type != VLS_LOAD && first_stmt == stmt) | |
2205 | { | |
2206 | /* STMT is the leader of the group. Check the operands of all the | |
2207 | stmts of the group. */ | |
2208 | gimple *next_stmt = GROUP_NEXT_ELEMENT (stmt_info); | |
2209 | while (next_stmt) | |
2210 | { | |
7e11fc7f | 2211 | tree op = vect_get_store_rhs (next_stmt); |
2de001ee RS |
2212 | gimple *def_stmt; |
2213 | enum vect_def_type dt; | |
2214 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt)) | |
2215 | { | |
2216 | if (dump_enabled_p ()) | |
2217 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2218 | "use not simple.\n"); | |
2219 | return false; | |
2220 | } | |
2221 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); | |
2222 | } | |
2223 | } | |
2224 | ||
2225 | if (overrun_p) | |
2226 | { | |
2227 | gcc_assert (can_overrun_p); | |
2228 | if (dump_enabled_p ()) | |
2229 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2230 | "Data access with gaps requires scalar " | |
2231 | "epilogue loop\n"); | |
2232 | LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true; | |
2233 | } | |
2234 | ||
2235 | return true; | |
2236 | } | |
2237 | ||
62da9e14 RS |
2238 | /* A subroutine of get_load_store_type, with a subset of the same |
2239 | arguments. Handle the case where STMT is a load or store that | |
2240 | accesses consecutive elements with a negative step. */ | |
2241 | ||
2242 | static vect_memory_access_type | |
2243 | get_negative_load_store_type (gimple *stmt, tree vectype, | |
2244 | vec_load_store_type vls_type, | |
2245 | unsigned int ncopies) | |
2246 | { | |
2247 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2248 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
2249 | dr_alignment_support alignment_support_scheme; | |
2250 | ||
2251 | if (ncopies > 1) | |
2252 | { | |
2253 | if (dump_enabled_p ()) | |
2254 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2255 | "multiple types with negative step.\n"); | |
2256 | return VMAT_ELEMENTWISE; | |
2257 | } | |
2258 | ||
2259 | alignment_support_scheme = vect_supportable_dr_alignment (dr, false); | |
2260 | if (alignment_support_scheme != dr_aligned | |
2261 | && alignment_support_scheme != dr_unaligned_supported) | |
2262 | { | |
2263 | if (dump_enabled_p ()) | |
2264 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2265 | "negative step but alignment required.\n"); | |
2266 | return VMAT_ELEMENTWISE; | |
2267 | } | |
2268 | ||
2269 | if (vls_type == VLS_STORE_INVARIANT) | |
2270 | { | |
2271 | if (dump_enabled_p ()) | |
2272 | dump_printf_loc (MSG_NOTE, vect_location, | |
2273 | "negative step with invariant source;" | |
2274 | " no permute needed.\n"); | |
2275 | return VMAT_CONTIGUOUS_DOWN; | |
2276 | } | |
2277 | ||
2278 | if (!perm_mask_for_reverse (vectype)) | |
2279 | { | |
2280 | if (dump_enabled_p ()) | |
2281 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2282 | "negative step and reversing not supported.\n"); | |
2283 | return VMAT_ELEMENTWISE; | |
2284 | } | |
2285 | ||
2286 | return VMAT_CONTIGUOUS_REVERSE; | |
2287 | } | |
2288 | ||
2de001ee RS |
2289 | /* Analyze load or store statement STMT of type VLS_TYPE. Return true |
2290 | if there is a memory access type that the vectorized form can use, | |
2291 | storing it in *MEMORY_ACCESS_TYPE if so. If we decide to use gathers | |
2292 | or scatters, fill in GS_INFO accordingly. | |
2293 | ||
2294 | SLP says whether we're performing SLP rather than loop vectorization. | |
7e11fc7f | 2295 | MASKED_P is true if the statement is conditional on a vectorized mask. |
62da9e14 RS |
2296 | VECTYPE is the vector type that the vectorized statements will use. |
2297 | NCOPIES is the number of vector statements that will be needed. */ | |
2de001ee RS |
2298 | |
2299 | static bool | |
7e11fc7f | 2300 | get_load_store_type (gimple *stmt, tree vectype, bool slp, bool masked_p, |
62da9e14 | 2301 | vec_load_store_type vls_type, unsigned int ncopies, |
2de001ee RS |
2302 | vect_memory_access_type *memory_access_type, |
2303 | gather_scatter_info *gs_info) | |
2304 | { | |
2305 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2306 | vec_info *vinfo = stmt_info->vinfo; | |
2307 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4d694b27 | 2308 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
2de001ee RS |
2309 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
2310 | { | |
2311 | *memory_access_type = VMAT_GATHER_SCATTER; | |
2312 | gimple *def_stmt; | |
2313 | if (!vect_check_gather_scatter (stmt, loop_vinfo, gs_info)) | |
2314 | gcc_unreachable (); | |
2315 | else if (!vect_is_simple_use (gs_info->offset, vinfo, &def_stmt, | |
2316 | &gs_info->offset_dt, | |
2317 | &gs_info->offset_vectype)) | |
2318 | { | |
2319 | if (dump_enabled_p ()) | |
2320 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2321 | "%s index use not simple.\n", | |
2322 | vls_type == VLS_LOAD ? "gather" : "scatter"); | |
2323 | return false; | |
2324 | } | |
2325 | } | |
2326 | else if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) | |
2327 | { | |
7e11fc7f | 2328 | if (!get_group_load_store_type (stmt, vectype, slp, masked_p, vls_type, |
429ef523 | 2329 | memory_access_type, gs_info)) |
2de001ee RS |
2330 | return false; |
2331 | } | |
2332 | else if (STMT_VINFO_STRIDED_P (stmt_info)) | |
2333 | { | |
2334 | gcc_assert (!slp); | |
ab2fc782 | 2335 | if (loop_vinfo |
429ef523 RS |
2336 | && vect_use_strided_gather_scatters_p (stmt, loop_vinfo, |
2337 | masked_p, gs_info)) | |
ab2fc782 RS |
2338 | *memory_access_type = VMAT_GATHER_SCATTER; |
2339 | else | |
2340 | *memory_access_type = VMAT_ELEMENTWISE; | |
2de001ee RS |
2341 | } |
2342 | else | |
62da9e14 RS |
2343 | { |
2344 | int cmp = compare_step_with_zero (stmt); | |
2345 | if (cmp < 0) | |
2346 | *memory_access_type = get_negative_load_store_type | |
2347 | (stmt, vectype, vls_type, ncopies); | |
2348 | else if (cmp == 0) | |
2349 | { | |
2350 | gcc_assert (vls_type == VLS_LOAD); | |
2351 | *memory_access_type = VMAT_INVARIANT; | |
2352 | } | |
2353 | else | |
2354 | *memory_access_type = VMAT_CONTIGUOUS; | |
2355 | } | |
2de001ee | 2356 | |
4d694b27 RS |
2357 | if ((*memory_access_type == VMAT_ELEMENTWISE |
2358 | || *memory_access_type == VMAT_STRIDED_SLP) | |
2359 | && !nunits.is_constant ()) | |
2360 | { | |
2361 | if (dump_enabled_p ()) | |
2362 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2363 | "Not using elementwise accesses due to variable " | |
2364 | "vectorization factor.\n"); | |
2365 | return false; | |
2366 | } | |
2367 | ||
2de001ee RS |
2368 | /* FIXME: At the moment the cost model seems to underestimate the |
2369 | cost of using elementwise accesses. This check preserves the | |
2370 | traditional behavior until that can be fixed. */ | |
2371 | if (*memory_access_type == VMAT_ELEMENTWISE | |
4aa157e8 RS |
2372 | && !STMT_VINFO_STRIDED_P (stmt_info) |
2373 | && !(stmt == GROUP_FIRST_ELEMENT (stmt_info) | |
2374 | && !GROUP_NEXT_ELEMENT (stmt_info) | |
2375 | && !pow2p_hwi (GROUP_SIZE (stmt_info)))) | |
2de001ee RS |
2376 | { |
2377 | if (dump_enabled_p ()) | |
2378 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2379 | "not falling back to elementwise accesses\n"); | |
2380 | return false; | |
2381 | } | |
2382 | return true; | |
2383 | } | |
2384 | ||
aaeefd88 | 2385 | /* Return true if boolean argument MASK is suitable for vectorizing |
929b4411 RS |
2386 | conditional load or store STMT. When returning true, store the type |
2387 | of the definition in *MASK_DT_OUT and the type of the vectorized mask | |
2388 | in *MASK_VECTYPE_OUT. */ | |
aaeefd88 RS |
2389 | |
2390 | static bool | |
929b4411 RS |
2391 | vect_check_load_store_mask (gimple *stmt, tree mask, |
2392 | vect_def_type *mask_dt_out, | |
2393 | tree *mask_vectype_out) | |
aaeefd88 RS |
2394 | { |
2395 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (mask))) | |
2396 | { | |
2397 | if (dump_enabled_p ()) | |
2398 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2399 | "mask argument is not a boolean.\n"); | |
2400 | return false; | |
2401 | } | |
2402 | ||
2403 | if (TREE_CODE (mask) != SSA_NAME) | |
2404 | { | |
2405 | if (dump_enabled_p ()) | |
2406 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2407 | "mask argument is not an SSA name.\n"); | |
2408 | return false; | |
2409 | } | |
2410 | ||
2411 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2412 | gimple *def_stmt; | |
929b4411 | 2413 | enum vect_def_type mask_dt; |
aaeefd88 | 2414 | tree mask_vectype; |
929b4411 | 2415 | if (!vect_is_simple_use (mask, stmt_info->vinfo, &def_stmt, &mask_dt, |
aaeefd88 RS |
2416 | &mask_vectype)) |
2417 | { | |
2418 | if (dump_enabled_p ()) | |
2419 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2420 | "mask use not simple.\n"); | |
2421 | return false; | |
2422 | } | |
2423 | ||
2424 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2425 | if (!mask_vectype) | |
2426 | mask_vectype = get_mask_type_for_scalar_type (TREE_TYPE (vectype)); | |
2427 | ||
2428 | if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype)) | |
2429 | { | |
2430 | if (dump_enabled_p ()) | |
2431 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2432 | "could not find an appropriate vector mask type.\n"); | |
2433 | return false; | |
2434 | } | |
2435 | ||
2436 | if (maybe_ne (TYPE_VECTOR_SUBPARTS (mask_vectype), | |
2437 | TYPE_VECTOR_SUBPARTS (vectype))) | |
2438 | { | |
2439 | if (dump_enabled_p ()) | |
2440 | { | |
2441 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2442 | "vector mask type "); | |
2443 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, mask_vectype); | |
2444 | dump_printf (MSG_MISSED_OPTIMIZATION, | |
2445 | " does not match vector data type "); | |
2446 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, vectype); | |
2447 | dump_printf (MSG_MISSED_OPTIMIZATION, ".\n"); | |
2448 | } | |
2449 | return false; | |
2450 | } | |
2451 | ||
929b4411 | 2452 | *mask_dt_out = mask_dt; |
aaeefd88 RS |
2453 | *mask_vectype_out = mask_vectype; |
2454 | return true; | |
2455 | } | |
2456 | ||
3133c3b6 RS |
2457 | /* Return true if stored value RHS is suitable for vectorizing store |
2458 | statement STMT. When returning true, store the type of the | |
929b4411 RS |
2459 | definition in *RHS_DT_OUT, the type of the vectorized store value in |
2460 | *RHS_VECTYPE_OUT and the type of the store in *VLS_TYPE_OUT. */ | |
3133c3b6 RS |
2461 | |
2462 | static bool | |
929b4411 RS |
2463 | vect_check_store_rhs (gimple *stmt, tree rhs, vect_def_type *rhs_dt_out, |
2464 | tree *rhs_vectype_out, vec_load_store_type *vls_type_out) | |
3133c3b6 RS |
2465 | { |
2466 | /* In the case this is a store from a constant make sure | |
2467 | native_encode_expr can handle it. */ | |
2468 | if (CONSTANT_CLASS_P (rhs) && native_encode_expr (rhs, NULL, 64) == 0) | |
2469 | { | |
2470 | if (dump_enabled_p ()) | |
2471 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2472 | "cannot encode constant as a byte sequence.\n"); | |
2473 | return false; | |
2474 | } | |
2475 | ||
2476 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2477 | gimple *def_stmt; | |
929b4411 | 2478 | enum vect_def_type rhs_dt; |
3133c3b6 | 2479 | tree rhs_vectype; |
929b4411 | 2480 | if (!vect_is_simple_use (rhs, stmt_info->vinfo, &def_stmt, &rhs_dt, |
3133c3b6 RS |
2481 | &rhs_vectype)) |
2482 | { | |
2483 | if (dump_enabled_p ()) | |
2484 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2485 | "use not simple.\n"); | |
2486 | return false; | |
2487 | } | |
2488 | ||
2489 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2490 | if (rhs_vectype && !useless_type_conversion_p (vectype, rhs_vectype)) | |
2491 | { | |
2492 | if (dump_enabled_p ()) | |
2493 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
2494 | "incompatible vector types.\n"); | |
2495 | return false; | |
2496 | } | |
2497 | ||
929b4411 | 2498 | *rhs_dt_out = rhs_dt; |
3133c3b6 | 2499 | *rhs_vectype_out = rhs_vectype; |
929b4411 | 2500 | if (rhs_dt == vect_constant_def || rhs_dt == vect_external_def) |
3133c3b6 RS |
2501 | *vls_type_out = VLS_STORE_INVARIANT; |
2502 | else | |
2503 | *vls_type_out = VLS_STORE; | |
2504 | return true; | |
2505 | } | |
2506 | ||
bc9587eb RS |
2507 | /* Build an all-ones vector mask of type MASKTYPE while vectorizing STMT. |
2508 | Note that we support masks with floating-point type, in which case the | |
2509 | floats are interpreted as a bitmask. */ | |
2510 | ||
2511 | static tree | |
2512 | vect_build_all_ones_mask (gimple *stmt, tree masktype) | |
2513 | { | |
2514 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
2515 | return build_int_cst (masktype, -1); | |
2516 | else if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE) | |
2517 | { | |
2518 | tree mask = build_int_cst (TREE_TYPE (masktype), -1); | |
2519 | mask = build_vector_from_val (masktype, mask); | |
2520 | return vect_init_vector (stmt, mask, masktype, NULL); | |
2521 | } | |
2522 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype))) | |
2523 | { | |
2524 | REAL_VALUE_TYPE r; | |
2525 | long tmp[6]; | |
2526 | for (int j = 0; j < 6; ++j) | |
2527 | tmp[j] = -1; | |
2528 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype))); | |
2529 | tree mask = build_real (TREE_TYPE (masktype), r); | |
2530 | mask = build_vector_from_val (masktype, mask); | |
2531 | return vect_init_vector (stmt, mask, masktype, NULL); | |
2532 | } | |
2533 | gcc_unreachable (); | |
2534 | } | |
2535 | ||
2536 | /* Build an all-zero merge value of type VECTYPE while vectorizing | |
2537 | STMT as a gather load. */ | |
2538 | ||
2539 | static tree | |
2540 | vect_build_zero_merge_argument (gimple *stmt, tree vectype) | |
2541 | { | |
2542 | tree merge; | |
2543 | if (TREE_CODE (TREE_TYPE (vectype)) == INTEGER_TYPE) | |
2544 | merge = build_int_cst (TREE_TYPE (vectype), 0); | |
2545 | else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (vectype))) | |
2546 | { | |
2547 | REAL_VALUE_TYPE r; | |
2548 | long tmp[6]; | |
2549 | for (int j = 0; j < 6; ++j) | |
2550 | tmp[j] = 0; | |
2551 | real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (vectype))); | |
2552 | merge = build_real (TREE_TYPE (vectype), r); | |
2553 | } | |
2554 | else | |
2555 | gcc_unreachable (); | |
2556 | merge = build_vector_from_val (vectype, merge); | |
2557 | return vect_init_vector (stmt, merge, vectype, NULL); | |
2558 | } | |
2559 | ||
c48d2d35 RS |
2560 | /* Build a gather load call while vectorizing STMT. Insert new instructions |
2561 | before GSI and add them to VEC_STMT. GS_INFO describes the gather load | |
2562 | operation. If the load is conditional, MASK is the unvectorized | |
929b4411 | 2563 | condition and MASK_DT is its definition type, otherwise MASK is null. */ |
c48d2d35 RS |
2564 | |
2565 | static void | |
2566 | vect_build_gather_load_calls (gimple *stmt, gimple_stmt_iterator *gsi, | |
2567 | gimple **vec_stmt, gather_scatter_info *gs_info, | |
929b4411 | 2568 | tree mask, vect_def_type mask_dt) |
c48d2d35 RS |
2569 | { |
2570 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2571 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2572 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2573 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2574 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
2575 | int ncopies = vect_get_num_copies (loop_vinfo, vectype); | |
2576 | edge pe = loop_preheader_edge (loop); | |
2577 | enum { NARROW, NONE, WIDEN } modifier; | |
2578 | poly_uint64 gather_off_nunits | |
2579 | = TYPE_VECTOR_SUBPARTS (gs_info->offset_vectype); | |
2580 | ||
2581 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info->decl)); | |
2582 | tree rettype = TREE_TYPE (TREE_TYPE (gs_info->decl)); | |
2583 | tree srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2584 | tree ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2585 | tree idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2586 | tree masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
2587 | tree scaletype = TREE_VALUE (arglist); | |
2588 | gcc_checking_assert (types_compatible_p (srctype, rettype) | |
2589 | && (!mask || types_compatible_p (srctype, masktype))); | |
2590 | ||
2591 | tree perm_mask = NULL_TREE; | |
2592 | tree mask_perm_mask = NULL_TREE; | |
2593 | if (known_eq (nunits, gather_off_nunits)) | |
2594 | modifier = NONE; | |
2595 | else if (known_eq (nunits * 2, gather_off_nunits)) | |
2596 | { | |
2597 | modifier = WIDEN; | |
2598 | ||
2599 | /* Currently widening gathers and scatters are only supported for | |
2600 | fixed-length vectors. */ | |
2601 | int count = gather_off_nunits.to_constant (); | |
2602 | vec_perm_builder sel (count, count, 1); | |
2603 | for (int i = 0; i < count; ++i) | |
2604 | sel.quick_push (i | (count / 2)); | |
2605 | ||
2606 | vec_perm_indices indices (sel, 1, count); | |
2607 | perm_mask = vect_gen_perm_mask_checked (gs_info->offset_vectype, | |
2608 | indices); | |
2609 | } | |
2610 | else if (known_eq (nunits, gather_off_nunits * 2)) | |
2611 | { | |
2612 | modifier = NARROW; | |
2613 | ||
2614 | /* Currently narrowing gathers and scatters are only supported for | |
2615 | fixed-length vectors. */ | |
2616 | int count = nunits.to_constant (); | |
2617 | vec_perm_builder sel (count, count, 1); | |
2618 | sel.quick_grow (count); | |
2619 | for (int i = 0; i < count; ++i) | |
2620 | sel[i] = i < count / 2 ? i : i + count / 2; | |
2621 | vec_perm_indices indices (sel, 2, count); | |
2622 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); | |
2623 | ||
2624 | ncopies *= 2; | |
2625 | ||
2626 | if (mask) | |
2627 | { | |
2628 | for (int i = 0; i < count; ++i) | |
2629 | sel[i] = i | (count / 2); | |
2630 | indices.new_vector (sel, 2, count); | |
2631 | mask_perm_mask = vect_gen_perm_mask_checked (masktype, indices); | |
2632 | } | |
2633 | } | |
2634 | else | |
2635 | gcc_unreachable (); | |
2636 | ||
2637 | tree vec_dest = vect_create_destination_var (gimple_get_lhs (stmt), | |
2638 | vectype); | |
2639 | ||
2640 | tree ptr = fold_convert (ptrtype, gs_info->base); | |
2641 | if (!is_gimple_min_invariant (ptr)) | |
2642 | { | |
2643 | gimple_seq seq; | |
2644 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
2645 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
2646 | gcc_assert (!new_bb); | |
2647 | } | |
2648 | ||
2649 | tree scale = build_int_cst (scaletype, gs_info->scale); | |
2650 | ||
2651 | tree vec_oprnd0 = NULL_TREE; | |
2652 | tree vec_mask = NULL_TREE; | |
2653 | tree src_op = NULL_TREE; | |
2654 | tree mask_op = NULL_TREE; | |
2655 | tree prev_res = NULL_TREE; | |
2656 | stmt_vec_info prev_stmt_info = NULL; | |
2657 | ||
2658 | if (!mask) | |
2659 | { | |
2660 | src_op = vect_build_zero_merge_argument (stmt, rettype); | |
2661 | mask_op = vect_build_all_ones_mask (stmt, masktype); | |
2662 | } | |
2663 | ||
2664 | for (int j = 0; j < ncopies; ++j) | |
2665 | { | |
2666 | tree op, var; | |
2667 | gimple *new_stmt; | |
2668 | if (modifier == WIDEN && (j & 1)) | |
2669 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, | |
2670 | perm_mask, stmt, gsi); | |
2671 | else if (j == 0) | |
2672 | op = vec_oprnd0 | |
2673 | = vect_get_vec_def_for_operand (gs_info->offset, stmt); | |
2674 | else | |
2675 | op = vec_oprnd0 | |
2676 | = vect_get_vec_def_for_stmt_copy (gs_info->offset_dt, vec_oprnd0); | |
2677 | ||
2678 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
2679 | { | |
2680 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)), | |
2681 | TYPE_VECTOR_SUBPARTS (idxtype))); | |
2682 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); | |
2683 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); | |
2684 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
2685 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2686 | op = var; | |
2687 | } | |
2688 | ||
2689 | if (mask) | |
2690 | { | |
2691 | if (mask_perm_mask && (j & 1)) | |
2692 | mask_op = permute_vec_elements (mask_op, mask_op, | |
2693 | mask_perm_mask, stmt, gsi); | |
2694 | else | |
2695 | { | |
2696 | if (j == 0) | |
2697 | vec_mask = vect_get_vec_def_for_operand (mask, stmt); | |
2698 | else | |
929b4411 | 2699 | vec_mask = vect_get_vec_def_for_stmt_copy (mask_dt, vec_mask); |
c48d2d35 RS |
2700 | |
2701 | mask_op = vec_mask; | |
2702 | if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask))) | |
2703 | { | |
2704 | gcc_assert | |
2705 | (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)), | |
2706 | TYPE_VECTOR_SUBPARTS (masktype))); | |
2707 | var = vect_get_new_ssa_name (masktype, vect_simple_var); | |
2708 | mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op); | |
2709 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, | |
2710 | mask_op); | |
2711 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2712 | mask_op = var; | |
2713 | } | |
2714 | } | |
2715 | src_op = mask_op; | |
2716 | } | |
2717 | ||
2718 | new_stmt = gimple_build_call (gs_info->decl, 5, src_op, ptr, op, | |
2719 | mask_op, scale); | |
2720 | ||
2721 | if (!useless_type_conversion_p (vectype, rettype)) | |
2722 | { | |
2723 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (vectype), | |
2724 | TYPE_VECTOR_SUBPARTS (rettype))); | |
2725 | op = vect_get_new_ssa_name (rettype, vect_simple_var); | |
2726 | gimple_call_set_lhs (new_stmt, op); | |
2727 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2728 | var = make_ssa_name (vec_dest); | |
2729 | op = build1 (VIEW_CONVERT_EXPR, vectype, op); | |
2730 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
2731 | } | |
2732 | else | |
2733 | { | |
2734 | var = make_ssa_name (vec_dest, new_stmt); | |
2735 | gimple_call_set_lhs (new_stmt, var); | |
2736 | } | |
2737 | ||
2738 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2739 | ||
2740 | if (modifier == NARROW) | |
2741 | { | |
2742 | if ((j & 1) == 0) | |
2743 | { | |
2744 | prev_res = var; | |
2745 | continue; | |
2746 | } | |
2747 | var = permute_vec_elements (prev_res, var, perm_mask, stmt, gsi); | |
2748 | new_stmt = SSA_NAME_DEF_STMT (var); | |
2749 | } | |
2750 | ||
2751 | if (prev_stmt_info == NULL) | |
2752 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2753 | else | |
2754 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2755 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2756 | } | |
2757 | } | |
2758 | ||
bfaa08b7 RS |
2759 | /* Prepare the base and offset in GS_INFO for vectorization. |
2760 | Set *DATAREF_PTR to the loop-invariant base address and *VEC_OFFSET | |
2761 | to the vectorized offset argument for the first copy of STMT. STMT | |
2762 | is the statement described by GS_INFO and LOOP is the containing loop. */ | |
2763 | ||
2764 | static void | |
2765 | vect_get_gather_scatter_ops (struct loop *loop, gimple *stmt, | |
2766 | gather_scatter_info *gs_info, | |
2767 | tree *dataref_ptr, tree *vec_offset) | |
2768 | { | |
2769 | gimple_seq stmts = NULL; | |
2770 | *dataref_ptr = force_gimple_operand (gs_info->base, &stmts, true, NULL_TREE); | |
2771 | if (stmts != NULL) | |
2772 | { | |
2773 | basic_block new_bb; | |
2774 | edge pe = loop_preheader_edge (loop); | |
2775 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
2776 | gcc_assert (!new_bb); | |
2777 | } | |
2778 | tree offset_type = TREE_TYPE (gs_info->offset); | |
2779 | tree offset_vectype = get_vectype_for_scalar_type (offset_type); | |
2780 | *vec_offset = vect_get_vec_def_for_operand (gs_info->offset, stmt, | |
2781 | offset_vectype); | |
2782 | } | |
2783 | ||
ab2fc782 RS |
2784 | /* Prepare to implement a grouped or strided load or store using |
2785 | the gather load or scatter store operation described by GS_INFO. | |
2786 | STMT is the load or store statement. | |
2787 | ||
2788 | Set *DATAREF_BUMP to the amount that should be added to the base | |
2789 | address after each copy of the vectorized statement. Set *VEC_OFFSET | |
2790 | to an invariant offset vector in which element I has the value | |
2791 | I * DR_STEP / SCALE. */ | |
2792 | ||
2793 | static void | |
2794 | vect_get_strided_load_store_ops (gimple *stmt, loop_vec_info loop_vinfo, | |
2795 | gather_scatter_info *gs_info, | |
2796 | tree *dataref_bump, tree *vec_offset) | |
2797 | { | |
2798 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2799 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
2800 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2801 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2802 | gimple_seq stmts; | |
2803 | ||
2804 | tree bump = size_binop (MULT_EXPR, | |
2805 | fold_convert (sizetype, DR_STEP (dr)), | |
2806 | size_int (TYPE_VECTOR_SUBPARTS (vectype))); | |
2807 | *dataref_bump = force_gimple_operand (bump, &stmts, true, NULL_TREE); | |
2808 | if (stmts) | |
2809 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
2810 | ||
2811 | /* The offset given in GS_INFO can have pointer type, so use the element | |
2812 | type of the vector instead. */ | |
2813 | tree offset_type = TREE_TYPE (gs_info->offset); | |
2814 | tree offset_vectype = get_vectype_for_scalar_type (offset_type); | |
2815 | offset_type = TREE_TYPE (offset_vectype); | |
2816 | ||
2817 | /* Calculate X = DR_STEP / SCALE and convert it to the appropriate type. */ | |
2818 | tree step = size_binop (EXACT_DIV_EXPR, DR_STEP (dr), | |
2819 | ssize_int (gs_info->scale)); | |
2820 | step = fold_convert (offset_type, step); | |
2821 | step = force_gimple_operand (step, &stmts, true, NULL_TREE); | |
2822 | ||
2823 | /* Create {0, X, X*2, X*3, ...}. */ | |
2824 | *vec_offset = gimple_build (&stmts, VEC_SERIES_EXPR, offset_vectype, | |
2825 | build_zero_cst (offset_type), step); | |
2826 | if (stmts) | |
2827 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
2828 | } | |
2829 | ||
2830 | /* Return the amount that should be added to a vector pointer to move | |
2831 | to the next or previous copy of AGGR_TYPE. DR is the data reference | |
2832 | being vectorized and MEMORY_ACCESS_TYPE describes the type of | |
2833 | vectorization. */ | |
2834 | ||
2835 | static tree | |
2836 | vect_get_data_ptr_increment (data_reference *dr, tree aggr_type, | |
2837 | vect_memory_access_type memory_access_type) | |
2838 | { | |
2839 | if (memory_access_type == VMAT_INVARIANT) | |
2840 | return size_zero_node; | |
2841 | ||
2842 | tree iv_step = TYPE_SIZE_UNIT (aggr_type); | |
2843 | tree step = vect_dr_behavior (dr)->step; | |
2844 | if (tree_int_cst_sgn (step) == -1) | |
2845 | iv_step = fold_build1 (NEGATE_EXPR, TREE_TYPE (iv_step), iv_step); | |
2846 | return iv_step; | |
2847 | } | |
2848 | ||
37b14185 RB |
2849 | /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64}. */ |
2850 | ||
2851 | static bool | |
2852 | vectorizable_bswap (gimple *stmt, gimple_stmt_iterator *gsi, | |
2853 | gimple **vec_stmt, slp_tree slp_node, | |
2854 | tree vectype_in, enum vect_def_type *dt) | |
2855 | { | |
2856 | tree op, vectype; | |
2857 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2858 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
928686b1 RS |
2859 | unsigned ncopies; |
2860 | unsigned HOST_WIDE_INT nunits, num_bytes; | |
37b14185 RB |
2861 | |
2862 | op = gimple_call_arg (stmt, 0); | |
2863 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
928686b1 RS |
2864 | |
2865 | if (!TYPE_VECTOR_SUBPARTS (vectype).is_constant (&nunits)) | |
2866 | return false; | |
37b14185 RB |
2867 | |
2868 | /* Multiple types in SLP are handled by creating the appropriate number of | |
2869 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
2870 | case of SLP. */ | |
2871 | if (slp_node) | |
2872 | ncopies = 1; | |
2873 | else | |
e8f142e2 | 2874 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
37b14185 RB |
2875 | |
2876 | gcc_assert (ncopies >= 1); | |
2877 | ||
2878 | tree char_vectype = get_same_sized_vectype (char_type_node, vectype_in); | |
2879 | if (! char_vectype) | |
2880 | return false; | |
2881 | ||
928686b1 RS |
2882 | if (!TYPE_VECTOR_SUBPARTS (char_vectype).is_constant (&num_bytes)) |
2883 | return false; | |
2884 | ||
794e3180 | 2885 | unsigned word_bytes = num_bytes / nunits; |
908a1a16 | 2886 | |
d980067b RS |
2887 | /* The encoding uses one stepped pattern for each byte in the word. */ |
2888 | vec_perm_builder elts (num_bytes, word_bytes, 3); | |
2889 | for (unsigned i = 0; i < 3; ++i) | |
37b14185 | 2890 | for (unsigned j = 0; j < word_bytes; ++j) |
908a1a16 | 2891 | elts.quick_push ((i + 1) * word_bytes - j - 1); |
37b14185 | 2892 | |
e3342de4 RS |
2893 | vec_perm_indices indices (elts, 1, num_bytes); |
2894 | if (!can_vec_perm_const_p (TYPE_MODE (char_vectype), indices)) | |
37b14185 RB |
2895 | return false; |
2896 | ||
2897 | if (! vec_stmt) | |
2898 | { | |
2899 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
2900 | if (dump_enabled_p ()) | |
2901 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_bswap ===" | |
2902 | "\n"); | |
78604de0 | 2903 | if (! slp_node) |
37b14185 RB |
2904 | { |
2905 | add_stmt_cost (stmt_info->vinfo->target_cost_data, | |
2906 | 1, vector_stmt, stmt_info, 0, vect_prologue); | |
2907 | add_stmt_cost (stmt_info->vinfo->target_cost_data, | |
2908 | ncopies, vec_perm, stmt_info, 0, vect_body); | |
2909 | } | |
2910 | return true; | |
2911 | } | |
2912 | ||
736d0f28 | 2913 | tree bswap_vconst = vec_perm_indices_to_tree (char_vectype, indices); |
37b14185 RB |
2914 | |
2915 | /* Transform. */ | |
2916 | vec<tree> vec_oprnds = vNULL; | |
2917 | gimple *new_stmt = NULL; | |
2918 | stmt_vec_info prev_stmt_info = NULL; | |
2919 | for (unsigned j = 0; j < ncopies; j++) | |
2920 | { | |
2921 | /* Handle uses. */ | |
2922 | if (j == 0) | |
306b0c92 | 2923 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); |
37b14185 RB |
2924 | else |
2925 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
2926 | ||
2927 | /* Arguments are ready. create the new vector stmt. */ | |
2928 | unsigned i; | |
2929 | tree vop; | |
2930 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) | |
2931 | { | |
2932 | tree tem = make_ssa_name (char_vectype); | |
2933 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
2934 | char_vectype, vop)); | |
2935 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2936 | tree tem2 = make_ssa_name (char_vectype); | |
2937 | new_stmt = gimple_build_assign (tem2, VEC_PERM_EXPR, | |
2938 | tem, tem, bswap_vconst); | |
2939 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2940 | tem = make_ssa_name (vectype); | |
2941 | new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
2942 | vectype, tem2)); | |
2943 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
2944 | if (slp_node) | |
2945 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
2946 | } | |
2947 | ||
2948 | if (slp_node) | |
2949 | continue; | |
2950 | ||
2951 | if (j == 0) | |
2952 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
2953 | else | |
2954 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
2955 | ||
2956 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
2957 | } | |
2958 | ||
2959 | vec_oprnds.release (); | |
2960 | return true; | |
2961 | } | |
2962 | ||
b1b6836e RS |
2963 | /* Return true if vector types VECTYPE_IN and VECTYPE_OUT have |
2964 | integer elements and if we can narrow VECTYPE_IN to VECTYPE_OUT | |
2965 | in a single step. On success, store the binary pack code in | |
2966 | *CONVERT_CODE. */ | |
2967 | ||
2968 | static bool | |
2969 | simple_integer_narrowing (tree vectype_out, tree vectype_in, | |
2970 | tree_code *convert_code) | |
2971 | { | |
2972 | if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_out)) | |
2973 | || !INTEGRAL_TYPE_P (TREE_TYPE (vectype_in))) | |
2974 | return false; | |
2975 | ||
2976 | tree_code code; | |
2977 | int multi_step_cvt = 0; | |
2978 | auto_vec <tree, 8> interm_types; | |
2979 | if (!supportable_narrowing_operation (NOP_EXPR, vectype_out, vectype_in, | |
2980 | &code, &multi_step_cvt, | |
2981 | &interm_types) | |
2982 | || multi_step_cvt) | |
2983 | return false; | |
2984 | ||
2985 | *convert_code = code; | |
2986 | return true; | |
2987 | } | |
5ce9450f | 2988 | |
ebfd146a IR |
2989 | /* Function vectorizable_call. |
2990 | ||
538dd0b7 | 2991 | Check if GS performs a function call that can be vectorized. |
b8698a0f | 2992 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
2993 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
2994 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
2995 | ||
2996 | static bool | |
355fe088 | 2997 | vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
190c2236 | 2998 | slp_tree slp_node) |
ebfd146a | 2999 | { |
538dd0b7 | 3000 | gcall *stmt; |
ebfd146a IR |
3001 | tree vec_dest; |
3002 | tree scalar_dest; | |
3003 | tree op, type; | |
3004 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; | |
538dd0b7 | 3005 | stmt_vec_info stmt_info = vinfo_for_stmt (gs), prev_stmt_info; |
ebfd146a | 3006 | tree vectype_out, vectype_in; |
c7bda0f4 RS |
3007 | poly_uint64 nunits_in; |
3008 | poly_uint64 nunits_out; | |
ebfd146a | 3009 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
190c2236 | 3010 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 3011 | vec_info *vinfo = stmt_info->vinfo; |
81c40241 | 3012 | tree fndecl, new_temp, rhs_type; |
355fe088 | 3013 | gimple *def_stmt; |
0502fb85 UB |
3014 | enum vect_def_type dt[3] |
3015 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 3016 | int ndts = 3; |
355fe088 | 3017 | gimple *new_stmt = NULL; |
ebfd146a | 3018 | int ncopies, j; |
6e1aa848 | 3019 | vec<tree> vargs = vNULL; |
ebfd146a IR |
3020 | enum { NARROW, NONE, WIDEN } modifier; |
3021 | size_t i, nargs; | |
9d5e7640 | 3022 | tree lhs; |
ebfd146a | 3023 | |
190c2236 | 3024 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
3025 | return false; |
3026 | ||
66c16fd9 RB |
3027 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3028 | && ! vec_stmt) | |
ebfd146a IR |
3029 | return false; |
3030 | ||
538dd0b7 DM |
3031 | /* Is GS a vectorizable call? */ |
3032 | stmt = dyn_cast <gcall *> (gs); | |
3033 | if (!stmt) | |
ebfd146a IR |
3034 | return false; |
3035 | ||
5ce9450f | 3036 | if (gimple_call_internal_p (stmt) |
bfaa08b7 | 3037 | && (internal_load_fn_p (gimple_call_internal_fn (stmt)) |
f307441a | 3038 | || internal_store_fn_p (gimple_call_internal_fn (stmt)))) |
c3a8f964 RS |
3039 | /* Handled by vectorizable_load and vectorizable_store. */ |
3040 | return false; | |
5ce9450f | 3041 | |
0136f8f0 AH |
3042 | if (gimple_call_lhs (stmt) == NULL_TREE |
3043 | || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
ebfd146a IR |
3044 | return false; |
3045 | ||
0136f8f0 | 3046 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); |
5a2c1986 | 3047 | |
b690cc0f RG |
3048 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
3049 | ||
ebfd146a IR |
3050 | /* Process function arguments. */ |
3051 | rhs_type = NULL_TREE; | |
b690cc0f | 3052 | vectype_in = NULL_TREE; |
ebfd146a IR |
3053 | nargs = gimple_call_num_args (stmt); |
3054 | ||
1b1562a5 MM |
3055 | /* Bail out if the function has more than three arguments, we do not have |
3056 | interesting builtin functions to vectorize with more than two arguments | |
3057 | except for fma. No arguments is also not good. */ | |
3058 | if (nargs == 0 || nargs > 3) | |
ebfd146a IR |
3059 | return false; |
3060 | ||
74bf76ed JJ |
3061 | /* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */ |
3062 | if (gimple_call_internal_p (stmt) | |
3063 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
3064 | { | |
3065 | nargs = 0; | |
3066 | rhs_type = unsigned_type_node; | |
3067 | } | |
3068 | ||
ebfd146a IR |
3069 | for (i = 0; i < nargs; i++) |
3070 | { | |
b690cc0f RG |
3071 | tree opvectype; |
3072 | ||
ebfd146a IR |
3073 | op = gimple_call_arg (stmt, i); |
3074 | ||
3075 | /* We can only handle calls with arguments of the same type. */ | |
3076 | if (rhs_type | |
8533c9d8 | 3077 | && !types_compatible_p (rhs_type, TREE_TYPE (op))) |
ebfd146a | 3078 | { |
73fbfcad | 3079 | if (dump_enabled_p ()) |
78c60e3d | 3080 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3081 | "argument types differ.\n"); |
ebfd146a IR |
3082 | return false; |
3083 | } | |
b690cc0f RG |
3084 | if (!rhs_type) |
3085 | rhs_type = TREE_TYPE (op); | |
ebfd146a | 3086 | |
81c40241 | 3087 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[i], &opvectype)) |
ebfd146a | 3088 | { |
73fbfcad | 3089 | if (dump_enabled_p ()) |
78c60e3d | 3090 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3091 | "use not simple.\n"); |
ebfd146a IR |
3092 | return false; |
3093 | } | |
ebfd146a | 3094 | |
b690cc0f RG |
3095 | if (!vectype_in) |
3096 | vectype_in = opvectype; | |
3097 | else if (opvectype | |
3098 | && opvectype != vectype_in) | |
3099 | { | |
73fbfcad | 3100 | if (dump_enabled_p ()) |
78c60e3d | 3101 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 3102 | "argument vector types differ.\n"); |
b690cc0f RG |
3103 | return false; |
3104 | } | |
3105 | } | |
3106 | /* If all arguments are external or constant defs use a vector type with | |
3107 | the same size as the output vector type. */ | |
ebfd146a | 3108 | if (!vectype_in) |
b690cc0f | 3109 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
3110 | if (vec_stmt) |
3111 | gcc_assert (vectype_in); | |
3112 | if (!vectype_in) | |
3113 | { | |
73fbfcad | 3114 | if (dump_enabled_p ()) |
7d8930a0 | 3115 | { |
78c60e3d SS |
3116 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
3117 | "no vectype for scalar type "); | |
3118 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 3119 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
3120 | } |
3121 | ||
3122 | return false; | |
3123 | } | |
ebfd146a IR |
3124 | |
3125 | /* FORNOW */ | |
b690cc0f RG |
3126 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
3127 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
c7bda0f4 | 3128 | if (known_eq (nunits_in * 2, nunits_out)) |
ebfd146a | 3129 | modifier = NARROW; |
c7bda0f4 | 3130 | else if (known_eq (nunits_out, nunits_in)) |
ebfd146a | 3131 | modifier = NONE; |
c7bda0f4 | 3132 | else if (known_eq (nunits_out * 2, nunits_in)) |
ebfd146a IR |
3133 | modifier = WIDEN; |
3134 | else | |
3135 | return false; | |
3136 | ||
70439f0d RS |
3137 | /* We only handle functions that do not read or clobber memory. */ |
3138 | if (gimple_vuse (stmt)) | |
3139 | { | |
3140 | if (dump_enabled_p ()) | |
3141 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3142 | "function reads from or writes to memory.\n"); | |
3143 | return false; | |
3144 | } | |
3145 | ||
ebfd146a IR |
3146 | /* For now, we only vectorize functions if a target specific builtin |
3147 | is available. TODO -- in some cases, it might be profitable to | |
3148 | insert the calls for pieces of the vector, in order to be able | |
3149 | to vectorize other operations in the loop. */ | |
70439f0d RS |
3150 | fndecl = NULL_TREE; |
3151 | internal_fn ifn = IFN_LAST; | |
3152 | combined_fn cfn = gimple_call_combined_fn (stmt); | |
3153 | tree callee = gimple_call_fndecl (stmt); | |
3154 | ||
3155 | /* First try using an internal function. */ | |
b1b6836e RS |
3156 | tree_code convert_code = ERROR_MARK; |
3157 | if (cfn != CFN_LAST | |
3158 | && (modifier == NONE | |
3159 | || (modifier == NARROW | |
3160 | && simple_integer_narrowing (vectype_out, vectype_in, | |
3161 | &convert_code)))) | |
70439f0d RS |
3162 | ifn = vectorizable_internal_function (cfn, callee, vectype_out, |
3163 | vectype_in); | |
3164 | ||
3165 | /* If that fails, try asking for a target-specific built-in function. */ | |
3166 | if (ifn == IFN_LAST) | |
3167 | { | |
3168 | if (cfn != CFN_LAST) | |
3169 | fndecl = targetm.vectorize.builtin_vectorized_function | |
3170 | (cfn, vectype_out, vectype_in); | |
7672aa9b | 3171 | else if (callee) |
70439f0d RS |
3172 | fndecl = targetm.vectorize.builtin_md_vectorized_function |
3173 | (callee, vectype_out, vectype_in); | |
3174 | } | |
3175 | ||
3176 | if (ifn == IFN_LAST && !fndecl) | |
ebfd146a | 3177 | { |
70439f0d | 3178 | if (cfn == CFN_GOMP_SIMD_LANE |
74bf76ed JJ |
3179 | && !slp_node |
3180 | && loop_vinfo | |
3181 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
3182 | && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME | |
3183 | && LOOP_VINFO_LOOP (loop_vinfo)->simduid | |
3184 | == SSA_NAME_VAR (gimple_call_arg (stmt, 0))) | |
3185 | { | |
3186 | /* We can handle IFN_GOMP_SIMD_LANE by returning a | |
3187 | { 0, 1, 2, ... vf - 1 } vector. */ | |
3188 | gcc_assert (nargs == 0); | |
3189 | } | |
37b14185 RB |
3190 | else if (modifier == NONE |
3191 | && (gimple_call_builtin_p (stmt, BUILT_IN_BSWAP16) | |
3192 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP32) | |
3193 | || gimple_call_builtin_p (stmt, BUILT_IN_BSWAP64))) | |
3194 | return vectorizable_bswap (stmt, gsi, vec_stmt, slp_node, | |
3195 | vectype_in, dt); | |
74bf76ed JJ |
3196 | else |
3197 | { | |
3198 | if (dump_enabled_p ()) | |
3199 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
e645e942 | 3200 | "function is not vectorizable.\n"); |
74bf76ed JJ |
3201 | return false; |
3202 | } | |
ebfd146a IR |
3203 | } |
3204 | ||
fce57248 | 3205 | if (slp_node) |
190c2236 | 3206 | ncopies = 1; |
b1b6836e | 3207 | else if (modifier == NARROW && ifn == IFN_LAST) |
e8f142e2 | 3208 | ncopies = vect_get_num_copies (loop_vinfo, vectype_out); |
ebfd146a | 3209 | else |
e8f142e2 | 3210 | ncopies = vect_get_num_copies (loop_vinfo, vectype_in); |
ebfd146a IR |
3211 | |
3212 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
3213 | needs to be generated. */ | |
3214 | gcc_assert (ncopies >= 1); | |
3215 | ||
3216 | if (!vec_stmt) /* transformation not required. */ | |
3217 | { | |
3218 | STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; | |
73fbfcad | 3219 | if (dump_enabled_p ()) |
e645e942 TJ |
3220 | dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_call ===" |
3221 | "\n"); | |
78604de0 RB |
3222 | if (!slp_node) |
3223 | { | |
3224 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); | |
3225 | if (ifn != IFN_LAST && modifier == NARROW && !slp_node) | |
3226 | add_stmt_cost (stmt_info->vinfo->target_cost_data, ncopies / 2, | |
3227 | vec_promote_demote, stmt_info, 0, vect_body); | |
3228 | } | |
b1b6836e | 3229 | |
ebfd146a IR |
3230 | return true; |
3231 | } | |
3232 | ||
67b8dbac | 3233 | /* Transform. */ |
ebfd146a | 3234 | |
73fbfcad | 3235 | if (dump_enabled_p ()) |
e645e942 | 3236 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); |
ebfd146a IR |
3237 | |
3238 | /* Handle def. */ | |
3239 | scalar_dest = gimple_call_lhs (stmt); | |
3240 | vec_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
3241 | ||
3242 | prev_stmt_info = NULL; | |
b1b6836e | 3243 | if (modifier == NONE || ifn != IFN_LAST) |
ebfd146a | 3244 | { |
b1b6836e | 3245 | tree prev_res = NULL_TREE; |
ebfd146a IR |
3246 | for (j = 0; j < ncopies; ++j) |
3247 | { | |
3248 | /* Build argument list for the vectorized call. */ | |
3249 | if (j == 0) | |
9771b263 | 3250 | vargs.create (nargs); |
ebfd146a | 3251 | else |
9771b263 | 3252 | vargs.truncate (0); |
ebfd146a | 3253 | |
190c2236 JJ |
3254 | if (slp_node) |
3255 | { | |
ef062b13 | 3256 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 3257 | vec<tree> vec_oprnds0; |
190c2236 JJ |
3258 | |
3259 | for (i = 0; i < nargs; i++) | |
9771b263 | 3260 | vargs.quick_push (gimple_call_arg (stmt, i)); |
306b0c92 | 3261 | vect_get_slp_defs (vargs, slp_node, &vec_defs); |
37b5ec8f | 3262 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
3263 | |
3264 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3265 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) |
190c2236 JJ |
3266 | { |
3267 | size_t k; | |
3268 | for (k = 0; k < nargs; k++) | |
3269 | { | |
37b5ec8f | 3270 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 | 3271 | vargs[k] = vec_oprndsk[i]; |
190c2236 | 3272 | } |
b1b6836e RS |
3273 | if (modifier == NARROW) |
3274 | { | |
3275 | tree half_res = make_ssa_name (vectype_in); | |
a844293d RS |
3276 | gcall *call |
3277 | = gimple_build_call_internal_vec (ifn, vargs); | |
3278 | gimple_call_set_lhs (call, half_res); | |
3279 | gimple_call_set_nothrow (call, true); | |
3280 | new_stmt = call; | |
b1b6836e RS |
3281 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3282 | if ((i & 1) == 0) | |
3283 | { | |
3284 | prev_res = half_res; | |
3285 | continue; | |
3286 | } | |
3287 | new_temp = make_ssa_name (vec_dest); | |
3288 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
3289 | prev_res, half_res); | |
3290 | } | |
70439f0d | 3291 | else |
b1b6836e | 3292 | { |
a844293d | 3293 | gcall *call; |
b1b6836e | 3294 | if (ifn != IFN_LAST) |
a844293d | 3295 | call = gimple_build_call_internal_vec (ifn, vargs); |
b1b6836e | 3296 | else |
a844293d RS |
3297 | call = gimple_build_call_vec (fndecl, vargs); |
3298 | new_temp = make_ssa_name (vec_dest, call); | |
3299 | gimple_call_set_lhs (call, new_temp); | |
3300 | gimple_call_set_nothrow (call, true); | |
3301 | new_stmt = call; | |
b1b6836e | 3302 | } |
190c2236 | 3303 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 3304 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
3305 | } |
3306 | ||
3307 | for (i = 0; i < nargs; i++) | |
3308 | { | |
37b5ec8f | 3309 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 3310 | vec_oprndsi.release (); |
190c2236 | 3311 | } |
190c2236 JJ |
3312 | continue; |
3313 | } | |
3314 | ||
ebfd146a IR |
3315 | for (i = 0; i < nargs; i++) |
3316 | { | |
3317 | op = gimple_call_arg (stmt, i); | |
3318 | if (j == 0) | |
3319 | vec_oprnd0 | |
81c40241 | 3320 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 3321 | else |
63827fb8 IR |
3322 | { |
3323 | vec_oprnd0 = gimple_call_arg (new_stmt, i); | |
3324 | vec_oprnd0 | |
3325 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); | |
3326 | } | |
ebfd146a | 3327 | |
9771b263 | 3328 | vargs.quick_push (vec_oprnd0); |
ebfd146a IR |
3329 | } |
3330 | ||
74bf76ed JJ |
3331 | if (gimple_call_internal_p (stmt) |
3332 | && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) | |
3333 | { | |
c7bda0f4 | 3334 | tree cst = build_index_vector (vectype_out, j * nunits_out, 1); |
74bf76ed | 3335 | tree new_var |
0e22bb5a | 3336 | = vect_get_new_ssa_name (vectype_out, vect_simple_var, "cst_"); |
355fe088 | 3337 | gimple *init_stmt = gimple_build_assign (new_var, cst); |
74bf76ed | 3338 | vect_init_vector_1 (stmt, init_stmt, NULL); |
b731b390 | 3339 | new_temp = make_ssa_name (vec_dest); |
0e22bb5a | 3340 | new_stmt = gimple_build_assign (new_temp, new_var); |
74bf76ed | 3341 | } |
b1b6836e RS |
3342 | else if (modifier == NARROW) |
3343 | { | |
3344 | tree half_res = make_ssa_name (vectype_in); | |
a844293d RS |
3345 | gcall *call = gimple_build_call_internal_vec (ifn, vargs); |
3346 | gimple_call_set_lhs (call, half_res); | |
3347 | gimple_call_set_nothrow (call, true); | |
3348 | new_stmt = call; | |
b1b6836e RS |
3349 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3350 | if ((j & 1) == 0) | |
3351 | { | |
3352 | prev_res = half_res; | |
3353 | continue; | |
3354 | } | |
3355 | new_temp = make_ssa_name (vec_dest); | |
3356 | new_stmt = gimple_build_assign (new_temp, convert_code, | |
3357 | prev_res, half_res); | |
3358 | } | |
74bf76ed JJ |
3359 | else |
3360 | { | |
a844293d | 3361 | gcall *call; |
70439f0d | 3362 | if (ifn != IFN_LAST) |
a844293d | 3363 | call = gimple_build_call_internal_vec (ifn, vargs); |
70439f0d | 3364 | else |
a844293d | 3365 | call = gimple_build_call_vec (fndecl, vargs); |
74bf76ed | 3366 | new_temp = make_ssa_name (vec_dest, new_stmt); |
a844293d RS |
3367 | gimple_call_set_lhs (call, new_temp); |
3368 | gimple_call_set_nothrow (call, true); | |
3369 | new_stmt = call; | |
74bf76ed | 3370 | } |
ebfd146a IR |
3371 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3372 | ||
b1b6836e | 3373 | if (j == (modifier == NARROW ? 1 : 0)) |
ebfd146a IR |
3374 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; |
3375 | else | |
3376 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3377 | ||
3378 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3379 | } | |
b1b6836e RS |
3380 | } |
3381 | else if (modifier == NARROW) | |
3382 | { | |
ebfd146a IR |
3383 | for (j = 0; j < ncopies; ++j) |
3384 | { | |
3385 | /* Build argument list for the vectorized call. */ | |
3386 | if (j == 0) | |
9771b263 | 3387 | vargs.create (nargs * 2); |
ebfd146a | 3388 | else |
9771b263 | 3389 | vargs.truncate (0); |
ebfd146a | 3390 | |
190c2236 JJ |
3391 | if (slp_node) |
3392 | { | |
ef062b13 | 3393 | auto_vec<vec<tree> > vec_defs (nargs); |
9771b263 | 3394 | vec<tree> vec_oprnds0; |
190c2236 JJ |
3395 | |
3396 | for (i = 0; i < nargs; i++) | |
9771b263 | 3397 | vargs.quick_push (gimple_call_arg (stmt, i)); |
306b0c92 | 3398 | vect_get_slp_defs (vargs, slp_node, &vec_defs); |
37b5ec8f | 3399 | vec_oprnds0 = vec_defs[0]; |
190c2236 JJ |
3400 | |
3401 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 3402 | for (i = 0; vec_oprnds0.iterate (i, &vec_oprnd0); i += 2) |
190c2236 JJ |
3403 | { |
3404 | size_t k; | |
9771b263 | 3405 | vargs.truncate (0); |
190c2236 JJ |
3406 | for (k = 0; k < nargs; k++) |
3407 | { | |
37b5ec8f | 3408 | vec<tree> vec_oprndsk = vec_defs[k]; |
9771b263 DN |
3409 | vargs.quick_push (vec_oprndsk[i]); |
3410 | vargs.quick_push (vec_oprndsk[i + 1]); | |
190c2236 | 3411 | } |
a844293d | 3412 | gcall *call; |
70439f0d | 3413 | if (ifn != IFN_LAST) |
a844293d | 3414 | call = gimple_build_call_internal_vec (ifn, vargs); |
70439f0d | 3415 | else |
a844293d RS |
3416 | call = gimple_build_call_vec (fndecl, vargs); |
3417 | new_temp = make_ssa_name (vec_dest, call); | |
3418 | gimple_call_set_lhs (call, new_temp); | |
3419 | gimple_call_set_nothrow (call, true); | |
3420 | new_stmt = call; | |
190c2236 | 3421 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 3422 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
190c2236 JJ |
3423 | } |
3424 | ||
3425 | for (i = 0; i < nargs; i++) | |
3426 | { | |
37b5ec8f | 3427 | vec<tree> vec_oprndsi = vec_defs[i]; |
9771b263 | 3428 | vec_oprndsi.release (); |
190c2236 | 3429 | } |
190c2236 JJ |
3430 | continue; |
3431 | } | |
3432 | ||
ebfd146a IR |
3433 | for (i = 0; i < nargs; i++) |
3434 | { | |
3435 | op = gimple_call_arg (stmt, i); | |
3436 | if (j == 0) | |
3437 | { | |
3438 | vec_oprnd0 | |
81c40241 | 3439 | = vect_get_vec_def_for_operand (op, stmt); |
ebfd146a | 3440 | vec_oprnd1 |
63827fb8 | 3441 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
3442 | } |
3443 | else | |
3444 | { | |
336ecb65 | 3445 | vec_oprnd1 = gimple_call_arg (new_stmt, 2*i + 1); |
ebfd146a | 3446 | vec_oprnd0 |
63827fb8 | 3447 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd1); |
ebfd146a | 3448 | vec_oprnd1 |
63827fb8 | 3449 | = vect_get_vec_def_for_stmt_copy (dt[i], vec_oprnd0); |
ebfd146a IR |
3450 | } |
3451 | ||
9771b263 DN |
3452 | vargs.quick_push (vec_oprnd0); |
3453 | vargs.quick_push (vec_oprnd1); | |
ebfd146a IR |
3454 | } |
3455 | ||
b1b6836e | 3456 | new_stmt = gimple_build_call_vec (fndecl, vargs); |
ebfd146a IR |
3457 | new_temp = make_ssa_name (vec_dest, new_stmt); |
3458 | gimple_call_set_lhs (new_stmt, new_temp); | |
ebfd146a IR |
3459 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3460 | ||
3461 | if (j == 0) | |
3462 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
3463 | else | |
3464 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
3465 | ||
3466 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
3467 | } | |
3468 | ||
3469 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a | 3470 | } |
b1b6836e RS |
3471 | else |
3472 | /* No current target implements this case. */ | |
3473 | return false; | |
ebfd146a | 3474 | |
9771b263 | 3475 | vargs.release (); |
ebfd146a | 3476 | |
ebfd146a IR |
3477 | /* The call in STMT might prevent it from being removed in dce. |
3478 | We however cannot remove it here, due to the way the ssa name | |
3479 | it defines is mapped to the new definition. So just replace | |
3480 | rhs of the statement with something harmless. */ | |
3481 | ||
dd34c087 JJ |
3482 | if (slp_node) |
3483 | return true; | |
3484 | ||
ebfd146a | 3485 | type = TREE_TYPE (scalar_dest); |
9d5e7640 IR |
3486 | if (is_pattern_stmt_p (stmt_info)) |
3487 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
3488 | else | |
3489 | lhs = gimple_call_lhs (stmt); | |
3cc2fa2a | 3490 | |
9d5e7640 | 3491 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); |
ebfd146a | 3492 | set_vinfo_for_stmt (new_stmt, stmt_info); |
dd34c087 | 3493 | set_vinfo_for_stmt (stmt, NULL); |
ebfd146a IR |
3494 | STMT_VINFO_STMT (stmt_info) = new_stmt; |
3495 | gsi_replace (gsi, new_stmt, false); | |
ebfd146a IR |
3496 | |
3497 | return true; | |
3498 | } | |
3499 | ||
3500 | ||
0136f8f0 AH |
3501 | struct simd_call_arg_info |
3502 | { | |
3503 | tree vectype; | |
3504 | tree op; | |
0136f8f0 | 3505 | HOST_WIDE_INT linear_step; |
34e82342 | 3506 | enum vect_def_type dt; |
0136f8f0 | 3507 | unsigned int align; |
17b658af | 3508 | bool simd_lane_linear; |
0136f8f0 AH |
3509 | }; |
3510 | ||
17b658af JJ |
3511 | /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME, |
3512 | is linear within simd lane (but not within whole loop), note it in | |
3513 | *ARGINFO. */ | |
3514 | ||
3515 | static void | |
3516 | vect_simd_lane_linear (tree op, struct loop *loop, | |
3517 | struct simd_call_arg_info *arginfo) | |
3518 | { | |
355fe088 | 3519 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
17b658af JJ |
3520 | |
3521 | if (!is_gimple_assign (def_stmt) | |
3522 | || gimple_assign_rhs_code (def_stmt) != POINTER_PLUS_EXPR | |
3523 | || !is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))) | |
3524 | return; | |
3525 | ||
3526 | tree base = gimple_assign_rhs1 (def_stmt); | |
3527 | HOST_WIDE_INT linear_step = 0; | |
3528 | tree v = gimple_assign_rhs2 (def_stmt); | |
3529 | while (TREE_CODE (v) == SSA_NAME) | |
3530 | { | |
3531 | tree t; | |
3532 | def_stmt = SSA_NAME_DEF_STMT (v); | |
3533 | if (is_gimple_assign (def_stmt)) | |
3534 | switch (gimple_assign_rhs_code (def_stmt)) | |
3535 | { | |
3536 | case PLUS_EXPR: | |
3537 | t = gimple_assign_rhs2 (def_stmt); | |
3538 | if (linear_step || TREE_CODE (t) != INTEGER_CST) | |
3539 | return; | |
3540 | base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base), base, t); | |
3541 | v = gimple_assign_rhs1 (def_stmt); | |
3542 | continue; | |
3543 | case MULT_EXPR: | |
3544 | t = gimple_assign_rhs2 (def_stmt); | |
3545 | if (linear_step || !tree_fits_shwi_p (t) || integer_zerop (t)) | |
3546 | return; | |
3547 | linear_step = tree_to_shwi (t); | |
3548 | v = gimple_assign_rhs1 (def_stmt); | |
3549 | continue; | |
3550 | CASE_CONVERT: | |
3551 | t = gimple_assign_rhs1 (def_stmt); | |
3552 | if (TREE_CODE (TREE_TYPE (t)) != INTEGER_TYPE | |
3553 | || (TYPE_PRECISION (TREE_TYPE (v)) | |
3554 | < TYPE_PRECISION (TREE_TYPE (t)))) | |
3555 | return; | |
3556 | if (!linear_step) | |
3557 | linear_step = 1; | |
3558 | v = t; | |
3559 | continue; | |
3560 | default: | |
3561 | return; | |
3562 | } | |
8e4284d0 | 3563 | else if (gimple_call_internal_p (def_stmt, IFN_GOMP_SIMD_LANE) |
17b658af JJ |
3564 | && loop->simduid |
3565 | && TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME | |
3566 | && (SSA_NAME_VAR (gimple_call_arg (def_stmt, 0)) | |
3567 | == loop->simduid)) | |
3568 | { | |
3569 | if (!linear_step) | |
3570 | linear_step = 1; | |
3571 | arginfo->linear_step = linear_step; | |
3572 | arginfo->op = base; | |
3573 | arginfo->simd_lane_linear = true; | |
3574 | return; | |
3575 | } | |
3576 | } | |
3577 | } | |
3578 | ||
cf1b2ba4 RS |
3579 | /* Return the number of elements in vector type VECTYPE, which is associated |
3580 | with a SIMD clone. At present these vectors always have a constant | |
3581 | length. */ | |
3582 | ||
3583 | static unsigned HOST_WIDE_INT | |
3584 | simd_clone_subparts (tree vectype) | |
3585 | { | |
928686b1 | 3586 | return TYPE_VECTOR_SUBPARTS (vectype).to_constant (); |
cf1b2ba4 RS |
3587 | } |
3588 | ||
0136f8f0 AH |
3589 | /* Function vectorizable_simd_clone_call. |
3590 | ||
3591 | Check if STMT performs a function call that can be vectorized | |
3592 | by calling a simd clone of the function. | |
3593 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
3594 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
3595 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
3596 | ||
3597 | static bool | |
355fe088 TS |
3598 | vectorizable_simd_clone_call (gimple *stmt, gimple_stmt_iterator *gsi, |
3599 | gimple **vec_stmt, slp_tree slp_node) | |
0136f8f0 AH |
3600 | { |
3601 | tree vec_dest; | |
3602 | tree scalar_dest; | |
3603 | tree op, type; | |
3604 | tree vec_oprnd0 = NULL_TREE; | |
3605 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info; | |
3606 | tree vectype; | |
3607 | unsigned int nunits; | |
3608 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
3609 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
310213d4 | 3610 | vec_info *vinfo = stmt_info->vinfo; |
0136f8f0 | 3611 | struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; |
81c40241 | 3612 | tree fndecl, new_temp; |
355fe088 TS |
3613 | gimple *def_stmt; |
3614 | gimple *new_stmt = NULL; | |
0136f8f0 | 3615 | int ncopies, j; |
00426f9a | 3616 | auto_vec<simd_call_arg_info> arginfo; |
0136f8f0 AH |
3617 | vec<tree> vargs = vNULL; |
3618 | size_t i, nargs; | |
3619 | tree lhs, rtype, ratype; | |
e7a74006 | 3620 | vec<constructor_elt, va_gc> *ret_ctor_elts = NULL; |
0136f8f0 AH |
3621 | |
3622 | /* Is STMT a vectorizable call? */ | |
3623 | if (!is_gimple_call (stmt)) | |
3624 | return false; | |
3625 | ||
3626 | fndecl = gimple_call_fndecl (stmt); | |
3627 | if (fndecl == NULL_TREE) | |
3628 | return false; | |
3629 | ||
d52f5295 | 3630 | struct cgraph_node *node = cgraph_node::get (fndecl); |
0136f8f0 AH |
3631 | if (node == NULL || node->simd_clones == NULL) |
3632 | return false; | |
3633 | ||
3634 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
3635 | return false; | |
3636 | ||
66c16fd9 RB |
3637 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
3638 | && ! vec_stmt) | |
0136f8f0 AH |
3639 | return false; |
3640 | ||
3641 | if (gimple_call_lhs (stmt) | |
3642 | && TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME) | |
3643 | return false; | |
3644 | ||
3645 | gcc_checking_assert (!stmt_can_throw_internal (stmt)); | |
3646 | ||
3647 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3648 | ||
3649 | if (loop_vinfo && nested_in_vect_loop_p (loop, stmt)) | |
3650 | return false; | |
3651 | ||
3652 | /* FORNOW */ | |
fce57248 | 3653 | if (slp_node) |
0136f8f0 AH |
3654 | return false; |
3655 | ||
3656 | /* Process function arguments. */ | |
3657 | nargs = gimple_call_num_args (stmt); | |
3658 | ||
3659 | /* Bail out if the function has zero arguments. */ | |
3660 | if (nargs == 0) | |
3661 | return false; | |
3662 | ||
00426f9a | 3663 | arginfo.reserve (nargs, true); |
0136f8f0 AH |
3664 | |
3665 | for (i = 0; i < nargs; i++) | |
3666 | { | |
3667 | simd_call_arg_info thisarginfo; | |
3668 | affine_iv iv; | |
3669 | ||
3670 | thisarginfo.linear_step = 0; | |
3671 | thisarginfo.align = 0; | |
3672 | thisarginfo.op = NULL_TREE; | |
17b658af | 3673 | thisarginfo.simd_lane_linear = false; |
0136f8f0 AH |
3674 | |
3675 | op = gimple_call_arg (stmt, i); | |
81c40241 RB |
3676 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &thisarginfo.dt, |
3677 | &thisarginfo.vectype) | |
0136f8f0 AH |
3678 | || thisarginfo.dt == vect_uninitialized_def) |
3679 | { | |
3680 | if (dump_enabled_p ()) | |
3681 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3682 | "use not simple.\n"); | |
0136f8f0 AH |
3683 | return false; |
3684 | } | |
3685 | ||
3686 | if (thisarginfo.dt == vect_constant_def | |
3687 | || thisarginfo.dt == vect_external_def) | |
3688 | gcc_assert (thisarginfo.vectype == NULL_TREE); | |
3689 | else | |
3690 | gcc_assert (thisarginfo.vectype != NULL_TREE); | |
3691 | ||
6c9e85fb JJ |
3692 | /* For linear arguments, the analyze phase should have saved |
3693 | the base and step in STMT_VINFO_SIMD_CLONE_INFO. */ | |
17b658af JJ |
3694 | if (i * 3 + 4 <= STMT_VINFO_SIMD_CLONE_INFO (stmt_info).length () |
3695 | && STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]) | |
6c9e85fb JJ |
3696 | { |
3697 | gcc_assert (vec_stmt); | |
3698 | thisarginfo.linear_step | |
17b658af | 3699 | = tree_to_shwi (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]); |
6c9e85fb | 3700 | thisarginfo.op |
17b658af JJ |
3701 | = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 1]; |
3702 | thisarginfo.simd_lane_linear | |
3703 | = (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 3] | |
3704 | == boolean_true_node); | |
6c9e85fb JJ |
3705 | /* If loop has been peeled for alignment, we need to adjust it. */ |
3706 | tree n1 = LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo); | |
3707 | tree n2 = LOOP_VINFO_NITERS (loop_vinfo); | |
17b658af | 3708 | if (n1 != n2 && !thisarginfo.simd_lane_linear) |
6c9e85fb JJ |
3709 | { |
3710 | tree bias = fold_build2 (MINUS_EXPR, TREE_TYPE (n1), n1, n2); | |
17b658af | 3711 | tree step = STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[i * 3 + 2]; |
6c9e85fb JJ |
3712 | tree opt = TREE_TYPE (thisarginfo.op); |
3713 | bias = fold_convert (TREE_TYPE (step), bias); | |
3714 | bias = fold_build2 (MULT_EXPR, TREE_TYPE (step), bias, step); | |
3715 | thisarginfo.op | |
3716 | = fold_build2 (POINTER_TYPE_P (opt) | |
3717 | ? POINTER_PLUS_EXPR : PLUS_EXPR, opt, | |
3718 | thisarginfo.op, bias); | |
3719 | } | |
3720 | } | |
3721 | else if (!vec_stmt | |
3722 | && thisarginfo.dt != vect_constant_def | |
3723 | && thisarginfo.dt != vect_external_def | |
3724 | && loop_vinfo | |
3725 | && TREE_CODE (op) == SSA_NAME | |
3726 | && simple_iv (loop, loop_containing_stmt (stmt), op, | |
3727 | &iv, false) | |
3728 | && tree_fits_shwi_p (iv.step)) | |
0136f8f0 AH |
3729 | { |
3730 | thisarginfo.linear_step = tree_to_shwi (iv.step); | |
3731 | thisarginfo.op = iv.base; | |
3732 | } | |
3733 | else if ((thisarginfo.dt == vect_constant_def | |
3734 | || thisarginfo.dt == vect_external_def) | |
3735 | && POINTER_TYPE_P (TREE_TYPE (op))) | |
3736 | thisarginfo.align = get_pointer_alignment (op) / BITS_PER_UNIT; | |
17b658af JJ |
3737 | /* Addresses of array elements indexed by GOMP_SIMD_LANE are |
3738 | linear too. */ | |
3739 | if (POINTER_TYPE_P (TREE_TYPE (op)) | |
3740 | && !thisarginfo.linear_step | |
3741 | && !vec_stmt | |
3742 | && thisarginfo.dt != vect_constant_def | |
3743 | && thisarginfo.dt != vect_external_def | |
3744 | && loop_vinfo | |
3745 | && !slp_node | |
3746 | && TREE_CODE (op) == SSA_NAME) | |
3747 | vect_simd_lane_linear (op, loop, &thisarginfo); | |
0136f8f0 AH |
3748 | |
3749 | arginfo.quick_push (thisarginfo); | |
3750 | } | |
3751 | ||
d9f21f6a RS |
3752 | unsigned HOST_WIDE_INT vf; |
3753 | if (!LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (&vf)) | |
3754 | { | |
3755 | if (dump_enabled_p ()) | |
3756 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
3757 | "not considering SIMD clones; not yet supported" | |
3758 | " for variable-width vectors.\n"); | |
3759 | return NULL; | |
3760 | } | |
3761 | ||
0136f8f0 AH |
3762 | unsigned int badness = 0; |
3763 | struct cgraph_node *bestn = NULL; | |
6c9e85fb JJ |
3764 | if (STMT_VINFO_SIMD_CLONE_INFO (stmt_info).exists ()) |
3765 | bestn = cgraph_node::get (STMT_VINFO_SIMD_CLONE_INFO (stmt_info)[0]); | |
0136f8f0 AH |
3766 | else |
3767 | for (struct cgraph_node *n = node->simd_clones; n != NULL; | |
3768 | n = n->simdclone->next_clone) | |
3769 | { | |
3770 | unsigned int this_badness = 0; | |
d9f21f6a | 3771 | if (n->simdclone->simdlen > vf |
0136f8f0 AH |
3772 | || n->simdclone->nargs != nargs) |
3773 | continue; | |
d9f21f6a RS |
3774 | if (n->simdclone->simdlen < vf) |
3775 | this_badness += (exact_log2 (vf) | |
0136f8f0 AH |
3776 | - exact_log2 (n->simdclone->simdlen)) * 1024; |
3777 | if (n->simdclone->inbranch) | |
3778 | this_badness += 2048; | |
3779 | int target_badness = targetm.simd_clone.usable (n); | |
3780 | if (target_badness < 0) | |
3781 | continue; | |
3782 | this_badness += target_badness * 512; | |
3783 | /* FORNOW: Have to add code to add the mask argument. */ | |
3784 | if (n->simdclone->inbranch) | |
3785 | continue; | |
3786 | for (i = 0; i < nargs; i++) | |
3787 | { | |
3788 | switch (n->simdclone->args[i].arg_type) | |
3789 | { | |
3790 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3791 | if (!useless_type_conversion_p | |
3792 | (n->simdclone->args[i].orig_type, | |
3793 | TREE_TYPE (gimple_call_arg (stmt, i)))) | |
3794 | i = -1; | |
3795 | else if (arginfo[i].dt == vect_constant_def | |
3796 | || arginfo[i].dt == vect_external_def | |
3797 | || arginfo[i].linear_step) | |
3798 | this_badness += 64; | |
3799 | break; | |
3800 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
3801 | if (arginfo[i].dt != vect_constant_def | |
3802 | && arginfo[i].dt != vect_external_def) | |
3803 | i = -1; | |
3804 | break; | |
3805 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
d9a6bd32 | 3806 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
3807 | if (arginfo[i].dt == vect_constant_def |
3808 | || arginfo[i].dt == vect_external_def | |
3809 | || (arginfo[i].linear_step | |
3810 | != n->simdclone->args[i].linear_step)) | |
3811 | i = -1; | |
3812 | break; | |
3813 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
d9a6bd32 JJ |
3814 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
3815 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
e01d41e5 JJ |
3816 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
3817 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
3818 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
3819 | /* FORNOW */ |
3820 | i = -1; | |
3821 | break; | |
3822 | case SIMD_CLONE_ARG_TYPE_MASK: | |
3823 | gcc_unreachable (); | |
3824 | } | |
3825 | if (i == (size_t) -1) | |
3826 | break; | |
3827 | if (n->simdclone->args[i].alignment > arginfo[i].align) | |
3828 | { | |
3829 | i = -1; | |
3830 | break; | |
3831 | } | |
3832 | if (arginfo[i].align) | |
3833 | this_badness += (exact_log2 (arginfo[i].align) | |
3834 | - exact_log2 (n->simdclone->args[i].alignment)); | |
3835 | } | |
3836 | if (i == (size_t) -1) | |
3837 | continue; | |
3838 | if (bestn == NULL || this_badness < badness) | |
3839 | { | |
3840 | bestn = n; | |
3841 | badness = this_badness; | |
3842 | } | |
3843 | } | |
3844 | ||
3845 | if (bestn == NULL) | |
00426f9a | 3846 | return false; |
0136f8f0 AH |
3847 | |
3848 | for (i = 0; i < nargs; i++) | |
3849 | if ((arginfo[i].dt == vect_constant_def | |
3850 | || arginfo[i].dt == vect_external_def) | |
3851 | && bestn->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
3852 | { | |
3853 | arginfo[i].vectype | |
3854 | = get_vectype_for_scalar_type (TREE_TYPE (gimple_call_arg (stmt, | |
3855 | i))); | |
3856 | if (arginfo[i].vectype == NULL | |
cf1b2ba4 | 3857 | || (simd_clone_subparts (arginfo[i].vectype) |
0136f8f0 | 3858 | > bestn->simdclone->simdlen)) |
00426f9a | 3859 | return false; |
0136f8f0 AH |
3860 | } |
3861 | ||
3862 | fndecl = bestn->decl; | |
3863 | nunits = bestn->simdclone->simdlen; | |
d9f21f6a | 3864 | ncopies = vf / nunits; |
0136f8f0 AH |
3865 | |
3866 | /* If the function isn't const, only allow it in simd loops where user | |
3867 | has asserted that at least nunits consecutive iterations can be | |
3868 | performed using SIMD instructions. */ | |
3869 | if ((loop == NULL || (unsigned) loop->safelen < nunits) | |
3870 | && gimple_vuse (stmt)) | |
00426f9a | 3871 | return false; |
0136f8f0 AH |
3872 | |
3873 | /* Sanity check: make sure that at least one copy of the vectorized stmt | |
3874 | needs to be generated. */ | |
3875 | gcc_assert (ncopies >= 1); | |
3876 | ||
3877 | if (!vec_stmt) /* transformation not required. */ | |
3878 | { | |
6c9e85fb JJ |
3879 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (bestn->decl); |
3880 | for (i = 0; i < nargs; i++) | |
7adb26f2 JJ |
3881 | if ((bestn->simdclone->args[i].arg_type |
3882 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
3883 | || (bestn->simdclone->args[i].arg_type | |
3884 | == SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP)) | |
6c9e85fb | 3885 | { |
17b658af | 3886 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_grow_cleared (i * 3 |
6c9e85fb JJ |
3887 | + 1); |
3888 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (arginfo[i].op); | |
3889 | tree lst = POINTER_TYPE_P (TREE_TYPE (arginfo[i].op)) | |
3890 | ? size_type_node : TREE_TYPE (arginfo[i].op); | |
3891 | tree ls = build_int_cst (lst, arginfo[i].linear_step); | |
3892 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (ls); | |
17b658af JJ |
3893 | tree sll = arginfo[i].simd_lane_linear |
3894 | ? boolean_true_node : boolean_false_node; | |
3895 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).safe_push (sll); | |
6c9e85fb | 3896 | } |
0136f8f0 AH |
3897 | STMT_VINFO_TYPE (stmt_info) = call_simd_clone_vec_info_type; |
3898 | if (dump_enabled_p ()) | |
3899 | dump_printf_loc (MSG_NOTE, vect_location, | |
3900 | "=== vectorizable_simd_clone_call ===\n"); | |
3901 | /* vect_model_simple_cost (stmt_info, ncopies, dt, NULL, NULL); */ | |
0136f8f0 AH |
3902 | return true; |
3903 | } | |
3904 | ||
67b8dbac | 3905 | /* Transform. */ |
0136f8f0 AH |
3906 | |
3907 | if (dump_enabled_p ()) | |
3908 | dump_printf_loc (MSG_NOTE, vect_location, "transform call.\n"); | |
3909 | ||
3910 | /* Handle def. */ | |
3911 | scalar_dest = gimple_call_lhs (stmt); | |
3912 | vec_dest = NULL_TREE; | |
3913 | rtype = NULL_TREE; | |
3914 | ratype = NULL_TREE; | |
3915 | if (scalar_dest) | |
3916 | { | |
3917 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3918 | rtype = TREE_TYPE (TREE_TYPE (fndecl)); | |
3919 | if (TREE_CODE (rtype) == ARRAY_TYPE) | |
3920 | { | |
3921 | ratype = rtype; | |
3922 | rtype = TREE_TYPE (ratype); | |
3923 | } | |
3924 | } | |
3925 | ||
3926 | prev_stmt_info = NULL; | |
3927 | for (j = 0; j < ncopies; ++j) | |
3928 | { | |
3929 | /* Build argument list for the vectorized call. */ | |
3930 | if (j == 0) | |
3931 | vargs.create (nargs); | |
3932 | else | |
3933 | vargs.truncate (0); | |
3934 | ||
3935 | for (i = 0; i < nargs; i++) | |
3936 | { | |
3937 | unsigned int k, l, m, o; | |
3938 | tree atype; | |
3939 | op = gimple_call_arg (stmt, i); | |
3940 | switch (bestn->simdclone->args[i].arg_type) | |
3941 | { | |
3942 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
3943 | atype = bestn->simdclone->args[i].vector_type; | |
cf1b2ba4 | 3944 | o = nunits / simd_clone_subparts (atype); |
0136f8f0 AH |
3945 | for (m = j * o; m < (j + 1) * o; m++) |
3946 | { | |
cf1b2ba4 RS |
3947 | if (simd_clone_subparts (atype) |
3948 | < simd_clone_subparts (arginfo[i].vectype)) | |
0136f8f0 | 3949 | { |
73a699ae | 3950 | poly_uint64 prec = GET_MODE_BITSIZE (TYPE_MODE (atype)); |
cf1b2ba4 RS |
3951 | k = (simd_clone_subparts (arginfo[i].vectype) |
3952 | / simd_clone_subparts (atype)); | |
0136f8f0 AH |
3953 | gcc_assert ((k & (k - 1)) == 0); |
3954 | if (m == 0) | |
3955 | vec_oprnd0 | |
81c40241 | 3956 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3957 | else |
3958 | { | |
3959 | vec_oprnd0 = arginfo[i].op; | |
3960 | if ((m & (k - 1)) == 0) | |
3961 | vec_oprnd0 | |
3962 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3963 | vec_oprnd0); | |
3964 | } | |
3965 | arginfo[i].op = vec_oprnd0; | |
3966 | vec_oprnd0 | |
3967 | = build3 (BIT_FIELD_REF, atype, vec_oprnd0, | |
92e29a5e | 3968 | bitsize_int (prec), |
0136f8f0 AH |
3969 | bitsize_int ((m & (k - 1)) * prec)); |
3970 | new_stmt | |
b731b390 | 3971 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
3972 | vec_oprnd0); |
3973 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
3974 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
3975 | } | |
3976 | else | |
3977 | { | |
cf1b2ba4 RS |
3978 | k = (simd_clone_subparts (atype) |
3979 | / simd_clone_subparts (arginfo[i].vectype)); | |
0136f8f0 AH |
3980 | gcc_assert ((k & (k - 1)) == 0); |
3981 | vec<constructor_elt, va_gc> *ctor_elts; | |
3982 | if (k != 1) | |
3983 | vec_alloc (ctor_elts, k); | |
3984 | else | |
3985 | ctor_elts = NULL; | |
3986 | for (l = 0; l < k; l++) | |
3987 | { | |
3988 | if (m == 0 && l == 0) | |
3989 | vec_oprnd0 | |
81c40241 | 3990 | = vect_get_vec_def_for_operand (op, stmt); |
0136f8f0 AH |
3991 | else |
3992 | vec_oprnd0 | |
3993 | = vect_get_vec_def_for_stmt_copy (arginfo[i].dt, | |
3994 | arginfo[i].op); | |
3995 | arginfo[i].op = vec_oprnd0; | |
3996 | if (k == 1) | |
3997 | break; | |
3998 | CONSTRUCTOR_APPEND_ELT (ctor_elts, NULL_TREE, | |
3999 | vec_oprnd0); | |
4000 | } | |
4001 | if (k == 1) | |
4002 | vargs.safe_push (vec_oprnd0); | |
4003 | else | |
4004 | { | |
4005 | vec_oprnd0 = build_constructor (atype, ctor_elts); | |
4006 | new_stmt | |
b731b390 | 4007 | = gimple_build_assign (make_ssa_name (atype), |
0136f8f0 AH |
4008 | vec_oprnd0); |
4009 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4010 | vargs.safe_push (gimple_assign_lhs (new_stmt)); | |
4011 | } | |
4012 | } | |
4013 | } | |
4014 | break; | |
4015 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
4016 | vargs.safe_push (op); | |
4017 | break; | |
4018 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
7adb26f2 | 4019 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: |
0136f8f0 AH |
4020 | if (j == 0) |
4021 | { | |
4022 | gimple_seq stmts; | |
4023 | arginfo[i].op | |
4024 | = force_gimple_operand (arginfo[i].op, &stmts, true, | |
4025 | NULL_TREE); | |
4026 | if (stmts != NULL) | |
4027 | { | |
4028 | basic_block new_bb; | |
4029 | edge pe = loop_preheader_edge (loop); | |
4030 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
4031 | gcc_assert (!new_bb); | |
4032 | } | |
17b658af JJ |
4033 | if (arginfo[i].simd_lane_linear) |
4034 | { | |
4035 | vargs.safe_push (arginfo[i].op); | |
4036 | break; | |
4037 | } | |
b731b390 | 4038 | tree phi_res = copy_ssa_name (op); |
538dd0b7 | 4039 | gphi *new_phi = create_phi_node (phi_res, loop->header); |
0136f8f0 | 4040 | set_vinfo_for_stmt (new_phi, |
310213d4 | 4041 | new_stmt_vec_info (new_phi, loop_vinfo)); |
0136f8f0 AH |
4042 | add_phi_arg (new_phi, arginfo[i].op, |
4043 | loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
4044 | enum tree_code code | |
4045 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
4046 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
4047 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
4048 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
4049 | widest_int cst |
4050 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
4051 | ncopies * nunits); | |
4052 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 4053 | tree phi_arg = copy_ssa_name (op); |
0d0e4a03 JJ |
4054 | new_stmt |
4055 | = gimple_build_assign (phi_arg, code, phi_res, tcst); | |
0136f8f0 AH |
4056 | gimple_stmt_iterator si = gsi_after_labels (loop->header); |
4057 | gsi_insert_after (&si, new_stmt, GSI_NEW_STMT); | |
4058 | set_vinfo_for_stmt (new_stmt, | |
310213d4 | 4059 | new_stmt_vec_info (new_stmt, loop_vinfo)); |
0136f8f0 AH |
4060 | add_phi_arg (new_phi, phi_arg, loop_latch_edge (loop), |
4061 | UNKNOWN_LOCATION); | |
4062 | arginfo[i].op = phi_res; | |
4063 | vargs.safe_push (phi_res); | |
4064 | } | |
4065 | else | |
4066 | { | |
4067 | enum tree_code code | |
4068 | = POINTER_TYPE_P (TREE_TYPE (op)) | |
4069 | ? POINTER_PLUS_EXPR : PLUS_EXPR; | |
4070 | tree type = POINTER_TYPE_P (TREE_TYPE (op)) | |
4071 | ? sizetype : TREE_TYPE (op); | |
807e902e KZ |
4072 | widest_int cst |
4073 | = wi::mul (bestn->simdclone->args[i].linear_step, | |
4074 | j * nunits); | |
4075 | tree tcst = wide_int_to_tree (type, cst); | |
b731b390 | 4076 | new_temp = make_ssa_name (TREE_TYPE (op)); |
0d0e4a03 JJ |
4077 | new_stmt = gimple_build_assign (new_temp, code, |
4078 | arginfo[i].op, tcst); | |
0136f8f0 AH |
4079 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4080 | vargs.safe_push (new_temp); | |
4081 | } | |
4082 | break; | |
7adb26f2 JJ |
4083 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: |
4084 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
0136f8f0 | 4085 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: |
e01d41e5 JJ |
4086 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: |
4087 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
4088 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
0136f8f0 AH |
4089 | default: |
4090 | gcc_unreachable (); | |
4091 | } | |
4092 | } | |
4093 | ||
4094 | new_stmt = gimple_build_call_vec (fndecl, vargs); | |
4095 | if (vec_dest) | |
4096 | { | |
cf1b2ba4 | 4097 | gcc_assert (ratype || simd_clone_subparts (rtype) == nunits); |
0136f8f0 | 4098 | if (ratype) |
b731b390 | 4099 | new_temp = create_tmp_var (ratype); |
cf1b2ba4 RS |
4100 | else if (simd_clone_subparts (vectype) |
4101 | == simd_clone_subparts (rtype)) | |
0136f8f0 AH |
4102 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4103 | else | |
4104 | new_temp = make_ssa_name (rtype, new_stmt); | |
4105 | gimple_call_set_lhs (new_stmt, new_temp); | |
4106 | } | |
4107 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4108 | ||
4109 | if (vec_dest) | |
4110 | { | |
cf1b2ba4 | 4111 | if (simd_clone_subparts (vectype) < nunits) |
0136f8f0 AH |
4112 | { |
4113 | unsigned int k, l; | |
73a699ae RS |
4114 | poly_uint64 prec = GET_MODE_BITSIZE (TYPE_MODE (vectype)); |
4115 | poly_uint64 bytes = GET_MODE_SIZE (TYPE_MODE (vectype)); | |
cf1b2ba4 | 4116 | k = nunits / simd_clone_subparts (vectype); |
0136f8f0 AH |
4117 | gcc_assert ((k & (k - 1)) == 0); |
4118 | for (l = 0; l < k; l++) | |
4119 | { | |
4120 | tree t; | |
4121 | if (ratype) | |
4122 | { | |
4123 | t = build_fold_addr_expr (new_temp); | |
4124 | t = build2 (MEM_REF, vectype, t, | |
73a699ae | 4125 | build_int_cst (TREE_TYPE (t), l * bytes)); |
0136f8f0 AH |
4126 | } |
4127 | else | |
4128 | t = build3 (BIT_FIELD_REF, vectype, new_temp, | |
92e29a5e | 4129 | bitsize_int (prec), bitsize_int (l * prec)); |
0136f8f0 | 4130 | new_stmt |
b731b390 | 4131 | = gimple_build_assign (make_ssa_name (vectype), t); |
0136f8f0 AH |
4132 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4133 | if (j == 0 && l == 0) | |
4134 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4135 | else | |
4136 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4137 | ||
4138 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4139 | } | |
4140 | ||
4141 | if (ratype) | |
3ba4ff41 | 4142 | vect_clobber_variable (stmt, gsi, new_temp); |
0136f8f0 AH |
4143 | continue; |
4144 | } | |
cf1b2ba4 | 4145 | else if (simd_clone_subparts (vectype) > nunits) |
0136f8f0 | 4146 | { |
cf1b2ba4 RS |
4147 | unsigned int k = (simd_clone_subparts (vectype) |
4148 | / simd_clone_subparts (rtype)); | |
0136f8f0 AH |
4149 | gcc_assert ((k & (k - 1)) == 0); |
4150 | if ((j & (k - 1)) == 0) | |
4151 | vec_alloc (ret_ctor_elts, k); | |
4152 | if (ratype) | |
4153 | { | |
cf1b2ba4 | 4154 | unsigned int m, o = nunits / simd_clone_subparts (rtype); |
0136f8f0 AH |
4155 | for (m = 0; m < o; m++) |
4156 | { | |
4157 | tree tem = build4 (ARRAY_REF, rtype, new_temp, | |
4158 | size_int (m), NULL_TREE, NULL_TREE); | |
4159 | new_stmt | |
b731b390 | 4160 | = gimple_build_assign (make_ssa_name (rtype), tem); |
0136f8f0 AH |
4161 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4162 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, | |
4163 | gimple_assign_lhs (new_stmt)); | |
4164 | } | |
3ba4ff41 | 4165 | vect_clobber_variable (stmt, gsi, new_temp); |
0136f8f0 AH |
4166 | } |
4167 | else | |
4168 | CONSTRUCTOR_APPEND_ELT (ret_ctor_elts, NULL_TREE, new_temp); | |
4169 | if ((j & (k - 1)) != k - 1) | |
4170 | continue; | |
4171 | vec_oprnd0 = build_constructor (vectype, ret_ctor_elts); | |
4172 | new_stmt | |
b731b390 | 4173 | = gimple_build_assign (make_ssa_name (vec_dest), vec_oprnd0); |
0136f8f0 AH |
4174 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4175 | ||
4176 | if ((unsigned) j == k - 1) | |
4177 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4178 | else | |
4179 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4180 | ||
4181 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4182 | continue; | |
4183 | } | |
4184 | else if (ratype) | |
4185 | { | |
4186 | tree t = build_fold_addr_expr (new_temp); | |
4187 | t = build2 (MEM_REF, vectype, t, | |
4188 | build_int_cst (TREE_TYPE (t), 0)); | |
4189 | new_stmt | |
b731b390 | 4190 | = gimple_build_assign (make_ssa_name (vec_dest), t); |
0136f8f0 | 4191 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3ba4ff41 | 4192 | vect_clobber_variable (stmt, gsi, new_temp); |
0136f8f0 AH |
4193 | } |
4194 | } | |
4195 | ||
4196 | if (j == 0) | |
4197 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4198 | else | |
4199 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4200 | ||
4201 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4202 | } | |
4203 | ||
4204 | vargs.release (); | |
4205 | ||
4206 | /* The call in STMT might prevent it from being removed in dce. | |
4207 | We however cannot remove it here, due to the way the ssa name | |
4208 | it defines is mapped to the new definition. So just replace | |
4209 | rhs of the statement with something harmless. */ | |
4210 | ||
4211 | if (slp_node) | |
4212 | return true; | |
4213 | ||
4214 | if (scalar_dest) | |
4215 | { | |
4216 | type = TREE_TYPE (scalar_dest); | |
4217 | if (is_pattern_stmt_p (stmt_info)) | |
4218 | lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info)); | |
4219 | else | |
4220 | lhs = gimple_call_lhs (stmt); | |
4221 | new_stmt = gimple_build_assign (lhs, build_zero_cst (type)); | |
4222 | } | |
4223 | else | |
4224 | new_stmt = gimple_build_nop (); | |
4225 | set_vinfo_for_stmt (new_stmt, stmt_info); | |
4226 | set_vinfo_for_stmt (stmt, NULL); | |
4227 | STMT_VINFO_STMT (stmt_info) = new_stmt; | |
2865f32a | 4228 | gsi_replace (gsi, new_stmt, true); |
0136f8f0 AH |
4229 | unlink_stmt_vdef (stmt); |
4230 | ||
4231 | return true; | |
4232 | } | |
4233 | ||
4234 | ||
ebfd146a IR |
4235 | /* Function vect_gen_widened_results_half |
4236 | ||
4237 | Create a vector stmt whose code, type, number of arguments, and result | |
b8698a0f | 4238 | variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are |
ff802fa1 | 4239 | VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. |
ebfd146a IR |
4240 | In the case that CODE is a CALL_EXPR, this means that a call to DECL |
4241 | needs to be created (DECL is a function-decl of a target-builtin). | |
4242 | STMT is the original scalar stmt that we are vectorizing. */ | |
4243 | ||
355fe088 | 4244 | static gimple * |
ebfd146a IR |
4245 | vect_gen_widened_results_half (enum tree_code code, |
4246 | tree decl, | |
4247 | tree vec_oprnd0, tree vec_oprnd1, int op_type, | |
4248 | tree vec_dest, gimple_stmt_iterator *gsi, | |
355fe088 | 4249 | gimple *stmt) |
b8698a0f | 4250 | { |
355fe088 | 4251 | gimple *new_stmt; |
b8698a0f L |
4252 | tree new_temp; |
4253 | ||
4254 | /* Generate half of the widened result: */ | |
4255 | if (code == CALL_EXPR) | |
4256 | { | |
4257 | /* Target specific support */ | |
ebfd146a IR |
4258 | if (op_type == binary_op) |
4259 | new_stmt = gimple_build_call (decl, 2, vec_oprnd0, vec_oprnd1); | |
4260 | else | |
4261 | new_stmt = gimple_build_call (decl, 1, vec_oprnd0); | |
4262 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4263 | gimple_call_set_lhs (new_stmt, new_temp); | |
b8698a0f L |
4264 | } |
4265 | else | |
ebfd146a | 4266 | { |
b8698a0f L |
4267 | /* Generic support */ |
4268 | gcc_assert (op_type == TREE_CODE_LENGTH (code)); | |
ebfd146a IR |
4269 | if (op_type != binary_op) |
4270 | vec_oprnd1 = NULL; | |
0d0e4a03 | 4271 | new_stmt = gimple_build_assign (vec_dest, code, vec_oprnd0, vec_oprnd1); |
ebfd146a IR |
4272 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4273 | gimple_assign_set_lhs (new_stmt, new_temp); | |
b8698a0f | 4274 | } |
ebfd146a IR |
4275 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
4276 | ||
ebfd146a IR |
4277 | return new_stmt; |
4278 | } | |
4279 | ||
4a00c761 JJ |
4280 | |
4281 | /* Get vectorized definitions for loop-based vectorization. For the first | |
4282 | operand we call vect_get_vec_def_for_operand() (with OPRND containing | |
4283 | scalar operand), and for the rest we get a copy with | |
4284 | vect_get_vec_def_for_stmt_copy() using the previous vector definition | |
4285 | (stored in OPRND). See vect_get_vec_def_for_stmt_copy() for details. | |
4286 | The vectors are collected into VEC_OPRNDS. */ | |
4287 | ||
4288 | static void | |
355fe088 | 4289 | vect_get_loop_based_defs (tree *oprnd, gimple *stmt, enum vect_def_type dt, |
9771b263 | 4290 | vec<tree> *vec_oprnds, int multi_step_cvt) |
4a00c761 JJ |
4291 | { |
4292 | tree vec_oprnd; | |
4293 | ||
4294 | /* Get first vector operand. */ | |
4295 | /* All the vector operands except the very first one (that is scalar oprnd) | |
4296 | are stmt copies. */ | |
4297 | if (TREE_CODE (TREE_TYPE (*oprnd)) != VECTOR_TYPE) | |
81c40241 | 4298 | vec_oprnd = vect_get_vec_def_for_operand (*oprnd, stmt); |
4a00c761 JJ |
4299 | else |
4300 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, *oprnd); | |
4301 | ||
9771b263 | 4302 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
4303 | |
4304 | /* Get second vector operand. */ | |
4305 | vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd); | |
9771b263 | 4306 | vec_oprnds->quick_push (vec_oprnd); |
4a00c761 JJ |
4307 | |
4308 | *oprnd = vec_oprnd; | |
4309 | ||
4310 | /* For conversion in multiple steps, continue to get operands | |
4311 | recursively. */ | |
4312 | if (multi_step_cvt) | |
4313 | vect_get_loop_based_defs (oprnd, stmt, dt, vec_oprnds, multi_step_cvt - 1); | |
4314 | } | |
4315 | ||
4316 | ||
4317 | /* Create vectorized demotion statements for vector operands from VEC_OPRNDS. | |
4318 | For multi-step conversions store the resulting vectors and call the function | |
4319 | recursively. */ | |
4320 | ||
4321 | static void | |
9771b263 | 4322 | vect_create_vectorized_demotion_stmts (vec<tree> *vec_oprnds, |
355fe088 | 4323 | int multi_step_cvt, gimple *stmt, |
9771b263 | 4324 | vec<tree> vec_dsts, |
4a00c761 JJ |
4325 | gimple_stmt_iterator *gsi, |
4326 | slp_tree slp_node, enum tree_code code, | |
4327 | stmt_vec_info *prev_stmt_info) | |
4328 | { | |
4329 | unsigned int i; | |
4330 | tree vop0, vop1, new_tmp, vec_dest; | |
355fe088 | 4331 | gimple *new_stmt; |
4a00c761 JJ |
4332 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
4333 | ||
9771b263 | 4334 | vec_dest = vec_dsts.pop (); |
4a00c761 | 4335 | |
9771b263 | 4336 | for (i = 0; i < vec_oprnds->length (); i += 2) |
4a00c761 JJ |
4337 | { |
4338 | /* Create demotion operation. */ | |
9771b263 DN |
4339 | vop0 = (*vec_oprnds)[i]; |
4340 | vop1 = (*vec_oprnds)[i + 1]; | |
0d0e4a03 | 4341 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
4a00c761 JJ |
4342 | new_tmp = make_ssa_name (vec_dest, new_stmt); |
4343 | gimple_assign_set_lhs (new_stmt, new_tmp); | |
4344 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4345 | ||
4346 | if (multi_step_cvt) | |
4347 | /* Store the resulting vector for next recursive call. */ | |
9771b263 | 4348 | (*vec_oprnds)[i/2] = new_tmp; |
4a00c761 JJ |
4349 | else |
4350 | { | |
4351 | /* This is the last step of the conversion sequence. Store the | |
4352 | vectors in SLP_NODE or in vector info of the scalar statement | |
4353 | (or in STMT_VINFO_RELATED_STMT chain). */ | |
4354 | if (slp_node) | |
9771b263 | 4355 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 4356 | else |
c689ce1e RB |
4357 | { |
4358 | if (!*prev_stmt_info) | |
4359 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
4360 | else | |
4361 | STMT_VINFO_RELATED_STMT (*prev_stmt_info) = new_stmt; | |
4a00c761 | 4362 | |
c689ce1e RB |
4363 | *prev_stmt_info = vinfo_for_stmt (new_stmt); |
4364 | } | |
4a00c761 JJ |
4365 | } |
4366 | } | |
4367 | ||
4368 | /* For multi-step demotion operations we first generate demotion operations | |
4369 | from the source type to the intermediate types, and then combine the | |
4370 | results (stored in VEC_OPRNDS) in demotion operation to the destination | |
4371 | type. */ | |
4372 | if (multi_step_cvt) | |
4373 | { | |
4374 | /* At each level of recursion we have half of the operands we had at the | |
4375 | previous level. */ | |
9771b263 | 4376 | vec_oprnds->truncate ((i+1)/2); |
4a00c761 JJ |
4377 | vect_create_vectorized_demotion_stmts (vec_oprnds, multi_step_cvt - 1, |
4378 | stmt, vec_dsts, gsi, slp_node, | |
4379 | VEC_PACK_TRUNC_EXPR, | |
4380 | prev_stmt_info); | |
4381 | } | |
4382 | ||
9771b263 | 4383 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4384 | } |
4385 | ||
4386 | ||
4387 | /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0 | |
4388 | and VEC_OPRNDS1 (for binary operations). For multi-step conversions store | |
4389 | the resulting vectors and call the function recursively. */ | |
4390 | ||
4391 | static void | |
9771b263 DN |
4392 | vect_create_vectorized_promotion_stmts (vec<tree> *vec_oprnds0, |
4393 | vec<tree> *vec_oprnds1, | |
355fe088 | 4394 | gimple *stmt, tree vec_dest, |
4a00c761 JJ |
4395 | gimple_stmt_iterator *gsi, |
4396 | enum tree_code code1, | |
4397 | enum tree_code code2, tree decl1, | |
4398 | tree decl2, int op_type) | |
4399 | { | |
4400 | int i; | |
4401 | tree vop0, vop1, new_tmp1, new_tmp2; | |
355fe088 | 4402 | gimple *new_stmt1, *new_stmt2; |
6e1aa848 | 4403 | vec<tree> vec_tmp = vNULL; |
4a00c761 | 4404 | |
9771b263 DN |
4405 | vec_tmp.create (vec_oprnds0->length () * 2); |
4406 | FOR_EACH_VEC_ELT (*vec_oprnds0, i, vop0) | |
4a00c761 JJ |
4407 | { |
4408 | if (op_type == binary_op) | |
9771b263 | 4409 | vop1 = (*vec_oprnds1)[i]; |
4a00c761 JJ |
4410 | else |
4411 | vop1 = NULL_TREE; | |
4412 | ||
4413 | /* Generate the two halves of promotion operation. */ | |
4414 | new_stmt1 = vect_gen_widened_results_half (code1, decl1, vop0, vop1, | |
4415 | op_type, vec_dest, gsi, stmt); | |
4416 | new_stmt2 = vect_gen_widened_results_half (code2, decl2, vop0, vop1, | |
4417 | op_type, vec_dest, gsi, stmt); | |
4418 | if (is_gimple_call (new_stmt1)) | |
4419 | { | |
4420 | new_tmp1 = gimple_call_lhs (new_stmt1); | |
4421 | new_tmp2 = gimple_call_lhs (new_stmt2); | |
4422 | } | |
4423 | else | |
4424 | { | |
4425 | new_tmp1 = gimple_assign_lhs (new_stmt1); | |
4426 | new_tmp2 = gimple_assign_lhs (new_stmt2); | |
4427 | } | |
4428 | ||
4429 | /* Store the results for the next step. */ | |
9771b263 DN |
4430 | vec_tmp.quick_push (new_tmp1); |
4431 | vec_tmp.quick_push (new_tmp2); | |
4a00c761 JJ |
4432 | } |
4433 | ||
689eaba3 | 4434 | vec_oprnds0->release (); |
4a00c761 JJ |
4435 | *vec_oprnds0 = vec_tmp; |
4436 | } | |
4437 | ||
4438 | ||
b8698a0f L |
4439 | /* Check if STMT performs a conversion operation, that can be vectorized. |
4440 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
4a00c761 | 4441 | stmt to replace it, put it in VEC_STMT, and insert it at GSI. |
ebfd146a IR |
4442 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ |
4443 | ||
4444 | static bool | |
355fe088 TS |
4445 | vectorizable_conversion (gimple *stmt, gimple_stmt_iterator *gsi, |
4446 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
4447 | { |
4448 | tree vec_dest; | |
4449 | tree scalar_dest; | |
4a00c761 | 4450 | tree op0, op1 = NULL_TREE; |
ebfd146a IR |
4451 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; |
4452 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
4453 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4454 | enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK; | |
4a00c761 | 4455 | enum tree_code codecvt1 = ERROR_MARK, codecvt2 = ERROR_MARK; |
ebfd146a IR |
4456 | tree decl1 = NULL_TREE, decl2 = NULL_TREE; |
4457 | tree new_temp; | |
355fe088 | 4458 | gimple *def_stmt; |
ebfd146a | 4459 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 4460 | int ndts = 2; |
355fe088 | 4461 | gimple *new_stmt = NULL; |
ebfd146a | 4462 | stmt_vec_info prev_stmt_info; |
062d5ccc RS |
4463 | poly_uint64 nunits_in; |
4464 | poly_uint64 nunits_out; | |
ebfd146a | 4465 | tree vectype_out, vectype_in; |
4a00c761 JJ |
4466 | int ncopies, i, j; |
4467 | tree lhs_type, rhs_type; | |
ebfd146a | 4468 | enum { NARROW, NONE, WIDEN } modifier; |
6e1aa848 DN |
4469 | vec<tree> vec_oprnds0 = vNULL; |
4470 | vec<tree> vec_oprnds1 = vNULL; | |
ebfd146a | 4471 | tree vop0; |
4a00c761 | 4472 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 4473 | vec_info *vinfo = stmt_info->vinfo; |
4a00c761 | 4474 | int multi_step_cvt = 0; |
6e1aa848 | 4475 | vec<tree> interm_types = vNULL; |
4a00c761 JJ |
4476 | tree last_oprnd, intermediate_type, cvt_type = NULL_TREE; |
4477 | int op_type; | |
4a00c761 | 4478 | unsigned short fltsz; |
ebfd146a IR |
4479 | |
4480 | /* Is STMT a vectorizable conversion? */ | |
4481 | ||
4a00c761 | 4482 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
4483 | return false; |
4484 | ||
66c16fd9 RB |
4485 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
4486 | && ! vec_stmt) | |
ebfd146a IR |
4487 | return false; |
4488 | ||
4489 | if (!is_gimple_assign (stmt)) | |
4490 | return false; | |
4491 | ||
4492 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
4493 | return false; | |
4494 | ||
4495 | code = gimple_assign_rhs_code (stmt); | |
4a00c761 JJ |
4496 | if (!CONVERT_EXPR_CODE_P (code) |
4497 | && code != FIX_TRUNC_EXPR | |
4498 | && code != FLOAT_EXPR | |
4499 | && code != WIDEN_MULT_EXPR | |
4500 | && code != WIDEN_LSHIFT_EXPR) | |
ebfd146a IR |
4501 | return false; |
4502 | ||
4a00c761 JJ |
4503 | op_type = TREE_CODE_LENGTH (code); |
4504 | ||
ebfd146a | 4505 | /* Check types of lhs and rhs. */ |
b690cc0f | 4506 | scalar_dest = gimple_assign_lhs (stmt); |
4a00c761 | 4507 | lhs_type = TREE_TYPE (scalar_dest); |
b690cc0f RG |
4508 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); |
4509 | ||
ebfd146a IR |
4510 | op0 = gimple_assign_rhs1 (stmt); |
4511 | rhs_type = TREE_TYPE (op0); | |
4a00c761 JJ |
4512 | |
4513 | if ((code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4514 | && !((INTEGRAL_TYPE_P (lhs_type) | |
4515 | && INTEGRAL_TYPE_P (rhs_type)) | |
4516 | || (SCALAR_FLOAT_TYPE_P (lhs_type) | |
4517 | && SCALAR_FLOAT_TYPE_P (rhs_type)))) | |
4518 | return false; | |
4519 | ||
e6f5c25d IE |
4520 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4521 | && ((INTEGRAL_TYPE_P (lhs_type) | |
2be65d9e | 4522 | && !type_has_mode_precision_p (lhs_type)) |
e6f5c25d | 4523 | || (INTEGRAL_TYPE_P (rhs_type) |
2be65d9e | 4524 | && !type_has_mode_precision_p (rhs_type)))) |
4a00c761 | 4525 | { |
73fbfcad | 4526 | if (dump_enabled_p ()) |
78c60e3d | 4527 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 TJ |
4528 | "type conversion to/from bit-precision unsupported." |
4529 | "\n"); | |
4a00c761 JJ |
4530 | return false; |
4531 | } | |
4532 | ||
b690cc0f | 4533 | /* Check the operands of the operation. */ |
81c40241 | 4534 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype_in)) |
b690cc0f | 4535 | { |
73fbfcad | 4536 | if (dump_enabled_p ()) |
78c60e3d | 4537 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4538 | "use not simple.\n"); |
b690cc0f RG |
4539 | return false; |
4540 | } | |
4a00c761 JJ |
4541 | if (op_type == binary_op) |
4542 | { | |
4543 | bool ok; | |
4544 | ||
4545 | op1 = gimple_assign_rhs2 (stmt); | |
4546 | gcc_assert (code == WIDEN_MULT_EXPR || code == WIDEN_LSHIFT_EXPR); | |
4547 | /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of | |
4548 | OP1. */ | |
4549 | if (CONSTANT_CLASS_P (op0)) | |
81c40241 | 4550 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &vectype_in); |
4a00c761 | 4551 | else |
81c40241 | 4552 | ok = vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1]); |
4a00c761 JJ |
4553 | |
4554 | if (!ok) | |
4555 | { | |
73fbfcad | 4556 | if (dump_enabled_p ()) |
78c60e3d | 4557 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4558 | "use not simple.\n"); |
4a00c761 JJ |
4559 | return false; |
4560 | } | |
4561 | } | |
4562 | ||
b690cc0f RG |
4563 | /* If op0 is an external or constant defs use a vector type of |
4564 | the same size as the output vector type. */ | |
ebfd146a | 4565 | if (!vectype_in) |
b690cc0f | 4566 | vectype_in = get_same_sized_vectype (rhs_type, vectype_out); |
7d8930a0 IR |
4567 | if (vec_stmt) |
4568 | gcc_assert (vectype_in); | |
4569 | if (!vectype_in) | |
4570 | { | |
73fbfcad | 4571 | if (dump_enabled_p ()) |
4a00c761 | 4572 | { |
78c60e3d SS |
4573 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
4574 | "no vectype for scalar type "); | |
4575 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
e645e942 | 4576 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
4a00c761 | 4577 | } |
7d8930a0 IR |
4578 | |
4579 | return false; | |
4580 | } | |
ebfd146a | 4581 | |
e6f5c25d IE |
4582 | if (VECTOR_BOOLEAN_TYPE_P (vectype_out) |
4583 | && !VECTOR_BOOLEAN_TYPE_P (vectype_in)) | |
4584 | { | |
4585 | if (dump_enabled_p ()) | |
4586 | { | |
4587 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
4588 | "can't convert between boolean and non " | |
4589 | "boolean vectors"); | |
4590 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type); | |
4591 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); | |
4592 | } | |
4593 | ||
4594 | return false; | |
4595 | } | |
4596 | ||
b690cc0f RG |
4597 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); |
4598 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
062d5ccc | 4599 | if (known_eq (nunits_out, nunits_in)) |
ebfd146a | 4600 | modifier = NONE; |
062d5ccc RS |
4601 | else if (multiple_p (nunits_out, nunits_in)) |
4602 | modifier = NARROW; | |
ebfd146a | 4603 | else |
062d5ccc RS |
4604 | { |
4605 | gcc_checking_assert (multiple_p (nunits_in, nunits_out)); | |
4606 | modifier = WIDEN; | |
4607 | } | |
ebfd146a | 4608 | |
ff802fa1 IR |
4609 | /* Multiple types in SLP are handled by creating the appropriate number of |
4610 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
4611 | case of SLP. */ | |
fce57248 | 4612 | if (slp_node) |
ebfd146a | 4613 | ncopies = 1; |
4a00c761 | 4614 | else if (modifier == NARROW) |
e8f142e2 | 4615 | ncopies = vect_get_num_copies (loop_vinfo, vectype_out); |
4a00c761 | 4616 | else |
e8f142e2 | 4617 | ncopies = vect_get_num_copies (loop_vinfo, vectype_in); |
b8698a0f | 4618 | |
ebfd146a IR |
4619 | /* Sanity check: make sure that at least one copy of the vectorized stmt |
4620 | needs to be generated. */ | |
4621 | gcc_assert (ncopies >= 1); | |
4622 | ||
16d22000 RS |
4623 | bool found_mode = false; |
4624 | scalar_mode lhs_mode = SCALAR_TYPE_MODE (lhs_type); | |
4625 | scalar_mode rhs_mode = SCALAR_TYPE_MODE (rhs_type); | |
4626 | opt_scalar_mode rhs_mode_iter; | |
b397965c | 4627 | |
ebfd146a | 4628 | /* Supportable by target? */ |
4a00c761 | 4629 | switch (modifier) |
ebfd146a | 4630 | { |
4a00c761 JJ |
4631 | case NONE: |
4632 | if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR) | |
4633 | return false; | |
4634 | if (supportable_convert_operation (code, vectype_out, vectype_in, | |
4635 | &decl1, &code1)) | |
4636 | break; | |
4637 | /* FALLTHRU */ | |
4638 | unsupported: | |
73fbfcad | 4639 | if (dump_enabled_p ()) |
78c60e3d | 4640 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 4641 | "conversion not supported by target.\n"); |
ebfd146a | 4642 | return false; |
ebfd146a | 4643 | |
4a00c761 JJ |
4644 | case WIDEN: |
4645 | if (supportable_widening_operation (code, stmt, vectype_out, vectype_in, | |
a86ec597 RH |
4646 | &code1, &code2, &multi_step_cvt, |
4647 | &interm_types)) | |
4a00c761 JJ |
4648 | { |
4649 | /* Binary widening operation can only be supported directly by the | |
4650 | architecture. */ | |
4651 | gcc_assert (!(multi_step_cvt && op_type == binary_op)); | |
4652 | break; | |
4653 | } | |
4654 | ||
4655 | if (code != FLOAT_EXPR | |
b397965c | 4656 | || GET_MODE_SIZE (lhs_mode) <= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4657 | goto unsupported; |
4658 | ||
b397965c | 4659 | fltsz = GET_MODE_SIZE (lhs_mode); |
16d22000 | 4660 | FOR_EACH_2XWIDER_MODE (rhs_mode_iter, rhs_mode) |
4a00c761 | 4661 | { |
16d22000 | 4662 | rhs_mode = rhs_mode_iter.require (); |
c94843d2 RS |
4663 | if (GET_MODE_SIZE (rhs_mode) > fltsz) |
4664 | break; | |
4665 | ||
4a00c761 JJ |
4666 | cvt_type |
4667 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4668 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4669 | if (cvt_type == NULL_TREE) | |
4670 | goto unsupported; | |
4671 | ||
4672 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4673 | { | |
4674 | if (!supportable_convert_operation (code, vectype_out, | |
4675 | cvt_type, &decl1, &codecvt1)) | |
4676 | goto unsupported; | |
4677 | } | |
4678 | else if (!supportable_widening_operation (code, stmt, vectype_out, | |
a86ec597 RH |
4679 | cvt_type, &codecvt1, |
4680 | &codecvt2, &multi_step_cvt, | |
4a00c761 JJ |
4681 | &interm_types)) |
4682 | continue; | |
4683 | else | |
4684 | gcc_assert (multi_step_cvt == 0); | |
4685 | ||
4686 | if (supportable_widening_operation (NOP_EXPR, stmt, cvt_type, | |
a86ec597 RH |
4687 | vectype_in, &code1, &code2, |
4688 | &multi_step_cvt, &interm_types)) | |
16d22000 RS |
4689 | { |
4690 | found_mode = true; | |
4691 | break; | |
4692 | } | |
4a00c761 JJ |
4693 | } |
4694 | ||
16d22000 | 4695 | if (!found_mode) |
4a00c761 JJ |
4696 | goto unsupported; |
4697 | ||
4698 | if (GET_MODE_SIZE (rhs_mode) == fltsz) | |
4699 | codecvt2 = ERROR_MARK; | |
4700 | else | |
4701 | { | |
4702 | multi_step_cvt++; | |
9771b263 | 4703 | interm_types.safe_push (cvt_type); |
4a00c761 JJ |
4704 | cvt_type = NULL_TREE; |
4705 | } | |
4706 | break; | |
4707 | ||
4708 | case NARROW: | |
4709 | gcc_assert (op_type == unary_op); | |
4710 | if (supportable_narrowing_operation (code, vectype_out, vectype_in, | |
4711 | &code1, &multi_step_cvt, | |
4712 | &interm_types)) | |
4713 | break; | |
4714 | ||
4715 | if (code != FIX_TRUNC_EXPR | |
b397965c | 4716 | || GET_MODE_SIZE (lhs_mode) >= GET_MODE_SIZE (rhs_mode)) |
4a00c761 JJ |
4717 | goto unsupported; |
4718 | ||
4a00c761 JJ |
4719 | cvt_type |
4720 | = build_nonstandard_integer_type (GET_MODE_BITSIZE (rhs_mode), 0); | |
4721 | cvt_type = get_same_sized_vectype (cvt_type, vectype_in); | |
4722 | if (cvt_type == NULL_TREE) | |
4723 | goto unsupported; | |
4724 | if (!supportable_convert_operation (code, cvt_type, vectype_in, | |
4725 | &decl1, &codecvt1)) | |
4726 | goto unsupported; | |
4727 | if (supportable_narrowing_operation (NOP_EXPR, vectype_out, cvt_type, | |
4728 | &code1, &multi_step_cvt, | |
4729 | &interm_types)) | |
4730 | break; | |
4731 | goto unsupported; | |
4732 | ||
4733 | default: | |
4734 | gcc_unreachable (); | |
ebfd146a IR |
4735 | } |
4736 | ||
4737 | if (!vec_stmt) /* transformation not required. */ | |
4738 | { | |
73fbfcad | 4739 | if (dump_enabled_p ()) |
78c60e3d | 4740 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4741 | "=== vectorizable_conversion ===\n"); |
4a00c761 | 4742 | if (code == FIX_TRUNC_EXPR || code == FLOAT_EXPR) |
8bd37302 BS |
4743 | { |
4744 | STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type; | |
78604de0 RB |
4745 | if (!slp_node) |
4746 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); | |
8bd37302 | 4747 | } |
4a00c761 JJ |
4748 | else if (modifier == NARROW) |
4749 | { | |
4750 | STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; | |
78604de0 RB |
4751 | if (!slp_node) |
4752 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); | |
4a00c761 JJ |
4753 | } |
4754 | else | |
4755 | { | |
4756 | STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; | |
78604de0 RB |
4757 | if (!slp_node) |
4758 | vect_model_promotion_demotion_cost (stmt_info, dt, multi_step_cvt); | |
4a00c761 | 4759 | } |
9771b263 | 4760 | interm_types.release (); |
ebfd146a IR |
4761 | return true; |
4762 | } | |
4763 | ||
67b8dbac | 4764 | /* Transform. */ |
73fbfcad | 4765 | if (dump_enabled_p ()) |
78c60e3d | 4766 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 4767 | "transform conversion. ncopies = %d.\n", ncopies); |
ebfd146a | 4768 | |
4a00c761 JJ |
4769 | if (op_type == binary_op) |
4770 | { | |
4771 | if (CONSTANT_CLASS_P (op0)) | |
4772 | op0 = fold_convert (TREE_TYPE (op1), op0); | |
4773 | else if (CONSTANT_CLASS_P (op1)) | |
4774 | op1 = fold_convert (TREE_TYPE (op0), op1); | |
4775 | } | |
4776 | ||
4777 | /* In case of multi-step conversion, we first generate conversion operations | |
4778 | to the intermediate types, and then from that types to the final one. | |
4779 | We create vector destinations for the intermediate type (TYPES) received | |
4780 | from supportable_*_operation, and store them in the correct order | |
4781 | for future use in vect_create_vectorized_*_stmts (). */ | |
8c681247 | 4782 | auto_vec<tree> vec_dsts (multi_step_cvt + 1); |
82294ec1 JJ |
4783 | vec_dest = vect_create_destination_var (scalar_dest, |
4784 | (cvt_type && modifier == WIDEN) | |
4785 | ? cvt_type : vectype_out); | |
9771b263 | 4786 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4787 | |
4788 | if (multi_step_cvt) | |
4789 | { | |
9771b263 DN |
4790 | for (i = interm_types.length () - 1; |
4791 | interm_types.iterate (i, &intermediate_type); i--) | |
4a00c761 JJ |
4792 | { |
4793 | vec_dest = vect_create_destination_var (scalar_dest, | |
4794 | intermediate_type); | |
9771b263 | 4795 | vec_dsts.quick_push (vec_dest); |
4a00c761 JJ |
4796 | } |
4797 | } | |
ebfd146a | 4798 | |
4a00c761 | 4799 | if (cvt_type) |
82294ec1 JJ |
4800 | vec_dest = vect_create_destination_var (scalar_dest, |
4801 | modifier == WIDEN | |
4802 | ? vectype_out : cvt_type); | |
4a00c761 JJ |
4803 | |
4804 | if (!slp_node) | |
4805 | { | |
30862efc | 4806 | if (modifier == WIDEN) |
4a00c761 | 4807 | { |
c3284718 | 4808 | vec_oprnds0.create (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1); |
4a00c761 | 4809 | if (op_type == binary_op) |
9771b263 | 4810 | vec_oprnds1.create (1); |
4a00c761 | 4811 | } |
30862efc | 4812 | else if (modifier == NARROW) |
9771b263 DN |
4813 | vec_oprnds0.create ( |
4814 | 2 * (multi_step_cvt ? vect_pow2 (multi_step_cvt) : 1)); | |
4a00c761 JJ |
4815 | } |
4816 | else if (code == WIDEN_LSHIFT_EXPR) | |
9771b263 | 4817 | vec_oprnds1.create (slp_node->vec_stmts_size); |
ebfd146a | 4818 | |
4a00c761 | 4819 | last_oprnd = op0; |
ebfd146a IR |
4820 | prev_stmt_info = NULL; |
4821 | switch (modifier) | |
4822 | { | |
4823 | case NONE: | |
4824 | for (j = 0; j < ncopies; j++) | |
4825 | { | |
ebfd146a | 4826 | if (j == 0) |
306b0c92 | 4827 | vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node); |
ebfd146a IR |
4828 | else |
4829 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL); | |
4830 | ||
9771b263 | 4831 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4832 | { |
4833 | /* Arguments are ready, create the new vector stmt. */ | |
4834 | if (code1 == CALL_EXPR) | |
4835 | { | |
4836 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4837 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4838 | gimple_call_set_lhs (new_stmt, new_temp); | |
4839 | } | |
4840 | else | |
4841 | { | |
4842 | gcc_assert (TREE_CODE_LENGTH (code1) == unary_op); | |
0d0e4a03 | 4843 | new_stmt = gimple_build_assign (vec_dest, code1, vop0); |
4a00c761 JJ |
4844 | new_temp = make_ssa_name (vec_dest, new_stmt); |
4845 | gimple_assign_set_lhs (new_stmt, new_temp); | |
4846 | } | |
4847 | ||
4848 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4849 | if (slp_node) | |
9771b263 | 4850 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
225ce44b RB |
4851 | else |
4852 | { | |
4853 | if (!prev_stmt_info) | |
4854 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
4855 | else | |
4856 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4857 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4858 | } | |
4a00c761 | 4859 | } |
ebfd146a IR |
4860 | } |
4861 | break; | |
4862 | ||
4863 | case WIDEN: | |
4864 | /* In case the vectorization factor (VF) is bigger than the number | |
4865 | of elements that we can fit in a vectype (nunits), we have to | |
4866 | generate more than one vector stmt - i.e - we need to "unroll" | |
4867 | the vector stmt by a factor VF/nunits. */ | |
4868 | for (j = 0; j < ncopies; j++) | |
4869 | { | |
4a00c761 | 4870 | /* Handle uses. */ |
ebfd146a | 4871 | if (j == 0) |
4a00c761 JJ |
4872 | { |
4873 | if (slp_node) | |
4874 | { | |
4875 | if (code == WIDEN_LSHIFT_EXPR) | |
4876 | { | |
4877 | unsigned int k; | |
ebfd146a | 4878 | |
4a00c761 JJ |
4879 | vec_oprnd1 = op1; |
4880 | /* Store vec_oprnd1 for every vector stmt to be created | |
4881 | for SLP_NODE. We check during the analysis that all | |
4882 | the shift arguments are the same. */ | |
4883 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 4884 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4885 | |
4886 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 4887 | slp_node); |
4a00c761 JJ |
4888 | } |
4889 | else | |
4890 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, | |
306b0c92 | 4891 | &vec_oprnds1, slp_node); |
4a00c761 JJ |
4892 | } |
4893 | else | |
4894 | { | |
81c40241 | 4895 | vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt); |
9771b263 | 4896 | vec_oprnds0.quick_push (vec_oprnd0); |
4a00c761 JJ |
4897 | if (op_type == binary_op) |
4898 | { | |
4899 | if (code == WIDEN_LSHIFT_EXPR) | |
4900 | vec_oprnd1 = op1; | |
4901 | else | |
81c40241 | 4902 | vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt); |
9771b263 | 4903 | vec_oprnds1.quick_push (vec_oprnd1); |
4a00c761 JJ |
4904 | } |
4905 | } | |
4906 | } | |
ebfd146a | 4907 | else |
4a00c761 JJ |
4908 | { |
4909 | vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0); | |
9771b263 DN |
4910 | vec_oprnds0.truncate (0); |
4911 | vec_oprnds0.quick_push (vec_oprnd0); | |
4a00c761 JJ |
4912 | if (op_type == binary_op) |
4913 | { | |
4914 | if (code == WIDEN_LSHIFT_EXPR) | |
4915 | vec_oprnd1 = op1; | |
4916 | else | |
4917 | vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], | |
4918 | vec_oprnd1); | |
9771b263 DN |
4919 | vec_oprnds1.truncate (0); |
4920 | vec_oprnds1.quick_push (vec_oprnd1); | |
4a00c761 JJ |
4921 | } |
4922 | } | |
ebfd146a | 4923 | |
4a00c761 JJ |
4924 | /* Arguments are ready. Create the new vector stmts. */ |
4925 | for (i = multi_step_cvt; i >= 0; i--) | |
4926 | { | |
9771b263 | 4927 | tree this_dest = vec_dsts[i]; |
4a00c761 JJ |
4928 | enum tree_code c1 = code1, c2 = code2; |
4929 | if (i == 0 && codecvt2 != ERROR_MARK) | |
4930 | { | |
4931 | c1 = codecvt1; | |
4932 | c2 = codecvt2; | |
4933 | } | |
4934 | vect_create_vectorized_promotion_stmts (&vec_oprnds0, | |
4935 | &vec_oprnds1, | |
4936 | stmt, this_dest, gsi, | |
4937 | c1, c2, decl1, decl2, | |
4938 | op_type); | |
4939 | } | |
4940 | ||
9771b263 | 4941 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
4942 | { |
4943 | if (cvt_type) | |
4944 | { | |
4945 | if (codecvt1 == CALL_EXPR) | |
4946 | { | |
4947 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
4948 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
4949 | gimple_call_set_lhs (new_stmt, new_temp); | |
4950 | } | |
4951 | else | |
4952 | { | |
4953 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 4954 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
4955 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
4956 | vop0); | |
4a00c761 JJ |
4957 | } |
4958 | ||
4959 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
4960 | } | |
4961 | else | |
4962 | new_stmt = SSA_NAME_DEF_STMT (vop0); | |
4963 | ||
4964 | if (slp_node) | |
9771b263 | 4965 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
4a00c761 | 4966 | else |
c689ce1e RB |
4967 | { |
4968 | if (!prev_stmt_info) | |
4969 | STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; | |
4970 | else | |
4971 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
4972 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
4973 | } | |
4a00c761 | 4974 | } |
ebfd146a | 4975 | } |
4a00c761 JJ |
4976 | |
4977 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
ebfd146a IR |
4978 | break; |
4979 | ||
4980 | case NARROW: | |
4981 | /* In case the vectorization factor (VF) is bigger than the number | |
4982 | of elements that we can fit in a vectype (nunits), we have to | |
4983 | generate more than one vector stmt - i.e - we need to "unroll" | |
4984 | the vector stmt by a factor VF/nunits. */ | |
4985 | for (j = 0; j < ncopies; j++) | |
4986 | { | |
4987 | /* Handle uses. */ | |
4a00c761 JJ |
4988 | if (slp_node) |
4989 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 4990 | slp_node); |
ebfd146a IR |
4991 | else |
4992 | { | |
9771b263 | 4993 | vec_oprnds0.truncate (0); |
4a00c761 JJ |
4994 | vect_get_loop_based_defs (&last_oprnd, stmt, dt[0], &vec_oprnds0, |
4995 | vect_pow2 (multi_step_cvt) - 1); | |
ebfd146a IR |
4996 | } |
4997 | ||
4a00c761 JJ |
4998 | /* Arguments are ready. Create the new vector stmts. */ |
4999 | if (cvt_type) | |
9771b263 | 5000 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
4a00c761 JJ |
5001 | { |
5002 | if (codecvt1 == CALL_EXPR) | |
5003 | { | |
5004 | new_stmt = gimple_build_call (decl1, 1, vop0); | |
5005 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5006 | gimple_call_set_lhs (new_stmt, new_temp); | |
5007 | } | |
5008 | else | |
5009 | { | |
5010 | gcc_assert (TREE_CODE_LENGTH (codecvt1) == unary_op); | |
b731b390 | 5011 | new_temp = make_ssa_name (vec_dest); |
0d0e4a03 JJ |
5012 | new_stmt = gimple_build_assign (new_temp, codecvt1, |
5013 | vop0); | |
4a00c761 | 5014 | } |
ebfd146a | 5015 | |
4a00c761 | 5016 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
9771b263 | 5017 | vec_oprnds0[i] = new_temp; |
4a00c761 | 5018 | } |
ebfd146a | 5019 | |
4a00c761 JJ |
5020 | vect_create_vectorized_demotion_stmts (&vec_oprnds0, multi_step_cvt, |
5021 | stmt, vec_dsts, gsi, | |
5022 | slp_node, code1, | |
5023 | &prev_stmt_info); | |
ebfd146a IR |
5024 | } |
5025 | ||
5026 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
4a00c761 | 5027 | break; |
ebfd146a IR |
5028 | } |
5029 | ||
9771b263 DN |
5030 | vec_oprnds0.release (); |
5031 | vec_oprnds1.release (); | |
9771b263 | 5032 | interm_types.release (); |
ebfd146a IR |
5033 | |
5034 | return true; | |
5035 | } | |
ff802fa1 IR |
5036 | |
5037 | ||
ebfd146a IR |
5038 | /* Function vectorizable_assignment. |
5039 | ||
b8698a0f L |
5040 | Check if STMT performs an assignment (copy) that can be vectorized. |
5041 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
5042 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5043 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5044 | ||
5045 | static bool | |
355fe088 TS |
5046 | vectorizable_assignment (gimple *stmt, gimple_stmt_iterator *gsi, |
5047 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a IR |
5048 | { |
5049 | tree vec_dest; | |
5050 | tree scalar_dest; | |
5051 | tree op; | |
5052 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
ebfd146a IR |
5053 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
5054 | tree new_temp; | |
355fe088 | 5055 | gimple *def_stmt; |
4fc5ebf1 JG |
5056 | enum vect_def_type dt[1] = {vect_unknown_def_type}; |
5057 | int ndts = 1; | |
ebfd146a | 5058 | int ncopies; |
f18b55bd | 5059 | int i, j; |
6e1aa848 | 5060 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 5061 | tree vop; |
a70d6342 | 5062 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5063 | vec_info *vinfo = stmt_info->vinfo; |
355fe088 | 5064 | gimple *new_stmt = NULL; |
f18b55bd | 5065 | stmt_vec_info prev_stmt_info = NULL; |
fde9c428 RG |
5066 | enum tree_code code; |
5067 | tree vectype_in; | |
ebfd146a | 5068 | |
a70d6342 | 5069 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5070 | return false; |
5071 | ||
66c16fd9 RB |
5072 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5073 | && ! vec_stmt) | |
ebfd146a IR |
5074 | return false; |
5075 | ||
5076 | /* Is vectorizable assignment? */ | |
5077 | if (!is_gimple_assign (stmt)) | |
5078 | return false; | |
5079 | ||
5080 | scalar_dest = gimple_assign_lhs (stmt); | |
5081 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
5082 | return false; | |
5083 | ||
fde9c428 | 5084 | code = gimple_assign_rhs_code (stmt); |
ebfd146a | 5085 | if (gimple_assign_single_p (stmt) |
fde9c428 RG |
5086 | || code == PAREN_EXPR |
5087 | || CONVERT_EXPR_CODE_P (code)) | |
ebfd146a IR |
5088 | op = gimple_assign_rhs1 (stmt); |
5089 | else | |
5090 | return false; | |
5091 | ||
7b7ec6c5 RG |
5092 | if (code == VIEW_CONVERT_EXPR) |
5093 | op = TREE_OPERAND (op, 0); | |
5094 | ||
465c8c19 | 5095 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
928686b1 | 5096 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 JJ |
5097 | |
5098 | /* Multiple types in SLP are handled by creating the appropriate number of | |
5099 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5100 | case of SLP. */ | |
fce57248 | 5101 | if (slp_node) |
465c8c19 JJ |
5102 | ncopies = 1; |
5103 | else | |
e8f142e2 | 5104 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
5105 | |
5106 | gcc_assert (ncopies >= 1); | |
5107 | ||
81c40241 | 5108 | if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[0], &vectype_in)) |
ebfd146a | 5109 | { |
73fbfcad | 5110 | if (dump_enabled_p ()) |
78c60e3d | 5111 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5112 | "use not simple.\n"); |
ebfd146a IR |
5113 | return false; |
5114 | } | |
5115 | ||
fde9c428 RG |
5116 | /* We can handle NOP_EXPR conversions that do not change the number |
5117 | of elements or the vector size. */ | |
7b7ec6c5 RG |
5118 | if ((CONVERT_EXPR_CODE_P (code) |
5119 | || code == VIEW_CONVERT_EXPR) | |
fde9c428 | 5120 | && (!vectype_in |
928686b1 | 5121 | || maybe_ne (TYPE_VECTOR_SUBPARTS (vectype_in), nunits) |
cf098191 RS |
5122 | || maybe_ne (GET_MODE_SIZE (TYPE_MODE (vectype)), |
5123 | GET_MODE_SIZE (TYPE_MODE (vectype_in))))) | |
fde9c428 RG |
5124 | return false; |
5125 | ||
7b7b1813 RG |
5126 | /* We do not handle bit-precision changes. */ |
5127 | if ((CONVERT_EXPR_CODE_P (code) | |
5128 | || code == VIEW_CONVERT_EXPR) | |
5129 | && INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) | |
2be65d9e RS |
5130 | && (!type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
5131 | || !type_has_mode_precision_p (TREE_TYPE (op))) | |
7b7b1813 RG |
5132 | /* But a conversion that does not change the bit-pattern is ok. */ |
5133 | && !((TYPE_PRECISION (TREE_TYPE (scalar_dest)) | |
5134 | > TYPE_PRECISION (TREE_TYPE (op))) | |
2dab46d5 IE |
5135 | && TYPE_UNSIGNED (TREE_TYPE (op))) |
5136 | /* Conversion between boolean types of different sizes is | |
5137 | a simple assignment in case their vectypes are same | |
5138 | boolean vectors. */ | |
5139 | && (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
5140 | || !VECTOR_BOOLEAN_TYPE_P (vectype_in))) | |
7b7b1813 | 5141 | { |
73fbfcad | 5142 | if (dump_enabled_p ()) |
78c60e3d SS |
5143 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5144 | "type conversion to/from bit-precision " | |
e645e942 | 5145 | "unsupported.\n"); |
7b7b1813 RG |
5146 | return false; |
5147 | } | |
5148 | ||
ebfd146a IR |
5149 | if (!vec_stmt) /* transformation not required. */ |
5150 | { | |
5151 | STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; | |
73fbfcad | 5152 | if (dump_enabled_p ()) |
78c60e3d | 5153 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5154 | "=== vectorizable_assignment ===\n"); |
78604de0 RB |
5155 | if (!slp_node) |
5156 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); | |
ebfd146a IR |
5157 | return true; |
5158 | } | |
5159 | ||
67b8dbac | 5160 | /* Transform. */ |
73fbfcad | 5161 | if (dump_enabled_p ()) |
e645e942 | 5162 | dump_printf_loc (MSG_NOTE, vect_location, "transform assignment.\n"); |
ebfd146a IR |
5163 | |
5164 | /* Handle def. */ | |
5165 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5166 | ||
5167 | /* Handle use. */ | |
f18b55bd | 5168 | for (j = 0; j < ncopies; j++) |
ebfd146a | 5169 | { |
f18b55bd IR |
5170 | /* Handle uses. */ |
5171 | if (j == 0) | |
306b0c92 | 5172 | vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node); |
f18b55bd IR |
5173 | else |
5174 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL); | |
5175 | ||
5176 | /* Arguments are ready. create the new vector stmt. */ | |
9771b263 | 5177 | FOR_EACH_VEC_ELT (vec_oprnds, i, vop) |
f18b55bd | 5178 | { |
7b7ec6c5 RG |
5179 | if (CONVERT_EXPR_CODE_P (code) |
5180 | || code == VIEW_CONVERT_EXPR) | |
4a73490d | 5181 | vop = build1 (VIEW_CONVERT_EXPR, vectype, vop); |
f18b55bd IR |
5182 | new_stmt = gimple_build_assign (vec_dest, vop); |
5183 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
5184 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5185 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5186 | if (slp_node) | |
9771b263 | 5187 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f18b55bd | 5188 | } |
ebfd146a IR |
5189 | |
5190 | if (slp_node) | |
f18b55bd IR |
5191 | continue; |
5192 | ||
5193 | if (j == 0) | |
5194 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5195 | else | |
5196 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5197 | ||
5198 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5199 | } | |
b8698a0f | 5200 | |
9771b263 | 5201 | vec_oprnds.release (); |
ebfd146a IR |
5202 | return true; |
5203 | } | |
5204 | ||
9dc3f7de | 5205 | |
1107f3ae IR |
5206 | /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE |
5207 | either as shift by a scalar or by a vector. */ | |
5208 | ||
5209 | bool | |
5210 | vect_supportable_shift (enum tree_code code, tree scalar_type) | |
5211 | { | |
5212 | ||
ef4bddc2 | 5213 | machine_mode vec_mode; |
1107f3ae IR |
5214 | optab optab; |
5215 | int icode; | |
5216 | tree vectype; | |
5217 | ||
5218 | vectype = get_vectype_for_scalar_type (scalar_type); | |
5219 | if (!vectype) | |
5220 | return false; | |
5221 | ||
5222 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
5223 | if (!optab | |
5224 | || optab_handler (optab, TYPE_MODE (vectype)) == CODE_FOR_nothing) | |
5225 | { | |
5226 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
5227 | if (!optab | |
5228 | || (optab_handler (optab, TYPE_MODE (vectype)) | |
5229 | == CODE_FOR_nothing)) | |
5230 | return false; | |
5231 | } | |
5232 | ||
5233 | vec_mode = TYPE_MODE (vectype); | |
5234 | icode = (int) optab_handler (optab, vec_mode); | |
5235 | if (icode == CODE_FOR_nothing) | |
5236 | return false; | |
5237 | ||
5238 | return true; | |
5239 | } | |
5240 | ||
5241 | ||
9dc3f7de IR |
5242 | /* Function vectorizable_shift. |
5243 | ||
5244 | Check if STMT performs a shift operation that can be vectorized. | |
5245 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
5246 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. | |
5247 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5248 | ||
5249 | static bool | |
355fe088 TS |
5250 | vectorizable_shift (gimple *stmt, gimple_stmt_iterator *gsi, |
5251 | gimple **vec_stmt, slp_tree slp_node) | |
9dc3f7de IR |
5252 | { |
5253 | tree vec_dest; | |
5254 | tree scalar_dest; | |
5255 | tree op0, op1 = NULL; | |
5256 | tree vec_oprnd1 = NULL_TREE; | |
5257 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
5258 | tree vectype; | |
5259 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
5260 | enum tree_code code; | |
ef4bddc2 | 5261 | machine_mode vec_mode; |
9dc3f7de IR |
5262 | tree new_temp; |
5263 | optab optab; | |
5264 | int icode; | |
ef4bddc2 | 5265 | machine_mode optab_op2_mode; |
355fe088 | 5266 | gimple *def_stmt; |
9dc3f7de | 5267 | enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type}; |
4fc5ebf1 | 5268 | int ndts = 2; |
355fe088 | 5269 | gimple *new_stmt = NULL; |
9dc3f7de | 5270 | stmt_vec_info prev_stmt_info; |
928686b1 RS |
5271 | poly_uint64 nunits_in; |
5272 | poly_uint64 nunits_out; | |
9dc3f7de | 5273 | tree vectype_out; |
cede2577 | 5274 | tree op1_vectype; |
9dc3f7de IR |
5275 | int ncopies; |
5276 | int j, i; | |
6e1aa848 DN |
5277 | vec<tree> vec_oprnds0 = vNULL; |
5278 | vec<tree> vec_oprnds1 = vNULL; | |
9dc3f7de IR |
5279 | tree vop0, vop1; |
5280 | unsigned int k; | |
49eab32e | 5281 | bool scalar_shift_arg = true; |
9dc3f7de | 5282 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5283 | vec_info *vinfo = stmt_info->vinfo; |
9dc3f7de IR |
5284 | |
5285 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
5286 | return false; | |
5287 | ||
66c16fd9 RB |
5288 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5289 | && ! vec_stmt) | |
9dc3f7de IR |
5290 | return false; |
5291 | ||
5292 | /* Is STMT a vectorizable binary/unary operation? */ | |
5293 | if (!is_gimple_assign (stmt)) | |
5294 | return false; | |
5295 | ||
5296 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
5297 | return false; | |
5298 | ||
5299 | code = gimple_assign_rhs_code (stmt); | |
5300 | ||
5301 | if (!(code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR | |
5302 | || code == RROTATE_EXPR)) | |
5303 | return false; | |
5304 | ||
5305 | scalar_dest = gimple_assign_lhs (stmt); | |
5306 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
2be65d9e | 5307 | if (!type_has_mode_precision_p (TREE_TYPE (scalar_dest))) |
7b7b1813 | 5308 | { |
73fbfcad | 5309 | if (dump_enabled_p ()) |
78c60e3d | 5310 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5311 | "bit-precision shifts not supported.\n"); |
7b7b1813 RG |
5312 | return false; |
5313 | } | |
9dc3f7de IR |
5314 | |
5315 | op0 = gimple_assign_rhs1 (stmt); | |
81c40241 | 5316 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
9dc3f7de | 5317 | { |
73fbfcad | 5318 | if (dump_enabled_p ()) |
78c60e3d | 5319 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5320 | "use not simple.\n"); |
9dc3f7de IR |
5321 | return false; |
5322 | } | |
5323 | /* If op0 is an external or constant def use a vector type with | |
5324 | the same size as the output vector type. */ | |
5325 | if (!vectype) | |
5326 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
5327 | if (vec_stmt) | |
5328 | gcc_assert (vectype); | |
5329 | if (!vectype) | |
5330 | { | |
73fbfcad | 5331 | if (dump_enabled_p ()) |
78c60e3d | 5332 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5333 | "no vectype for scalar type\n"); |
9dc3f7de IR |
5334 | return false; |
5335 | } | |
5336 | ||
5337 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
5338 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
928686b1 | 5339 | if (maybe_ne (nunits_out, nunits_in)) |
9dc3f7de IR |
5340 | return false; |
5341 | ||
5342 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 5343 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1], &op1_vectype)) |
9dc3f7de | 5344 | { |
73fbfcad | 5345 | if (dump_enabled_p ()) |
78c60e3d | 5346 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5347 | "use not simple.\n"); |
9dc3f7de IR |
5348 | return false; |
5349 | } | |
5350 | ||
9dc3f7de IR |
5351 | /* Multiple types in SLP are handled by creating the appropriate number of |
5352 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
5353 | case of SLP. */ | |
fce57248 | 5354 | if (slp_node) |
9dc3f7de IR |
5355 | ncopies = 1; |
5356 | else | |
e8f142e2 | 5357 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
9dc3f7de IR |
5358 | |
5359 | gcc_assert (ncopies >= 1); | |
5360 | ||
5361 | /* Determine whether the shift amount is a vector, or scalar. If the | |
5362 | shift/rotate amount is a vector, use the vector/vector shift optabs. */ | |
5363 | ||
dbfa87aa YR |
5364 | if ((dt[1] == vect_internal_def |
5365 | || dt[1] == vect_induction_def) | |
5366 | && !slp_node) | |
49eab32e JJ |
5367 | scalar_shift_arg = false; |
5368 | else if (dt[1] == vect_constant_def | |
5369 | || dt[1] == vect_external_def | |
5370 | || dt[1] == vect_internal_def) | |
5371 | { | |
5372 | /* In SLP, need to check whether the shift count is the same, | |
5373 | in loops if it is a constant or invariant, it is always | |
5374 | a scalar shift. */ | |
5375 | if (slp_node) | |
5376 | { | |
355fe088 TS |
5377 | vec<gimple *> stmts = SLP_TREE_SCALAR_STMTS (slp_node); |
5378 | gimple *slpstmt; | |
49eab32e | 5379 | |
9771b263 | 5380 | FOR_EACH_VEC_ELT (stmts, k, slpstmt) |
49eab32e JJ |
5381 | if (!operand_equal_p (gimple_assign_rhs2 (slpstmt), op1, 0)) |
5382 | scalar_shift_arg = false; | |
5383 | } | |
60d393e8 RB |
5384 | |
5385 | /* If the shift amount is computed by a pattern stmt we cannot | |
5386 | use the scalar amount directly thus give up and use a vector | |
5387 | shift. */ | |
5388 | if (dt[1] == vect_internal_def) | |
5389 | { | |
5390 | gimple *def = SSA_NAME_DEF_STMT (op1); | |
5391 | if (is_pattern_stmt_p (vinfo_for_stmt (def))) | |
5392 | scalar_shift_arg = false; | |
5393 | } | |
49eab32e JJ |
5394 | } |
5395 | else | |
5396 | { | |
73fbfcad | 5397 | if (dump_enabled_p ()) |
78c60e3d | 5398 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5399 | "operand mode requires invariant argument.\n"); |
49eab32e JJ |
5400 | return false; |
5401 | } | |
5402 | ||
9dc3f7de | 5403 | /* Vector shifted by vector. */ |
49eab32e | 5404 | if (!scalar_shift_arg) |
9dc3f7de IR |
5405 | { |
5406 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
73fbfcad | 5407 | if (dump_enabled_p ()) |
78c60e3d | 5408 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5409 | "vector/vector shift/rotate found.\n"); |
78c60e3d | 5410 | |
aa948027 JJ |
5411 | if (!op1_vectype) |
5412 | op1_vectype = get_same_sized_vectype (TREE_TYPE (op1), vectype_out); | |
5413 | if (op1_vectype == NULL_TREE | |
5414 | || TYPE_MODE (op1_vectype) != TYPE_MODE (vectype)) | |
cede2577 | 5415 | { |
73fbfcad | 5416 | if (dump_enabled_p ()) |
78c60e3d SS |
5417 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5418 | "unusable type for last operand in" | |
e645e942 | 5419 | " vector/vector shift/rotate.\n"); |
cede2577 JJ |
5420 | return false; |
5421 | } | |
9dc3f7de IR |
5422 | } |
5423 | /* See if the machine has a vector shifted by scalar insn and if not | |
5424 | then see if it has a vector shifted by vector insn. */ | |
49eab32e | 5425 | else |
9dc3f7de IR |
5426 | { |
5427 | optab = optab_for_tree_code (code, vectype, optab_scalar); | |
5428 | if (optab | |
5429 | && optab_handler (optab, TYPE_MODE (vectype)) != CODE_FOR_nothing) | |
5430 | { | |
73fbfcad | 5431 | if (dump_enabled_p ()) |
78c60e3d | 5432 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5433 | "vector/scalar shift/rotate found.\n"); |
9dc3f7de IR |
5434 | } |
5435 | else | |
5436 | { | |
5437 | optab = optab_for_tree_code (code, vectype, optab_vector); | |
5438 | if (optab | |
5439 | && (optab_handler (optab, TYPE_MODE (vectype)) | |
5440 | != CODE_FOR_nothing)) | |
5441 | { | |
49eab32e JJ |
5442 | scalar_shift_arg = false; |
5443 | ||
73fbfcad | 5444 | if (dump_enabled_p ()) |
78c60e3d | 5445 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5446 | "vector/vector shift/rotate found.\n"); |
9dc3f7de IR |
5447 | |
5448 | /* Unlike the other binary operators, shifts/rotates have | |
5449 | the rhs being int, instead of the same type as the lhs, | |
5450 | so make sure the scalar is the right type if we are | |
aa948027 | 5451 | dealing with vectors of long long/long/short/char. */ |
9dc3f7de IR |
5452 | if (dt[1] == vect_constant_def) |
5453 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
aa948027 JJ |
5454 | else if (!useless_type_conversion_p (TREE_TYPE (vectype), |
5455 | TREE_TYPE (op1))) | |
5456 | { | |
5457 | if (slp_node | |
5458 | && TYPE_MODE (TREE_TYPE (vectype)) | |
5459 | != TYPE_MODE (TREE_TYPE (op1))) | |
5460 | { | |
73fbfcad | 5461 | if (dump_enabled_p ()) |
78c60e3d SS |
5462 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5463 | "unusable type for last operand in" | |
e645e942 | 5464 | " vector/vector shift/rotate.\n"); |
21c0a521 | 5465 | return false; |
aa948027 JJ |
5466 | } |
5467 | if (vec_stmt && !slp_node) | |
5468 | { | |
5469 | op1 = fold_convert (TREE_TYPE (vectype), op1); | |
5470 | op1 = vect_init_vector (stmt, op1, | |
5471 | TREE_TYPE (vectype), NULL); | |
5472 | } | |
5473 | } | |
9dc3f7de IR |
5474 | } |
5475 | } | |
5476 | } | |
9dc3f7de IR |
5477 | |
5478 | /* Supportable by target? */ | |
5479 | if (!optab) | |
5480 | { | |
73fbfcad | 5481 | if (dump_enabled_p ()) |
78c60e3d | 5482 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5483 | "no optab.\n"); |
9dc3f7de IR |
5484 | return false; |
5485 | } | |
5486 | vec_mode = TYPE_MODE (vectype); | |
5487 | icode = (int) optab_handler (optab, vec_mode); | |
5488 | if (icode == CODE_FOR_nothing) | |
5489 | { | |
73fbfcad | 5490 | if (dump_enabled_p ()) |
78c60e3d | 5491 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5492 | "op not supported by target.\n"); |
9dc3f7de | 5493 | /* Check only during analysis. */ |
cf098191 | 5494 | if (maybe_ne (GET_MODE_SIZE (vec_mode), UNITS_PER_WORD) |
ca09abcb RS |
5495 | || (!vec_stmt |
5496 | && !vect_worthwhile_without_simd_p (vinfo, code))) | |
9dc3f7de | 5497 | return false; |
73fbfcad | 5498 | if (dump_enabled_p ()) |
e645e942 TJ |
5499 | dump_printf_loc (MSG_NOTE, vect_location, |
5500 | "proceeding using word mode.\n"); | |
9dc3f7de IR |
5501 | } |
5502 | ||
5503 | /* Worthwhile without SIMD support? Check only during analysis. */ | |
ca09abcb RS |
5504 | if (!vec_stmt |
5505 | && !VECTOR_MODE_P (TYPE_MODE (vectype)) | |
5506 | && !vect_worthwhile_without_simd_p (vinfo, code)) | |
9dc3f7de | 5507 | { |
73fbfcad | 5508 | if (dump_enabled_p ()) |
78c60e3d | 5509 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5510 | "not worthwhile without SIMD support.\n"); |
9dc3f7de IR |
5511 | return false; |
5512 | } | |
5513 | ||
5514 | if (!vec_stmt) /* transformation not required. */ | |
5515 | { | |
5516 | STMT_VINFO_TYPE (stmt_info) = shift_vec_info_type; | |
73fbfcad | 5517 | if (dump_enabled_p ()) |
e645e942 TJ |
5518 | dump_printf_loc (MSG_NOTE, vect_location, |
5519 | "=== vectorizable_shift ===\n"); | |
78604de0 RB |
5520 | if (!slp_node) |
5521 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); | |
9dc3f7de IR |
5522 | return true; |
5523 | } | |
5524 | ||
67b8dbac | 5525 | /* Transform. */ |
9dc3f7de | 5526 | |
73fbfcad | 5527 | if (dump_enabled_p ()) |
78c60e3d | 5528 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5529 | "transform binary/unary operation.\n"); |
9dc3f7de IR |
5530 | |
5531 | /* Handle def. */ | |
5532 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
5533 | ||
9dc3f7de IR |
5534 | prev_stmt_info = NULL; |
5535 | for (j = 0; j < ncopies; j++) | |
5536 | { | |
5537 | /* Handle uses. */ | |
5538 | if (j == 0) | |
5539 | { | |
5540 | if (scalar_shift_arg) | |
5541 | { | |
5542 | /* Vector shl and shr insn patterns can be defined with scalar | |
5543 | operand 2 (shift operand). In this case, use constant or loop | |
5544 | invariant op1 directly, without extending it to vector mode | |
5545 | first. */ | |
5546 | optab_op2_mode = insn_data[icode].operand[2].mode; | |
5547 | if (!VECTOR_MODE_P (optab_op2_mode)) | |
5548 | { | |
73fbfcad | 5549 | if (dump_enabled_p ()) |
78c60e3d | 5550 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5551 | "operand 1 using scalar mode.\n"); |
9dc3f7de | 5552 | vec_oprnd1 = op1; |
8930f723 | 5553 | vec_oprnds1.create (slp_node ? slp_node->vec_stmts_size : 1); |
9771b263 | 5554 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5555 | if (slp_node) |
5556 | { | |
5557 | /* Store vec_oprnd1 for every vector stmt to be created | |
5558 | for SLP_NODE. We check during the analysis that all | |
5559 | the shift arguments are the same. | |
5560 | TODO: Allow different constants for different vector | |
5561 | stmts generated for an SLP instance. */ | |
5562 | for (k = 0; k < slp_node->vec_stmts_size - 1; k++) | |
9771b263 | 5563 | vec_oprnds1.quick_push (vec_oprnd1); |
9dc3f7de IR |
5564 | } |
5565 | } | |
5566 | } | |
5567 | ||
5568 | /* vec_oprnd1 is available if operand 1 should be of a scalar-type | |
5569 | (a special case for certain kind of vector shifts); otherwise, | |
5570 | operand 1 should be of a vector type (the usual case). */ | |
5571 | if (vec_oprnd1) | |
5572 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5573 | slp_node); |
9dc3f7de IR |
5574 | else |
5575 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
306b0c92 | 5576 | slp_node); |
9dc3f7de IR |
5577 | } |
5578 | else | |
5579 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5580 | ||
5581 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 5582 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
9dc3f7de | 5583 | { |
9771b263 | 5584 | vop1 = vec_oprnds1[i]; |
0d0e4a03 | 5585 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1); |
9dc3f7de IR |
5586 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5587 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5588 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5589 | if (slp_node) | |
9771b263 | 5590 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
9dc3f7de IR |
5591 | } |
5592 | ||
5593 | if (slp_node) | |
5594 | continue; | |
5595 | ||
5596 | if (j == 0) | |
5597 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5598 | else | |
5599 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5600 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
5601 | } | |
5602 | ||
9771b263 DN |
5603 | vec_oprnds0.release (); |
5604 | vec_oprnds1.release (); | |
9dc3f7de IR |
5605 | |
5606 | return true; | |
5607 | } | |
5608 | ||
5609 | ||
ebfd146a IR |
5610 | /* Function vectorizable_operation. |
5611 | ||
16949072 RG |
5612 | Check if STMT performs a binary, unary or ternary operation that can |
5613 | be vectorized. | |
b8698a0f | 5614 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized |
ebfd146a IR |
5615 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
5616 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
5617 | ||
5618 | static bool | |
355fe088 TS |
5619 | vectorizable_operation (gimple *stmt, gimple_stmt_iterator *gsi, |
5620 | gimple **vec_stmt, slp_tree slp_node) | |
ebfd146a | 5621 | { |
00f07b86 | 5622 | tree vec_dest; |
ebfd146a | 5623 | tree scalar_dest; |
16949072 | 5624 | tree op0, op1 = NULL_TREE, op2 = NULL_TREE; |
ebfd146a | 5625 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
00f07b86 | 5626 | tree vectype; |
ebfd146a | 5627 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
0eb952ea | 5628 | enum tree_code code, orig_code; |
ef4bddc2 | 5629 | machine_mode vec_mode; |
ebfd146a IR |
5630 | tree new_temp; |
5631 | int op_type; | |
00f07b86 | 5632 | optab optab; |
523ba738 | 5633 | bool target_support_p; |
355fe088 | 5634 | gimple *def_stmt; |
16949072 RG |
5635 | enum vect_def_type dt[3] |
5636 | = {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 5637 | int ndts = 3; |
355fe088 | 5638 | gimple *new_stmt = NULL; |
ebfd146a | 5639 | stmt_vec_info prev_stmt_info; |
928686b1 RS |
5640 | poly_uint64 nunits_in; |
5641 | poly_uint64 nunits_out; | |
ebfd146a IR |
5642 | tree vectype_out; |
5643 | int ncopies; | |
5644 | int j, i; | |
6e1aa848 DN |
5645 | vec<tree> vec_oprnds0 = vNULL; |
5646 | vec<tree> vec_oprnds1 = vNULL; | |
5647 | vec<tree> vec_oprnds2 = vNULL; | |
16949072 | 5648 | tree vop0, vop1, vop2; |
a70d6342 | 5649 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 5650 | vec_info *vinfo = stmt_info->vinfo; |
a70d6342 | 5651 | |
a70d6342 | 5652 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
5653 | return false; |
5654 | ||
66c16fd9 RB |
5655 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
5656 | && ! vec_stmt) | |
ebfd146a IR |
5657 | return false; |
5658 | ||
5659 | /* Is STMT a vectorizable binary/unary operation? */ | |
5660 | if (!is_gimple_assign (stmt)) | |
5661 | return false; | |
5662 | ||
5663 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
5664 | return false; | |
5665 | ||
0eb952ea | 5666 | orig_code = code = gimple_assign_rhs_code (stmt); |
ebfd146a | 5667 | |
1af4ebf5 MG |
5668 | /* For pointer addition and subtraction, we should use the normal |
5669 | plus and minus for the vector operation. */ | |
ebfd146a IR |
5670 | if (code == POINTER_PLUS_EXPR) |
5671 | code = PLUS_EXPR; | |
1af4ebf5 MG |
5672 | if (code == POINTER_DIFF_EXPR) |
5673 | code = MINUS_EXPR; | |
ebfd146a IR |
5674 | |
5675 | /* Support only unary or binary operations. */ | |
5676 | op_type = TREE_CODE_LENGTH (code); | |
16949072 | 5677 | if (op_type != unary_op && op_type != binary_op && op_type != ternary_op) |
ebfd146a | 5678 | { |
73fbfcad | 5679 | if (dump_enabled_p ()) |
78c60e3d | 5680 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5681 | "num. args = %d (not unary/binary/ternary op).\n", |
78c60e3d | 5682 | op_type); |
ebfd146a IR |
5683 | return false; |
5684 | } | |
5685 | ||
b690cc0f RG |
5686 | scalar_dest = gimple_assign_lhs (stmt); |
5687 | vectype_out = STMT_VINFO_VECTYPE (stmt_info); | |
5688 | ||
7b7b1813 RG |
5689 | /* Most operations cannot handle bit-precision types without extra |
5690 | truncations. */ | |
045c1278 | 5691 | if (!VECTOR_BOOLEAN_TYPE_P (vectype_out) |
2be65d9e | 5692 | && !type_has_mode_precision_p (TREE_TYPE (scalar_dest)) |
7b7b1813 RG |
5693 | /* Exception are bitwise binary operations. */ |
5694 | && code != BIT_IOR_EXPR | |
5695 | && code != BIT_XOR_EXPR | |
5696 | && code != BIT_AND_EXPR) | |
5697 | { | |
73fbfcad | 5698 | if (dump_enabled_p ()) |
78c60e3d | 5699 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5700 | "bit-precision arithmetic not supported.\n"); |
7b7b1813 RG |
5701 | return false; |
5702 | } | |
5703 | ||
ebfd146a | 5704 | op0 = gimple_assign_rhs1 (stmt); |
81c40241 | 5705 | if (!vect_is_simple_use (op0, vinfo, &def_stmt, &dt[0], &vectype)) |
ebfd146a | 5706 | { |
73fbfcad | 5707 | if (dump_enabled_p ()) |
78c60e3d | 5708 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5709 | "use not simple.\n"); |
ebfd146a IR |
5710 | return false; |
5711 | } | |
b690cc0f RG |
5712 | /* If op0 is an external or constant def use a vector type with |
5713 | the same size as the output vector type. */ | |
5714 | if (!vectype) | |
b036c6c5 IE |
5715 | { |
5716 | /* For boolean type we cannot determine vectype by | |
5717 | invariant value (don't know whether it is a vector | |
5718 | of booleans or vector of integers). We use output | |
5719 | vectype because operations on boolean don't change | |
5720 | type. */ | |
2568d8a1 | 5721 | if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op0))) |
b036c6c5 | 5722 | { |
2568d8a1 | 5723 | if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (scalar_dest))) |
b036c6c5 IE |
5724 | { |
5725 | if (dump_enabled_p ()) | |
5726 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
5727 | "not supported operation on bool value.\n"); | |
5728 | return false; | |
5729 | } | |
5730 | vectype = vectype_out; | |
5731 | } | |
5732 | else | |
5733 | vectype = get_same_sized_vectype (TREE_TYPE (op0), vectype_out); | |
5734 | } | |
7d8930a0 IR |
5735 | if (vec_stmt) |
5736 | gcc_assert (vectype); | |
5737 | if (!vectype) | |
5738 | { | |
73fbfcad | 5739 | if (dump_enabled_p ()) |
7d8930a0 | 5740 | { |
78c60e3d SS |
5741 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
5742 | "no vectype for scalar type "); | |
5743 | dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, | |
5744 | TREE_TYPE (op0)); | |
e645e942 | 5745 | dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); |
7d8930a0 IR |
5746 | } |
5747 | ||
5748 | return false; | |
5749 | } | |
b690cc0f RG |
5750 | |
5751 | nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); | |
5752 | nunits_in = TYPE_VECTOR_SUBPARTS (vectype); | |
928686b1 | 5753 | if (maybe_ne (nunits_out, nunits_in)) |
b690cc0f | 5754 | return false; |
ebfd146a | 5755 | |
16949072 | 5756 | if (op_type == binary_op || op_type == ternary_op) |
ebfd146a IR |
5757 | { |
5758 | op1 = gimple_assign_rhs2 (stmt); | |
81c40241 | 5759 | if (!vect_is_simple_use (op1, vinfo, &def_stmt, &dt[1])) |
ebfd146a | 5760 | { |
73fbfcad | 5761 | if (dump_enabled_p ()) |
78c60e3d | 5762 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5763 | "use not simple.\n"); |
ebfd146a IR |
5764 | return false; |
5765 | } | |
5766 | } | |
16949072 RG |
5767 | if (op_type == ternary_op) |
5768 | { | |
5769 | op2 = gimple_assign_rhs3 (stmt); | |
81c40241 | 5770 | if (!vect_is_simple_use (op2, vinfo, &def_stmt, &dt[2])) |
16949072 | 5771 | { |
73fbfcad | 5772 | if (dump_enabled_p ()) |
78c60e3d | 5773 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5774 | "use not simple.\n"); |
16949072 RG |
5775 | return false; |
5776 | } | |
5777 | } | |
ebfd146a | 5778 | |
b690cc0f | 5779 | /* Multiple types in SLP are handled by creating the appropriate number of |
ff802fa1 | 5780 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
b690cc0f | 5781 | case of SLP. */ |
fce57248 | 5782 | if (slp_node) |
b690cc0f RG |
5783 | ncopies = 1; |
5784 | else | |
e8f142e2 | 5785 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
b690cc0f RG |
5786 | |
5787 | gcc_assert (ncopies >= 1); | |
5788 | ||
9dc3f7de | 5789 | /* Shifts are handled in vectorizable_shift (). */ |
ebfd146a IR |
5790 | if (code == LSHIFT_EXPR || code == RSHIFT_EXPR || code == LROTATE_EXPR |
5791 | || code == RROTATE_EXPR) | |
9dc3f7de | 5792 | return false; |
ebfd146a | 5793 | |
ebfd146a | 5794 | /* Supportable by target? */ |
00f07b86 RH |
5795 | |
5796 | vec_mode = TYPE_MODE (vectype); | |
5797 | if (code == MULT_HIGHPART_EXPR) | |
523ba738 | 5798 | target_support_p = can_mult_highpart_p (vec_mode, TYPE_UNSIGNED (vectype)); |
00f07b86 RH |
5799 | else |
5800 | { | |
5801 | optab = optab_for_tree_code (code, vectype, optab_default); | |
5802 | if (!optab) | |
5deb57cb | 5803 | { |
73fbfcad | 5804 | if (dump_enabled_p ()) |
78c60e3d | 5805 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5806 | "no optab.\n"); |
00f07b86 | 5807 | return false; |
5deb57cb | 5808 | } |
523ba738 RS |
5809 | target_support_p = (optab_handler (optab, vec_mode) |
5810 | != CODE_FOR_nothing); | |
5deb57cb JJ |
5811 | } |
5812 | ||
523ba738 | 5813 | if (!target_support_p) |
ebfd146a | 5814 | { |
73fbfcad | 5815 | if (dump_enabled_p ()) |
78c60e3d | 5816 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5817 | "op not supported by target.\n"); |
ebfd146a | 5818 | /* Check only during analysis. */ |
cf098191 | 5819 | if (maybe_ne (GET_MODE_SIZE (vec_mode), UNITS_PER_WORD) |
ca09abcb | 5820 | || (!vec_stmt && !vect_worthwhile_without_simd_p (vinfo, code))) |
ebfd146a | 5821 | return false; |
73fbfcad | 5822 | if (dump_enabled_p ()) |
e645e942 TJ |
5823 | dump_printf_loc (MSG_NOTE, vect_location, |
5824 | "proceeding using word mode.\n"); | |
383d9c83 IR |
5825 | } |
5826 | ||
4a00c761 | 5827 | /* Worthwhile without SIMD support? Check only during analysis. */ |
5deb57cb JJ |
5828 | if (!VECTOR_MODE_P (vec_mode) |
5829 | && !vec_stmt | |
ca09abcb | 5830 | && !vect_worthwhile_without_simd_p (vinfo, code)) |
7d8930a0 | 5831 | { |
73fbfcad | 5832 | if (dump_enabled_p ()) |
78c60e3d | 5833 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 5834 | "not worthwhile without SIMD support.\n"); |
e34842c6 | 5835 | return false; |
7d8930a0 | 5836 | } |
ebfd146a | 5837 | |
ebfd146a IR |
5838 | if (!vec_stmt) /* transformation not required. */ |
5839 | { | |
4a00c761 | 5840 | STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; |
73fbfcad | 5841 | if (dump_enabled_p ()) |
78c60e3d | 5842 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5843 | "=== vectorizable_operation ===\n"); |
78604de0 RB |
5844 | if (!slp_node) |
5845 | vect_model_simple_cost (stmt_info, ncopies, dt, ndts, NULL, NULL); | |
ebfd146a IR |
5846 | return true; |
5847 | } | |
5848 | ||
67b8dbac | 5849 | /* Transform. */ |
ebfd146a | 5850 | |
73fbfcad | 5851 | if (dump_enabled_p ()) |
78c60e3d | 5852 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 5853 | "transform binary/unary operation.\n"); |
383d9c83 | 5854 | |
ebfd146a | 5855 | /* Handle def. */ |
00f07b86 | 5856 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
b8698a0f | 5857 | |
0eb952ea JJ |
5858 | /* POINTER_DIFF_EXPR has pointer arguments which are vectorized as |
5859 | vectors with unsigned elements, but the result is signed. So, we | |
5860 | need to compute the MINUS_EXPR into vectype temporary and | |
5861 | VIEW_CONVERT_EXPR it into the final vectype_out result. */ | |
5862 | tree vec_cvt_dest = NULL_TREE; | |
5863 | if (orig_code == POINTER_DIFF_EXPR) | |
5864 | vec_cvt_dest = vect_create_destination_var (scalar_dest, vectype_out); | |
5865 | ||
ebfd146a IR |
5866 | /* In case the vectorization factor (VF) is bigger than the number |
5867 | of elements that we can fit in a vectype (nunits), we have to generate | |
5868 | more than one vector stmt - i.e - we need to "unroll" the | |
4a00c761 JJ |
5869 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
5870 | from one copy of the vector stmt to the next, in the field | |
5871 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following | |
5872 | stages to find the correct vector defs to be used when vectorizing | |
5873 | stmts that use the defs of the current stmt. The example below | |
5874 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., | |
5875 | we need to create 4 vectorized stmts): | |
5876 | ||
5877 | before vectorization: | |
5878 | RELATED_STMT VEC_STMT | |
5879 | S1: x = memref - - | |
5880 | S2: z = x + 1 - - | |
5881 | ||
5882 | step 1: vectorize stmt S1 (done in vectorizable_load. See more details | |
5883 | there): | |
5884 | RELATED_STMT VEC_STMT | |
5885 | VS1_0: vx0 = memref0 VS1_1 - | |
5886 | VS1_1: vx1 = memref1 VS1_2 - | |
5887 | VS1_2: vx2 = memref2 VS1_3 - | |
5888 | VS1_3: vx3 = memref3 - - | |
5889 | S1: x = load - VS1_0 | |
5890 | S2: z = x + 1 - - | |
5891 | ||
5892 | step2: vectorize stmt S2 (done here): | |
5893 | To vectorize stmt S2 we first need to find the relevant vector | |
5894 | def for the first operand 'x'. This is, as usual, obtained from | |
5895 | the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt | |
5896 | that defines 'x' (S1). This way we find the stmt VS1_0, and the | |
5897 | relevant vector def 'vx0'. Having found 'vx0' we can generate | |
5898 | the vector stmt VS2_0, and as usual, record it in the | |
5899 | STMT_VINFO_VEC_STMT of stmt S2. | |
5900 | When creating the second copy (VS2_1), we obtain the relevant vector | |
5901 | def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of | |
5902 | stmt VS1_0. This way we find the stmt VS1_1 and the relevant | |
5903 | vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a | |
5904 | pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. | |
5905 | Similarly when creating stmts VS2_2 and VS2_3. This is the resulting | |
5906 | chain of stmts and pointers: | |
5907 | RELATED_STMT VEC_STMT | |
5908 | VS1_0: vx0 = memref0 VS1_1 - | |
5909 | VS1_1: vx1 = memref1 VS1_2 - | |
5910 | VS1_2: vx2 = memref2 VS1_3 - | |
5911 | VS1_3: vx3 = memref3 - - | |
5912 | S1: x = load - VS1_0 | |
5913 | VS2_0: vz0 = vx0 + v1 VS2_1 - | |
5914 | VS2_1: vz1 = vx1 + v1 VS2_2 - | |
5915 | VS2_2: vz2 = vx2 + v1 VS2_3 - | |
5916 | VS2_3: vz3 = vx3 + v1 - - | |
5917 | S2: z = x + 1 - VS2_0 */ | |
ebfd146a IR |
5918 | |
5919 | prev_stmt_info = NULL; | |
5920 | for (j = 0; j < ncopies; j++) | |
5921 | { | |
5922 | /* Handle uses. */ | |
5923 | if (j == 0) | |
4a00c761 | 5924 | { |
d6476f90 | 5925 | if (op_type == binary_op) |
4a00c761 | 5926 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, |
306b0c92 | 5927 | slp_node); |
d6476f90 RB |
5928 | else if (op_type == ternary_op) |
5929 | { | |
5930 | if (slp_node) | |
5931 | { | |
5932 | auto_vec<tree> ops(3); | |
5933 | ops.quick_push (op0); | |
5934 | ops.quick_push (op1); | |
5935 | ops.quick_push (op2); | |
5936 | auto_vec<vec<tree> > vec_defs(3); | |
5937 | vect_get_slp_defs (ops, slp_node, &vec_defs); | |
5938 | vec_oprnds0 = vec_defs[0]; | |
5939 | vec_oprnds1 = vec_defs[1]; | |
5940 | vec_oprnds2 = vec_defs[2]; | |
5941 | } | |
5942 | else | |
5943 | { | |
5944 | vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, | |
5945 | NULL); | |
5946 | vect_get_vec_defs (op2, NULL_TREE, stmt, &vec_oprnds2, NULL, | |
5947 | NULL); | |
5948 | } | |
5949 | } | |
4a00c761 JJ |
5950 | else |
5951 | vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL, | |
306b0c92 | 5952 | slp_node); |
4a00c761 | 5953 | } |
ebfd146a | 5954 | else |
4a00c761 JJ |
5955 | { |
5956 | vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1); | |
5957 | if (op_type == ternary_op) | |
5958 | { | |
9771b263 DN |
5959 | tree vec_oprnd = vec_oprnds2.pop (); |
5960 | vec_oprnds2.quick_push (vect_get_vec_def_for_stmt_copy (dt[2], | |
5961 | vec_oprnd)); | |
4a00c761 JJ |
5962 | } |
5963 | } | |
5964 | ||
5965 | /* Arguments are ready. Create the new vector stmt. */ | |
9771b263 | 5966 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0) |
ebfd146a | 5967 | { |
4a00c761 | 5968 | vop1 = ((op_type == binary_op || op_type == ternary_op) |
9771b263 | 5969 | ? vec_oprnds1[i] : NULL_TREE); |
4a00c761 | 5970 | vop2 = ((op_type == ternary_op) |
9771b263 | 5971 | ? vec_oprnds2[i] : NULL_TREE); |
0d0e4a03 | 5972 | new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1, vop2); |
4a00c761 JJ |
5973 | new_temp = make_ssa_name (vec_dest, new_stmt); |
5974 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5975 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
0eb952ea JJ |
5976 | if (vec_cvt_dest) |
5977 | { | |
5978 | new_temp = build1 (VIEW_CONVERT_EXPR, vectype_out, new_temp); | |
5979 | new_stmt = gimple_build_assign (vec_cvt_dest, VIEW_CONVERT_EXPR, | |
5980 | new_temp); | |
5981 | new_temp = make_ssa_name (vec_cvt_dest, new_stmt); | |
5982 | gimple_assign_set_lhs (new_stmt, new_temp); | |
5983 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
5984 | } | |
4a00c761 | 5985 | if (slp_node) |
9771b263 | 5986 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
ebfd146a IR |
5987 | } |
5988 | ||
4a00c761 JJ |
5989 | if (slp_node) |
5990 | continue; | |
5991 | ||
5992 | if (j == 0) | |
5993 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
5994 | else | |
5995 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
5996 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
ebfd146a IR |
5997 | } |
5998 | ||
9771b263 DN |
5999 | vec_oprnds0.release (); |
6000 | vec_oprnds1.release (); | |
6001 | vec_oprnds2.release (); | |
ebfd146a | 6002 | |
ebfd146a IR |
6003 | return true; |
6004 | } | |
6005 | ||
f702e7d4 | 6006 | /* A helper function to ensure data reference DR's base alignment. */ |
c716e67f XDL |
6007 | |
6008 | static void | |
f702e7d4 | 6009 | ensure_base_align (struct data_reference *dr) |
c716e67f XDL |
6010 | { |
6011 | if (!dr->aux) | |
6012 | return; | |
6013 | ||
52639a61 | 6014 | if (DR_VECT_AUX (dr)->base_misaligned) |
c716e67f | 6015 | { |
52639a61 | 6016 | tree base_decl = DR_VECT_AUX (dr)->base_decl; |
c716e67f | 6017 | |
f702e7d4 RS |
6018 | unsigned int align_base_to = DR_TARGET_ALIGNMENT (dr) * BITS_PER_UNIT; |
6019 | ||
428f0c67 | 6020 | if (decl_in_symtab_p (base_decl)) |
f702e7d4 | 6021 | symtab_node::get (base_decl)->increase_alignment (align_base_to); |
428f0c67 JH |
6022 | else |
6023 | { | |
f702e7d4 | 6024 | SET_DECL_ALIGN (base_decl, align_base_to); |
428f0c67 JH |
6025 | DECL_USER_ALIGN (base_decl) = 1; |
6026 | } | |
52639a61 | 6027 | DR_VECT_AUX (dr)->base_misaligned = false; |
c716e67f XDL |
6028 | } |
6029 | } | |
6030 | ||
ebfd146a | 6031 | |
44fc7854 BE |
6032 | /* Function get_group_alias_ptr_type. |
6033 | ||
6034 | Return the alias type for the group starting at FIRST_STMT. */ | |
6035 | ||
6036 | static tree | |
6037 | get_group_alias_ptr_type (gimple *first_stmt) | |
6038 | { | |
6039 | struct data_reference *first_dr, *next_dr; | |
6040 | gimple *next_stmt; | |
6041 | ||
6042 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
6043 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (first_stmt)); | |
6044 | while (next_stmt) | |
6045 | { | |
6046 | next_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (next_stmt)); | |
6047 | if (get_alias_set (DR_REF (first_dr)) | |
6048 | != get_alias_set (DR_REF (next_dr))) | |
6049 | { | |
6050 | if (dump_enabled_p ()) | |
6051 | dump_printf_loc (MSG_NOTE, vect_location, | |
6052 | "conflicting alias set types.\n"); | |
6053 | return ptr_type_node; | |
6054 | } | |
6055 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); | |
6056 | } | |
6057 | return reference_alias_ptr_type (DR_REF (first_dr)); | |
6058 | } | |
6059 | ||
6060 | ||
ebfd146a IR |
6061 | /* Function vectorizable_store. |
6062 | ||
b8698a0f L |
6063 | Check if STMT defines a non scalar data-ref (array/pointer/structure) that |
6064 | can be vectorized. | |
6065 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
6066 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
6067 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
6068 | ||
6069 | static bool | |
355fe088 | 6070 | vectorizable_store (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 6071 | slp_tree slp_node) |
ebfd146a | 6072 | { |
ebfd146a IR |
6073 | tree data_ref; |
6074 | tree op; | |
6075 | tree vec_oprnd = NULL_TREE; | |
6076 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
6077 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; | |
272c6793 | 6078 | tree elem_type; |
ebfd146a | 6079 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 6080 | struct loop *loop = NULL; |
ef4bddc2 | 6081 | machine_mode vec_mode; |
ebfd146a IR |
6082 | tree dummy; |
6083 | enum dr_alignment_support alignment_support_scheme; | |
355fe088 | 6084 | gimple *def_stmt; |
929b4411 RS |
6085 | enum vect_def_type rhs_dt = vect_unknown_def_type; |
6086 | enum vect_def_type mask_dt = vect_unknown_def_type; | |
ebfd146a IR |
6087 | stmt_vec_info prev_stmt_info = NULL; |
6088 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 6089 | tree dataref_offset = NULL_TREE; |
355fe088 | 6090 | gimple *ptr_incr = NULL; |
ebfd146a IR |
6091 | int ncopies; |
6092 | int j; | |
2de001ee RS |
6093 | gimple *next_stmt, *first_stmt; |
6094 | bool grouped_store; | |
ebfd146a | 6095 | unsigned int group_size, i; |
6e1aa848 DN |
6096 | vec<tree> oprnds = vNULL; |
6097 | vec<tree> result_chain = vNULL; | |
ebfd146a | 6098 | bool inv_p; |
09dfa495 | 6099 | tree offset = NULL_TREE; |
6e1aa848 | 6100 | vec<tree> vec_oprnds = vNULL; |
ebfd146a | 6101 | bool slp = (slp_node != NULL); |
ebfd146a | 6102 | unsigned int vec_num; |
a70d6342 | 6103 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
310213d4 | 6104 | vec_info *vinfo = stmt_info->vinfo; |
272c6793 | 6105 | tree aggr_type; |
134c85ca | 6106 | gather_scatter_info gs_info; |
355fe088 | 6107 | gimple *new_stmt; |
d9f21f6a | 6108 | poly_uint64 vf; |
2de001ee | 6109 | vec_load_store_type vls_type; |
44fc7854 | 6110 | tree ref_type; |
a70d6342 | 6111 | |
a70d6342 | 6112 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
ebfd146a IR |
6113 | return false; |
6114 | ||
66c16fd9 RB |
6115 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
6116 | && ! vec_stmt) | |
ebfd146a IR |
6117 | return false; |
6118 | ||
6119 | /* Is vectorizable store? */ | |
6120 | ||
c3a8f964 RS |
6121 | tree mask = NULL_TREE, mask_vectype = NULL_TREE; |
6122 | if (is_gimple_assign (stmt)) | |
6123 | { | |
6124 | tree scalar_dest = gimple_assign_lhs (stmt); | |
6125 | if (TREE_CODE (scalar_dest) == VIEW_CONVERT_EXPR | |
6126 | && is_pattern_stmt_p (stmt_info)) | |
6127 | scalar_dest = TREE_OPERAND (scalar_dest, 0); | |
6128 | if (TREE_CODE (scalar_dest) != ARRAY_REF | |
6129 | && TREE_CODE (scalar_dest) != BIT_FIELD_REF | |
6130 | && TREE_CODE (scalar_dest) != INDIRECT_REF | |
6131 | && TREE_CODE (scalar_dest) != COMPONENT_REF | |
6132 | && TREE_CODE (scalar_dest) != IMAGPART_EXPR | |
6133 | && TREE_CODE (scalar_dest) != REALPART_EXPR | |
6134 | && TREE_CODE (scalar_dest) != MEM_REF) | |
6135 | return false; | |
6136 | } | |
6137 | else | |
6138 | { | |
6139 | gcall *call = dyn_cast <gcall *> (stmt); | |
f307441a RS |
6140 | if (!call || !gimple_call_internal_p (call)) |
6141 | return false; | |
6142 | ||
6143 | internal_fn ifn = gimple_call_internal_fn (call); | |
6144 | if (!internal_store_fn_p (ifn)) | |
c3a8f964 | 6145 | return false; |
ebfd146a | 6146 | |
c3a8f964 RS |
6147 | if (slp_node != NULL) |
6148 | { | |
6149 | if (dump_enabled_p ()) | |
6150 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6151 | "SLP of masked stores not supported.\n"); | |
6152 | return false; | |
6153 | } | |
6154 | ||
f307441a RS |
6155 | int mask_index = internal_fn_mask_index (ifn); |
6156 | if (mask_index >= 0) | |
6157 | { | |
6158 | mask = gimple_call_arg (call, mask_index); | |
929b4411 RS |
6159 | if (!vect_check_load_store_mask (stmt, mask, &mask_dt, |
6160 | &mask_vectype)) | |
f307441a RS |
6161 | return false; |
6162 | } | |
c3a8f964 RS |
6163 | } |
6164 | ||
6165 | op = vect_get_store_rhs (stmt); | |
ebfd146a | 6166 | |
fce57248 RS |
6167 | /* Cannot have hybrid store SLP -- that would mean storing to the |
6168 | same location twice. */ | |
6169 | gcc_assert (slp == PURE_SLP_STMT (stmt_info)); | |
6170 | ||
f4d09712 | 6171 | tree vectype = STMT_VINFO_VECTYPE (stmt_info), rhs_vectype = NULL_TREE; |
4d694b27 | 6172 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 JJ |
6173 | |
6174 | if (loop_vinfo) | |
b17dc4d4 RB |
6175 | { |
6176 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
6177 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
6178 | } | |
6179 | else | |
6180 | vf = 1; | |
465c8c19 JJ |
6181 | |
6182 | /* Multiple types in SLP are handled by creating the appropriate number of | |
6183 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in | |
6184 | case of SLP. */ | |
fce57248 | 6185 | if (slp) |
465c8c19 JJ |
6186 | ncopies = 1; |
6187 | else | |
e8f142e2 | 6188 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
6189 | |
6190 | gcc_assert (ncopies >= 1); | |
6191 | ||
6192 | /* FORNOW. This restriction should be relaxed. */ | |
6193 | if (loop && nested_in_vect_loop_p (loop, stmt) && ncopies > 1) | |
6194 | { | |
6195 | if (dump_enabled_p ()) | |
6196 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6197 | "multiple types in nested loop.\n"); | |
6198 | return false; | |
6199 | } | |
6200 | ||
929b4411 | 6201 | if (!vect_check_store_rhs (stmt, op, &rhs_dt, &rhs_vectype, &vls_type)) |
f4d09712 KY |
6202 | return false; |
6203 | ||
272c6793 | 6204 | elem_type = TREE_TYPE (vectype); |
ebfd146a | 6205 | vec_mode = TYPE_MODE (vectype); |
7b7b1813 | 6206 | |
ebfd146a IR |
6207 | if (!STMT_VINFO_DATA_REF (stmt_info)) |
6208 | return false; | |
6209 | ||
2de001ee | 6210 | vect_memory_access_type memory_access_type; |
7e11fc7f | 6211 | if (!get_load_store_type (stmt, vectype, slp, mask, vls_type, ncopies, |
2de001ee RS |
6212 | &memory_access_type, &gs_info)) |
6213 | return false; | |
3bab6342 | 6214 | |
c3a8f964 RS |
6215 | if (mask) |
6216 | { | |
7e11fc7f RS |
6217 | if (memory_access_type == VMAT_CONTIGUOUS) |
6218 | { | |
6219 | if (!VECTOR_MODE_P (vec_mode) | |
6220 | || !can_vec_mask_load_store_p (vec_mode, | |
6221 | TYPE_MODE (mask_vectype), false)) | |
6222 | return false; | |
6223 | } | |
f307441a RS |
6224 | else if (memory_access_type != VMAT_LOAD_STORE_LANES |
6225 | && (memory_access_type != VMAT_GATHER_SCATTER || gs_info.decl)) | |
c3a8f964 RS |
6226 | { |
6227 | if (dump_enabled_p ()) | |
6228 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
6229 | "unsupported access type for masked store.\n"); | |
6230 | return false; | |
6231 | } | |
c3a8f964 RS |
6232 | } |
6233 | else | |
6234 | { | |
6235 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
6236 | (e.g. - array initialization with 0). */ | |
6237 | if (optab_handler (mov_optab, vec_mode) == CODE_FOR_nothing) | |
6238 | return false; | |
6239 | } | |
6240 | ||
f307441a | 6241 | grouped_store = (STMT_VINFO_GROUPED_ACCESS (stmt_info) |
b5ec4de7 RS |
6242 | && memory_access_type != VMAT_GATHER_SCATTER |
6243 | && (slp || memory_access_type != VMAT_CONTIGUOUS)); | |
7cfb4d93 RS |
6244 | if (grouped_store) |
6245 | { | |
6246 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); | |
6247 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
6248 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
6249 | } | |
6250 | else | |
6251 | { | |
6252 | first_stmt = stmt; | |
6253 | first_dr = dr; | |
6254 | group_size = vec_num = 1; | |
6255 | } | |
6256 | ||
ebfd146a IR |
6257 | if (!vec_stmt) /* transformation not required. */ |
6258 | { | |
2de001ee | 6259 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; |
7cfb4d93 RS |
6260 | |
6261 | if (loop_vinfo | |
6262 | && LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo)) | |
6263 | check_load_store_masking (loop_vinfo, vectype, vls_type, group_size, | |
bfaa08b7 | 6264 | memory_access_type, &gs_info); |
7cfb4d93 | 6265 | |
ebfd146a | 6266 | STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; |
2e8ab70c | 6267 | /* The SLP costs are calculated during SLP analysis. */ |
78604de0 | 6268 | if (!slp_node) |
9ce4345a RS |
6269 | vect_model_store_cost (stmt_info, ncopies, memory_access_type, |
6270 | vls_type, NULL, NULL, NULL); | |
ebfd146a IR |
6271 | return true; |
6272 | } | |
2de001ee | 6273 | gcc_assert (memory_access_type == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); |
ebfd146a | 6274 | |
67b8dbac | 6275 | /* Transform. */ |
ebfd146a | 6276 | |
f702e7d4 | 6277 | ensure_base_align (dr); |
c716e67f | 6278 | |
f307441a | 6279 | if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.decl) |
3bab6342 | 6280 | { |
c3a8f964 | 6281 | tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE, src; |
134c85ca | 6282 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
6283 | tree rettype, srctype, ptrtype, idxtype, masktype, scaletype; |
6284 | tree ptr, mask, var, scale, perm_mask = NULL_TREE; | |
6285 | edge pe = loop_preheader_edge (loop); | |
6286 | gimple_seq seq; | |
6287 | basic_block new_bb; | |
6288 | enum { NARROW, NONE, WIDEN } modifier; | |
4d694b27 RS |
6289 | poly_uint64 scatter_off_nunits |
6290 | = TYPE_VECTOR_SUBPARTS (gs_info.offset_vectype); | |
3bab6342 | 6291 | |
4d694b27 | 6292 | if (known_eq (nunits, scatter_off_nunits)) |
3bab6342 | 6293 | modifier = NONE; |
4d694b27 | 6294 | else if (known_eq (nunits * 2, scatter_off_nunits)) |
3bab6342 | 6295 | { |
3bab6342 AT |
6296 | modifier = WIDEN; |
6297 | ||
4d694b27 RS |
6298 | /* Currently gathers and scatters are only supported for |
6299 | fixed-length vectors. */ | |
6300 | unsigned int count = scatter_off_nunits.to_constant (); | |
6301 | vec_perm_builder sel (count, count, 1); | |
6302 | for (i = 0; i < (unsigned int) count; ++i) | |
6303 | sel.quick_push (i | (count / 2)); | |
3bab6342 | 6304 | |
4d694b27 | 6305 | vec_perm_indices indices (sel, 1, count); |
e3342de4 RS |
6306 | perm_mask = vect_gen_perm_mask_checked (gs_info.offset_vectype, |
6307 | indices); | |
3bab6342 AT |
6308 | gcc_assert (perm_mask != NULL_TREE); |
6309 | } | |
4d694b27 | 6310 | else if (known_eq (nunits, scatter_off_nunits * 2)) |
3bab6342 | 6311 | { |
3bab6342 AT |
6312 | modifier = NARROW; |
6313 | ||
4d694b27 RS |
6314 | /* Currently gathers and scatters are only supported for |
6315 | fixed-length vectors. */ | |
6316 | unsigned int count = nunits.to_constant (); | |
6317 | vec_perm_builder sel (count, count, 1); | |
6318 | for (i = 0; i < (unsigned int) count; ++i) | |
6319 | sel.quick_push (i | (count / 2)); | |
3bab6342 | 6320 | |
4d694b27 | 6321 | vec_perm_indices indices (sel, 2, count); |
e3342de4 | 6322 | perm_mask = vect_gen_perm_mask_checked (vectype, indices); |
3bab6342 AT |
6323 | gcc_assert (perm_mask != NULL_TREE); |
6324 | ncopies *= 2; | |
6325 | } | |
6326 | else | |
6327 | gcc_unreachable (); | |
6328 | ||
134c85ca | 6329 | rettype = TREE_TYPE (TREE_TYPE (gs_info.decl)); |
3bab6342 AT |
6330 | ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); |
6331 | masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6332 | idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6333 | srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist); | |
6334 | scaletype = TREE_VALUE (arglist); | |
6335 | ||
6336 | gcc_checking_assert (TREE_CODE (masktype) == INTEGER_TYPE | |
6337 | && TREE_CODE (rettype) == VOID_TYPE); | |
6338 | ||
134c85ca | 6339 | ptr = fold_convert (ptrtype, gs_info.base); |
3bab6342 AT |
6340 | if (!is_gimple_min_invariant (ptr)) |
6341 | { | |
6342 | ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE); | |
6343 | new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
6344 | gcc_assert (!new_bb); | |
6345 | } | |
6346 | ||
6347 | /* Currently we support only unconditional scatter stores, | |
6348 | so mask should be all ones. */ | |
6349 | mask = build_int_cst (masktype, -1); | |
6350 | mask = vect_init_vector (stmt, mask, masktype, NULL); | |
6351 | ||
134c85ca | 6352 | scale = build_int_cst (scaletype, gs_info.scale); |
3bab6342 AT |
6353 | |
6354 | prev_stmt_info = NULL; | |
6355 | for (j = 0; j < ncopies; ++j) | |
6356 | { | |
6357 | if (j == 0) | |
6358 | { | |
6359 | src = vec_oprnd1 | |
c3a8f964 | 6360 | = vect_get_vec_def_for_operand (op, stmt); |
3bab6342 | 6361 | op = vec_oprnd0 |
134c85ca | 6362 | = vect_get_vec_def_for_operand (gs_info.offset, stmt); |
3bab6342 AT |
6363 | } |
6364 | else if (modifier != NONE && (j & 1)) | |
6365 | { | |
6366 | if (modifier == WIDEN) | |
6367 | { | |
6368 | src = vec_oprnd1 | |
929b4411 | 6369 | = vect_get_vec_def_for_stmt_copy (rhs_dt, vec_oprnd1); |
3bab6342 AT |
6370 | op = permute_vec_elements (vec_oprnd0, vec_oprnd0, perm_mask, |
6371 | stmt, gsi); | |
6372 | } | |
6373 | else if (modifier == NARROW) | |
6374 | { | |
6375 | src = permute_vec_elements (vec_oprnd1, vec_oprnd1, perm_mask, | |
6376 | stmt, gsi); | |
6377 | op = vec_oprnd0 | |
134c85ca RS |
6378 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
6379 | vec_oprnd0); | |
3bab6342 AT |
6380 | } |
6381 | else | |
6382 | gcc_unreachable (); | |
6383 | } | |
6384 | else | |
6385 | { | |
6386 | src = vec_oprnd1 | |
929b4411 | 6387 | = vect_get_vec_def_for_stmt_copy (rhs_dt, vec_oprnd1); |
3bab6342 | 6388 | op = vec_oprnd0 |
134c85ca RS |
6389 | = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
6390 | vec_oprnd0); | |
3bab6342 AT |
6391 | } |
6392 | ||
6393 | if (!useless_type_conversion_p (srctype, TREE_TYPE (src))) | |
6394 | { | |
928686b1 RS |
6395 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (src)), |
6396 | TYPE_VECTOR_SUBPARTS (srctype))); | |
0e22bb5a | 6397 | var = vect_get_new_ssa_name (srctype, vect_simple_var); |
3bab6342 AT |
6398 | src = build1 (VIEW_CONVERT_EXPR, srctype, src); |
6399 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, src); | |
6400 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6401 | src = var; | |
6402 | } | |
6403 | ||
6404 | if (!useless_type_conversion_p (idxtype, TREE_TYPE (op))) | |
6405 | { | |
928686b1 RS |
6406 | gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)), |
6407 | TYPE_VECTOR_SUBPARTS (idxtype))); | |
0e22bb5a | 6408 | var = vect_get_new_ssa_name (idxtype, vect_simple_var); |
3bab6342 AT |
6409 | op = build1 (VIEW_CONVERT_EXPR, idxtype, op); |
6410 | new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op); | |
6411 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6412 | op = var; | |
6413 | } | |
6414 | ||
6415 | new_stmt | |
134c85ca | 6416 | = gimple_build_call (gs_info.decl, 5, ptr, mask, op, src, scale); |
3bab6342 AT |
6417 | |
6418 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
6419 | ||
6420 | if (prev_stmt_info == NULL) | |
6421 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
6422 | else | |
6423 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
6424 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
6425 | } | |
6426 | return true; | |
6427 | } | |
6428 | ||
f307441a | 6429 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 6430 | { |
f307441a RS |
6431 | gimple *group_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
6432 | GROUP_STORE_COUNT (vinfo_for_stmt (group_stmt))++; | |
6433 | } | |
ebfd146a | 6434 | |
f307441a RS |
6435 | if (grouped_store) |
6436 | { | |
ebfd146a | 6437 | /* FORNOW */ |
a70d6342 | 6438 | gcc_assert (!loop || !nested_in_vect_loop_p (loop, stmt)); |
ebfd146a IR |
6439 | |
6440 | /* We vectorize all the stmts of the interleaving group when we | |
6441 | reach the last stmt in the group. */ | |
e14c1050 IR |
6442 | if (GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) |
6443 | < GROUP_SIZE (vinfo_for_stmt (first_stmt)) | |
ebfd146a IR |
6444 | && !slp) |
6445 | { | |
6446 | *vec_stmt = NULL; | |
6447 | return true; | |
6448 | } | |
6449 | ||
6450 | if (slp) | |
4b5caab7 | 6451 | { |
0d0293ac | 6452 | grouped_store = false; |
4b5caab7 IR |
6453 | /* VEC_NUM is the number of vect stmts to be created for this |
6454 | group. */ | |
6455 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
9771b263 | 6456 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
52eab378 | 6457 | gcc_assert (GROUP_FIRST_ELEMENT (vinfo_for_stmt (first_stmt)) == first_stmt); |
4b5caab7 | 6458 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); |
c3a8f964 | 6459 | op = vect_get_store_rhs (first_stmt); |
4b5caab7 | 6460 | } |
ebfd146a | 6461 | else |
4b5caab7 IR |
6462 | /* VEC_NUM is the number of vect stmts to be created for this |
6463 | group. */ | |
ebfd146a | 6464 | vec_num = group_size; |
44fc7854 BE |
6465 | |
6466 | ref_type = get_group_alias_ptr_type (first_stmt); | |
ebfd146a | 6467 | } |
b8698a0f | 6468 | else |
7cfb4d93 | 6469 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); |
b8698a0f | 6470 | |
73fbfcad | 6471 | if (dump_enabled_p ()) |
78c60e3d | 6472 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 6473 | "transform store. ncopies = %d\n", ncopies); |
ebfd146a | 6474 | |
2de001ee RS |
6475 | if (memory_access_type == VMAT_ELEMENTWISE |
6476 | || memory_access_type == VMAT_STRIDED_SLP) | |
f2e2a985 MM |
6477 | { |
6478 | gimple_stmt_iterator incr_gsi; | |
6479 | bool insert_after; | |
355fe088 | 6480 | gimple *incr; |
f2e2a985 MM |
6481 | tree offvar; |
6482 | tree ivstep; | |
6483 | tree running_off; | |
f2e2a985 MM |
6484 | tree stride_base, stride_step, alias_off; |
6485 | tree vec_oprnd; | |
f502d50e | 6486 | unsigned int g; |
4d694b27 RS |
6487 | /* Checked by get_load_store_type. */ |
6488 | unsigned int const_nunits = nunits.to_constant (); | |
f2e2a985 | 6489 | |
7cfb4d93 | 6490 | gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); |
f2e2a985 MM |
6491 | gcc_assert (!nested_in_vect_loop_p (loop, stmt)); |
6492 | ||
6493 | stride_base | |
6494 | = fold_build_pointer_plus | |
b210f45f | 6495 | (DR_BASE_ADDRESS (first_dr), |
f2e2a985 | 6496 | size_binop (PLUS_EXPR, |
b210f45f | 6497 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
44fc7854 | 6498 | convert_to_ptrofftype (DR_INIT (first_dr)))); |
b210f45f | 6499 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); |
f2e2a985 MM |
6500 | |
6501 | /* For a store with loop-invariant (but other than power-of-2) | |
6502 | stride (i.e. not a grouped access) like so: | |
6503 | ||
6504 | for (i = 0; i < n; i += stride) | |
6505 | array[i] = ...; | |
6506 | ||
6507 | we generate a new induction variable and new stores from | |
6508 | the components of the (vectorized) rhs: | |
6509 | ||
6510 | for (j = 0; ; j += VF*stride) | |
6511 | vectemp = ...; | |
6512 | tmp1 = vectemp[0]; | |
6513 | array[j] = tmp1; | |
6514 | tmp2 = vectemp[1]; | |
6515 | array[j + stride] = tmp2; | |
6516 | ... | |
6517 | */ | |
6518 | ||
4d694b27 | 6519 | unsigned nstores = const_nunits; |
b17dc4d4 | 6520 | unsigned lnel = 1; |
cee62fee | 6521 | tree ltype = elem_type; |
04199738 | 6522 | tree lvectype = vectype; |
cee62fee MM |
6523 | if (slp) |
6524 | { | |
4d694b27 RS |
6525 | if (group_size < const_nunits |
6526 | && const_nunits % group_size == 0) | |
b17dc4d4 | 6527 | { |
4d694b27 | 6528 | nstores = const_nunits / group_size; |
b17dc4d4 RB |
6529 | lnel = group_size; |
6530 | ltype = build_vector_type (elem_type, group_size); | |
04199738 RB |
6531 | lvectype = vectype; |
6532 | ||
6533 | /* First check if vec_extract optab doesn't support extraction | |
6534 | of vector elts directly. */ | |
b397965c | 6535 | scalar_mode elmode = SCALAR_TYPE_MODE (elem_type); |
9da15d40 RS |
6536 | machine_mode vmode; |
6537 | if (!mode_for_vector (elmode, group_size).exists (&vmode) | |
6538 | || !VECTOR_MODE_P (vmode) | |
414fef4e | 6539 | || !targetm.vector_mode_supported_p (vmode) |
04199738 RB |
6540 | || (convert_optab_handler (vec_extract_optab, |
6541 | TYPE_MODE (vectype), vmode) | |
6542 | == CODE_FOR_nothing)) | |
6543 | { | |
6544 | /* Try to avoid emitting an extract of vector elements | |
6545 | by performing the extracts using an integer type of the | |
6546 | same size, extracting from a vector of those and then | |
6547 | re-interpreting it as the original vector type if | |
6548 | supported. */ | |
6549 | unsigned lsize | |
6550 | = group_size * GET_MODE_BITSIZE (elmode); | |
fffbab82 | 6551 | elmode = int_mode_for_size (lsize, 0).require (); |
4d694b27 | 6552 | unsigned int lnunits = const_nunits / group_size; |
04199738 RB |
6553 | /* If we can't construct such a vector fall back to |
6554 | element extracts from the original vector type and | |
6555 | element size stores. */ | |
4d694b27 | 6556 | if (mode_for_vector (elmode, lnunits).exists (&vmode) |
9da15d40 | 6557 | && VECTOR_MODE_P (vmode) |
414fef4e | 6558 | && targetm.vector_mode_supported_p (vmode) |
04199738 RB |
6559 | && (convert_optab_handler (vec_extract_optab, |
6560 | vmode, elmode) | |
6561 | != CODE_FOR_nothing)) | |
6562 | { | |
4d694b27 | 6563 | nstores = lnunits; |
04199738 RB |
6564 | lnel = group_size; |
6565 | ltype = build_nonstandard_integer_type (lsize, 1); | |
6566 | lvectype = build_vector_type (ltype, nstores); | |
6567 | } | |
6568 | /* Else fall back to vector extraction anyway. | |
6569 | Fewer stores are more important than avoiding spilling | |
6570 | of the vector we extract from. Compared to the | |
6571 | construction case in vectorizable_load no store-forwarding | |
6572 | issue exists here for reasonable archs. */ | |
6573 | } | |
b17dc4d4 | 6574 | } |
4d694b27 RS |
6575 | else if (group_size >= const_nunits |
6576 | && group_size % const_nunits == 0) | |
b17dc4d4 RB |
6577 | { |
6578 | nstores = 1; | |
4d694b27 | 6579 | lnel = const_nunits; |
b17dc4d4 | 6580 | ltype = vectype; |
04199738 | 6581 | lvectype = vectype; |
b17dc4d4 | 6582 | } |
cee62fee MM |
6583 | ltype = build_aligned_type (ltype, TYPE_ALIGN (elem_type)); |
6584 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
6585 | } | |
6586 | ||
f2e2a985 MM |
6587 | ivstep = stride_step; |
6588 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (ivstep), ivstep, | |
b17dc4d4 | 6589 | build_int_cst (TREE_TYPE (ivstep), vf)); |
f2e2a985 MM |
6590 | |
6591 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
6592 | ||
b210f45f RB |
6593 | stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base); |
6594 | ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep); | |
f2e2a985 MM |
6595 | create_iv (stride_base, ivstep, NULL, |
6596 | loop, &incr_gsi, insert_after, | |
6597 | &offvar, NULL); | |
6598 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 6599 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
f2e2a985 | 6600 | |
b210f45f | 6601 | stride_step = cse_and_gimplify_to_preheader (loop_vinfo, stride_step); |
f2e2a985 MM |
6602 | |
6603 | prev_stmt_info = NULL; | |
44fc7854 | 6604 | alias_off = build_int_cst (ref_type, 0); |
f502d50e MM |
6605 | next_stmt = first_stmt; |
6606 | for (g = 0; g < group_size; g++) | |
f2e2a985 | 6607 | { |
f502d50e MM |
6608 | running_off = offvar; |
6609 | if (g) | |
f2e2a985 | 6610 | { |
f502d50e MM |
6611 | tree size = TYPE_SIZE_UNIT (ltype); |
6612 | tree pos = fold_build2 (MULT_EXPR, sizetype, size_int (g), | |
f2e2a985 | 6613 | size); |
f502d50e | 6614 | tree newoff = copy_ssa_name (running_off, NULL); |
f2e2a985 | 6615 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, |
f502d50e | 6616 | running_off, pos); |
f2e2a985 | 6617 | vect_finish_stmt_generation (stmt, incr, gsi); |
f2e2a985 | 6618 | running_off = newoff; |
f502d50e | 6619 | } |
b17dc4d4 RB |
6620 | unsigned int group_el = 0; |
6621 | unsigned HOST_WIDE_INT | |
6622 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
f502d50e MM |
6623 | for (j = 0; j < ncopies; j++) |
6624 | { | |
c3a8f964 | 6625 | /* We've set op and dt above, from vect_get_store_rhs, |
f502d50e MM |
6626 | and first_stmt == stmt. */ |
6627 | if (j == 0) | |
6628 | { | |
6629 | if (slp) | |
6630 | { | |
6631 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, NULL, | |
306b0c92 | 6632 | slp_node); |
f502d50e MM |
6633 | vec_oprnd = vec_oprnds[0]; |
6634 | } | |
6635 | else | |
6636 | { | |
c3a8f964 | 6637 | op = vect_get_store_rhs (next_stmt); |
81c40241 | 6638 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
f502d50e MM |
6639 | } |
6640 | } | |
f2e2a985 | 6641 | else |
f502d50e MM |
6642 | { |
6643 | if (slp) | |
6644 | vec_oprnd = vec_oprnds[j]; | |
6645 | else | |
c079cbac | 6646 | { |
929b4411 RS |
6647 | vect_is_simple_use (op, vinfo, &def_stmt, &rhs_dt); |
6648 | vec_oprnd = vect_get_vec_def_for_stmt_copy (rhs_dt, | |
6649 | vec_oprnd); | |
c079cbac | 6650 | } |
f502d50e | 6651 | } |
04199738 RB |
6652 | /* Pun the vector to extract from if necessary. */ |
6653 | if (lvectype != vectype) | |
6654 | { | |
6655 | tree tem = make_ssa_name (lvectype); | |
6656 | gimple *pun | |
6657 | = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR, | |
6658 | lvectype, vec_oprnd)); | |
6659 | vect_finish_stmt_generation (stmt, pun, gsi); | |
6660 | vec_oprnd = tem; | |
6661 | } | |
f502d50e MM |
6662 | for (i = 0; i < nstores; i++) |
6663 | { | |
6664 | tree newref, newoff; | |
355fe088 | 6665 | gimple *incr, *assign; |
f502d50e MM |
6666 | tree size = TYPE_SIZE (ltype); |
6667 | /* Extract the i'th component. */ | |
6668 | tree pos = fold_build2 (MULT_EXPR, bitsizetype, | |
6669 | bitsize_int (i), size); | |
6670 | tree elem = fold_build3 (BIT_FIELD_REF, ltype, vec_oprnd, | |
6671 | size, pos); | |
6672 | ||
6673 | elem = force_gimple_operand_gsi (gsi, elem, true, | |
6674 | NULL_TREE, true, | |
6675 | GSI_SAME_STMT); | |
6676 | ||
b17dc4d4 RB |
6677 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
6678 | group_el * elsz); | |
f502d50e | 6679 | newref = build2 (MEM_REF, ltype, |
b17dc4d4 | 6680 | running_off, this_off); |
19986382 | 6681 | vect_copy_ref_info (newref, DR_REF (first_dr)); |
f502d50e MM |
6682 | |
6683 | /* And store it to *running_off. */ | |
6684 | assign = gimple_build_assign (newref, elem); | |
6685 | vect_finish_stmt_generation (stmt, assign, gsi); | |
6686 | ||
b17dc4d4 RB |
6687 | group_el += lnel; |
6688 | if (! slp | |
6689 | || group_el == group_size) | |
6690 | { | |
6691 | newoff = copy_ssa_name (running_off, NULL); | |
6692 | incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
6693 | running_off, stride_step); | |
6694 | vect_finish_stmt_generation (stmt, incr, gsi); | |
f502d50e | 6695 | |
b17dc4d4 RB |
6696 | running_off = newoff; |
6697 | group_el = 0; | |
6698 | } | |
225ce44b RB |
6699 | if (g == group_size - 1 |
6700 | && !slp) | |
f502d50e MM |
6701 | { |
6702 | if (j == 0 && i == 0) | |
225ce44b RB |
6703 | STMT_VINFO_VEC_STMT (stmt_info) |
6704 | = *vec_stmt = assign; | |
f502d50e MM |
6705 | else |
6706 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = assign; | |
6707 | prev_stmt_info = vinfo_for_stmt (assign); | |
6708 | } | |
6709 | } | |
f2e2a985 | 6710 | } |
f502d50e | 6711 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
b17dc4d4 RB |
6712 | if (slp) |
6713 | break; | |
f2e2a985 | 6714 | } |
778dd3b6 RB |
6715 | |
6716 | vec_oprnds.release (); | |
f2e2a985 MM |
6717 | return true; |
6718 | } | |
6719 | ||
8c681247 | 6720 | auto_vec<tree> dr_chain (group_size); |
9771b263 | 6721 | oprnds.create (group_size); |
ebfd146a | 6722 | |
720f5239 | 6723 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 6724 | gcc_assert (alignment_support_scheme); |
70088b95 RS |
6725 | vec_loop_masks *loop_masks |
6726 | = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) | |
6727 | ? &LOOP_VINFO_MASKS (loop_vinfo) | |
6728 | : NULL); | |
272c6793 | 6729 | /* Targets with store-lane instructions must not require explicit |
c3a8f964 RS |
6730 | realignment. vect_supportable_dr_alignment always returns either |
6731 | dr_aligned or dr_unaligned_supported for masked operations. */ | |
7cfb4d93 RS |
6732 | gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES |
6733 | && !mask | |
70088b95 | 6734 | && !loop_masks) |
272c6793 RS |
6735 | || alignment_support_scheme == dr_aligned |
6736 | || alignment_support_scheme == dr_unaligned_supported); | |
6737 | ||
62da9e14 RS |
6738 | if (memory_access_type == VMAT_CONTIGUOUS_DOWN |
6739 | || memory_access_type == VMAT_CONTIGUOUS_REVERSE) | |
09dfa495 BM |
6740 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
6741 | ||
f307441a RS |
6742 | tree bump; |
6743 | tree vec_offset = NULL_TREE; | |
6744 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
6745 | { | |
6746 | aggr_type = NULL_TREE; | |
6747 | bump = NULL_TREE; | |
6748 | } | |
6749 | else if (memory_access_type == VMAT_GATHER_SCATTER) | |
6750 | { | |
6751 | aggr_type = elem_type; | |
6752 | vect_get_strided_load_store_ops (stmt, loop_vinfo, &gs_info, | |
6753 | &bump, &vec_offset); | |
6754 | } | |
272c6793 | 6755 | else |
f307441a RS |
6756 | { |
6757 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
6758 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
6759 | else | |
6760 | aggr_type = vectype; | |
6761 | bump = vect_get_data_ptr_increment (dr, aggr_type, memory_access_type); | |
6762 | } | |
ebfd146a | 6763 | |
c3a8f964 RS |
6764 | if (mask) |
6765 | LOOP_VINFO_HAS_MASK_STORE (loop_vinfo) = true; | |
6766 | ||
ebfd146a IR |
6767 | /* In case the vectorization factor (VF) is bigger than the number |
6768 | of elements that we can fit in a vectype (nunits), we have to generate | |
6769 | more than one vector stmt - i.e - we need to "unroll" the | |
b8698a0f | 6770 | vector stmt by a factor VF/nunits. For more details see documentation in |
ebfd146a IR |
6771 | vect_get_vec_def_for_copy_stmt. */ |
6772 | ||
0d0293ac | 6773 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
6774 | |
6775 | S1: &base + 2 = x2 | |
6776 | S2: &base = x0 | |
6777 | S3: &base + 1 = x1 | |
6778 | S4: &base + 3 = x3 | |
6779 | ||
6780 | We create vectorized stores starting from base address (the access of the | |
6781 | first stmt in the chain (S2 in the above example), when the last store stmt | |
6782 | of the chain (S4) is reached: | |
6783 | ||
6784 | VS1: &base = vx2 | |
6785 | VS2: &base + vec_size*1 = vx0 | |
6786 | VS3: &base + vec_size*2 = vx1 | |
6787 | VS4: &base + vec_size*3 = vx3 | |
6788 | ||
6789 | Then permutation statements are generated: | |
6790 | ||
3fcc1b55 JJ |
6791 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} > |
6792 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} > | |
ebfd146a | 6793 | ... |
b8698a0f | 6794 | |
ebfd146a IR |
6795 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts |
6796 | (the order of the data-refs in the output of vect_permute_store_chain | |
6797 | corresponds to the order of scalar stmts in the interleaving chain - see | |
6798 | the documentation of vect_permute_store_chain()). | |
6799 | ||
6800 | In case of both multiple types and interleaving, above vector stores and | |
ff802fa1 | 6801 | permutation stmts are created for every copy. The result vector stmts are |
ebfd146a | 6802 | put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding |
b8698a0f | 6803 | STMT_VINFO_RELATED_STMT for the next copies. |
ebfd146a IR |
6804 | */ |
6805 | ||
6806 | prev_stmt_info = NULL; | |
c3a8f964 | 6807 | tree vec_mask = NULL_TREE; |
ebfd146a IR |
6808 | for (j = 0; j < ncopies; j++) |
6809 | { | |
ebfd146a IR |
6810 | |
6811 | if (j == 0) | |
6812 | { | |
6813 | if (slp) | |
6814 | { | |
6815 | /* Get vectorized arguments for SLP_NODE. */ | |
d092494c | 6816 | vect_get_vec_defs (op, NULL_TREE, stmt, &vec_oprnds, |
306b0c92 | 6817 | NULL, slp_node); |
ebfd146a | 6818 | |
9771b263 | 6819 | vec_oprnd = vec_oprnds[0]; |
ebfd146a IR |
6820 | } |
6821 | else | |
6822 | { | |
b8698a0f L |
6823 | /* For interleaved stores we collect vectorized defs for all the |
6824 | stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then | |
6825 | used as an input to vect_permute_store_chain(), and OPRNDS as | |
ebfd146a IR |
6826 | an input to vect_get_vec_def_for_stmt_copy() for the next copy. |
6827 | ||
0d0293ac | 6828 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a | 6829 | OPRNDS are of size 1. */ |
b8698a0f | 6830 | next_stmt = first_stmt; |
ebfd146a IR |
6831 | for (i = 0; i < group_size; i++) |
6832 | { | |
b8698a0f L |
6833 | /* Since gaps are not supported for interleaved stores, |
6834 | GROUP_SIZE is the exact number of stmts in the chain. | |
6835 | Therefore, NEXT_STMT can't be NULL_TREE. In case that | |
6836 | there is no interleaving, GROUP_SIZE is 1, and only one | |
ebfd146a | 6837 | iteration of the loop will be executed. */ |
c3a8f964 | 6838 | op = vect_get_store_rhs (next_stmt); |
81c40241 | 6839 | vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt); |
9771b263 DN |
6840 | dr_chain.quick_push (vec_oprnd); |
6841 | oprnds.quick_push (vec_oprnd); | |
e14c1050 | 6842 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
ebfd146a | 6843 | } |
c3a8f964 RS |
6844 | if (mask) |
6845 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, | |
6846 | mask_vectype); | |
ebfd146a IR |
6847 | } |
6848 | ||
6849 | /* We should have catched mismatched types earlier. */ | |
6850 | gcc_assert (useless_type_conversion_p (vectype, | |
6851 | TREE_TYPE (vec_oprnd))); | |
74bf76ed JJ |
6852 | bool simd_lane_access_p |
6853 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
6854 | if (simd_lane_access_p | |
6855 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
6856 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
6857 | && integer_zerop (DR_OFFSET (first_dr)) | |
6858 | && integer_zerop (DR_INIT (first_dr)) | |
6859 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
44fc7854 | 6860 | get_alias_set (TREE_TYPE (ref_type)))) |
74bf76ed JJ |
6861 | { |
6862 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
44fc7854 | 6863 | dataref_offset = build_int_cst (ref_type, 0); |
8928eff3 | 6864 | inv_p = false; |
74bf76ed | 6865 | } |
f307441a RS |
6866 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
6867 | { | |
6868 | vect_get_gather_scatter_ops (loop, stmt, &gs_info, | |
6869 | &dataref_ptr, &vec_offset); | |
6870 | inv_p = false; | |
6871 | } | |
74bf76ed JJ |
6872 | else |
6873 | dataref_ptr | |
6874 | = vect_create_data_ref_ptr (first_stmt, aggr_type, | |
6875 | simd_lane_access_p ? loop : NULL, | |
09dfa495 | 6876 | offset, &dummy, gsi, &ptr_incr, |
f307441a RS |
6877 | simd_lane_access_p, &inv_p, |
6878 | NULL_TREE, bump); | |
a70d6342 | 6879 | gcc_assert (bb_vinfo || !inv_p); |
ebfd146a | 6880 | } |
b8698a0f | 6881 | else |
ebfd146a | 6882 | { |
b8698a0f L |
6883 | /* For interleaved stores we created vectorized defs for all the |
6884 | defs stored in OPRNDS in the previous iteration (previous copy). | |
6885 | DR_CHAIN is then used as an input to vect_permute_store_chain(), | |
ebfd146a IR |
6886 | and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the |
6887 | next copy. | |
0d0293ac | 6888 | If the store is not grouped, GROUP_SIZE is 1, and DR_CHAIN and |
ebfd146a IR |
6889 | OPRNDS are of size 1. */ |
6890 | for (i = 0; i < group_size; i++) | |
6891 | { | |
9771b263 | 6892 | op = oprnds[i]; |
929b4411 RS |
6893 | vect_is_simple_use (op, vinfo, &def_stmt, &rhs_dt); |
6894 | vec_oprnd = vect_get_vec_def_for_stmt_copy (rhs_dt, op); | |
9771b263 DN |
6895 | dr_chain[i] = vec_oprnd; |
6896 | oprnds[i] = vec_oprnd; | |
ebfd146a | 6897 | } |
c3a8f964 | 6898 | if (mask) |
929b4411 | 6899 | vec_mask = vect_get_vec_def_for_stmt_copy (mask_dt, vec_mask); |
74bf76ed JJ |
6900 | if (dataref_offset) |
6901 | dataref_offset | |
f307441a RS |
6902 | = int_const_binop (PLUS_EXPR, dataref_offset, bump); |
6903 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
929b4411 RS |
6904 | vec_offset = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
6905 | vec_offset); | |
74bf76ed JJ |
6906 | else |
6907 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, stmt, | |
f307441a | 6908 | bump); |
ebfd146a IR |
6909 | } |
6910 | ||
2de001ee | 6911 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
ebfd146a | 6912 | { |
272c6793 | 6913 | tree vec_array; |
267d3070 | 6914 | |
3ba4ff41 | 6915 | /* Get an array into which we can store the individual vectors. */ |
272c6793 | 6916 | vec_array = create_vector_array (vectype, vec_num); |
3ba4ff41 RS |
6917 | |
6918 | /* Invalidate the current contents of VEC_ARRAY. This should | |
6919 | become an RTL clobber too, which prevents the vector registers | |
6920 | from being upward-exposed. */ | |
6921 | vect_clobber_variable (stmt, gsi, vec_array); | |
6922 | ||
6923 | /* Store the individual vectors into the array. */ | |
272c6793 | 6924 | for (i = 0; i < vec_num; i++) |
c2d7ab2a | 6925 | { |
9771b263 | 6926 | vec_oprnd = dr_chain[i]; |
272c6793 | 6927 | write_vector_array (stmt, gsi, vec_oprnd, vec_array, i); |
267d3070 | 6928 | } |
b8698a0f | 6929 | |
7cfb4d93 | 6930 | tree final_mask = NULL; |
70088b95 RS |
6931 | if (loop_masks) |
6932 | final_mask = vect_get_loop_mask (gsi, loop_masks, ncopies, | |
6933 | vectype, j); | |
7cfb4d93 RS |
6934 | if (vec_mask) |
6935 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
6936 | vec_mask, gsi); | |
6937 | ||
7e11fc7f | 6938 | gcall *call; |
7cfb4d93 | 6939 | if (final_mask) |
7e11fc7f RS |
6940 | { |
6941 | /* Emit: | |
6942 | MASK_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK, | |
6943 | VEC_ARRAY). */ | |
6944 | unsigned int align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype)); | |
6945 | tree alias_ptr = build_int_cst (ref_type, align); | |
6946 | call = gimple_build_call_internal (IFN_MASK_STORE_LANES, 4, | |
6947 | dataref_ptr, alias_ptr, | |
7cfb4d93 | 6948 | final_mask, vec_array); |
7e11fc7f RS |
6949 | } |
6950 | else | |
6951 | { | |
6952 | /* Emit: | |
6953 | MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */ | |
6954 | data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type); | |
6955 | call = gimple_build_call_internal (IFN_STORE_LANES, 1, | |
6956 | vec_array); | |
6957 | gimple_call_set_lhs (call, data_ref); | |
6958 | } | |
a844293d RS |
6959 | gimple_call_set_nothrow (call, true); |
6960 | new_stmt = call; | |
267d3070 | 6961 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
3ba4ff41 RS |
6962 | |
6963 | /* Record that VEC_ARRAY is now dead. */ | |
6964 | vect_clobber_variable (stmt, gsi, vec_array); | |
272c6793 RS |
6965 | } |
6966 | else | |
6967 | { | |
6968 | new_stmt = NULL; | |
0d0293ac | 6969 | if (grouped_store) |
272c6793 | 6970 | { |
b6b9227d JJ |
6971 | if (j == 0) |
6972 | result_chain.create (group_size); | |
272c6793 RS |
6973 | /* Permute. */ |
6974 | vect_permute_store_chain (dr_chain, group_size, stmt, gsi, | |
6975 | &result_chain); | |
6976 | } | |
c2d7ab2a | 6977 | |
272c6793 RS |
6978 | next_stmt = first_stmt; |
6979 | for (i = 0; i < vec_num; i++) | |
6980 | { | |
644ffefd | 6981 | unsigned align, misalign; |
272c6793 | 6982 | |
7cfb4d93 | 6983 | tree final_mask = NULL_TREE; |
70088b95 RS |
6984 | if (loop_masks) |
6985 | final_mask = vect_get_loop_mask (gsi, loop_masks, | |
6986 | vec_num * ncopies, | |
7cfb4d93 RS |
6987 | vectype, vec_num * j + i); |
6988 | if (vec_mask) | |
6989 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
6990 | vec_mask, gsi); | |
6991 | ||
f307441a RS |
6992 | if (memory_access_type == VMAT_GATHER_SCATTER) |
6993 | { | |
6994 | tree scale = size_int (gs_info.scale); | |
6995 | gcall *call; | |
70088b95 | 6996 | if (loop_masks) |
f307441a RS |
6997 | call = gimple_build_call_internal |
6998 | (IFN_MASK_SCATTER_STORE, 5, dataref_ptr, vec_offset, | |
6999 | scale, vec_oprnd, final_mask); | |
7000 | else | |
7001 | call = gimple_build_call_internal | |
7002 | (IFN_SCATTER_STORE, 4, dataref_ptr, vec_offset, | |
7003 | scale, vec_oprnd); | |
7004 | gimple_call_set_nothrow (call, true); | |
7005 | new_stmt = call; | |
7006 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7007 | break; | |
7008 | } | |
7009 | ||
272c6793 RS |
7010 | if (i > 0) |
7011 | /* Bump the vector pointer. */ | |
7012 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
f307441a | 7013 | stmt, bump); |
272c6793 RS |
7014 | |
7015 | if (slp) | |
9771b263 | 7016 | vec_oprnd = vec_oprnds[i]; |
0d0293ac MM |
7017 | else if (grouped_store) |
7018 | /* For grouped stores vectorized defs are interleaved in | |
272c6793 | 7019 | vect_permute_store_chain(). */ |
9771b263 | 7020 | vec_oprnd = result_chain[i]; |
272c6793 | 7021 | |
f702e7d4 | 7022 | align = DR_TARGET_ALIGNMENT (first_dr); |
272c6793 | 7023 | if (aligned_access_p (first_dr)) |
644ffefd | 7024 | misalign = 0; |
272c6793 RS |
7025 | else if (DR_MISALIGNMENT (first_dr) == -1) |
7026 | { | |
25f68d90 | 7027 | align = dr_alignment (vect_dr_behavior (first_dr)); |
52639a61 | 7028 | misalign = 0; |
272c6793 RS |
7029 | } |
7030 | else | |
c3a8f964 | 7031 | misalign = DR_MISALIGNMENT (first_dr); |
aed93b23 RB |
7032 | if (dataref_offset == NULL_TREE |
7033 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
7034 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), align, |
7035 | misalign); | |
c2d7ab2a | 7036 | |
62da9e14 | 7037 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
09dfa495 BM |
7038 | { |
7039 | tree perm_mask = perm_mask_for_reverse (vectype); | |
7040 | tree perm_dest | |
c3a8f964 | 7041 | = vect_create_destination_var (vect_get_store_rhs (stmt), |
09dfa495 | 7042 | vectype); |
b731b390 | 7043 | tree new_temp = make_ssa_name (perm_dest); |
09dfa495 BM |
7044 | |
7045 | /* Generate the permute statement. */ | |
355fe088 | 7046 | gimple *perm_stmt |
0d0e4a03 JJ |
7047 | = gimple_build_assign (new_temp, VEC_PERM_EXPR, vec_oprnd, |
7048 | vec_oprnd, perm_mask); | |
09dfa495 BM |
7049 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
7050 | ||
7051 | perm_stmt = SSA_NAME_DEF_STMT (new_temp); | |
7052 | vec_oprnd = new_temp; | |
7053 | } | |
7054 | ||
272c6793 | 7055 | /* Arguments are ready. Create the new vector stmt. */ |
7cfb4d93 | 7056 | if (final_mask) |
c3a8f964 RS |
7057 | { |
7058 | align = least_bit_hwi (misalign | align); | |
7059 | tree ptr = build_int_cst (ref_type, align); | |
7060 | gcall *call | |
7061 | = gimple_build_call_internal (IFN_MASK_STORE, 4, | |
7062 | dataref_ptr, ptr, | |
7cfb4d93 | 7063 | final_mask, vec_oprnd); |
c3a8f964 RS |
7064 | gimple_call_set_nothrow (call, true); |
7065 | new_stmt = call; | |
7066 | } | |
7067 | else | |
7068 | { | |
7069 | data_ref = fold_build2 (MEM_REF, vectype, | |
7070 | dataref_ptr, | |
7071 | dataref_offset | |
7072 | ? dataref_offset | |
7073 | : build_int_cst (ref_type, 0)); | |
7074 | if (aligned_access_p (first_dr)) | |
7075 | ; | |
7076 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
7077 | TREE_TYPE (data_ref) | |
7078 | = build_aligned_type (TREE_TYPE (data_ref), | |
7079 | align * BITS_PER_UNIT); | |
7080 | else | |
7081 | TREE_TYPE (data_ref) | |
7082 | = build_aligned_type (TREE_TYPE (data_ref), | |
7083 | TYPE_ALIGN (elem_type)); | |
19986382 | 7084 | vect_copy_ref_info (data_ref, DR_REF (first_dr)); |
c3a8f964 RS |
7085 | new_stmt = gimple_build_assign (data_ref, vec_oprnd); |
7086 | } | |
272c6793 | 7087 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
272c6793 RS |
7088 | |
7089 | if (slp) | |
7090 | continue; | |
7091 | ||
e14c1050 | 7092 | next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); |
272c6793 RS |
7093 | if (!next_stmt) |
7094 | break; | |
7095 | } | |
ebfd146a | 7096 | } |
1da0876c RS |
7097 | if (!slp) |
7098 | { | |
7099 | if (j == 0) | |
7100 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7101 | else | |
7102 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7103 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7104 | } | |
ebfd146a IR |
7105 | } |
7106 | ||
9771b263 DN |
7107 | oprnds.release (); |
7108 | result_chain.release (); | |
7109 | vec_oprnds.release (); | |
ebfd146a IR |
7110 | |
7111 | return true; | |
7112 | } | |
7113 | ||
557be5a8 AL |
7114 | /* Given a vector type VECTYPE, turns permutation SEL into the equivalent |
7115 | VECTOR_CST mask. No checks are made that the target platform supports the | |
7ac7e286 | 7116 | mask, so callers may wish to test can_vec_perm_const_p separately, or use |
557be5a8 | 7117 | vect_gen_perm_mask_checked. */ |
a1e53f3f | 7118 | |
3fcc1b55 | 7119 | tree |
4aae3cb3 | 7120 | vect_gen_perm_mask_any (tree vectype, const vec_perm_indices &sel) |
a1e53f3f | 7121 | { |
b00cb3bf | 7122 | tree mask_type; |
a1e53f3f | 7123 | |
0ecc2b7d RS |
7124 | poly_uint64 nunits = sel.length (); |
7125 | gcc_assert (known_eq (nunits, TYPE_VECTOR_SUBPARTS (vectype))); | |
b00cb3bf RS |
7126 | |
7127 | mask_type = build_vector_type (ssizetype, nunits); | |
736d0f28 | 7128 | return vec_perm_indices_to_tree (mask_type, sel); |
a1e53f3f L |
7129 | } |
7130 | ||
7ac7e286 | 7131 | /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_const_p, |
cf7aa6a3 | 7132 | i.e. that the target supports the pattern _for arbitrary input vectors_. */ |
557be5a8 AL |
7133 | |
7134 | tree | |
4aae3cb3 | 7135 | vect_gen_perm_mask_checked (tree vectype, const vec_perm_indices &sel) |
557be5a8 | 7136 | { |
7ac7e286 | 7137 | gcc_assert (can_vec_perm_const_p (TYPE_MODE (vectype), sel)); |
557be5a8 AL |
7138 | return vect_gen_perm_mask_any (vectype, sel); |
7139 | } | |
7140 | ||
aec7ae7d JJ |
7141 | /* Given a vector variable X and Y, that was generated for the scalar |
7142 | STMT, generate instructions to permute the vector elements of X and Y | |
7143 | using permutation mask MASK_VEC, insert them at *GSI and return the | |
7144 | permuted vector variable. */ | |
a1e53f3f L |
7145 | |
7146 | static tree | |
355fe088 | 7147 | permute_vec_elements (tree x, tree y, tree mask_vec, gimple *stmt, |
aec7ae7d | 7148 | gimple_stmt_iterator *gsi) |
a1e53f3f L |
7149 | { |
7150 | tree vectype = TREE_TYPE (x); | |
aec7ae7d | 7151 | tree perm_dest, data_ref; |
355fe088 | 7152 | gimple *perm_stmt; |
a1e53f3f | 7153 | |
7ad429a4 RS |
7154 | tree scalar_dest = gimple_get_lhs (stmt); |
7155 | if (TREE_CODE (scalar_dest) == SSA_NAME) | |
7156 | perm_dest = vect_create_destination_var (scalar_dest, vectype); | |
7157 | else | |
7158 | perm_dest = vect_get_new_vect_var (vectype, vect_simple_var, NULL); | |
b731b390 | 7159 | data_ref = make_ssa_name (perm_dest); |
a1e53f3f L |
7160 | |
7161 | /* Generate the permute statement. */ | |
0d0e4a03 | 7162 | perm_stmt = gimple_build_assign (data_ref, VEC_PERM_EXPR, x, y, mask_vec); |
a1e53f3f L |
7163 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); |
7164 | ||
7165 | return data_ref; | |
7166 | } | |
7167 | ||
6b916b36 RB |
7168 | /* Hoist the definitions of all SSA uses on STMT out of the loop LOOP, |
7169 | inserting them on the loops preheader edge. Returns true if we | |
7170 | were successful in doing so (and thus STMT can be moved then), | |
7171 | otherwise returns false. */ | |
7172 | ||
7173 | static bool | |
355fe088 | 7174 | hoist_defs_of_uses (gimple *stmt, struct loop *loop) |
6b916b36 RB |
7175 | { |
7176 | ssa_op_iter i; | |
7177 | tree op; | |
7178 | bool any = false; | |
7179 | ||
7180 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
7181 | { | |
355fe088 | 7182 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
7183 | if (!gimple_nop_p (def_stmt) |
7184 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
7185 | { | |
7186 | /* Make sure we don't need to recurse. While we could do | |
7187 | so in simple cases when there are more complex use webs | |
7188 | we don't have an easy way to preserve stmt order to fulfil | |
7189 | dependencies within them. */ | |
7190 | tree op2; | |
7191 | ssa_op_iter i2; | |
d1417442 JJ |
7192 | if (gimple_code (def_stmt) == GIMPLE_PHI) |
7193 | return false; | |
6b916b36 RB |
7194 | FOR_EACH_SSA_TREE_OPERAND (op2, def_stmt, i2, SSA_OP_USE) |
7195 | { | |
355fe088 | 7196 | gimple *def_stmt2 = SSA_NAME_DEF_STMT (op2); |
6b916b36 RB |
7197 | if (!gimple_nop_p (def_stmt2) |
7198 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2))) | |
7199 | return false; | |
7200 | } | |
7201 | any = true; | |
7202 | } | |
7203 | } | |
7204 | ||
7205 | if (!any) | |
7206 | return true; | |
7207 | ||
7208 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE) | |
7209 | { | |
355fe088 | 7210 | gimple *def_stmt = SSA_NAME_DEF_STMT (op); |
6b916b36 RB |
7211 | if (!gimple_nop_p (def_stmt) |
7212 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) | |
7213 | { | |
7214 | gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); | |
7215 | gsi_remove (&gsi, false); | |
7216 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt); | |
7217 | } | |
7218 | } | |
7219 | ||
7220 | return true; | |
7221 | } | |
7222 | ||
ebfd146a IR |
7223 | /* vectorizable_load. |
7224 | ||
b8698a0f L |
7225 | Check if STMT reads a non scalar data-ref (array/pointer/structure) that |
7226 | can be vectorized. | |
7227 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
ebfd146a IR |
7228 | stmt to replace it, put it in VEC_STMT, and insert it at BSI. |
7229 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
7230 | ||
7231 | static bool | |
355fe088 | 7232 | vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, |
c716e67f | 7233 | slp_tree slp_node, slp_instance slp_node_instance) |
ebfd146a IR |
7234 | { |
7235 | tree scalar_dest; | |
7236 | tree vec_dest = NULL; | |
7237 | tree data_ref = NULL; | |
7238 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
b8698a0f | 7239 | stmt_vec_info prev_stmt_info; |
ebfd146a | 7240 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 7241 | struct loop *loop = NULL; |
ebfd146a | 7242 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; |
a70d6342 | 7243 | bool nested_in_vect_loop = false; |
c716e67f | 7244 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; |
272c6793 | 7245 | tree elem_type; |
ebfd146a | 7246 | tree new_temp; |
ef4bddc2 | 7247 | machine_mode mode; |
355fe088 | 7248 | gimple *new_stmt = NULL; |
ebfd146a IR |
7249 | tree dummy; |
7250 | enum dr_alignment_support alignment_support_scheme; | |
7251 | tree dataref_ptr = NULL_TREE; | |
74bf76ed | 7252 | tree dataref_offset = NULL_TREE; |
355fe088 | 7253 | gimple *ptr_incr = NULL; |
ebfd146a | 7254 | int ncopies; |
4d694b27 RS |
7255 | int i, j; |
7256 | unsigned int group_size; | |
7257 | poly_uint64 group_gap_adj; | |
ebfd146a IR |
7258 | tree msq = NULL_TREE, lsq; |
7259 | tree offset = NULL_TREE; | |
356bbc4c | 7260 | tree byte_offset = NULL_TREE; |
ebfd146a | 7261 | tree realignment_token = NULL_TREE; |
538dd0b7 | 7262 | gphi *phi = NULL; |
6e1aa848 | 7263 | vec<tree> dr_chain = vNULL; |
0d0293ac | 7264 | bool grouped_load = false; |
355fe088 | 7265 | gimple *first_stmt; |
4f0a0218 | 7266 | gimple *first_stmt_for_drptr = NULL; |
ebfd146a IR |
7267 | bool inv_p; |
7268 | bool compute_in_loop = false; | |
7269 | struct loop *at_loop; | |
7270 | int vec_num; | |
7271 | bool slp = (slp_node != NULL); | |
7272 | bool slp_perm = false; | |
a70d6342 | 7273 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
d9f21f6a | 7274 | poly_uint64 vf; |
272c6793 | 7275 | tree aggr_type; |
134c85ca | 7276 | gather_scatter_info gs_info; |
310213d4 | 7277 | vec_info *vinfo = stmt_info->vinfo; |
44fc7854 | 7278 | tree ref_type; |
929b4411 | 7279 | enum vect_def_type mask_dt = vect_unknown_def_type; |
a70d6342 | 7280 | |
465c8c19 JJ |
7281 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
7282 | return false; | |
7283 | ||
66c16fd9 RB |
7284 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
7285 | && ! vec_stmt) | |
465c8c19 JJ |
7286 | return false; |
7287 | ||
c3a8f964 RS |
7288 | tree mask = NULL_TREE, mask_vectype = NULL_TREE; |
7289 | if (is_gimple_assign (stmt)) | |
7290 | { | |
7291 | scalar_dest = gimple_assign_lhs (stmt); | |
7292 | if (TREE_CODE (scalar_dest) != SSA_NAME) | |
7293 | return false; | |
465c8c19 | 7294 | |
c3a8f964 RS |
7295 | tree_code code = gimple_assign_rhs_code (stmt); |
7296 | if (code != ARRAY_REF | |
7297 | && code != BIT_FIELD_REF | |
7298 | && code != INDIRECT_REF | |
7299 | && code != COMPONENT_REF | |
7300 | && code != IMAGPART_EXPR | |
7301 | && code != REALPART_EXPR | |
7302 | && code != MEM_REF | |
7303 | && TREE_CODE_CLASS (code) != tcc_declaration) | |
7304 | return false; | |
7305 | } | |
7306 | else | |
7307 | { | |
7308 | gcall *call = dyn_cast <gcall *> (stmt); | |
bfaa08b7 RS |
7309 | if (!call || !gimple_call_internal_p (call)) |
7310 | return false; | |
7311 | ||
7312 | internal_fn ifn = gimple_call_internal_fn (call); | |
7313 | if (!internal_load_fn_p (ifn)) | |
c3a8f964 | 7314 | return false; |
465c8c19 | 7315 | |
c3a8f964 RS |
7316 | scalar_dest = gimple_call_lhs (call); |
7317 | if (!scalar_dest) | |
7318 | return false; | |
7319 | ||
7320 | if (slp_node != NULL) | |
7321 | { | |
7322 | if (dump_enabled_p ()) | |
7323 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7324 | "SLP of masked loads not supported.\n"); | |
7325 | return false; | |
7326 | } | |
7327 | ||
bfaa08b7 RS |
7328 | int mask_index = internal_fn_mask_index (ifn); |
7329 | if (mask_index >= 0) | |
7330 | { | |
7331 | mask = gimple_call_arg (call, mask_index); | |
929b4411 RS |
7332 | if (!vect_check_load_store_mask (stmt, mask, &mask_dt, |
7333 | &mask_vectype)) | |
bfaa08b7 RS |
7334 | return false; |
7335 | } | |
c3a8f964 | 7336 | } |
465c8c19 JJ |
7337 | |
7338 | if (!STMT_VINFO_DATA_REF (stmt_info)) | |
7339 | return false; | |
7340 | ||
7341 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
4d694b27 | 7342 | poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); |
465c8c19 | 7343 | |
a70d6342 IR |
7344 | if (loop_vinfo) |
7345 | { | |
7346 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
7347 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
7348 | vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
7349 | } | |
7350 | else | |
3533e503 | 7351 | vf = 1; |
ebfd146a IR |
7352 | |
7353 | /* Multiple types in SLP are handled by creating the appropriate number of | |
ff802fa1 | 7354 | vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in |
ebfd146a | 7355 | case of SLP. */ |
fce57248 | 7356 | if (slp) |
ebfd146a IR |
7357 | ncopies = 1; |
7358 | else | |
e8f142e2 | 7359 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
ebfd146a IR |
7360 | |
7361 | gcc_assert (ncopies >= 1); | |
7362 | ||
7363 | /* FORNOW. This restriction should be relaxed. */ | |
7364 | if (nested_in_vect_loop && ncopies > 1) | |
7365 | { | |
73fbfcad | 7366 | if (dump_enabled_p ()) |
78c60e3d | 7367 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7368 | "multiple types in nested loop.\n"); |
ebfd146a IR |
7369 | return false; |
7370 | } | |
7371 | ||
f2556b68 RB |
7372 | /* Invalidate assumptions made by dependence analysis when vectorization |
7373 | on the unrolled body effectively re-orders stmts. */ | |
7374 | if (ncopies > 1 | |
7375 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
d9f21f6a RS |
7376 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
7377 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
f2556b68 RB |
7378 | { |
7379 | if (dump_enabled_p ()) | |
7380 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7381 | "cannot perform implicit CSE when unrolling " | |
7382 | "with negative dependence distance\n"); | |
7383 | return false; | |
7384 | } | |
7385 | ||
7b7b1813 | 7386 | elem_type = TREE_TYPE (vectype); |
947131ba | 7387 | mode = TYPE_MODE (vectype); |
ebfd146a IR |
7388 | |
7389 | /* FORNOW. In some cases can vectorize even if data-type not supported | |
7390 | (e.g. - data copies). */ | |
947131ba | 7391 | if (optab_handler (mov_optab, mode) == CODE_FOR_nothing) |
ebfd146a | 7392 | { |
73fbfcad | 7393 | if (dump_enabled_p ()) |
78c60e3d | 7394 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 7395 | "Aligned load, but unsupported type.\n"); |
ebfd146a IR |
7396 | return false; |
7397 | } | |
7398 | ||
ebfd146a | 7399 | /* Check if the load is a part of an interleaving chain. */ |
0d0293ac | 7400 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) |
ebfd146a | 7401 | { |
0d0293ac | 7402 | grouped_load = true; |
ebfd146a | 7403 | /* FORNOW */ |
2de001ee RS |
7404 | gcc_assert (!nested_in_vect_loop); |
7405 | gcc_assert (!STMT_VINFO_GATHER_SCATTER_P (stmt_info)); | |
ebfd146a | 7406 | |
e14c1050 | 7407 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
d3465d72 | 7408 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
d5f035ea | 7409 | |
b1af7da6 RB |
7410 | if (slp && SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) |
7411 | slp_perm = true; | |
7412 | ||
f2556b68 RB |
7413 | /* Invalidate assumptions made by dependence analysis when vectorization |
7414 | on the unrolled body effectively re-orders stmts. */ | |
7415 | if (!PURE_SLP_STMT (stmt_info) | |
7416 | && STMT_VINFO_MIN_NEG_DIST (stmt_info) != 0 | |
d9f21f6a RS |
7417 | && maybe_gt (LOOP_VINFO_VECT_FACTOR (loop_vinfo), |
7418 | STMT_VINFO_MIN_NEG_DIST (stmt_info))) | |
f2556b68 RB |
7419 | { |
7420 | if (dump_enabled_p ()) | |
7421 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7422 | "cannot perform implicit CSE when performing " | |
7423 | "group loads with negative dependence distance\n"); | |
7424 | return false; | |
7425 | } | |
96bb56b2 RB |
7426 | |
7427 | /* Similarly when the stmt is a load that is both part of a SLP | |
7428 | instance and a loop vectorized stmt via the same-dr mechanism | |
7429 | we have to give up. */ | |
7430 | if (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info) | |
7431 | && (STMT_SLP_TYPE (stmt_info) | |
7432 | != STMT_SLP_TYPE (vinfo_for_stmt | |
7433 | (STMT_VINFO_GROUP_SAME_DR_STMT (stmt_info))))) | |
7434 | { | |
7435 | if (dump_enabled_p ()) | |
7436 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7437 | "conflicting SLP types for CSEd load\n"); | |
7438 | return false; | |
7439 | } | |
ebfd146a | 7440 | } |
7cfb4d93 RS |
7441 | else |
7442 | group_size = 1; | |
ebfd146a | 7443 | |
2de001ee | 7444 | vect_memory_access_type memory_access_type; |
7e11fc7f | 7445 | if (!get_load_store_type (stmt, vectype, slp, mask, VLS_LOAD, ncopies, |
2de001ee RS |
7446 | &memory_access_type, &gs_info)) |
7447 | return false; | |
a1e53f3f | 7448 | |
c3a8f964 RS |
7449 | if (mask) |
7450 | { | |
7451 | if (memory_access_type == VMAT_CONTIGUOUS) | |
7452 | { | |
7e11fc7f RS |
7453 | machine_mode vec_mode = TYPE_MODE (vectype); |
7454 | if (!VECTOR_MODE_P (vec_mode) | |
7455 | || !can_vec_mask_load_store_p (vec_mode, | |
c3a8f964 RS |
7456 | TYPE_MODE (mask_vectype), true)) |
7457 | return false; | |
7458 | } | |
bfaa08b7 | 7459 | else if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.decl) |
c3a8f964 RS |
7460 | { |
7461 | tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info.decl)); | |
7462 | tree masktype | |
7463 | = TREE_VALUE (TREE_CHAIN (TREE_CHAIN (TREE_CHAIN (arglist)))); | |
7464 | if (TREE_CODE (masktype) == INTEGER_TYPE) | |
7465 | { | |
7466 | if (dump_enabled_p ()) | |
7467 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7468 | "masked gather with integer mask not" | |
7469 | " supported."); | |
7470 | return false; | |
7471 | } | |
7472 | } | |
bfaa08b7 RS |
7473 | else if (memory_access_type != VMAT_LOAD_STORE_LANES |
7474 | && memory_access_type != VMAT_GATHER_SCATTER) | |
c3a8f964 RS |
7475 | { |
7476 | if (dump_enabled_p ()) | |
7477 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
7478 | "unsupported access type for masked load.\n"); | |
7479 | return false; | |
7480 | } | |
7481 | } | |
7482 | ||
ebfd146a IR |
7483 | if (!vec_stmt) /* transformation not required. */ |
7484 | { | |
2de001ee RS |
7485 | if (!slp) |
7486 | STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) = memory_access_type; | |
7cfb4d93 RS |
7487 | |
7488 | if (loop_vinfo | |
7489 | && LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo)) | |
7490 | check_load_store_masking (loop_vinfo, vectype, VLS_LOAD, group_size, | |
bfaa08b7 | 7491 | memory_access_type, &gs_info); |
7cfb4d93 | 7492 | |
ebfd146a | 7493 | STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; |
2e8ab70c | 7494 | /* The SLP costs are calculated during SLP analysis. */ |
78604de0 | 7495 | if (! slp_node) |
2de001ee | 7496 | vect_model_load_cost (stmt_info, ncopies, memory_access_type, |
2e8ab70c | 7497 | NULL, NULL, NULL); |
ebfd146a IR |
7498 | return true; |
7499 | } | |
7500 | ||
2de001ee RS |
7501 | if (!slp) |
7502 | gcc_assert (memory_access_type | |
7503 | == STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)); | |
7504 | ||
73fbfcad | 7505 | if (dump_enabled_p ()) |
78c60e3d | 7506 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 7507 | "transform load. ncopies = %d\n", ncopies); |
ebfd146a | 7508 | |
67b8dbac | 7509 | /* Transform. */ |
ebfd146a | 7510 | |
f702e7d4 | 7511 | ensure_base_align (dr); |
c716e67f | 7512 | |
bfaa08b7 | 7513 | if (memory_access_type == VMAT_GATHER_SCATTER && gs_info.decl) |
aec7ae7d | 7514 | { |
929b4411 RS |
7515 | vect_build_gather_load_calls (stmt, gsi, vec_stmt, &gs_info, mask, |
7516 | mask_dt); | |
aec7ae7d JJ |
7517 | return true; |
7518 | } | |
2de001ee RS |
7519 | |
7520 | if (memory_access_type == VMAT_ELEMENTWISE | |
7521 | || memory_access_type == VMAT_STRIDED_SLP) | |
7d75abc8 MM |
7522 | { |
7523 | gimple_stmt_iterator incr_gsi; | |
7524 | bool insert_after; | |
355fe088 | 7525 | gimple *incr; |
7d75abc8 | 7526 | tree offvar; |
7d75abc8 MM |
7527 | tree ivstep; |
7528 | tree running_off; | |
9771b263 | 7529 | vec<constructor_elt, va_gc> *v = NULL; |
14ac6aa2 | 7530 | tree stride_base, stride_step, alias_off; |
4d694b27 RS |
7531 | /* Checked by get_load_store_type. */ |
7532 | unsigned int const_nunits = nunits.to_constant (); | |
b210f45f | 7533 | unsigned HOST_WIDE_INT cst_offset = 0; |
14ac6aa2 | 7534 | |
7cfb4d93 | 7535 | gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); |
14ac6aa2 | 7536 | gcc_assert (!nested_in_vect_loop); |
7d75abc8 | 7537 | |
b210f45f | 7538 | if (grouped_load) |
44fc7854 BE |
7539 | { |
7540 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); | |
7541 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
44fc7854 | 7542 | } |
ab313a8c | 7543 | else |
44fc7854 BE |
7544 | { |
7545 | first_stmt = stmt; | |
7546 | first_dr = dr; | |
b210f45f RB |
7547 | } |
7548 | if (slp && grouped_load) | |
7549 | { | |
7550 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); | |
7551 | ref_type = get_group_alias_ptr_type (first_stmt); | |
7552 | } | |
7553 | else | |
7554 | { | |
7555 | if (grouped_load) | |
7556 | cst_offset | |
7557 | = (tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))) | |
7558 | * vect_get_place_in_interleaving_chain (stmt, first_stmt)); | |
44fc7854 | 7559 | group_size = 1; |
b210f45f | 7560 | ref_type = reference_alias_ptr_type (DR_REF (dr)); |
44fc7854 | 7561 | } |
ab313a8c | 7562 | |
14ac6aa2 RB |
7563 | stride_base |
7564 | = fold_build_pointer_plus | |
ab313a8c | 7565 | (DR_BASE_ADDRESS (first_dr), |
14ac6aa2 | 7566 | size_binop (PLUS_EXPR, |
ab313a8c RB |
7567 | convert_to_ptrofftype (DR_OFFSET (first_dr)), |
7568 | convert_to_ptrofftype (DR_INIT (first_dr)))); | |
7569 | stride_step = fold_convert (sizetype, DR_STEP (first_dr)); | |
7d75abc8 MM |
7570 | |
7571 | /* For a load with loop-invariant (but other than power-of-2) | |
7572 | stride (i.e. not a grouped access) like so: | |
7573 | ||
7574 | for (i = 0; i < n; i += stride) | |
7575 | ... = array[i]; | |
7576 | ||
7577 | we generate a new induction variable and new accesses to | |
7578 | form a new vector (or vectors, depending on ncopies): | |
7579 | ||
7580 | for (j = 0; ; j += VF*stride) | |
7581 | tmp1 = array[j]; | |
7582 | tmp2 = array[j + stride]; | |
7583 | ... | |
7584 | vectemp = {tmp1, tmp2, ...} | |
7585 | */ | |
7586 | ||
ab313a8c RB |
7587 | ivstep = fold_build2 (MULT_EXPR, TREE_TYPE (stride_step), stride_step, |
7588 | build_int_cst (TREE_TYPE (stride_step), vf)); | |
7d75abc8 MM |
7589 | |
7590 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
7591 | ||
b210f45f RB |
7592 | stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base); |
7593 | ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep); | |
7594 | create_iv (stride_base, ivstep, NULL, | |
7d75abc8 MM |
7595 | loop, &incr_gsi, insert_after, |
7596 | &offvar, NULL); | |
7597 | incr = gsi_stmt (incr_gsi); | |
310213d4 | 7598 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo)); |
7d75abc8 | 7599 | |
b210f45f | 7600 | stride_step = cse_and_gimplify_to_preheader (loop_vinfo, stride_step); |
7d75abc8 MM |
7601 | |
7602 | prev_stmt_info = NULL; | |
7603 | running_off = offvar; | |
44fc7854 | 7604 | alias_off = build_int_cst (ref_type, 0); |
4d694b27 | 7605 | int nloads = const_nunits; |
e09b4c37 | 7606 | int lnel = 1; |
7b5fc413 | 7607 | tree ltype = TREE_TYPE (vectype); |
ea60dd34 | 7608 | tree lvectype = vectype; |
b266b968 | 7609 | auto_vec<tree> dr_chain; |
2de001ee | 7610 | if (memory_access_type == VMAT_STRIDED_SLP) |
7b5fc413 | 7611 | { |
4d694b27 | 7612 | if (group_size < const_nunits) |
e09b4c37 | 7613 | { |
ff03930a JJ |
7614 | /* First check if vec_init optab supports construction from |
7615 | vector elts directly. */ | |
b397965c | 7616 | scalar_mode elmode = SCALAR_TYPE_MODE (TREE_TYPE (vectype)); |
9da15d40 RS |
7617 | machine_mode vmode; |
7618 | if (mode_for_vector (elmode, group_size).exists (&vmode) | |
7619 | && VECTOR_MODE_P (vmode) | |
414fef4e | 7620 | && targetm.vector_mode_supported_p (vmode) |
ff03930a JJ |
7621 | && (convert_optab_handler (vec_init_optab, |
7622 | TYPE_MODE (vectype), vmode) | |
7623 | != CODE_FOR_nothing)) | |
ea60dd34 | 7624 | { |
4d694b27 | 7625 | nloads = const_nunits / group_size; |
ea60dd34 | 7626 | lnel = group_size; |
ff03930a JJ |
7627 | ltype = build_vector_type (TREE_TYPE (vectype), group_size); |
7628 | } | |
7629 | else | |
7630 | { | |
7631 | /* Otherwise avoid emitting a constructor of vector elements | |
7632 | by performing the loads using an integer type of the same | |
7633 | size, constructing a vector of those and then | |
7634 | re-interpreting it as the original vector type. | |
7635 | This avoids a huge runtime penalty due to the general | |
7636 | inability to perform store forwarding from smaller stores | |
7637 | to a larger load. */ | |
7638 | unsigned lsize | |
7639 | = group_size * TYPE_PRECISION (TREE_TYPE (vectype)); | |
fffbab82 | 7640 | elmode = int_mode_for_size (lsize, 0).require (); |
4d694b27 | 7641 | unsigned int lnunits = const_nunits / group_size; |
ff03930a JJ |
7642 | /* If we can't construct such a vector fall back to |
7643 | element loads of the original vector type. */ | |
4d694b27 | 7644 | if (mode_for_vector (elmode, lnunits).exists (&vmode) |
9da15d40 | 7645 | && VECTOR_MODE_P (vmode) |
414fef4e | 7646 | && targetm.vector_mode_supported_p (vmode) |
ff03930a JJ |
7647 | && (convert_optab_handler (vec_init_optab, vmode, elmode) |
7648 | != CODE_FOR_nothing)) | |
7649 | { | |
4d694b27 | 7650 | nloads = lnunits; |
ff03930a JJ |
7651 | lnel = group_size; |
7652 | ltype = build_nonstandard_integer_type (lsize, 1); | |
7653 | lvectype = build_vector_type (ltype, nloads); | |
7654 | } | |
ea60dd34 | 7655 | } |
e09b4c37 | 7656 | } |
2de001ee | 7657 | else |
e09b4c37 | 7658 | { |
ea60dd34 | 7659 | nloads = 1; |
4d694b27 | 7660 | lnel = const_nunits; |
e09b4c37 | 7661 | ltype = vectype; |
e09b4c37 | 7662 | } |
2de001ee RS |
7663 | ltype = build_aligned_type (ltype, TYPE_ALIGN (TREE_TYPE (vectype))); |
7664 | } | |
7665 | if (slp) | |
7666 | { | |
66c16fd9 RB |
7667 | /* For SLP permutation support we need to load the whole group, |
7668 | not only the number of vector stmts the permutation result | |
7669 | fits in. */ | |
b266b968 | 7670 | if (slp_perm) |
66c16fd9 | 7671 | { |
d9f21f6a RS |
7672 | /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for |
7673 | variable VF. */ | |
7674 | unsigned int const_vf = vf.to_constant (); | |
4d694b27 | 7675 | ncopies = CEIL (group_size * const_vf, const_nunits); |
66c16fd9 RB |
7676 | dr_chain.create (ncopies); |
7677 | } | |
7678 | else | |
7679 | ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
7b5fc413 | 7680 | } |
4d694b27 | 7681 | unsigned int group_el = 0; |
e09b4c37 RB |
7682 | unsigned HOST_WIDE_INT |
7683 | elsz = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
7d75abc8 MM |
7684 | for (j = 0; j < ncopies; j++) |
7685 | { | |
7b5fc413 | 7686 | if (nloads > 1) |
e09b4c37 RB |
7687 | vec_alloc (v, nloads); |
7688 | for (i = 0; i < nloads; i++) | |
7b5fc413 | 7689 | { |
e09b4c37 | 7690 | tree this_off = build_int_cst (TREE_TYPE (alias_off), |
b210f45f | 7691 | group_el * elsz + cst_offset); |
19986382 RB |
7692 | tree data_ref = build2 (MEM_REF, ltype, running_off, this_off); |
7693 | vect_copy_ref_info (data_ref, DR_REF (first_dr)); | |
7694 | new_stmt = gimple_build_assign (make_ssa_name (ltype), data_ref); | |
e09b4c37 RB |
7695 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
7696 | if (nloads > 1) | |
7697 | CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, | |
7698 | gimple_assign_lhs (new_stmt)); | |
7699 | ||
7700 | group_el += lnel; | |
7701 | if (! slp | |
7702 | || group_el == group_size) | |
7b5fc413 | 7703 | { |
e09b4c37 RB |
7704 | tree newoff = copy_ssa_name (running_off); |
7705 | gimple *incr = gimple_build_assign (newoff, POINTER_PLUS_EXPR, | |
7706 | running_off, stride_step); | |
7b5fc413 RB |
7707 | vect_finish_stmt_generation (stmt, incr, gsi); |
7708 | ||
7709 | running_off = newoff; | |
e09b4c37 | 7710 | group_el = 0; |
7b5fc413 | 7711 | } |
7b5fc413 | 7712 | } |
e09b4c37 | 7713 | if (nloads > 1) |
7d75abc8 | 7714 | { |
ea60dd34 RB |
7715 | tree vec_inv = build_constructor (lvectype, v); |
7716 | new_temp = vect_init_vector (stmt, vec_inv, lvectype, gsi); | |
e09b4c37 | 7717 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
ea60dd34 RB |
7718 | if (lvectype != vectype) |
7719 | { | |
7720 | new_stmt = gimple_build_assign (make_ssa_name (vectype), | |
7721 | VIEW_CONVERT_EXPR, | |
7722 | build1 (VIEW_CONVERT_EXPR, | |
7723 | vectype, new_temp)); | |
7724 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
7725 | } | |
7d75abc8 MM |
7726 | } |
7727 | ||
7b5fc413 | 7728 | if (slp) |
b266b968 | 7729 | { |
b266b968 RB |
7730 | if (slp_perm) |
7731 | dr_chain.quick_push (gimple_assign_lhs (new_stmt)); | |
66c16fd9 RB |
7732 | else |
7733 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
b266b968 | 7734 | } |
7d75abc8 | 7735 | else |
225ce44b RB |
7736 | { |
7737 | if (j == 0) | |
7738 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
7739 | else | |
7740 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
7741 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
7742 | } | |
7d75abc8 | 7743 | } |
b266b968 | 7744 | if (slp_perm) |
29afecdf RB |
7745 | { |
7746 | unsigned n_perms; | |
7747 | vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, | |
7748 | slp_node_instance, false, &n_perms); | |
7749 | } | |
7d75abc8 MM |
7750 | return true; |
7751 | } | |
aec7ae7d | 7752 | |
b5ec4de7 RS |
7753 | if (memory_access_type == VMAT_GATHER_SCATTER |
7754 | || (!slp && memory_access_type == VMAT_CONTIGUOUS)) | |
ab2fc782 RS |
7755 | grouped_load = false; |
7756 | ||
0d0293ac | 7757 | if (grouped_load) |
ebfd146a | 7758 | { |
e14c1050 | 7759 | first_stmt = GROUP_FIRST_ELEMENT (stmt_info); |
44fc7854 | 7760 | group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); |
4f0a0218 | 7761 | /* For SLP vectorization we directly vectorize a subchain |
52eab378 RB |
7762 | without permutation. */ |
7763 | if (slp && ! SLP_TREE_LOAD_PERMUTATION (slp_node).exists ()) | |
4f0a0218 RB |
7764 | first_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[0]; |
7765 | /* For BB vectorization always use the first stmt to base | |
7766 | the data ref pointer on. */ | |
7767 | if (bb_vinfo) | |
7768 | first_stmt_for_drptr = SLP_TREE_SCALAR_STMTS (slp_node)[0]; | |
6aa904c4 | 7769 | |
ebfd146a | 7770 | /* Check if the chain of loads is already vectorized. */ |
01d8bf07 RB |
7771 | if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)) |
7772 | /* For SLP we would need to copy over SLP_TREE_VEC_STMTS. | |
7773 | ??? But we can only do so if there is exactly one | |
7774 | as we have no way to get at the rest. Leave the CSE | |
7775 | opportunity alone. | |
7776 | ??? With the group load eventually participating | |
7777 | in multiple different permutations (having multiple | |
7778 | slp nodes which refer to the same group) the CSE | |
7779 | is even wrong code. See PR56270. */ | |
7780 | && !slp) | |
ebfd146a IR |
7781 | { |
7782 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); | |
7783 | return true; | |
7784 | } | |
7785 | first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); | |
9b999e8c | 7786 | group_gap_adj = 0; |
ebfd146a IR |
7787 | |
7788 | /* VEC_NUM is the number of vect stmts to be created for this group. */ | |
7789 | if (slp) | |
7790 | { | |
0d0293ac | 7791 | grouped_load = false; |
91ff1504 RB |
7792 | /* For SLP permutation support we need to load the whole group, |
7793 | not only the number of vector stmts the permutation result | |
7794 | fits in. */ | |
7795 | if (slp_perm) | |
b267968e | 7796 | { |
d9f21f6a RS |
7797 | /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for |
7798 | variable VF. */ | |
7799 | unsigned int const_vf = vf.to_constant (); | |
4d694b27 RS |
7800 | unsigned int const_nunits = nunits.to_constant (); |
7801 | vec_num = CEIL (group_size * const_vf, const_nunits); | |
b267968e RB |
7802 | group_gap_adj = vf * group_size - nunits * vec_num; |
7803 | } | |
91ff1504 | 7804 | else |
b267968e RB |
7805 | { |
7806 | vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); | |
796bd467 RB |
7807 | group_gap_adj |
7808 | = group_size - SLP_INSTANCE_GROUP_SIZE (slp_node_instance); | |
b267968e | 7809 | } |
a70d6342 | 7810 | } |
ebfd146a | 7811 | else |
9b999e8c | 7812 | vec_num = group_size; |
44fc7854 BE |
7813 | |
7814 | ref_type = get_group_alias_ptr_type (first_stmt); | |
ebfd146a IR |
7815 | } |
7816 | else | |
7817 | { | |
7818 | first_stmt = stmt; | |
7819 | first_dr = dr; | |
7820 | group_size = vec_num = 1; | |
9b999e8c | 7821 | group_gap_adj = 0; |
44fc7854 | 7822 | ref_type = reference_alias_ptr_type (DR_REF (first_dr)); |
ebfd146a IR |
7823 | } |
7824 | ||
720f5239 | 7825 | alignment_support_scheme = vect_supportable_dr_alignment (first_dr, false); |
ebfd146a | 7826 | gcc_assert (alignment_support_scheme); |
70088b95 RS |
7827 | vec_loop_masks *loop_masks |
7828 | = (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) | |
7829 | ? &LOOP_VINFO_MASKS (loop_vinfo) | |
7830 | : NULL); | |
7cfb4d93 RS |
7831 | /* Targets with store-lane instructions must not require explicit |
7832 | realignment. vect_supportable_dr_alignment always returns either | |
7833 | dr_aligned or dr_unaligned_supported for masked operations. */ | |
7834 | gcc_assert ((memory_access_type != VMAT_LOAD_STORE_LANES | |
7835 | && !mask | |
70088b95 | 7836 | && !loop_masks) |
272c6793 RS |
7837 | || alignment_support_scheme == dr_aligned |
7838 | || alignment_support_scheme == dr_unaligned_supported); | |
ebfd146a IR |
7839 | |
7840 | /* In case the vectorization factor (VF) is bigger than the number | |
7841 | of elements that we can fit in a vectype (nunits), we have to generate | |
7842 | more than one vector stmt - i.e - we need to "unroll" the | |
ff802fa1 | 7843 | vector stmt by a factor VF/nunits. In doing so, we record a pointer |
ebfd146a | 7844 | from one copy of the vector stmt to the next, in the field |
ff802fa1 | 7845 | STMT_VINFO_RELATED_STMT. This is necessary in order to allow following |
ebfd146a | 7846 | stages to find the correct vector defs to be used when vectorizing |
ff802fa1 IR |
7847 | stmts that use the defs of the current stmt. The example below |
7848 | illustrates the vectorization process when VF=16 and nunits=4 (i.e., we | |
7849 | need to create 4 vectorized stmts): | |
ebfd146a IR |
7850 | |
7851 | before vectorization: | |
7852 | RELATED_STMT VEC_STMT | |
7853 | S1: x = memref - - | |
7854 | S2: z = x + 1 - - | |
7855 | ||
7856 | step 1: vectorize stmt S1: | |
7857 | We first create the vector stmt VS1_0, and, as usual, record a | |
7858 | pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. | |
7859 | Next, we create the vector stmt VS1_1, and record a pointer to | |
7860 | it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. | |
ff802fa1 | 7861 | Similarly, for VS1_2 and VS1_3. This is the resulting chain of |
ebfd146a IR |
7862 | stmts and pointers: |
7863 | RELATED_STMT VEC_STMT | |
7864 | VS1_0: vx0 = memref0 VS1_1 - | |
7865 | VS1_1: vx1 = memref1 VS1_2 - | |
7866 | VS1_2: vx2 = memref2 VS1_3 - | |
7867 | VS1_3: vx3 = memref3 - - | |
7868 | S1: x = load - VS1_0 | |
7869 | S2: z = x + 1 - - | |
7870 | ||
b8698a0f L |
7871 | See in documentation in vect_get_vec_def_for_stmt_copy for how the |
7872 | information we recorded in RELATED_STMT field is used to vectorize | |
ebfd146a IR |
7873 | stmt S2. */ |
7874 | ||
0d0293ac | 7875 | /* In case of interleaving (non-unit grouped access): |
ebfd146a IR |
7876 | |
7877 | S1: x2 = &base + 2 | |
7878 | S2: x0 = &base | |
7879 | S3: x1 = &base + 1 | |
7880 | S4: x3 = &base + 3 | |
7881 | ||
b8698a0f | 7882 | Vectorized loads are created in the order of memory accesses |
ebfd146a IR |
7883 | starting from the access of the first stmt of the chain: |
7884 | ||
7885 | VS1: vx0 = &base | |
7886 | VS2: vx1 = &base + vec_size*1 | |
7887 | VS3: vx3 = &base + vec_size*2 | |
7888 | VS4: vx4 = &base + vec_size*3 | |
7889 | ||
7890 | Then permutation statements are generated: | |
7891 | ||
e2c83630 RH |
7892 | VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } > |
7893 | VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } > | |
ebfd146a IR |
7894 | ... |
7895 | ||
7896 | And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts | |
7897 | (the order of the data-refs in the output of vect_permute_load_chain | |
7898 | corresponds to the order of scalar stmts in the interleaving chain - see | |
7899 | the documentation of vect_permute_load_chain()). | |
7900 | The generation of permutation stmts and recording them in | |
0d0293ac | 7901 | STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load(). |
ebfd146a | 7902 | |
b8698a0f | 7903 | In case of both multiple types and interleaving, the vector loads and |
ff802fa1 IR |
7904 | permutation stmts above are created for every copy. The result vector |
7905 | stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the | |
7906 | corresponding STMT_VINFO_RELATED_STMT for the next copies. */ | |
ebfd146a IR |
7907 | |
7908 | /* If the data reference is aligned (dr_aligned) or potentially unaligned | |
7909 | on a target that supports unaligned accesses (dr_unaligned_supported) | |
7910 | we generate the following code: | |
7911 | p = initial_addr; | |
7912 | indx = 0; | |
7913 | loop { | |
7914 | p = p + indx * vectype_size; | |
7915 | vec_dest = *(p); | |
7916 | indx = indx + 1; | |
7917 | } | |
7918 | ||
7919 | Otherwise, the data reference is potentially unaligned on a target that | |
b8698a0f | 7920 | does not support unaligned accesses (dr_explicit_realign_optimized) - |
ebfd146a IR |
7921 | then generate the following code, in which the data in each iteration is |
7922 | obtained by two vector loads, one from the previous iteration, and one | |
7923 | from the current iteration: | |
7924 | p1 = initial_addr; | |
7925 | msq_init = *(floor(p1)) | |
7926 | p2 = initial_addr + VS - 1; | |
7927 | realignment_token = call target_builtin; | |
7928 | indx = 0; | |
7929 | loop { | |
7930 | p2 = p2 + indx * vectype_size | |
7931 | lsq = *(floor(p2)) | |
7932 | vec_dest = realign_load (msq, lsq, realignment_token) | |
7933 | indx = indx + 1; | |
7934 | msq = lsq; | |
7935 | } */ | |
7936 | ||
7937 | /* If the misalignment remains the same throughout the execution of the | |
7938 | loop, we can create the init_addr and permutation mask at the loop | |
ff802fa1 | 7939 | preheader. Otherwise, it needs to be created inside the loop. |
ebfd146a IR |
7940 | This can only occur when vectorizing memory accesses in the inner-loop |
7941 | nested within an outer-loop that is being vectorized. */ | |
7942 | ||
d1e4b493 | 7943 | if (nested_in_vect_loop |
cf098191 RS |
7944 | && !multiple_p (DR_STEP_ALIGNMENT (dr), |
7945 | GET_MODE_SIZE (TYPE_MODE (vectype)))) | |
ebfd146a IR |
7946 | { |
7947 | gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized); | |
7948 | compute_in_loop = true; | |
7949 | } | |
7950 | ||
7951 | if ((alignment_support_scheme == dr_explicit_realign_optimized | |
7952 | || alignment_support_scheme == dr_explicit_realign) | |
59fd17e3 | 7953 | && !compute_in_loop) |
ebfd146a IR |
7954 | { |
7955 | msq = vect_setup_realignment (first_stmt, gsi, &realignment_token, | |
7956 | alignment_support_scheme, NULL_TREE, | |
7957 | &at_loop); | |
7958 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
7959 | { | |
538dd0b7 | 7960 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (msq)); |
356bbc4c JJ |
7961 | byte_offset = size_binop (MINUS_EXPR, TYPE_SIZE_UNIT (vectype), |
7962 | size_one_node); | |
ebfd146a IR |
7963 | } |
7964 | } | |
7965 | else | |
7966 | at_loop = loop; | |
7967 | ||
62da9e14 | 7968 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
a1e53f3f L |
7969 | offset = size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1); |
7970 | ||
ab2fc782 RS |
7971 | tree bump; |
7972 | tree vec_offset = NULL_TREE; | |
7973 | if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) | |
7974 | { | |
7975 | aggr_type = NULL_TREE; | |
7976 | bump = NULL_TREE; | |
7977 | } | |
7978 | else if (memory_access_type == VMAT_GATHER_SCATTER) | |
7979 | { | |
7980 | aggr_type = elem_type; | |
7981 | vect_get_strided_load_store_ops (stmt, loop_vinfo, &gs_info, | |
7982 | &bump, &vec_offset); | |
7983 | } | |
272c6793 | 7984 | else |
ab2fc782 RS |
7985 | { |
7986 | if (memory_access_type == VMAT_LOAD_STORE_LANES) | |
7987 | aggr_type = build_array_type_nelts (elem_type, vec_num * nunits); | |
7988 | else | |
7989 | aggr_type = vectype; | |
7990 | bump = vect_get_data_ptr_increment (dr, aggr_type, memory_access_type); | |
7991 | } | |
272c6793 | 7992 | |
c3a8f964 | 7993 | tree vec_mask = NULL_TREE; |
ebfd146a | 7994 | prev_stmt_info = NULL; |
4d694b27 | 7995 | poly_uint64 group_elt = 0; |
ebfd146a | 7996 | for (j = 0; j < ncopies; j++) |
b8698a0f | 7997 | { |
272c6793 | 7998 | /* 1. Create the vector or array pointer update chain. */ |
ebfd146a | 7999 | if (j == 0) |
74bf76ed JJ |
8000 | { |
8001 | bool simd_lane_access_p | |
8002 | = STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info); | |
8003 | if (simd_lane_access_p | |
8004 | && TREE_CODE (DR_BASE_ADDRESS (first_dr)) == ADDR_EXPR | |
8005 | && VAR_P (TREE_OPERAND (DR_BASE_ADDRESS (first_dr), 0)) | |
8006 | && integer_zerop (DR_OFFSET (first_dr)) | |
8007 | && integer_zerop (DR_INIT (first_dr)) | |
8008 | && alias_sets_conflict_p (get_alias_set (aggr_type), | |
44fc7854 | 8009 | get_alias_set (TREE_TYPE (ref_type))) |
74bf76ed JJ |
8010 | && (alignment_support_scheme == dr_aligned |
8011 | || alignment_support_scheme == dr_unaligned_supported)) | |
8012 | { | |
8013 | dataref_ptr = unshare_expr (DR_BASE_ADDRESS (first_dr)); | |
44fc7854 | 8014 | dataref_offset = build_int_cst (ref_type, 0); |
8928eff3 | 8015 | inv_p = false; |
74bf76ed | 8016 | } |
4f0a0218 RB |
8017 | else if (first_stmt_for_drptr |
8018 | && first_stmt != first_stmt_for_drptr) | |
8019 | { | |
8020 | dataref_ptr | |
8021 | = vect_create_data_ref_ptr (first_stmt_for_drptr, aggr_type, | |
8022 | at_loop, offset, &dummy, gsi, | |
8023 | &ptr_incr, simd_lane_access_p, | |
ab2fc782 | 8024 | &inv_p, byte_offset, bump); |
4f0a0218 RB |
8025 | /* Adjust the pointer by the difference to first_stmt. */ |
8026 | data_reference_p ptrdr | |
8027 | = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt_for_drptr)); | |
8028 | tree diff = fold_convert (sizetype, | |
8029 | size_binop (MINUS_EXPR, | |
8030 | DR_INIT (first_dr), | |
8031 | DR_INIT (ptrdr))); | |
8032 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
8033 | stmt, diff); | |
8034 | } | |
bfaa08b7 RS |
8035 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
8036 | { | |
8037 | vect_get_gather_scatter_ops (loop, stmt, &gs_info, | |
8038 | &dataref_ptr, &vec_offset); | |
8039 | inv_p = false; | |
8040 | } | |
74bf76ed JJ |
8041 | else |
8042 | dataref_ptr | |
8043 | = vect_create_data_ref_ptr (first_stmt, aggr_type, at_loop, | |
8044 | offset, &dummy, gsi, &ptr_incr, | |
356bbc4c | 8045 | simd_lane_access_p, &inv_p, |
ab2fc782 | 8046 | byte_offset, bump); |
c3a8f964 RS |
8047 | if (mask) |
8048 | vec_mask = vect_get_vec_def_for_operand (mask, stmt, | |
8049 | mask_vectype); | |
74bf76ed | 8050 | } |
ebfd146a | 8051 | else |
c3a8f964 RS |
8052 | { |
8053 | if (dataref_offset) | |
8054 | dataref_offset = int_const_binop (PLUS_EXPR, dataref_offset, | |
ab2fc782 | 8055 | bump); |
bfaa08b7 | 8056 | else if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)) |
929b4411 RS |
8057 | vec_offset = vect_get_vec_def_for_stmt_copy (gs_info.offset_dt, |
8058 | vec_offset); | |
c3a8f964 | 8059 | else |
ab2fc782 RS |
8060 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
8061 | stmt, bump); | |
c3a8f964 | 8062 | if (mask) |
929b4411 | 8063 | vec_mask = vect_get_vec_def_for_stmt_copy (mask_dt, vec_mask); |
c3a8f964 | 8064 | } |
ebfd146a | 8065 | |
0d0293ac | 8066 | if (grouped_load || slp_perm) |
9771b263 | 8067 | dr_chain.create (vec_num); |
5ce1ee7f | 8068 | |
2de001ee | 8069 | if (memory_access_type == VMAT_LOAD_STORE_LANES) |
ebfd146a | 8070 | { |
272c6793 RS |
8071 | tree vec_array; |
8072 | ||
8073 | vec_array = create_vector_array (vectype, vec_num); | |
8074 | ||
7cfb4d93 | 8075 | tree final_mask = NULL_TREE; |
70088b95 RS |
8076 | if (loop_masks) |
8077 | final_mask = vect_get_loop_mask (gsi, loop_masks, ncopies, | |
8078 | vectype, j); | |
7cfb4d93 RS |
8079 | if (vec_mask) |
8080 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
8081 | vec_mask, gsi); | |
8082 | ||
7e11fc7f | 8083 | gcall *call; |
7cfb4d93 | 8084 | if (final_mask) |
7e11fc7f RS |
8085 | { |
8086 | /* Emit: | |
8087 | VEC_ARRAY = MASK_LOAD_LANES (DATAREF_PTR, ALIAS_PTR, | |
8088 | VEC_MASK). */ | |
8089 | unsigned int align = TYPE_ALIGN_UNIT (TREE_TYPE (vectype)); | |
8090 | tree alias_ptr = build_int_cst (ref_type, align); | |
8091 | call = gimple_build_call_internal (IFN_MASK_LOAD_LANES, 3, | |
8092 | dataref_ptr, alias_ptr, | |
7cfb4d93 | 8093 | final_mask); |
7e11fc7f RS |
8094 | } |
8095 | else | |
8096 | { | |
8097 | /* Emit: | |
8098 | VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */ | |
8099 | data_ref = create_array_ref (aggr_type, dataref_ptr, ref_type); | |
8100 | call = gimple_build_call_internal (IFN_LOAD_LANES, 1, data_ref); | |
8101 | } | |
a844293d RS |
8102 | gimple_call_set_lhs (call, vec_array); |
8103 | gimple_call_set_nothrow (call, true); | |
8104 | new_stmt = call; | |
272c6793 | 8105 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
ebfd146a | 8106 | |
272c6793 RS |
8107 | /* Extract each vector into an SSA_NAME. */ |
8108 | for (i = 0; i < vec_num; i++) | |
ebfd146a | 8109 | { |
272c6793 RS |
8110 | new_temp = read_vector_array (stmt, gsi, scalar_dest, |
8111 | vec_array, i); | |
9771b263 | 8112 | dr_chain.quick_push (new_temp); |
272c6793 RS |
8113 | } |
8114 | ||
8115 | /* Record the mapping between SSA_NAMEs and statements. */ | |
0d0293ac | 8116 | vect_record_grouped_load_vectors (stmt, dr_chain); |
3ba4ff41 RS |
8117 | |
8118 | /* Record that VEC_ARRAY is now dead. */ | |
8119 | vect_clobber_variable (stmt, gsi, vec_array); | |
272c6793 RS |
8120 | } |
8121 | else | |
8122 | { | |
8123 | for (i = 0; i < vec_num; i++) | |
8124 | { | |
7cfb4d93 | 8125 | tree final_mask = NULL_TREE; |
70088b95 | 8126 | if (loop_masks |
7cfb4d93 | 8127 | && memory_access_type != VMAT_INVARIANT) |
70088b95 RS |
8128 | final_mask = vect_get_loop_mask (gsi, loop_masks, |
8129 | vec_num * ncopies, | |
7cfb4d93 RS |
8130 | vectype, vec_num * j + i); |
8131 | if (vec_mask) | |
8132 | final_mask = prepare_load_store_mask (mask_vectype, final_mask, | |
8133 | vec_mask, gsi); | |
8134 | ||
272c6793 RS |
8135 | if (i > 0) |
8136 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, | |
ab2fc782 | 8137 | stmt, bump); |
272c6793 RS |
8138 | |
8139 | /* 2. Create the vector-load in the loop. */ | |
8140 | switch (alignment_support_scheme) | |
8141 | { | |
8142 | case dr_aligned: | |
8143 | case dr_unaligned_supported: | |
be1ac4ec | 8144 | { |
644ffefd MJ |
8145 | unsigned int align, misalign; |
8146 | ||
bfaa08b7 RS |
8147 | if (memory_access_type == VMAT_GATHER_SCATTER) |
8148 | { | |
8149 | tree scale = size_int (gs_info.scale); | |
8150 | gcall *call; | |
70088b95 | 8151 | if (loop_masks) |
bfaa08b7 RS |
8152 | call = gimple_build_call_internal |
8153 | (IFN_MASK_GATHER_LOAD, 4, dataref_ptr, | |
8154 | vec_offset, scale, final_mask); | |
8155 | else | |
8156 | call = gimple_build_call_internal | |
8157 | (IFN_GATHER_LOAD, 3, dataref_ptr, | |
8158 | vec_offset, scale); | |
8159 | gimple_call_set_nothrow (call, true); | |
8160 | new_stmt = call; | |
8161 | data_ref = NULL_TREE; | |
8162 | break; | |
8163 | } | |
8164 | ||
f702e7d4 | 8165 | align = DR_TARGET_ALIGNMENT (dr); |
272c6793 RS |
8166 | if (alignment_support_scheme == dr_aligned) |
8167 | { | |
8168 | gcc_assert (aligned_access_p (first_dr)); | |
644ffefd | 8169 | misalign = 0; |
272c6793 RS |
8170 | } |
8171 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
8172 | { | |
25f68d90 | 8173 | align = dr_alignment (vect_dr_behavior (first_dr)); |
52639a61 | 8174 | misalign = 0; |
272c6793 RS |
8175 | } |
8176 | else | |
c3a8f964 | 8177 | misalign = DR_MISALIGNMENT (first_dr); |
aed93b23 RB |
8178 | if (dataref_offset == NULL_TREE |
8179 | && TREE_CODE (dataref_ptr) == SSA_NAME) | |
74bf76ed JJ |
8180 | set_ptr_info_alignment (get_ptr_info (dataref_ptr), |
8181 | align, misalign); | |
c3a8f964 | 8182 | |
7cfb4d93 | 8183 | if (final_mask) |
c3a8f964 RS |
8184 | { |
8185 | align = least_bit_hwi (misalign | align); | |
8186 | tree ptr = build_int_cst (ref_type, align); | |
8187 | gcall *call | |
8188 | = gimple_build_call_internal (IFN_MASK_LOAD, 3, | |
8189 | dataref_ptr, ptr, | |
7cfb4d93 | 8190 | final_mask); |
c3a8f964 RS |
8191 | gimple_call_set_nothrow (call, true); |
8192 | new_stmt = call; | |
8193 | data_ref = NULL_TREE; | |
8194 | } | |
8195 | else | |
8196 | { | |
8197 | data_ref | |
8198 | = fold_build2 (MEM_REF, vectype, dataref_ptr, | |
8199 | dataref_offset | |
8200 | ? dataref_offset | |
8201 | : build_int_cst (ref_type, 0)); | |
8202 | if (alignment_support_scheme == dr_aligned) | |
8203 | ; | |
8204 | else if (DR_MISALIGNMENT (first_dr) == -1) | |
8205 | TREE_TYPE (data_ref) | |
8206 | = build_aligned_type (TREE_TYPE (data_ref), | |
8207 | align * BITS_PER_UNIT); | |
8208 | else | |
8209 | TREE_TYPE (data_ref) | |
8210 | = build_aligned_type (TREE_TYPE (data_ref), | |
8211 | TYPE_ALIGN (elem_type)); | |
8212 | } | |
272c6793 | 8213 | break; |
be1ac4ec | 8214 | } |
272c6793 | 8215 | case dr_explicit_realign: |
267d3070 | 8216 | { |
272c6793 | 8217 | tree ptr, bump; |
272c6793 | 8218 | |
d88981fc | 8219 | tree vs = size_int (TYPE_VECTOR_SUBPARTS (vectype)); |
272c6793 RS |
8220 | |
8221 | if (compute_in_loop) | |
8222 | msq = vect_setup_realignment (first_stmt, gsi, | |
8223 | &realignment_token, | |
8224 | dr_explicit_realign, | |
8225 | dataref_ptr, NULL); | |
8226 | ||
aed93b23 RB |
8227 | if (TREE_CODE (dataref_ptr) == SSA_NAME) |
8228 | ptr = copy_ssa_name (dataref_ptr); | |
8229 | else | |
8230 | ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
f702e7d4 | 8231 | unsigned int align = DR_TARGET_ALIGNMENT (first_dr); |
0d0e4a03 JJ |
8232 | new_stmt = gimple_build_assign |
8233 | (ptr, BIT_AND_EXPR, dataref_ptr, | |
272c6793 RS |
8234 | build_int_cst |
8235 | (TREE_TYPE (dataref_ptr), | |
f702e7d4 | 8236 | -(HOST_WIDE_INT) align)); |
272c6793 RS |
8237 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
8238 | data_ref | |
8239 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 8240 | build_int_cst (ref_type, 0)); |
19986382 | 8241 | vect_copy_ref_info (data_ref, DR_REF (first_dr)); |
272c6793 RS |
8242 | vec_dest = vect_create_destination_var (scalar_dest, |
8243 | vectype); | |
8244 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
8245 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
8246 | gimple_assign_set_lhs (new_stmt, new_temp); | |
8247 | gimple_set_vdef (new_stmt, gimple_vdef (stmt)); | |
8248 | gimple_set_vuse (new_stmt, gimple_vuse (stmt)); | |
8249 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8250 | msq = new_temp; | |
8251 | ||
d88981fc | 8252 | bump = size_binop (MULT_EXPR, vs, |
7b7b1813 | 8253 | TYPE_SIZE_UNIT (elem_type)); |
d88981fc | 8254 | bump = size_binop (MINUS_EXPR, bump, size_one_node); |
272c6793 | 8255 | ptr = bump_vector_ptr (dataref_ptr, NULL, gsi, stmt, bump); |
0d0e4a03 JJ |
8256 | new_stmt = gimple_build_assign |
8257 | (NULL_TREE, BIT_AND_EXPR, ptr, | |
272c6793 | 8258 | build_int_cst |
f702e7d4 | 8259 | (TREE_TYPE (ptr), -(HOST_WIDE_INT) align)); |
aed93b23 | 8260 | ptr = copy_ssa_name (ptr, new_stmt); |
272c6793 RS |
8261 | gimple_assign_set_lhs (new_stmt, ptr); |
8262 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8263 | data_ref | |
8264 | = build2 (MEM_REF, vectype, ptr, | |
44fc7854 | 8265 | build_int_cst (ref_type, 0)); |
272c6793 | 8266 | break; |
267d3070 | 8267 | } |
272c6793 | 8268 | case dr_explicit_realign_optimized: |
f702e7d4 RS |
8269 | { |
8270 | if (TREE_CODE (dataref_ptr) == SSA_NAME) | |
8271 | new_temp = copy_ssa_name (dataref_ptr); | |
8272 | else | |
8273 | new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); | |
8274 | unsigned int align = DR_TARGET_ALIGNMENT (first_dr); | |
8275 | new_stmt = gimple_build_assign | |
8276 | (new_temp, BIT_AND_EXPR, dataref_ptr, | |
8277 | build_int_cst (TREE_TYPE (dataref_ptr), | |
8278 | -(HOST_WIDE_INT) align)); | |
8279 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8280 | data_ref | |
8281 | = build2 (MEM_REF, vectype, new_temp, | |
8282 | build_int_cst (ref_type, 0)); | |
8283 | break; | |
8284 | } | |
272c6793 RS |
8285 | default: |
8286 | gcc_unreachable (); | |
8287 | } | |
ebfd146a | 8288 | vec_dest = vect_create_destination_var (scalar_dest, vectype); |
c3a8f964 RS |
8289 | /* DATA_REF is null if we've already built the statement. */ |
8290 | if (data_ref) | |
19986382 RB |
8291 | { |
8292 | vect_copy_ref_info (data_ref, DR_REF (first_dr)); | |
8293 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
8294 | } | |
ebfd146a | 8295 | new_temp = make_ssa_name (vec_dest, new_stmt); |
c3a8f964 | 8296 | gimple_set_lhs (new_stmt, new_temp); |
ebfd146a IR |
8297 | vect_finish_stmt_generation (stmt, new_stmt, gsi); |
8298 | ||
272c6793 RS |
8299 | /* 3. Handle explicit realignment if necessary/supported. |
8300 | Create in loop: | |
8301 | vec_dest = realign_load (msq, lsq, realignment_token) */ | |
8302 | if (alignment_support_scheme == dr_explicit_realign_optimized | |
8303 | || alignment_support_scheme == dr_explicit_realign) | |
ebfd146a | 8304 | { |
272c6793 RS |
8305 | lsq = gimple_assign_lhs (new_stmt); |
8306 | if (!realignment_token) | |
8307 | realignment_token = dataref_ptr; | |
8308 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
0d0e4a03 JJ |
8309 | new_stmt = gimple_build_assign (vec_dest, REALIGN_LOAD_EXPR, |
8310 | msq, lsq, realignment_token); | |
272c6793 RS |
8311 | new_temp = make_ssa_name (vec_dest, new_stmt); |
8312 | gimple_assign_set_lhs (new_stmt, new_temp); | |
8313 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8314 | ||
8315 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
8316 | { | |
8317 | gcc_assert (phi); | |
8318 | if (i == vec_num - 1 && j == ncopies - 1) | |
8319 | add_phi_arg (phi, lsq, | |
8320 | loop_latch_edge (containing_loop), | |
9e227d60 | 8321 | UNKNOWN_LOCATION); |
272c6793 RS |
8322 | msq = lsq; |
8323 | } | |
ebfd146a | 8324 | } |
ebfd146a | 8325 | |
59fd17e3 RB |
8326 | /* 4. Handle invariant-load. */ |
8327 | if (inv_p && !bb_vinfo) | |
8328 | { | |
59fd17e3 | 8329 | gcc_assert (!grouped_load); |
d1417442 JJ |
8330 | /* If we have versioned for aliasing or the loop doesn't |
8331 | have any data dependencies that would preclude this, | |
8332 | then we are sure this is a loop invariant load and | |
8333 | thus we can insert it on the preheader edge. */ | |
8334 | if (LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) | |
8335 | && !nested_in_vect_loop | |
6b916b36 | 8336 | && hoist_defs_of_uses (stmt, loop)) |
a0e35eb0 RB |
8337 | { |
8338 | if (dump_enabled_p ()) | |
8339 | { | |
8340 | dump_printf_loc (MSG_NOTE, vect_location, | |
8341 | "hoisting out of the vectorized " | |
8342 | "loop: "); | |
8343 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
a0e35eb0 | 8344 | } |
b731b390 | 8345 | tree tem = copy_ssa_name (scalar_dest); |
a0e35eb0 RB |
8346 | gsi_insert_on_edge_immediate |
8347 | (loop_preheader_edge (loop), | |
8348 | gimple_build_assign (tem, | |
8349 | unshare_expr | |
8350 | (gimple_assign_rhs1 (stmt)))); | |
8351 | new_temp = vect_init_vector (stmt, tem, vectype, NULL); | |
34cd48e5 RB |
8352 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
8353 | set_vinfo_for_stmt (new_stmt, | |
8354 | new_stmt_vec_info (new_stmt, vinfo)); | |
a0e35eb0 RB |
8355 | } |
8356 | else | |
8357 | { | |
8358 | gimple_stmt_iterator gsi2 = *gsi; | |
8359 | gsi_next (&gsi2); | |
8360 | new_temp = vect_init_vector (stmt, scalar_dest, | |
8361 | vectype, &gsi2); | |
34cd48e5 | 8362 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
a0e35eb0 | 8363 | } |
59fd17e3 RB |
8364 | } |
8365 | ||
62da9e14 | 8366 | if (memory_access_type == VMAT_CONTIGUOUS_REVERSE) |
272c6793 | 8367 | { |
aec7ae7d JJ |
8368 | tree perm_mask = perm_mask_for_reverse (vectype); |
8369 | new_temp = permute_vec_elements (new_temp, new_temp, | |
8370 | perm_mask, stmt, gsi); | |
ebfd146a IR |
8371 | new_stmt = SSA_NAME_DEF_STMT (new_temp); |
8372 | } | |
267d3070 | 8373 | |
272c6793 | 8374 | /* Collect vector loads and later create their permutation in |
0d0293ac MM |
8375 | vect_transform_grouped_load (). */ |
8376 | if (grouped_load || slp_perm) | |
9771b263 | 8377 | dr_chain.quick_push (new_temp); |
267d3070 | 8378 | |
272c6793 RS |
8379 | /* Store vector loads in the corresponding SLP_NODE. */ |
8380 | if (slp && !slp_perm) | |
9771b263 | 8381 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
b267968e RB |
8382 | |
8383 | /* With SLP permutation we load the gaps as well, without | |
8384 | we need to skip the gaps after we manage to fully load | |
8385 | all elements. group_gap_adj is GROUP_SIZE here. */ | |
8386 | group_elt += nunits; | |
d9f21f6a RS |
8387 | if (maybe_ne (group_gap_adj, 0U) |
8388 | && !slp_perm | |
8389 | && known_eq (group_elt, group_size - group_gap_adj)) | |
b267968e | 8390 | { |
d9f21f6a RS |
8391 | poly_wide_int bump_val |
8392 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) | |
8393 | * group_gap_adj); | |
8e6cdc90 | 8394 | tree bump = wide_int_to_tree (sizetype, bump_val); |
b267968e RB |
8395 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
8396 | stmt, bump); | |
8397 | group_elt = 0; | |
8398 | } | |
272c6793 | 8399 | } |
9b999e8c RB |
8400 | /* Bump the vector pointer to account for a gap or for excess |
8401 | elements loaded for a permuted SLP load. */ | |
d9f21f6a | 8402 | if (maybe_ne (group_gap_adj, 0U) && slp_perm) |
a64b9c26 | 8403 | { |
d9f21f6a RS |
8404 | poly_wide_int bump_val |
8405 | = (wi::to_wide (TYPE_SIZE_UNIT (elem_type)) | |
8406 | * group_gap_adj); | |
8e6cdc90 | 8407 | tree bump = wide_int_to_tree (sizetype, bump_val); |
a64b9c26 RB |
8408 | dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, gsi, |
8409 | stmt, bump); | |
8410 | } | |
ebfd146a IR |
8411 | } |
8412 | ||
8413 | if (slp && !slp_perm) | |
8414 | continue; | |
8415 | ||
8416 | if (slp_perm) | |
8417 | { | |
29afecdf | 8418 | unsigned n_perms; |
01d8bf07 | 8419 | if (!vect_transform_slp_perm_load (slp_node, dr_chain, gsi, vf, |
29afecdf RB |
8420 | slp_node_instance, false, |
8421 | &n_perms)) | |
ebfd146a | 8422 | { |
9771b263 | 8423 | dr_chain.release (); |
ebfd146a IR |
8424 | return false; |
8425 | } | |
8426 | } | |
8427 | else | |
8428 | { | |
0d0293ac | 8429 | if (grouped_load) |
ebfd146a | 8430 | { |
2de001ee | 8431 | if (memory_access_type != VMAT_LOAD_STORE_LANES) |
0d0293ac | 8432 | vect_transform_grouped_load (stmt, dr_chain, group_size, gsi); |
ebfd146a | 8433 | *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
ebfd146a IR |
8434 | } |
8435 | else | |
8436 | { | |
8437 | if (j == 0) | |
8438 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
8439 | else | |
8440 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
8441 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
8442 | } | |
8443 | } | |
9771b263 | 8444 | dr_chain.release (); |
ebfd146a IR |
8445 | } |
8446 | ||
ebfd146a IR |
8447 | return true; |
8448 | } | |
8449 | ||
8450 | /* Function vect_is_simple_cond. | |
b8698a0f | 8451 | |
ebfd146a IR |
8452 | Input: |
8453 | LOOP - the loop that is being vectorized. | |
8454 | COND - Condition that is checked for simple use. | |
8455 | ||
e9e1d143 RG |
8456 | Output: |
8457 | *COMP_VECTYPE - the vector type for the comparison. | |
4fc5ebf1 | 8458 | *DTS - The def types for the arguments of the comparison |
e9e1d143 | 8459 | |
ebfd146a IR |
8460 | Returns whether a COND can be vectorized. Checks whether |
8461 | condition operands are supportable using vec_is_simple_use. */ | |
8462 | ||
87aab9b2 | 8463 | static bool |
4fc5ebf1 | 8464 | vect_is_simple_cond (tree cond, vec_info *vinfo, |
8da4c8d8 RB |
8465 | tree *comp_vectype, enum vect_def_type *dts, |
8466 | tree vectype) | |
ebfd146a IR |
8467 | { |
8468 | tree lhs, rhs; | |
e9e1d143 | 8469 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
ebfd146a | 8470 | |
a414c77f IE |
8471 | /* Mask case. */ |
8472 | if (TREE_CODE (cond) == SSA_NAME | |
2568d8a1 | 8473 | && VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (cond))) |
a414c77f IE |
8474 | { |
8475 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (cond); | |
8476 | if (!vect_is_simple_use (cond, vinfo, &lhs_def_stmt, | |
4fc5ebf1 | 8477 | &dts[0], comp_vectype) |
a414c77f IE |
8478 | || !*comp_vectype |
8479 | || !VECTOR_BOOLEAN_TYPE_P (*comp_vectype)) | |
8480 | return false; | |
8481 | return true; | |
8482 | } | |
8483 | ||
ebfd146a IR |
8484 | if (!COMPARISON_CLASS_P (cond)) |
8485 | return false; | |
8486 | ||
8487 | lhs = TREE_OPERAND (cond, 0); | |
8488 | rhs = TREE_OPERAND (cond, 1); | |
8489 | ||
8490 | if (TREE_CODE (lhs) == SSA_NAME) | |
8491 | { | |
355fe088 | 8492 | gimple *lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); |
4fc5ebf1 | 8493 | if (!vect_is_simple_use (lhs, vinfo, &lhs_def_stmt, &dts[0], &vectype1)) |
ebfd146a IR |
8494 | return false; |
8495 | } | |
4fc5ebf1 JG |
8496 | else if (TREE_CODE (lhs) == INTEGER_CST || TREE_CODE (lhs) == REAL_CST |
8497 | || TREE_CODE (lhs) == FIXED_CST) | |
8498 | dts[0] = vect_constant_def; | |
8499 | else | |
ebfd146a IR |
8500 | return false; |
8501 | ||
8502 | if (TREE_CODE (rhs) == SSA_NAME) | |
8503 | { | |
355fe088 | 8504 | gimple *rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); |
4fc5ebf1 | 8505 | if (!vect_is_simple_use (rhs, vinfo, &rhs_def_stmt, &dts[1], &vectype2)) |
ebfd146a IR |
8506 | return false; |
8507 | } | |
4fc5ebf1 JG |
8508 | else if (TREE_CODE (rhs) == INTEGER_CST || TREE_CODE (rhs) == REAL_CST |
8509 | || TREE_CODE (rhs) == FIXED_CST) | |
8510 | dts[1] = vect_constant_def; | |
8511 | else | |
ebfd146a IR |
8512 | return false; |
8513 | ||
28b33016 | 8514 | if (vectype1 && vectype2 |
928686b1 RS |
8515 | && maybe_ne (TYPE_VECTOR_SUBPARTS (vectype1), |
8516 | TYPE_VECTOR_SUBPARTS (vectype2))) | |
28b33016 IE |
8517 | return false; |
8518 | ||
e9e1d143 | 8519 | *comp_vectype = vectype1 ? vectype1 : vectype2; |
8da4c8d8 RB |
8520 | /* Invariant comparison. */ |
8521 | if (! *comp_vectype) | |
8522 | { | |
8523 | tree scalar_type = TREE_TYPE (lhs); | |
8524 | /* If we can widen the comparison to match vectype do so. */ | |
8525 | if (INTEGRAL_TYPE_P (scalar_type) | |
8526 | && tree_int_cst_lt (TYPE_SIZE (scalar_type), | |
8527 | TYPE_SIZE (TREE_TYPE (vectype)))) | |
8528 | scalar_type = build_nonstandard_integer_type | |
8529 | (tree_to_uhwi (TYPE_SIZE (TREE_TYPE (vectype))), | |
8530 | TYPE_UNSIGNED (scalar_type)); | |
8531 | *comp_vectype = get_vectype_for_scalar_type (scalar_type); | |
8532 | } | |
8533 | ||
ebfd146a IR |
8534 | return true; |
8535 | } | |
8536 | ||
8537 | /* vectorizable_condition. | |
8538 | ||
b8698a0f L |
8539 | Check if STMT is conditional modify expression that can be vectorized. |
8540 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
8541 | stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it | |
4bbe8262 IR |
8542 | at GSI. |
8543 | ||
8544 | When STMT is vectorized as nested cycle, REDUC_DEF is the vector variable | |
8545 | to be used at REDUC_INDEX (in then clause if REDUC_INDEX is 1, and in | |
0ad23163 | 8546 | else clause if it is 2). |
ebfd146a IR |
8547 | |
8548 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
8549 | ||
4bbe8262 | 8550 | bool |
355fe088 TS |
8551 | vectorizable_condition (gimple *stmt, gimple_stmt_iterator *gsi, |
8552 | gimple **vec_stmt, tree reduc_def, int reduc_index, | |
f7e531cf | 8553 | slp_tree slp_node) |
ebfd146a IR |
8554 | { |
8555 | tree scalar_dest = NULL_TREE; | |
8556 | tree vec_dest = NULL_TREE; | |
01216d27 JJ |
8557 | tree cond_expr, cond_expr0 = NULL_TREE, cond_expr1 = NULL_TREE; |
8558 | tree then_clause, else_clause; | |
ebfd146a | 8559 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
df11cc78 | 8560 | tree comp_vectype = NULL_TREE; |
ff802fa1 IR |
8561 | tree vec_cond_lhs = NULL_TREE, vec_cond_rhs = NULL_TREE; |
8562 | tree vec_then_clause = NULL_TREE, vec_else_clause = NULL_TREE; | |
5958f9e2 | 8563 | tree vec_compare; |
ebfd146a IR |
8564 | tree new_temp; |
8565 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4fc5ebf1 JG |
8566 | enum vect_def_type dts[4] |
8567 | = {vect_unknown_def_type, vect_unknown_def_type, | |
8568 | vect_unknown_def_type, vect_unknown_def_type}; | |
8569 | int ndts = 4; | |
f7e531cf | 8570 | int ncopies; |
01216d27 | 8571 | enum tree_code code, cond_code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
a855b1b1 | 8572 | stmt_vec_info prev_stmt_info = NULL; |
f7e531cf IR |
8573 | int i, j; |
8574 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
6e1aa848 DN |
8575 | vec<tree> vec_oprnds0 = vNULL; |
8576 | vec<tree> vec_oprnds1 = vNULL; | |
8577 | vec<tree> vec_oprnds2 = vNULL; | |
8578 | vec<tree> vec_oprnds3 = vNULL; | |
74946978 | 8579 | tree vec_cmp_type; |
a414c77f | 8580 | bool masked = false; |
b8698a0f | 8581 | |
f7e531cf IR |
8582 | if (reduc_index && STMT_SLP_TYPE (stmt_info)) |
8583 | return false; | |
8584 | ||
bb6c2b68 RS |
8585 | vect_reduction_type reduction_type |
8586 | = STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info); | |
8587 | if (reduction_type == TREE_CODE_REDUCTION) | |
af29617a AH |
8588 | { |
8589 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) | |
8590 | return false; | |
ebfd146a | 8591 | |
af29617a AH |
8592 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
8593 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
8594 | && reduc_def)) | |
8595 | return false; | |
ebfd146a | 8596 | |
af29617a AH |
8597 | /* FORNOW: not yet supported. */ |
8598 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
8599 | { | |
8600 | if (dump_enabled_p ()) | |
8601 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
8602 | "value used after loop.\n"); | |
8603 | return false; | |
8604 | } | |
ebfd146a IR |
8605 | } |
8606 | ||
8607 | /* Is vectorizable conditional operation? */ | |
8608 | if (!is_gimple_assign (stmt)) | |
8609 | return false; | |
8610 | ||
8611 | code = gimple_assign_rhs_code (stmt); | |
8612 | ||
8613 | if (code != COND_EXPR) | |
8614 | return false; | |
8615 | ||
465c8c19 | 8616 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
2947d3b2 | 8617 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; |
465c8c19 | 8618 | |
fce57248 | 8619 | if (slp_node) |
465c8c19 JJ |
8620 | ncopies = 1; |
8621 | else | |
e8f142e2 | 8622 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
465c8c19 JJ |
8623 | |
8624 | gcc_assert (ncopies >= 1); | |
8625 | if (reduc_index && ncopies > 1) | |
8626 | return false; /* FORNOW */ | |
8627 | ||
4e71066d RG |
8628 | cond_expr = gimple_assign_rhs1 (stmt); |
8629 | then_clause = gimple_assign_rhs2 (stmt); | |
8630 | else_clause = gimple_assign_rhs3 (stmt); | |
ebfd146a | 8631 | |
4fc5ebf1 | 8632 | if (!vect_is_simple_cond (cond_expr, stmt_info->vinfo, |
8da4c8d8 | 8633 | &comp_vectype, &dts[0], vectype) |
e9e1d143 | 8634 | || !comp_vectype) |
ebfd146a IR |
8635 | return false; |
8636 | ||
81c40241 | 8637 | gimple *def_stmt; |
4fc5ebf1 | 8638 | if (!vect_is_simple_use (then_clause, stmt_info->vinfo, &def_stmt, &dts[2], |
2947d3b2 IE |
8639 | &vectype1)) |
8640 | return false; | |
4fc5ebf1 | 8641 | if (!vect_is_simple_use (else_clause, stmt_info->vinfo, &def_stmt, &dts[3], |
2947d3b2 | 8642 | &vectype2)) |
ebfd146a | 8643 | return false; |
2947d3b2 IE |
8644 | |
8645 | if (vectype1 && !useless_type_conversion_p (vectype, vectype1)) | |
8646 | return false; | |
8647 | ||
8648 | if (vectype2 && !useless_type_conversion_p (vectype, vectype2)) | |
ebfd146a IR |
8649 | return false; |
8650 | ||
28b33016 IE |
8651 | masked = !COMPARISON_CLASS_P (cond_expr); |
8652 | vec_cmp_type = build_same_sized_truth_vector_type (comp_vectype); | |
8653 | ||
74946978 MP |
8654 | if (vec_cmp_type == NULL_TREE) |
8655 | return false; | |
784fb9b3 | 8656 | |
01216d27 JJ |
8657 | cond_code = TREE_CODE (cond_expr); |
8658 | if (!masked) | |
8659 | { | |
8660 | cond_expr0 = TREE_OPERAND (cond_expr, 0); | |
8661 | cond_expr1 = TREE_OPERAND (cond_expr, 1); | |
8662 | } | |
8663 | ||
8664 | if (!masked && VECTOR_BOOLEAN_TYPE_P (comp_vectype)) | |
8665 | { | |
8666 | /* Boolean values may have another representation in vectors | |
8667 | and therefore we prefer bit operations over comparison for | |
8668 | them (which also works for scalar masks). We store opcodes | |
8669 | to use in bitop1 and bitop2. Statement is vectorized as | |
8670 | BITOP2 (rhs1 BITOP1 rhs2) or rhs1 BITOP2 (BITOP1 rhs2) | |
8671 | depending on bitop1 and bitop2 arity. */ | |
8672 | switch (cond_code) | |
8673 | { | |
8674 | case GT_EXPR: | |
8675 | bitop1 = BIT_NOT_EXPR; | |
8676 | bitop2 = BIT_AND_EXPR; | |
8677 | break; | |
8678 | case GE_EXPR: | |
8679 | bitop1 = BIT_NOT_EXPR; | |
8680 | bitop2 = BIT_IOR_EXPR; | |
8681 | break; | |
8682 | case LT_EXPR: | |
8683 | bitop1 = BIT_NOT_EXPR; | |
8684 | bitop2 = BIT_AND_EXPR; | |
8685 | std::swap (cond_expr0, cond_expr1); | |
8686 | break; | |
8687 | case LE_EXPR: | |
8688 | bitop1 = BIT_NOT_EXPR; | |
8689 | bitop2 = BIT_IOR_EXPR; | |
8690 | std::swap (cond_expr0, cond_expr1); | |
8691 | break; | |
8692 | case NE_EXPR: | |
8693 | bitop1 = BIT_XOR_EXPR; | |
8694 | break; | |
8695 | case EQ_EXPR: | |
8696 | bitop1 = BIT_XOR_EXPR; | |
8697 | bitop2 = BIT_NOT_EXPR; | |
8698 | break; | |
8699 | default: | |
8700 | return false; | |
8701 | } | |
8702 | cond_code = SSA_NAME; | |
8703 | } | |
8704 | ||
b8698a0f | 8705 | if (!vec_stmt) |
ebfd146a IR |
8706 | { |
8707 | STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; | |
01216d27 JJ |
8708 | if (bitop1 != NOP_EXPR) |
8709 | { | |
8710 | machine_mode mode = TYPE_MODE (comp_vectype); | |
8711 | optab optab; | |
8712 | ||
8713 | optab = optab_for_tree_code (bitop1, comp_vectype, optab_default); | |
8714 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8715 | return false; | |
8716 | ||
8717 | if (bitop2 != NOP_EXPR) | |
8718 | { | |
8719 | optab = optab_for_tree_code (bitop2, comp_vectype, | |
8720 | optab_default); | |
8721 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
8722 | return false; | |
8723 | } | |
8724 | } | |
4fc5ebf1 JG |
8725 | if (expand_vec_cond_expr_p (vectype, comp_vectype, |
8726 | cond_code)) | |
8727 | { | |
78604de0 RB |
8728 | if (!slp_node) |
8729 | vect_model_simple_cost (stmt_info, ncopies, dts, ndts, NULL, NULL); | |
4fc5ebf1 JG |
8730 | return true; |
8731 | } | |
8732 | return false; | |
ebfd146a IR |
8733 | } |
8734 | ||
f7e531cf IR |
8735 | /* Transform. */ |
8736 | ||
8737 | if (!slp_node) | |
8738 | { | |
9771b263 DN |
8739 | vec_oprnds0.create (1); |
8740 | vec_oprnds1.create (1); | |
8741 | vec_oprnds2.create (1); | |
8742 | vec_oprnds3.create (1); | |
f7e531cf | 8743 | } |
ebfd146a IR |
8744 | |
8745 | /* Handle def. */ | |
8746 | scalar_dest = gimple_assign_lhs (stmt); | |
bb6c2b68 RS |
8747 | if (reduction_type != EXTRACT_LAST_REDUCTION) |
8748 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
ebfd146a IR |
8749 | |
8750 | /* Handle cond expr. */ | |
a855b1b1 MM |
8751 | for (j = 0; j < ncopies; j++) |
8752 | { | |
bb6c2b68 | 8753 | gimple *new_stmt = NULL; |
a855b1b1 MM |
8754 | if (j == 0) |
8755 | { | |
f7e531cf IR |
8756 | if (slp_node) |
8757 | { | |
00f96dc9 TS |
8758 | auto_vec<tree, 4> ops; |
8759 | auto_vec<vec<tree>, 4> vec_defs; | |
9771b263 | 8760 | |
a414c77f | 8761 | if (masked) |
01216d27 | 8762 | ops.safe_push (cond_expr); |
a414c77f IE |
8763 | else |
8764 | { | |
01216d27 JJ |
8765 | ops.safe_push (cond_expr0); |
8766 | ops.safe_push (cond_expr1); | |
a414c77f | 8767 | } |
9771b263 DN |
8768 | ops.safe_push (then_clause); |
8769 | ops.safe_push (else_clause); | |
306b0c92 | 8770 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
37b5ec8f JJ |
8771 | vec_oprnds3 = vec_defs.pop (); |
8772 | vec_oprnds2 = vec_defs.pop (); | |
a414c77f IE |
8773 | if (!masked) |
8774 | vec_oprnds1 = vec_defs.pop (); | |
37b5ec8f | 8775 | vec_oprnds0 = vec_defs.pop (); |
f7e531cf IR |
8776 | } |
8777 | else | |
8778 | { | |
355fe088 | 8779 | gimple *gtemp; |
a414c77f IE |
8780 | if (masked) |
8781 | { | |
8782 | vec_cond_lhs | |
8783 | = vect_get_vec_def_for_operand (cond_expr, stmt, | |
8784 | comp_vectype); | |
8785 | vect_is_simple_use (cond_expr, stmt_info->vinfo, | |
8786 | >emp, &dts[0]); | |
8787 | } | |
8788 | else | |
8789 | { | |
01216d27 JJ |
8790 | vec_cond_lhs |
8791 | = vect_get_vec_def_for_operand (cond_expr0, | |
8792 | stmt, comp_vectype); | |
8793 | vect_is_simple_use (cond_expr0, loop_vinfo, >emp, &dts[0]); | |
8794 | ||
8795 | vec_cond_rhs | |
8796 | = vect_get_vec_def_for_operand (cond_expr1, | |
8797 | stmt, comp_vectype); | |
8798 | vect_is_simple_use (cond_expr1, loop_vinfo, >emp, &dts[1]); | |
a414c77f | 8799 | } |
f7e531cf IR |
8800 | if (reduc_index == 1) |
8801 | vec_then_clause = reduc_def; | |
8802 | else | |
8803 | { | |
8804 | vec_then_clause = vect_get_vec_def_for_operand (then_clause, | |
81c40241 RB |
8805 | stmt); |
8806 | vect_is_simple_use (then_clause, loop_vinfo, | |
8807 | >emp, &dts[2]); | |
f7e531cf IR |
8808 | } |
8809 | if (reduc_index == 2) | |
8810 | vec_else_clause = reduc_def; | |
8811 | else | |
8812 | { | |
8813 | vec_else_clause = vect_get_vec_def_for_operand (else_clause, | |
81c40241 RB |
8814 | stmt); |
8815 | vect_is_simple_use (else_clause, loop_vinfo, >emp, &dts[3]); | |
f7e531cf | 8816 | } |
a855b1b1 MM |
8817 | } |
8818 | } | |
8819 | else | |
8820 | { | |
a414c77f IE |
8821 | vec_cond_lhs |
8822 | = vect_get_vec_def_for_stmt_copy (dts[0], | |
8823 | vec_oprnds0.pop ()); | |
8824 | if (!masked) | |
8825 | vec_cond_rhs | |
8826 | = vect_get_vec_def_for_stmt_copy (dts[1], | |
8827 | vec_oprnds1.pop ()); | |
8828 | ||
a855b1b1 | 8829 | vec_then_clause = vect_get_vec_def_for_stmt_copy (dts[2], |
9771b263 | 8830 | vec_oprnds2.pop ()); |
a855b1b1 | 8831 | vec_else_clause = vect_get_vec_def_for_stmt_copy (dts[3], |
9771b263 | 8832 | vec_oprnds3.pop ()); |
f7e531cf IR |
8833 | } |
8834 | ||
8835 | if (!slp_node) | |
8836 | { | |
9771b263 | 8837 | vec_oprnds0.quick_push (vec_cond_lhs); |
a414c77f IE |
8838 | if (!masked) |
8839 | vec_oprnds1.quick_push (vec_cond_rhs); | |
9771b263 DN |
8840 | vec_oprnds2.quick_push (vec_then_clause); |
8841 | vec_oprnds3.quick_push (vec_else_clause); | |
a855b1b1 MM |
8842 | } |
8843 | ||
9dc3f7de | 8844 | /* Arguments are ready. Create the new vector stmt. */ |
9771b263 | 8845 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_cond_lhs) |
f7e531cf | 8846 | { |
9771b263 DN |
8847 | vec_then_clause = vec_oprnds2[i]; |
8848 | vec_else_clause = vec_oprnds3[i]; | |
a855b1b1 | 8849 | |
a414c77f IE |
8850 | if (masked) |
8851 | vec_compare = vec_cond_lhs; | |
8852 | else | |
8853 | { | |
8854 | vec_cond_rhs = vec_oprnds1[i]; | |
01216d27 JJ |
8855 | if (bitop1 == NOP_EXPR) |
8856 | vec_compare = build2 (cond_code, vec_cmp_type, | |
8857 | vec_cond_lhs, vec_cond_rhs); | |
8858 | else | |
8859 | { | |
8860 | new_temp = make_ssa_name (vec_cmp_type); | |
8861 | if (bitop1 == BIT_NOT_EXPR) | |
8862 | new_stmt = gimple_build_assign (new_temp, bitop1, | |
8863 | vec_cond_rhs); | |
8864 | else | |
8865 | new_stmt | |
8866 | = gimple_build_assign (new_temp, bitop1, vec_cond_lhs, | |
8867 | vec_cond_rhs); | |
8868 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8869 | if (bitop2 == NOP_EXPR) | |
8870 | vec_compare = new_temp; | |
8871 | else if (bitop2 == BIT_NOT_EXPR) | |
8872 | { | |
8873 | /* Instead of doing ~x ? y : z do x ? z : y. */ | |
8874 | vec_compare = new_temp; | |
8875 | std::swap (vec_then_clause, vec_else_clause); | |
8876 | } | |
8877 | else | |
8878 | { | |
8879 | vec_compare = make_ssa_name (vec_cmp_type); | |
8880 | new_stmt | |
8881 | = gimple_build_assign (vec_compare, bitop2, | |
8882 | vec_cond_lhs, new_temp); | |
8883 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8884 | } | |
8885 | } | |
a414c77f | 8886 | } |
bb6c2b68 RS |
8887 | if (reduction_type == EXTRACT_LAST_REDUCTION) |
8888 | { | |
8889 | if (!is_gimple_val (vec_compare)) | |
8890 | { | |
8891 | tree vec_compare_name = make_ssa_name (vec_cmp_type); | |
8892 | new_stmt = gimple_build_assign (vec_compare_name, | |
8893 | vec_compare); | |
8894 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8895 | vec_compare = vec_compare_name; | |
8896 | } | |
8897 | gcc_assert (reduc_index == 2); | |
8898 | new_stmt = gimple_build_call_internal | |
8899 | (IFN_FOLD_EXTRACT_LAST, 3, else_clause, vec_compare, | |
8900 | vec_then_clause); | |
8901 | gimple_call_set_lhs (new_stmt, scalar_dest); | |
8902 | SSA_NAME_DEF_STMT (scalar_dest) = new_stmt; | |
8903 | if (stmt == gsi_stmt (*gsi)) | |
8904 | vect_finish_replace_stmt (stmt, new_stmt); | |
8905 | else | |
8906 | { | |
8907 | /* In this case we're moving the definition to later in the | |
8908 | block. That doesn't matter because the only uses of the | |
8909 | lhs are in phi statements. */ | |
8910 | gimple_stmt_iterator old_gsi = gsi_for_stmt (stmt); | |
8911 | gsi_remove (&old_gsi, true); | |
8912 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8913 | } | |
8914 | } | |
8915 | else | |
8916 | { | |
8917 | new_temp = make_ssa_name (vec_dest); | |
8918 | new_stmt = gimple_build_assign (new_temp, VEC_COND_EXPR, | |
8919 | vec_compare, vec_then_clause, | |
8920 | vec_else_clause); | |
8921 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
8922 | } | |
f7e531cf | 8923 | if (slp_node) |
9771b263 | 8924 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); |
f7e531cf IR |
8925 | } |
8926 | ||
8927 | if (slp_node) | |
8928 | continue; | |
8929 | ||
8930 | if (j == 0) | |
8931 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
8932 | else | |
8933 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
8934 | ||
8935 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
a855b1b1 | 8936 | } |
b8698a0f | 8937 | |
9771b263 DN |
8938 | vec_oprnds0.release (); |
8939 | vec_oprnds1.release (); | |
8940 | vec_oprnds2.release (); | |
8941 | vec_oprnds3.release (); | |
f7e531cf | 8942 | |
ebfd146a IR |
8943 | return true; |
8944 | } | |
8945 | ||
42fd8198 IE |
8946 | /* vectorizable_comparison. |
8947 | ||
8948 | Check if STMT is comparison expression that can be vectorized. | |
8949 | If VEC_STMT is also passed, vectorize the STMT: create a vectorized | |
8950 | comparison, put it in VEC_STMT, and insert it at GSI. | |
8951 | ||
8952 | Return FALSE if not a vectorizable STMT, TRUE otherwise. */ | |
8953 | ||
fce57248 | 8954 | static bool |
42fd8198 IE |
8955 | vectorizable_comparison (gimple *stmt, gimple_stmt_iterator *gsi, |
8956 | gimple **vec_stmt, tree reduc_def, | |
8957 | slp_tree slp_node) | |
8958 | { | |
8959 | tree lhs, rhs1, rhs2; | |
8960 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
8961 | tree vectype1 = NULL_TREE, vectype2 = NULL_TREE; | |
8962 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
8963 | tree vec_rhs1 = NULL_TREE, vec_rhs2 = NULL_TREE; | |
8964 | tree new_temp; | |
8965 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
8966 | enum vect_def_type dts[2] = {vect_unknown_def_type, vect_unknown_def_type}; | |
4fc5ebf1 | 8967 | int ndts = 2; |
928686b1 | 8968 | poly_uint64 nunits; |
42fd8198 | 8969 | int ncopies; |
49e76ff1 | 8970 | enum tree_code code, bitop1 = NOP_EXPR, bitop2 = NOP_EXPR; |
42fd8198 IE |
8971 | stmt_vec_info prev_stmt_info = NULL; |
8972 | int i, j; | |
8973 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); | |
8974 | vec<tree> vec_oprnds0 = vNULL; | |
8975 | vec<tree> vec_oprnds1 = vNULL; | |
8976 | gimple *def_stmt; | |
8977 | tree mask_type; | |
8978 | tree mask; | |
8979 | ||
c245362b IE |
8980 | if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo) |
8981 | return false; | |
8982 | ||
30480bcd | 8983 | if (!vectype || !VECTOR_BOOLEAN_TYPE_P (vectype)) |
42fd8198 IE |
8984 | return false; |
8985 | ||
8986 | mask_type = vectype; | |
8987 | nunits = TYPE_VECTOR_SUBPARTS (vectype); | |
8988 | ||
fce57248 | 8989 | if (slp_node) |
42fd8198 IE |
8990 | ncopies = 1; |
8991 | else | |
e8f142e2 | 8992 | ncopies = vect_get_num_copies (loop_vinfo, vectype); |
42fd8198 IE |
8993 | |
8994 | gcc_assert (ncopies >= 1); | |
42fd8198 IE |
8995 | if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def |
8996 | && !(STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle | |
8997 | && reduc_def)) | |
8998 | return false; | |
8999 | ||
9000 | if (STMT_VINFO_LIVE_P (stmt_info)) | |
9001 | { | |
9002 | if (dump_enabled_p ()) | |
9003 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
9004 | "value used after loop.\n"); | |
9005 | return false; | |
9006 | } | |
9007 | ||
9008 | if (!is_gimple_assign (stmt)) | |
9009 | return false; | |
9010 | ||
9011 | code = gimple_assign_rhs_code (stmt); | |
9012 | ||
9013 | if (TREE_CODE_CLASS (code) != tcc_comparison) | |
9014 | return false; | |
9015 | ||
9016 | rhs1 = gimple_assign_rhs1 (stmt); | |
9017 | rhs2 = gimple_assign_rhs2 (stmt); | |
9018 | ||
9019 | if (!vect_is_simple_use (rhs1, stmt_info->vinfo, &def_stmt, | |
9020 | &dts[0], &vectype1)) | |
9021 | return false; | |
9022 | ||
9023 | if (!vect_is_simple_use (rhs2, stmt_info->vinfo, &def_stmt, | |
9024 | &dts[1], &vectype2)) | |
9025 | return false; | |
9026 | ||
9027 | if (vectype1 && vectype2 | |
928686b1 RS |
9028 | && maybe_ne (TYPE_VECTOR_SUBPARTS (vectype1), |
9029 | TYPE_VECTOR_SUBPARTS (vectype2))) | |
42fd8198 IE |
9030 | return false; |
9031 | ||
9032 | vectype = vectype1 ? vectype1 : vectype2; | |
9033 | ||
9034 | /* Invariant comparison. */ | |
9035 | if (!vectype) | |
9036 | { | |
69a9a66f | 9037 | vectype = get_vectype_for_scalar_type (TREE_TYPE (rhs1)); |
928686b1 | 9038 | if (maybe_ne (TYPE_VECTOR_SUBPARTS (vectype), nunits)) |
42fd8198 IE |
9039 | return false; |
9040 | } | |
928686b1 | 9041 | else if (maybe_ne (nunits, TYPE_VECTOR_SUBPARTS (vectype))) |
42fd8198 IE |
9042 | return false; |
9043 | ||
49e76ff1 IE |
9044 | /* Can't compare mask and non-mask types. */ |
9045 | if (vectype1 && vectype2 | |
9046 | && (VECTOR_BOOLEAN_TYPE_P (vectype1) ^ VECTOR_BOOLEAN_TYPE_P (vectype2))) | |
9047 | return false; | |
9048 | ||
9049 | /* Boolean values may have another representation in vectors | |
9050 | and therefore we prefer bit operations over comparison for | |
9051 | them (which also works for scalar masks). We store opcodes | |
9052 | to use in bitop1 and bitop2. Statement is vectorized as | |
9053 | BITOP2 (rhs1 BITOP1 rhs2) or | |
9054 | rhs1 BITOP2 (BITOP1 rhs2) | |
9055 | depending on bitop1 and bitop2 arity. */ | |
9056 | if (VECTOR_BOOLEAN_TYPE_P (vectype)) | |
9057 | { | |
9058 | if (code == GT_EXPR) | |
9059 | { | |
9060 | bitop1 = BIT_NOT_EXPR; | |
9061 | bitop2 = BIT_AND_EXPR; | |
9062 | } | |
9063 | else if (code == GE_EXPR) | |
9064 | { | |
9065 | bitop1 = BIT_NOT_EXPR; | |
9066 | bitop2 = BIT_IOR_EXPR; | |
9067 | } | |
9068 | else if (code == LT_EXPR) | |
9069 | { | |
9070 | bitop1 = BIT_NOT_EXPR; | |
9071 | bitop2 = BIT_AND_EXPR; | |
9072 | std::swap (rhs1, rhs2); | |
264d951a | 9073 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
9074 | } |
9075 | else if (code == LE_EXPR) | |
9076 | { | |
9077 | bitop1 = BIT_NOT_EXPR; | |
9078 | bitop2 = BIT_IOR_EXPR; | |
9079 | std::swap (rhs1, rhs2); | |
264d951a | 9080 | std::swap (dts[0], dts[1]); |
49e76ff1 IE |
9081 | } |
9082 | else | |
9083 | { | |
9084 | bitop1 = BIT_XOR_EXPR; | |
9085 | if (code == EQ_EXPR) | |
9086 | bitop2 = BIT_NOT_EXPR; | |
9087 | } | |
9088 | } | |
9089 | ||
42fd8198 IE |
9090 | if (!vec_stmt) |
9091 | { | |
9092 | STMT_VINFO_TYPE (stmt_info) = comparison_vec_info_type; | |
78604de0 RB |
9093 | if (!slp_node) |
9094 | vect_model_simple_cost (stmt_info, ncopies * (1 + (bitop2 != NOP_EXPR)), | |
9095 | dts, ndts, NULL, NULL); | |
49e76ff1 | 9096 | if (bitop1 == NOP_EXPR) |
96592eed | 9097 | return expand_vec_cmp_expr_p (vectype, mask_type, code); |
49e76ff1 IE |
9098 | else |
9099 | { | |
9100 | machine_mode mode = TYPE_MODE (vectype); | |
9101 | optab optab; | |
9102 | ||
9103 | optab = optab_for_tree_code (bitop1, vectype, optab_default); | |
9104 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
9105 | return false; | |
9106 | ||
9107 | if (bitop2 != NOP_EXPR) | |
9108 | { | |
9109 | optab = optab_for_tree_code (bitop2, vectype, optab_default); | |
9110 | if (!optab || optab_handler (optab, mode) == CODE_FOR_nothing) | |
9111 | return false; | |
9112 | } | |
9113 | return true; | |
9114 | } | |
42fd8198 IE |
9115 | } |
9116 | ||
9117 | /* Transform. */ | |
9118 | if (!slp_node) | |
9119 | { | |
9120 | vec_oprnds0.create (1); | |
9121 | vec_oprnds1.create (1); | |
9122 | } | |
9123 | ||
9124 | /* Handle def. */ | |
9125 | lhs = gimple_assign_lhs (stmt); | |
9126 | mask = vect_create_destination_var (lhs, mask_type); | |
9127 | ||
9128 | /* Handle cmp expr. */ | |
9129 | for (j = 0; j < ncopies; j++) | |
9130 | { | |
9131 | gassign *new_stmt = NULL; | |
9132 | if (j == 0) | |
9133 | { | |
9134 | if (slp_node) | |
9135 | { | |
9136 | auto_vec<tree, 2> ops; | |
9137 | auto_vec<vec<tree>, 2> vec_defs; | |
9138 | ||
9139 | ops.safe_push (rhs1); | |
9140 | ops.safe_push (rhs2); | |
306b0c92 | 9141 | vect_get_slp_defs (ops, slp_node, &vec_defs); |
42fd8198 IE |
9142 | vec_oprnds1 = vec_defs.pop (); |
9143 | vec_oprnds0 = vec_defs.pop (); | |
9144 | } | |
9145 | else | |
9146 | { | |
e4af0bc4 IE |
9147 | vec_rhs1 = vect_get_vec_def_for_operand (rhs1, stmt, vectype); |
9148 | vec_rhs2 = vect_get_vec_def_for_operand (rhs2, stmt, vectype); | |
42fd8198 IE |
9149 | } |
9150 | } | |
9151 | else | |
9152 | { | |
9153 | vec_rhs1 = vect_get_vec_def_for_stmt_copy (dts[0], | |
9154 | vec_oprnds0.pop ()); | |
9155 | vec_rhs2 = vect_get_vec_def_for_stmt_copy (dts[1], | |
9156 | vec_oprnds1.pop ()); | |
9157 | } | |
9158 | ||
9159 | if (!slp_node) | |
9160 | { | |
9161 | vec_oprnds0.quick_push (vec_rhs1); | |
9162 | vec_oprnds1.quick_push (vec_rhs2); | |
9163 | } | |
9164 | ||
9165 | /* Arguments are ready. Create the new vector stmt. */ | |
9166 | FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_rhs1) | |
9167 | { | |
9168 | vec_rhs2 = vec_oprnds1[i]; | |
9169 | ||
9170 | new_temp = make_ssa_name (mask); | |
49e76ff1 IE |
9171 | if (bitop1 == NOP_EXPR) |
9172 | { | |
9173 | new_stmt = gimple_build_assign (new_temp, code, | |
9174 | vec_rhs1, vec_rhs2); | |
9175 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9176 | } | |
9177 | else | |
9178 | { | |
9179 | if (bitop1 == BIT_NOT_EXPR) | |
9180 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs2); | |
9181 | else | |
9182 | new_stmt = gimple_build_assign (new_temp, bitop1, vec_rhs1, | |
9183 | vec_rhs2); | |
9184 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9185 | if (bitop2 != NOP_EXPR) | |
9186 | { | |
9187 | tree res = make_ssa_name (mask); | |
9188 | if (bitop2 == BIT_NOT_EXPR) | |
9189 | new_stmt = gimple_build_assign (res, bitop2, new_temp); | |
9190 | else | |
9191 | new_stmt = gimple_build_assign (res, bitop2, vec_rhs1, | |
9192 | new_temp); | |
9193 | vect_finish_stmt_generation (stmt, new_stmt, gsi); | |
9194 | } | |
9195 | } | |
42fd8198 IE |
9196 | if (slp_node) |
9197 | SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); | |
9198 | } | |
9199 | ||
9200 | if (slp_node) | |
9201 | continue; | |
9202 | ||
9203 | if (j == 0) | |
9204 | STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; | |
9205 | else | |
9206 | STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; | |
9207 | ||
9208 | prev_stmt_info = vinfo_for_stmt (new_stmt); | |
9209 | } | |
9210 | ||
9211 | vec_oprnds0.release (); | |
9212 | vec_oprnds1.release (); | |
9213 | ||
9214 | return true; | |
9215 | } | |
ebfd146a | 9216 | |
68a0f2ff RS |
9217 | /* If SLP_NODE is nonnull, return true if vectorizable_live_operation |
9218 | can handle all live statements in the node. Otherwise return true | |
9219 | if STMT is not live or if vectorizable_live_operation can handle it. | |
9220 | GSI and VEC_STMT are as for vectorizable_live_operation. */ | |
9221 | ||
9222 | static bool | |
9223 | can_vectorize_live_stmts (gimple *stmt, gimple_stmt_iterator *gsi, | |
9224 | slp_tree slp_node, gimple **vec_stmt) | |
9225 | { | |
9226 | if (slp_node) | |
9227 | { | |
9228 | gimple *slp_stmt; | |
9229 | unsigned int i; | |
9230 | FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (slp_node), i, slp_stmt) | |
9231 | { | |
9232 | stmt_vec_info slp_stmt_info = vinfo_for_stmt (slp_stmt); | |
9233 | if (STMT_VINFO_LIVE_P (slp_stmt_info) | |
9234 | && !vectorizable_live_operation (slp_stmt, gsi, slp_node, i, | |
9235 | vec_stmt)) | |
9236 | return false; | |
9237 | } | |
9238 | } | |
9239 | else if (STMT_VINFO_LIVE_P (vinfo_for_stmt (stmt)) | |
9240 | && !vectorizable_live_operation (stmt, gsi, slp_node, -1, vec_stmt)) | |
9241 | return false; | |
9242 | ||
9243 | return true; | |
9244 | } | |
9245 | ||
8644a673 | 9246 | /* Make sure the statement is vectorizable. */ |
ebfd146a IR |
9247 | |
9248 | bool | |
891ad31c RB |
9249 | vect_analyze_stmt (gimple *stmt, bool *need_to_vectorize, slp_tree node, |
9250 | slp_instance node_instance) | |
ebfd146a | 9251 | { |
8644a673 | 9252 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
a70d6342 | 9253 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
b8698a0f | 9254 | enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info); |
ebfd146a | 9255 | bool ok; |
355fe088 | 9256 | gimple *pattern_stmt; |
363477c0 | 9257 | gimple_seq pattern_def_seq; |
ebfd146a | 9258 | |
73fbfcad | 9259 | if (dump_enabled_p ()) |
ebfd146a | 9260 | { |
78c60e3d SS |
9261 | dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: "); |
9262 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
8644a673 | 9263 | } |
ebfd146a | 9264 | |
1825a1f3 | 9265 | if (gimple_has_volatile_ops (stmt)) |
b8698a0f | 9266 | { |
73fbfcad | 9267 | if (dump_enabled_p ()) |
78c60e3d | 9268 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9269 | "not vectorized: stmt has volatile operands\n"); |
1825a1f3 IR |
9270 | |
9271 | return false; | |
9272 | } | |
b8698a0f L |
9273 | |
9274 | /* Skip stmts that do not need to be vectorized. In loops this is expected | |
8644a673 IR |
9275 | to include: |
9276 | - the COND_EXPR which is the loop exit condition | |
9277 | - any LABEL_EXPRs in the loop | |
b8698a0f | 9278 | - computations that are used only for array indexing or loop control. |
8644a673 | 9279 | In basic blocks we only analyze statements that are a part of some SLP |
83197f37 | 9280 | instance, therefore, all the statements are relevant. |
ebfd146a | 9281 | |
d092494c | 9282 | Pattern statement needs to be analyzed instead of the original statement |
83197f37 | 9283 | if the original statement is not relevant. Otherwise, we analyze both |
079c527f JJ |
9284 | statements. In basic blocks we are called from some SLP instance |
9285 | traversal, don't analyze pattern stmts instead, the pattern stmts | |
9286 | already will be part of SLP instance. */ | |
83197f37 IR |
9287 | |
9288 | pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info); | |
b8698a0f | 9289 | if (!STMT_VINFO_RELEVANT_P (stmt_info) |
8644a673 | 9290 | && !STMT_VINFO_LIVE_P (stmt_info)) |
ebfd146a | 9291 | { |
9d5e7640 | 9292 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
83197f37 | 9293 | && pattern_stmt |
9d5e7640 IR |
9294 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) |
9295 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
9296 | { | |
83197f37 | 9297 | /* Analyze PATTERN_STMT instead of the original stmt. */ |
9d5e7640 IR |
9298 | stmt = pattern_stmt; |
9299 | stmt_info = vinfo_for_stmt (pattern_stmt); | |
73fbfcad | 9300 | if (dump_enabled_p ()) |
9d5e7640 | 9301 | { |
78c60e3d SS |
9302 | dump_printf_loc (MSG_NOTE, vect_location, |
9303 | "==> examining pattern statement: "); | |
9304 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
9d5e7640 IR |
9305 | } |
9306 | } | |
9307 | else | |
9308 | { | |
73fbfcad | 9309 | if (dump_enabled_p ()) |
e645e942 | 9310 | dump_printf_loc (MSG_NOTE, vect_location, "irrelevant.\n"); |
ebfd146a | 9311 | |
9d5e7640 IR |
9312 | return true; |
9313 | } | |
8644a673 | 9314 | } |
83197f37 | 9315 | else if (STMT_VINFO_IN_PATTERN_P (stmt_info) |
079c527f | 9316 | && node == NULL |
83197f37 IR |
9317 | && pattern_stmt |
9318 | && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) | |
9319 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) | |
9320 | { | |
9321 | /* Analyze PATTERN_STMT too. */ | |
73fbfcad | 9322 | if (dump_enabled_p ()) |
83197f37 | 9323 | { |
78c60e3d SS |
9324 | dump_printf_loc (MSG_NOTE, vect_location, |
9325 | "==> examining pattern statement: "); | |
9326 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
83197f37 IR |
9327 | } |
9328 | ||
891ad31c RB |
9329 | if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node, |
9330 | node_instance)) | |
83197f37 IR |
9331 | return false; |
9332 | } | |
ebfd146a | 9333 | |
1107f3ae | 9334 | if (is_pattern_stmt_p (stmt_info) |
079c527f | 9335 | && node == NULL |
363477c0 | 9336 | && (pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))) |
1107f3ae | 9337 | { |
363477c0 | 9338 | gimple_stmt_iterator si; |
1107f3ae | 9339 | |
363477c0 JJ |
9340 | for (si = gsi_start (pattern_def_seq); !gsi_end_p (si); gsi_next (&si)) |
9341 | { | |
355fe088 | 9342 | gimple *pattern_def_stmt = gsi_stmt (si); |
363477c0 JJ |
9343 | if (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_def_stmt)) |
9344 | || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_def_stmt))) | |
9345 | { | |
9346 | /* Analyze def stmt of STMT if it's a pattern stmt. */ | |
73fbfcad | 9347 | if (dump_enabled_p ()) |
363477c0 | 9348 | { |
78c60e3d SS |
9349 | dump_printf_loc (MSG_NOTE, vect_location, |
9350 | "==> examining pattern def statement: "); | |
9351 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_def_stmt, 0); | |
363477c0 | 9352 | } |
1107f3ae | 9353 | |
363477c0 | 9354 | if (!vect_analyze_stmt (pattern_def_stmt, |
891ad31c | 9355 | need_to_vectorize, node, node_instance)) |
363477c0 JJ |
9356 | return false; |
9357 | } | |
9358 | } | |
9359 | } | |
1107f3ae | 9360 | |
8644a673 IR |
9361 | switch (STMT_VINFO_DEF_TYPE (stmt_info)) |
9362 | { | |
9363 | case vect_internal_def: | |
9364 | break; | |
ebfd146a | 9365 | |
8644a673 | 9366 | case vect_reduction_def: |
7c5222ff | 9367 | case vect_nested_cycle: |
14a61437 RB |
9368 | gcc_assert (!bb_vinfo |
9369 | && (relevance == vect_used_in_outer | |
9370 | || relevance == vect_used_in_outer_by_reduction | |
9371 | || relevance == vect_used_by_reduction | |
b28ead45 AH |
9372 | || relevance == vect_unused_in_scope |
9373 | || relevance == vect_used_only_live)); | |
8644a673 IR |
9374 | break; |
9375 | ||
9376 | case vect_induction_def: | |
e7baeb39 RB |
9377 | gcc_assert (!bb_vinfo); |
9378 | break; | |
9379 | ||
8644a673 IR |
9380 | case vect_constant_def: |
9381 | case vect_external_def: | |
9382 | case vect_unknown_def_type: | |
9383 | default: | |
9384 | gcc_unreachable (); | |
9385 | } | |
ebfd146a | 9386 | |
8644a673 | 9387 | if (STMT_VINFO_RELEVANT_P (stmt_info)) |
ebfd146a | 9388 | { |
8644a673 | 9389 | gcc_assert (!VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))); |
0136f8f0 AH |
9390 | gcc_assert (STMT_VINFO_VECTYPE (stmt_info) |
9391 | || (is_gimple_call (stmt) | |
9392 | && gimple_call_lhs (stmt) == NULL_TREE)); | |
8644a673 | 9393 | *need_to_vectorize = true; |
ebfd146a IR |
9394 | } |
9395 | ||
b1af7da6 RB |
9396 | if (PURE_SLP_STMT (stmt_info) && !node) |
9397 | { | |
9398 | dump_printf_loc (MSG_NOTE, vect_location, | |
9399 | "handled only by SLP analysis\n"); | |
9400 | return true; | |
9401 | } | |
9402 | ||
9403 | ok = true; | |
9404 | if (!bb_vinfo | |
9405 | && (STMT_VINFO_RELEVANT_P (stmt_info) | |
9406 | || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)) | |
9407 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
9408 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
9409 | || vectorizable_shift (stmt, NULL, NULL, node) | |
9410 | || vectorizable_operation (stmt, NULL, NULL, node) | |
9411 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
9412 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
9413 | || vectorizable_call (stmt, NULL, NULL, node) | |
9414 | || vectorizable_store (stmt, NULL, NULL, node) | |
891ad31c | 9415 | || vectorizable_reduction (stmt, NULL, NULL, node, node_instance) |
e7baeb39 | 9416 | || vectorizable_induction (stmt, NULL, NULL, node) |
42fd8198 IE |
9417 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
9418 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 RB |
9419 | else |
9420 | { | |
9421 | if (bb_vinfo) | |
9422 | ok = (vectorizable_simd_clone_call (stmt, NULL, NULL, node) | |
9423 | || vectorizable_conversion (stmt, NULL, NULL, node) | |
9424 | || vectorizable_shift (stmt, NULL, NULL, node) | |
9425 | || vectorizable_operation (stmt, NULL, NULL, node) | |
9426 | || vectorizable_assignment (stmt, NULL, NULL, node) | |
9427 | || vectorizable_load (stmt, NULL, NULL, node, NULL) | |
9428 | || vectorizable_call (stmt, NULL, NULL, node) | |
9429 | || vectorizable_store (stmt, NULL, NULL, node) | |
42fd8198 IE |
9430 | || vectorizable_condition (stmt, NULL, NULL, NULL, 0, node) |
9431 | || vectorizable_comparison (stmt, NULL, NULL, NULL, node)); | |
b1af7da6 | 9432 | } |
8644a673 IR |
9433 | |
9434 | if (!ok) | |
ebfd146a | 9435 | { |
73fbfcad | 9436 | if (dump_enabled_p ()) |
8644a673 | 9437 | { |
78c60e3d SS |
9438 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9439 | "not vectorized: relevant stmt not "); | |
9440 | dump_printf (MSG_MISSED_OPTIMIZATION, "supported: "); | |
9441 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); | |
8644a673 | 9442 | } |
b8698a0f | 9443 | |
ebfd146a IR |
9444 | return false; |
9445 | } | |
9446 | ||
a70d6342 IR |
9447 | if (bb_vinfo) |
9448 | return true; | |
9449 | ||
8644a673 IR |
9450 | /* Stmts that are (also) "live" (i.e. - that are used out of the loop) |
9451 | need extra handling, except for vectorizable reductions. */ | |
68a0f2ff RS |
9452 | if (STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
9453 | && !can_vectorize_live_stmts (stmt, NULL, node, NULL)) | |
ebfd146a | 9454 | { |
73fbfcad | 9455 | if (dump_enabled_p ()) |
8644a673 | 9456 | { |
78c60e3d | 9457 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
68a0f2ff | 9458 | "not vectorized: live stmt not supported: "); |
78c60e3d | 9459 | dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); |
8644a673 | 9460 | } |
b8698a0f | 9461 | |
8644a673 | 9462 | return false; |
ebfd146a IR |
9463 | } |
9464 | ||
ebfd146a IR |
9465 | return true; |
9466 | } | |
9467 | ||
9468 | ||
9469 | /* Function vect_transform_stmt. | |
9470 | ||
9471 | Create a vectorized stmt to replace STMT, and insert it at BSI. */ | |
9472 | ||
9473 | bool | |
355fe088 | 9474 | vect_transform_stmt (gimple *stmt, gimple_stmt_iterator *gsi, |
0d0293ac | 9475 | bool *grouped_store, slp_tree slp_node, |
ebfd146a IR |
9476 | slp_instance slp_node_instance) |
9477 | { | |
9478 | bool is_store = false; | |
355fe088 | 9479 | gimple *vec_stmt = NULL; |
ebfd146a | 9480 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
ebfd146a | 9481 | bool done; |
ebfd146a | 9482 | |
fce57248 | 9483 | gcc_assert (slp_node || !PURE_SLP_STMT (stmt_info)); |
355fe088 | 9484 | gimple *old_vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); |
225ce44b | 9485 | |
e57d9a82 RB |
9486 | bool nested_p = (STMT_VINFO_LOOP_VINFO (stmt_info) |
9487 | && nested_in_vect_loop_p | |
9488 | (LOOP_VINFO_LOOP (STMT_VINFO_LOOP_VINFO (stmt_info)), | |
9489 | stmt)); | |
9490 | ||
ebfd146a IR |
9491 | switch (STMT_VINFO_TYPE (stmt_info)) |
9492 | { | |
9493 | case type_demotion_vec_info_type: | |
ebfd146a | 9494 | case type_promotion_vec_info_type: |
ebfd146a IR |
9495 | case type_conversion_vec_info_type: |
9496 | done = vectorizable_conversion (stmt, gsi, &vec_stmt, slp_node); | |
9497 | gcc_assert (done); | |
9498 | break; | |
9499 | ||
9500 | case induc_vec_info_type: | |
e7baeb39 | 9501 | done = vectorizable_induction (stmt, gsi, &vec_stmt, slp_node); |
ebfd146a IR |
9502 | gcc_assert (done); |
9503 | break; | |
9504 | ||
9dc3f7de IR |
9505 | case shift_vec_info_type: |
9506 | done = vectorizable_shift (stmt, gsi, &vec_stmt, slp_node); | |
9507 | gcc_assert (done); | |
9508 | break; | |
9509 | ||
ebfd146a IR |
9510 | case op_vec_info_type: |
9511 | done = vectorizable_operation (stmt, gsi, &vec_stmt, slp_node); | |
9512 | gcc_assert (done); | |
9513 | break; | |
9514 | ||
9515 | case assignment_vec_info_type: | |
9516 | done = vectorizable_assignment (stmt, gsi, &vec_stmt, slp_node); | |
9517 | gcc_assert (done); | |
9518 | break; | |
9519 | ||
9520 | case load_vec_info_type: | |
b8698a0f | 9521 | done = vectorizable_load (stmt, gsi, &vec_stmt, slp_node, |
ebfd146a IR |
9522 | slp_node_instance); |
9523 | gcc_assert (done); | |
9524 | break; | |
9525 | ||
9526 | case store_vec_info_type: | |
9527 | done = vectorizable_store (stmt, gsi, &vec_stmt, slp_node); | |
9528 | gcc_assert (done); | |
0d0293ac | 9529 | if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && !slp_node) |
ebfd146a IR |
9530 | { |
9531 | /* In case of interleaving, the whole chain is vectorized when the | |
ff802fa1 | 9532 | last store in the chain is reached. Store stmts before the last |
ebfd146a IR |
9533 | one are skipped, and there vec_stmt_info shouldn't be freed |
9534 | meanwhile. */ | |
0d0293ac | 9535 | *grouped_store = true; |
f307441a RS |
9536 | stmt_vec_info group_info |
9537 | = vinfo_for_stmt (GROUP_FIRST_ELEMENT (stmt_info)); | |
9538 | if (GROUP_STORE_COUNT (group_info) == GROUP_SIZE (group_info)) | |
ebfd146a | 9539 | is_store = true; |
f307441a | 9540 | } |
ebfd146a IR |
9541 | else |
9542 | is_store = true; | |
9543 | break; | |
9544 | ||
9545 | case condition_vec_info_type: | |
f7e531cf | 9546 | done = vectorizable_condition (stmt, gsi, &vec_stmt, NULL, 0, slp_node); |
ebfd146a IR |
9547 | gcc_assert (done); |
9548 | break; | |
9549 | ||
42fd8198 IE |
9550 | case comparison_vec_info_type: |
9551 | done = vectorizable_comparison (stmt, gsi, &vec_stmt, NULL, slp_node); | |
9552 | gcc_assert (done); | |
9553 | break; | |
9554 | ||
ebfd146a | 9555 | case call_vec_info_type: |
190c2236 | 9556 | done = vectorizable_call (stmt, gsi, &vec_stmt, slp_node); |
039d9ea1 | 9557 | stmt = gsi_stmt (*gsi); |
ebfd146a IR |
9558 | break; |
9559 | ||
0136f8f0 AH |
9560 | case call_simd_clone_vec_info_type: |
9561 | done = vectorizable_simd_clone_call (stmt, gsi, &vec_stmt, slp_node); | |
9562 | stmt = gsi_stmt (*gsi); | |
9563 | break; | |
9564 | ||
ebfd146a | 9565 | case reduc_vec_info_type: |
891ad31c RB |
9566 | done = vectorizable_reduction (stmt, gsi, &vec_stmt, slp_node, |
9567 | slp_node_instance); | |
ebfd146a IR |
9568 | gcc_assert (done); |
9569 | break; | |
9570 | ||
9571 | default: | |
9572 | if (!STMT_VINFO_LIVE_P (stmt_info)) | |
9573 | { | |
73fbfcad | 9574 | if (dump_enabled_p ()) |
78c60e3d | 9575 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 9576 | "stmt not supported.\n"); |
ebfd146a IR |
9577 | gcc_unreachable (); |
9578 | } | |
9579 | } | |
9580 | ||
225ce44b RB |
9581 | /* Verify SLP vectorization doesn't mess with STMT_VINFO_VEC_STMT. |
9582 | This would break hybrid SLP vectorization. */ | |
9583 | if (slp_node) | |
d90f8440 RB |
9584 | gcc_assert (!vec_stmt |
9585 | && STMT_VINFO_VEC_STMT (stmt_info) == old_vec_stmt); | |
225ce44b | 9586 | |
ebfd146a IR |
9587 | /* Handle inner-loop stmts whose DEF is used in the loop-nest that |
9588 | is being vectorized, but outside the immediately enclosing loop. */ | |
9589 | if (vec_stmt | |
e57d9a82 | 9590 | && nested_p |
ebfd146a IR |
9591 | && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type |
9592 | && (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_outer | |
b8698a0f | 9593 | || STMT_VINFO_RELEVANT (stmt_info) == |
a70d6342 | 9594 | vect_used_in_outer_by_reduction)) |
ebfd146a | 9595 | { |
a70d6342 IR |
9596 | struct loop *innerloop = LOOP_VINFO_LOOP ( |
9597 | STMT_VINFO_LOOP_VINFO (stmt_info))->inner; | |
ebfd146a IR |
9598 | imm_use_iterator imm_iter; |
9599 | use_operand_p use_p; | |
9600 | tree scalar_dest; | |
355fe088 | 9601 | gimple *exit_phi; |
ebfd146a | 9602 | |
73fbfcad | 9603 | if (dump_enabled_p ()) |
78c60e3d | 9604 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 9605 | "Record the vdef for outer-loop vectorization.\n"); |
ebfd146a IR |
9606 | |
9607 | /* Find the relevant loop-exit phi-node, and reord the vec_stmt there | |
9608 | (to be used when vectorizing outer-loop stmts that use the DEF of | |
9609 | STMT). */ | |
9610 | if (gimple_code (stmt) == GIMPLE_PHI) | |
9611 | scalar_dest = PHI_RESULT (stmt); | |
9612 | else | |
9613 | scalar_dest = gimple_assign_lhs (stmt); | |
9614 | ||
9615 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) | |
9616 | { | |
9617 | if (!flow_bb_inside_loop_p (innerloop, gimple_bb (USE_STMT (use_p)))) | |
9618 | { | |
9619 | exit_phi = USE_STMT (use_p); | |
9620 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (exit_phi)) = vec_stmt; | |
9621 | } | |
9622 | } | |
9623 | } | |
9624 | ||
9625 | /* Handle stmts whose DEF is used outside the loop-nest that is | |
9626 | being vectorized. */ | |
68a0f2ff | 9627 | if (STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type) |
ebfd146a | 9628 | { |
68a0f2ff | 9629 | done = can_vectorize_live_stmts (stmt, gsi, slp_node, &vec_stmt); |
ebfd146a IR |
9630 | gcc_assert (done); |
9631 | } | |
9632 | ||
9633 | if (vec_stmt) | |
83197f37 | 9634 | STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; |
ebfd146a | 9635 | |
b8698a0f | 9636 | return is_store; |
ebfd146a IR |
9637 | } |
9638 | ||
9639 | ||
b8698a0f | 9640 | /* Remove a group of stores (for SLP or interleaving), free their |
ebfd146a IR |
9641 | stmt_vec_info. */ |
9642 | ||
9643 | void | |
355fe088 | 9644 | vect_remove_stores (gimple *first_stmt) |
ebfd146a | 9645 | { |
355fe088 TS |
9646 | gimple *next = first_stmt; |
9647 | gimple *tmp; | |
ebfd146a IR |
9648 | gimple_stmt_iterator next_si; |
9649 | ||
9650 | while (next) | |
9651 | { | |
78048b1c JJ |
9652 | stmt_vec_info stmt_info = vinfo_for_stmt (next); |
9653 | ||
9654 | tmp = GROUP_NEXT_ELEMENT (stmt_info); | |
9655 | if (is_pattern_stmt_p (stmt_info)) | |
9656 | next = STMT_VINFO_RELATED_STMT (stmt_info); | |
ebfd146a IR |
9657 | /* Free the attached stmt_vec_info and remove the stmt. */ |
9658 | next_si = gsi_for_stmt (next); | |
3d3f2249 | 9659 | unlink_stmt_vdef (next); |
ebfd146a | 9660 | gsi_remove (&next_si, true); |
3d3f2249 | 9661 | release_defs (next); |
ebfd146a IR |
9662 | free_stmt_vec_info (next); |
9663 | next = tmp; | |
9664 | } | |
9665 | } | |
9666 | ||
9667 | ||
9668 | /* Function new_stmt_vec_info. | |
9669 | ||
9670 | Create and initialize a new stmt_vec_info struct for STMT. */ | |
9671 | ||
9672 | stmt_vec_info | |
310213d4 | 9673 | new_stmt_vec_info (gimple *stmt, vec_info *vinfo) |
ebfd146a IR |
9674 | { |
9675 | stmt_vec_info res; | |
9676 | res = (stmt_vec_info) xcalloc (1, sizeof (struct _stmt_vec_info)); | |
9677 | ||
9678 | STMT_VINFO_TYPE (res) = undef_vec_info_type; | |
9679 | STMT_VINFO_STMT (res) = stmt; | |
310213d4 | 9680 | res->vinfo = vinfo; |
8644a673 | 9681 | STMT_VINFO_RELEVANT (res) = vect_unused_in_scope; |
ebfd146a IR |
9682 | STMT_VINFO_LIVE_P (res) = false; |
9683 | STMT_VINFO_VECTYPE (res) = NULL; | |
9684 | STMT_VINFO_VEC_STMT (res) = NULL; | |
4b5caab7 | 9685 | STMT_VINFO_VECTORIZABLE (res) = true; |
ebfd146a IR |
9686 | STMT_VINFO_IN_PATTERN_P (res) = false; |
9687 | STMT_VINFO_RELATED_STMT (res) = NULL; | |
363477c0 | 9688 | STMT_VINFO_PATTERN_DEF_SEQ (res) = NULL; |
ebfd146a | 9689 | STMT_VINFO_DATA_REF (res) = NULL; |
af29617a | 9690 | STMT_VINFO_VEC_REDUCTION_TYPE (res) = TREE_CODE_REDUCTION; |
7e16ce79 | 9691 | STMT_VINFO_VEC_CONST_COND_REDUC_CODE (res) = ERROR_MARK; |
ebfd146a | 9692 | |
ebfd146a IR |
9693 | if (gimple_code (stmt) == GIMPLE_PHI |
9694 | && is_loop_header_bb_p (gimple_bb (stmt))) | |
9695 | STMT_VINFO_DEF_TYPE (res) = vect_unknown_def_type; | |
9696 | else | |
8644a673 IR |
9697 | STMT_VINFO_DEF_TYPE (res) = vect_internal_def; |
9698 | ||
9771b263 | 9699 | STMT_VINFO_SAME_ALIGN_REFS (res).create (0); |
32e8bb8e | 9700 | STMT_SLP_TYPE (res) = loop_vect; |
78810bd3 RB |
9701 | STMT_VINFO_NUM_SLP_USES (res) = 0; |
9702 | ||
e14c1050 IR |
9703 | GROUP_FIRST_ELEMENT (res) = NULL; |
9704 | GROUP_NEXT_ELEMENT (res) = NULL; | |
9705 | GROUP_SIZE (res) = 0; | |
9706 | GROUP_STORE_COUNT (res) = 0; | |
9707 | GROUP_GAP (res) = 0; | |
9708 | GROUP_SAME_DR_STMT (res) = NULL; | |
ebfd146a IR |
9709 | |
9710 | return res; | |
9711 | } | |
9712 | ||
9713 | ||
9714 | /* Create a hash table for stmt_vec_info. */ | |
9715 | ||
9716 | void | |
9717 | init_stmt_vec_info_vec (void) | |
9718 | { | |
9771b263 DN |
9719 | gcc_assert (!stmt_vec_info_vec.exists ()); |
9720 | stmt_vec_info_vec.create (50); | |
ebfd146a IR |
9721 | } |
9722 | ||
9723 | ||
9724 | /* Free hash table for stmt_vec_info. */ | |
9725 | ||
9726 | void | |
9727 | free_stmt_vec_info_vec (void) | |
9728 | { | |
93675444 | 9729 | unsigned int i; |
3161455c | 9730 | stmt_vec_info info; |
93675444 JJ |
9731 | FOR_EACH_VEC_ELT (stmt_vec_info_vec, i, info) |
9732 | if (info != NULL) | |
3161455c | 9733 | free_stmt_vec_info (STMT_VINFO_STMT (info)); |
9771b263 DN |
9734 | gcc_assert (stmt_vec_info_vec.exists ()); |
9735 | stmt_vec_info_vec.release (); | |
ebfd146a IR |
9736 | } |
9737 | ||
9738 | ||
9739 | /* Free stmt vectorization related info. */ | |
9740 | ||
9741 | void | |
355fe088 | 9742 | free_stmt_vec_info (gimple *stmt) |
ebfd146a IR |
9743 | { |
9744 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
9745 | ||
9746 | if (!stmt_info) | |
9747 | return; | |
9748 | ||
78048b1c JJ |
9749 | /* Check if this statement has a related "pattern stmt" |
9750 | (introduced by the vectorizer during the pattern recognition | |
9751 | pass). Free pattern's stmt_vec_info and def stmt's stmt_vec_info | |
9752 | too. */ | |
9753 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) | |
9754 | { | |
9755 | stmt_vec_info patt_info | |
9756 | = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); | |
9757 | if (patt_info) | |
9758 | { | |
363477c0 | 9759 | gimple_seq seq = STMT_VINFO_PATTERN_DEF_SEQ (patt_info); |
355fe088 | 9760 | gimple *patt_stmt = STMT_VINFO_STMT (patt_info); |
f0281fde RB |
9761 | gimple_set_bb (patt_stmt, NULL); |
9762 | tree lhs = gimple_get_lhs (patt_stmt); | |
e6f5c25d | 9763 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde | 9764 | release_ssa_name (lhs); |
363477c0 JJ |
9765 | if (seq) |
9766 | { | |
9767 | gimple_stmt_iterator si; | |
9768 | for (si = gsi_start (seq); !gsi_end_p (si); gsi_next (&si)) | |
f0281fde | 9769 | { |
355fe088 | 9770 | gimple *seq_stmt = gsi_stmt (si); |
f0281fde | 9771 | gimple_set_bb (seq_stmt, NULL); |
7532abf2 | 9772 | lhs = gimple_get_lhs (seq_stmt); |
e6f5c25d | 9773 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
f0281fde RB |
9774 | release_ssa_name (lhs); |
9775 | free_stmt_vec_info (seq_stmt); | |
9776 | } | |
363477c0 | 9777 | } |
f0281fde | 9778 | free_stmt_vec_info (patt_stmt); |
78048b1c JJ |
9779 | } |
9780 | } | |
9781 | ||
9771b263 | 9782 | STMT_VINFO_SAME_ALIGN_REFS (stmt_info).release (); |
6c9e85fb | 9783 | STMT_VINFO_SIMD_CLONE_INFO (stmt_info).release (); |
ebfd146a IR |
9784 | set_vinfo_for_stmt (stmt, NULL); |
9785 | free (stmt_info); | |
9786 | } | |
9787 | ||
9788 | ||
bb67d9c7 | 9789 | /* Function get_vectype_for_scalar_type_and_size. |
ebfd146a | 9790 | |
bb67d9c7 | 9791 | Returns the vector type corresponding to SCALAR_TYPE and SIZE as supported |
ebfd146a IR |
9792 | by the target. */ |
9793 | ||
c803b2a9 | 9794 | tree |
86e36728 | 9795 | get_vectype_for_scalar_type_and_size (tree scalar_type, poly_uint64 size) |
ebfd146a | 9796 | { |
c7d97b28 | 9797 | tree orig_scalar_type = scalar_type; |
3bd8f481 | 9798 | scalar_mode inner_mode; |
ef4bddc2 | 9799 | machine_mode simd_mode; |
86e36728 | 9800 | poly_uint64 nunits; |
ebfd146a IR |
9801 | tree vectype; |
9802 | ||
3bd8f481 RS |
9803 | if (!is_int_mode (TYPE_MODE (scalar_type), &inner_mode) |
9804 | && !is_float_mode (TYPE_MODE (scalar_type), &inner_mode)) | |
ebfd146a IR |
9805 | return NULL_TREE; |
9806 | ||
3bd8f481 | 9807 | unsigned int nbytes = GET_MODE_SIZE (inner_mode); |
48f2e373 | 9808 | |
7b7b1813 RG |
9809 | /* For vector types of elements whose mode precision doesn't |
9810 | match their types precision we use a element type of mode | |
9811 | precision. The vectorization routines will have to make sure | |
48f2e373 RB |
9812 | they support the proper result truncation/extension. |
9813 | We also make sure to build vector types with INTEGER_TYPE | |
9814 | component type only. */ | |
6d7971b8 | 9815 | if (INTEGRAL_TYPE_P (scalar_type) |
48f2e373 RB |
9816 | && (GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type) |
9817 | || TREE_CODE (scalar_type) != INTEGER_TYPE)) | |
7b7b1813 RG |
9818 | scalar_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (inner_mode), |
9819 | TYPE_UNSIGNED (scalar_type)); | |
6d7971b8 | 9820 | |
ccbf5bb4 RG |
9821 | /* We shouldn't end up building VECTOR_TYPEs of non-scalar components. |
9822 | When the component mode passes the above test simply use a type | |
9823 | corresponding to that mode. The theory is that any use that | |
9824 | would cause problems with this will disable vectorization anyway. */ | |
dfc2e2ac | 9825 | else if (!SCALAR_FLOAT_TYPE_P (scalar_type) |
e67f39f7 | 9826 | && !INTEGRAL_TYPE_P (scalar_type)) |
60b95d28 RB |
9827 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1); |
9828 | ||
9829 | /* We can't build a vector type of elements with alignment bigger than | |
9830 | their size. */ | |
dfc2e2ac | 9831 | else if (nbytes < TYPE_ALIGN_UNIT (scalar_type)) |
aca43c6c JJ |
9832 | scalar_type = lang_hooks.types.type_for_mode (inner_mode, |
9833 | TYPE_UNSIGNED (scalar_type)); | |
ccbf5bb4 | 9834 | |
dfc2e2ac RB |
9835 | /* If we felt back to using the mode fail if there was |
9836 | no scalar type for it. */ | |
9837 | if (scalar_type == NULL_TREE) | |
9838 | return NULL_TREE; | |
9839 | ||
bb67d9c7 RG |
9840 | /* If no size was supplied use the mode the target prefers. Otherwise |
9841 | lookup a vector mode of the specified size. */ | |
86e36728 | 9842 | if (known_eq (size, 0U)) |
bb67d9c7 | 9843 | simd_mode = targetm.vectorize.preferred_simd_mode (inner_mode); |
86e36728 RS |
9844 | else if (!multiple_p (size, nbytes, &nunits) |
9845 | || !mode_for_vector (inner_mode, nunits).exists (&simd_mode)) | |
9da15d40 | 9846 | return NULL_TREE; |
4c8fd8ac | 9847 | /* NOTE: nunits == 1 is allowed to support single element vector types. */ |
86e36728 | 9848 | if (!multiple_p (GET_MODE_SIZE (simd_mode), nbytes, &nunits)) |
cc4b5170 | 9849 | return NULL_TREE; |
ebfd146a IR |
9850 | |
9851 | vectype = build_vector_type (scalar_type, nunits); | |
ebfd146a IR |
9852 | |
9853 | if (!VECTOR_MODE_P (TYPE_MODE (vectype)) | |
9854 | && !INTEGRAL_MODE_P (TYPE_MODE (vectype))) | |
451dabda | 9855 | return NULL_TREE; |
ebfd146a | 9856 | |
c7d97b28 RB |
9857 | /* Re-attach the address-space qualifier if we canonicalized the scalar |
9858 | type. */ | |
9859 | if (TYPE_ADDR_SPACE (orig_scalar_type) != TYPE_ADDR_SPACE (vectype)) | |
9860 | return build_qualified_type | |
9861 | (vectype, KEEP_QUAL_ADDR_SPACE (TYPE_QUALS (orig_scalar_type))); | |
9862 | ||
ebfd146a IR |
9863 | return vectype; |
9864 | } | |
9865 | ||
86e36728 | 9866 | poly_uint64 current_vector_size; |
bb67d9c7 RG |
9867 | |
9868 | /* Function get_vectype_for_scalar_type. | |
9869 | ||
9870 | Returns the vector type corresponding to SCALAR_TYPE as supported | |
9871 | by the target. */ | |
9872 | ||
9873 | tree | |
9874 | get_vectype_for_scalar_type (tree scalar_type) | |
9875 | { | |
9876 | tree vectype; | |
9877 | vectype = get_vectype_for_scalar_type_and_size (scalar_type, | |
9878 | current_vector_size); | |
9879 | if (vectype | |
86e36728 | 9880 | && known_eq (current_vector_size, 0U)) |
bb67d9c7 RG |
9881 | current_vector_size = GET_MODE_SIZE (TYPE_MODE (vectype)); |
9882 | return vectype; | |
9883 | } | |
9884 | ||
42fd8198 IE |
9885 | /* Function get_mask_type_for_scalar_type. |
9886 | ||
9887 | Returns the mask type corresponding to a result of comparison | |
9888 | of vectors of specified SCALAR_TYPE as supported by target. */ | |
9889 | ||
9890 | tree | |
9891 | get_mask_type_for_scalar_type (tree scalar_type) | |
9892 | { | |
9893 | tree vectype = get_vectype_for_scalar_type (scalar_type); | |
9894 | ||
9895 | if (!vectype) | |
9896 | return NULL; | |
9897 | ||
9898 | return build_truth_vector_type (TYPE_VECTOR_SUBPARTS (vectype), | |
9899 | current_vector_size); | |
9900 | } | |
9901 | ||
b690cc0f RG |
9902 | /* Function get_same_sized_vectype |
9903 | ||
9904 | Returns a vector type corresponding to SCALAR_TYPE of size | |
9905 | VECTOR_TYPE if supported by the target. */ | |
9906 | ||
9907 | tree | |
bb67d9c7 | 9908 | get_same_sized_vectype (tree scalar_type, tree vector_type) |
b690cc0f | 9909 | { |
2568d8a1 | 9910 | if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type)) |
9f47c7e5 IE |
9911 | return build_same_sized_truth_vector_type (vector_type); |
9912 | ||
bb67d9c7 RG |
9913 | return get_vectype_for_scalar_type_and_size |
9914 | (scalar_type, GET_MODE_SIZE (TYPE_MODE (vector_type))); | |
b690cc0f RG |
9915 | } |
9916 | ||
ebfd146a IR |
9917 | /* Function vect_is_simple_use. |
9918 | ||
9919 | Input: | |
81c40241 RB |
9920 | VINFO - the vect info of the loop or basic block that is being vectorized. |
9921 | OPERAND - operand in the loop or bb. | |
9922 | Output: | |
9923 | DEF_STMT - the defining stmt in case OPERAND is an SSA_NAME. | |
9924 | DT - the type of definition | |
ebfd146a IR |
9925 | |
9926 | Returns whether a stmt with OPERAND can be vectorized. | |
b8698a0f | 9927 | For loops, supportable operands are constants, loop invariants, and operands |
ff802fa1 | 9928 | that are defined by the current iteration of the loop. Unsupportable |
b8698a0f | 9929 | operands are those that are defined by a previous iteration of the loop (as |
a70d6342 IR |
9930 | is the case in reduction/induction computations). |
9931 | For basic blocks, supportable operands are constants and bb invariants. | |
9932 | For now, operands defined outside the basic block are not supported. */ | |
ebfd146a IR |
9933 | |
9934 | bool | |
81c40241 RB |
9935 | vect_is_simple_use (tree operand, vec_info *vinfo, |
9936 | gimple **def_stmt, enum vect_def_type *dt) | |
b8698a0f | 9937 | { |
ebfd146a | 9938 | *def_stmt = NULL; |
3fc356dc | 9939 | *dt = vect_unknown_def_type; |
b8698a0f | 9940 | |
73fbfcad | 9941 | if (dump_enabled_p ()) |
ebfd146a | 9942 | { |
78c60e3d SS |
9943 | dump_printf_loc (MSG_NOTE, vect_location, |
9944 | "vect_is_simple_use: operand "); | |
9945 | dump_generic_expr (MSG_NOTE, TDF_SLIM, operand); | |
e645e942 | 9946 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 9947 | } |
b8698a0f | 9948 | |
b758f602 | 9949 | if (CONSTANT_CLASS_P (operand)) |
ebfd146a IR |
9950 | { |
9951 | *dt = vect_constant_def; | |
9952 | return true; | |
9953 | } | |
b8698a0f | 9954 | |
ebfd146a IR |
9955 | if (is_gimple_min_invariant (operand)) |
9956 | { | |
8644a673 | 9957 | *dt = vect_external_def; |
ebfd146a IR |
9958 | return true; |
9959 | } | |
9960 | ||
ebfd146a IR |
9961 | if (TREE_CODE (operand) != SSA_NAME) |
9962 | { | |
73fbfcad | 9963 | if (dump_enabled_p ()) |
af29617a AH |
9964 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
9965 | "not ssa-name.\n"); | |
ebfd146a IR |
9966 | return false; |
9967 | } | |
b8698a0f | 9968 | |
3fc356dc | 9969 | if (SSA_NAME_IS_DEFAULT_DEF (operand)) |
ebfd146a | 9970 | { |
3fc356dc RB |
9971 | *dt = vect_external_def; |
9972 | return true; | |
ebfd146a IR |
9973 | } |
9974 | ||
3fc356dc | 9975 | *def_stmt = SSA_NAME_DEF_STMT (operand); |
73fbfcad | 9976 | if (dump_enabled_p ()) |
ebfd146a | 9977 | { |
78c60e3d SS |
9978 | dump_printf_loc (MSG_NOTE, vect_location, "def_stmt: "); |
9979 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, *def_stmt, 0); | |
ebfd146a IR |
9980 | } |
9981 | ||
61d371eb | 9982 | if (! vect_stmt_in_region_p (vinfo, *def_stmt)) |
8644a673 | 9983 | *dt = vect_external_def; |
ebfd146a IR |
9984 | else |
9985 | { | |
3fc356dc | 9986 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (*def_stmt); |
603cca93 | 9987 | *dt = STMT_VINFO_DEF_TYPE (stmt_vinfo); |
ebfd146a IR |
9988 | } |
9989 | ||
2e8ab70c RB |
9990 | if (dump_enabled_p ()) |
9991 | { | |
9992 | dump_printf_loc (MSG_NOTE, vect_location, "type of def: "); | |
9993 | switch (*dt) | |
9994 | { | |
9995 | case vect_uninitialized_def: | |
9996 | dump_printf (MSG_NOTE, "uninitialized\n"); | |
9997 | break; | |
9998 | case vect_constant_def: | |
9999 | dump_printf (MSG_NOTE, "constant\n"); | |
10000 | break; | |
10001 | case vect_external_def: | |
10002 | dump_printf (MSG_NOTE, "external\n"); | |
10003 | break; | |
10004 | case vect_internal_def: | |
10005 | dump_printf (MSG_NOTE, "internal\n"); | |
10006 | break; | |
10007 | case vect_induction_def: | |
10008 | dump_printf (MSG_NOTE, "induction\n"); | |
10009 | break; | |
10010 | case vect_reduction_def: | |
10011 | dump_printf (MSG_NOTE, "reduction\n"); | |
10012 | break; | |
10013 | case vect_double_reduction_def: | |
10014 | dump_printf (MSG_NOTE, "double reduction\n"); | |
10015 | break; | |
10016 | case vect_nested_cycle: | |
10017 | dump_printf (MSG_NOTE, "nested cycle\n"); | |
10018 | break; | |
10019 | case vect_unknown_def_type: | |
10020 | dump_printf (MSG_NOTE, "unknown\n"); | |
10021 | break; | |
10022 | } | |
10023 | } | |
10024 | ||
81c40241 | 10025 | if (*dt == vect_unknown_def_type) |
ebfd146a | 10026 | { |
73fbfcad | 10027 | if (dump_enabled_p ()) |
78c60e3d | 10028 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 10029 | "Unsupported pattern.\n"); |
ebfd146a IR |
10030 | return false; |
10031 | } | |
10032 | ||
ebfd146a IR |
10033 | switch (gimple_code (*def_stmt)) |
10034 | { | |
10035 | case GIMPLE_PHI: | |
ebfd146a | 10036 | case GIMPLE_ASSIGN: |
ebfd146a | 10037 | case GIMPLE_CALL: |
81c40241 | 10038 | break; |
ebfd146a | 10039 | default: |
73fbfcad | 10040 | if (dump_enabled_p ()) |
78c60e3d | 10041 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 10042 | "unsupported defining stmt:\n"); |
ebfd146a IR |
10043 | return false; |
10044 | } | |
10045 | ||
10046 | return true; | |
10047 | } | |
10048 | ||
81c40241 | 10049 | /* Function vect_is_simple_use. |
b690cc0f | 10050 | |
81c40241 | 10051 | Same as vect_is_simple_use but also determines the vector operand |
b690cc0f RG |
10052 | type of OPERAND and stores it to *VECTYPE. If the definition of |
10053 | OPERAND is vect_uninitialized_def, vect_constant_def or | |
10054 | vect_external_def *VECTYPE will be set to NULL_TREE and the caller | |
10055 | is responsible to compute the best suited vector type for the | |
10056 | scalar operand. */ | |
10057 | ||
10058 | bool | |
81c40241 RB |
10059 | vect_is_simple_use (tree operand, vec_info *vinfo, |
10060 | gimple **def_stmt, enum vect_def_type *dt, tree *vectype) | |
b690cc0f | 10061 | { |
81c40241 | 10062 | if (!vect_is_simple_use (operand, vinfo, def_stmt, dt)) |
b690cc0f RG |
10063 | return false; |
10064 | ||
10065 | /* Now get a vector type if the def is internal, otherwise supply | |
10066 | NULL_TREE and leave it up to the caller to figure out a proper | |
10067 | type for the use stmt. */ | |
10068 | if (*dt == vect_internal_def | |
10069 | || *dt == vect_induction_def | |
10070 | || *dt == vect_reduction_def | |
10071 | || *dt == vect_double_reduction_def | |
10072 | || *dt == vect_nested_cycle) | |
10073 | { | |
10074 | stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt); | |
83197f37 IR |
10075 | |
10076 | if (STMT_VINFO_IN_PATTERN_P (stmt_info) | |
10077 | && !STMT_VINFO_RELEVANT (stmt_info) | |
10078 | && !STMT_VINFO_LIVE_P (stmt_info)) | |
b690cc0f | 10079 | stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); |
83197f37 | 10080 | |
b690cc0f RG |
10081 | *vectype = STMT_VINFO_VECTYPE (stmt_info); |
10082 | gcc_assert (*vectype != NULL_TREE); | |
10083 | } | |
10084 | else if (*dt == vect_uninitialized_def | |
10085 | || *dt == vect_constant_def | |
10086 | || *dt == vect_external_def) | |
10087 | *vectype = NULL_TREE; | |
10088 | else | |
10089 | gcc_unreachable (); | |
10090 | ||
10091 | return true; | |
10092 | } | |
10093 | ||
ebfd146a IR |
10094 | |
10095 | /* Function supportable_widening_operation | |
10096 | ||
b8698a0f L |
10097 | Check whether an operation represented by the code CODE is a |
10098 | widening operation that is supported by the target platform in | |
b690cc0f RG |
10099 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
10100 | producing a result of type VECTYPE_OUT). | |
b8698a0f | 10101 | |
ebfd146a IR |
10102 | Widening operations we currently support are NOP (CONVERT), FLOAT |
10103 | and WIDEN_MULT. This function checks if these operations are supported | |
10104 | by the target platform either directly (via vector tree-codes), or via | |
10105 | target builtins. | |
10106 | ||
10107 | Output: | |
b8698a0f L |
10108 | - CODE1 and CODE2 are codes of vector operations to be used when |
10109 | vectorizing the operation, if available. | |
ebfd146a IR |
10110 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
10111 | case of multi-step conversion (like char->short->int - in that case | |
10112 | MULTI_STEP_CVT will be 1). | |
b8698a0f L |
10113 | - INTERM_TYPES contains the intermediate type required to perform the |
10114 | widening operation (short in the above example). */ | |
ebfd146a IR |
10115 | |
10116 | bool | |
355fe088 | 10117 | supportable_widening_operation (enum tree_code code, gimple *stmt, |
b690cc0f | 10118 | tree vectype_out, tree vectype_in, |
ebfd146a IR |
10119 | enum tree_code *code1, enum tree_code *code2, |
10120 | int *multi_step_cvt, | |
9771b263 | 10121 | vec<tree> *interm_types) |
ebfd146a IR |
10122 | { |
10123 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
10124 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4ef69dfc | 10125 | struct loop *vect_loop = NULL; |
ef4bddc2 | 10126 | machine_mode vec_mode; |
81f40b79 | 10127 | enum insn_code icode1, icode2; |
ebfd146a | 10128 | optab optab1, optab2; |
b690cc0f RG |
10129 | tree vectype = vectype_in; |
10130 | tree wide_vectype = vectype_out; | |
ebfd146a | 10131 | enum tree_code c1, c2; |
4a00c761 JJ |
10132 | int i; |
10133 | tree prev_type, intermediate_type; | |
ef4bddc2 | 10134 | machine_mode intermediate_mode, prev_mode; |
4a00c761 | 10135 | optab optab3, optab4; |
ebfd146a | 10136 | |
4a00c761 | 10137 | *multi_step_cvt = 0; |
4ef69dfc IR |
10138 | if (loop_info) |
10139 | vect_loop = LOOP_VINFO_LOOP (loop_info); | |
10140 | ||
ebfd146a IR |
10141 | switch (code) |
10142 | { | |
10143 | case WIDEN_MULT_EXPR: | |
6ae6116f RH |
10144 | /* The result of a vectorized widening operation usually requires |
10145 | two vectors (because the widened results do not fit into one vector). | |
10146 | The generated vector results would normally be expected to be | |
10147 | generated in the same order as in the original scalar computation, | |
10148 | i.e. if 8 results are generated in each vector iteration, they are | |
10149 | to be organized as follows: | |
10150 | vect1: [res1,res2,res3,res4], | |
10151 | vect2: [res5,res6,res7,res8]. | |
10152 | ||
10153 | However, in the special case that the result of the widening | |
10154 | operation is used in a reduction computation only, the order doesn't | |
10155 | matter (because when vectorizing a reduction we change the order of | |
10156 | the computation). Some targets can take advantage of this and | |
10157 | generate more efficient code. For example, targets like Altivec, | |
10158 | that support widen_mult using a sequence of {mult_even,mult_odd} | |
10159 | generate the following vectors: | |
10160 | vect1: [res1,res3,res5,res7], | |
10161 | vect2: [res2,res4,res6,res8]. | |
10162 | ||
10163 | When vectorizing outer-loops, we execute the inner-loop sequentially | |
10164 | (each vectorized inner-loop iteration contributes to VF outer-loop | |
10165 | iterations in parallel). We therefore don't allow to change the | |
10166 | order of the computation in the inner-loop during outer-loop | |
10167 | vectorization. */ | |
10168 | /* TODO: Another case in which order doesn't *really* matter is when we | |
10169 | widen and then contract again, e.g. (short)((int)x * y >> 8). | |
10170 | Normally, pack_trunc performs an even/odd permute, whereas the | |
10171 | repack from an even/odd expansion would be an interleave, which | |
10172 | would be significantly simpler for e.g. AVX2. */ | |
10173 | /* In any case, in order to avoid duplicating the code below, recurse | |
10174 | on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values | |
10175 | are properly set up for the caller. If we fail, we'll continue with | |
10176 | a VEC_WIDEN_MULT_LO/HI_EXPR check. */ | |
10177 | if (vect_loop | |
10178 | && STMT_VINFO_RELEVANT (stmt_info) == vect_used_by_reduction | |
10179 | && !nested_in_vect_loop_p (vect_loop, stmt) | |
10180 | && supportable_widening_operation (VEC_WIDEN_MULT_EVEN_EXPR, | |
10181 | stmt, vectype_out, vectype_in, | |
a86ec597 RH |
10182 | code1, code2, multi_step_cvt, |
10183 | interm_types)) | |
ebc047a2 CH |
10184 | { |
10185 | /* Elements in a vector with vect_used_by_reduction property cannot | |
10186 | be reordered if the use chain with this property does not have the | |
10187 | same operation. One such an example is s += a * b, where elements | |
10188 | in a and b cannot be reordered. Here we check if the vector defined | |
10189 | by STMT is only directly used in the reduction statement. */ | |
10190 | tree lhs = gimple_assign_lhs (stmt); | |
10191 | use_operand_p dummy; | |
355fe088 | 10192 | gimple *use_stmt; |
ebc047a2 CH |
10193 | stmt_vec_info use_stmt_info = NULL; |
10194 | if (single_imm_use (lhs, &dummy, &use_stmt) | |
10195 | && (use_stmt_info = vinfo_for_stmt (use_stmt)) | |
10196 | && STMT_VINFO_DEF_TYPE (use_stmt_info) == vect_reduction_def) | |
10197 | return true; | |
10198 | } | |
4a00c761 JJ |
10199 | c1 = VEC_WIDEN_MULT_LO_EXPR; |
10200 | c2 = VEC_WIDEN_MULT_HI_EXPR; | |
ebfd146a IR |
10201 | break; |
10202 | ||
81c40241 RB |
10203 | case DOT_PROD_EXPR: |
10204 | c1 = DOT_PROD_EXPR; | |
10205 | c2 = DOT_PROD_EXPR; | |
10206 | break; | |
10207 | ||
10208 | case SAD_EXPR: | |
10209 | c1 = SAD_EXPR; | |
10210 | c2 = SAD_EXPR; | |
10211 | break; | |
10212 | ||
6ae6116f RH |
10213 | case VEC_WIDEN_MULT_EVEN_EXPR: |
10214 | /* Support the recursion induced just above. */ | |
10215 | c1 = VEC_WIDEN_MULT_EVEN_EXPR; | |
10216 | c2 = VEC_WIDEN_MULT_ODD_EXPR; | |
10217 | break; | |
10218 | ||
36ba4aae | 10219 | case WIDEN_LSHIFT_EXPR: |
4a00c761 JJ |
10220 | c1 = VEC_WIDEN_LSHIFT_LO_EXPR; |
10221 | c2 = VEC_WIDEN_LSHIFT_HI_EXPR; | |
36ba4aae IR |
10222 | break; |
10223 | ||
ebfd146a | 10224 | CASE_CONVERT: |
4a00c761 JJ |
10225 | c1 = VEC_UNPACK_LO_EXPR; |
10226 | c2 = VEC_UNPACK_HI_EXPR; | |
ebfd146a IR |
10227 | break; |
10228 | ||
10229 | case FLOAT_EXPR: | |
4a00c761 JJ |
10230 | c1 = VEC_UNPACK_FLOAT_LO_EXPR; |
10231 | c2 = VEC_UNPACK_FLOAT_HI_EXPR; | |
ebfd146a IR |
10232 | break; |
10233 | ||
10234 | case FIX_TRUNC_EXPR: | |
10235 | /* ??? Not yet implemented due to missing VEC_UNPACK_FIX_TRUNC_HI_EXPR/ | |
10236 | VEC_UNPACK_FIX_TRUNC_LO_EXPR tree codes and optabs used for | |
10237 | computing the operation. */ | |
10238 | return false; | |
10239 | ||
10240 | default: | |
10241 | gcc_unreachable (); | |
10242 | } | |
10243 | ||
6ae6116f | 10244 | if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR) |
6b4db501 | 10245 | std::swap (c1, c2); |
4a00c761 | 10246 | |
ebfd146a IR |
10247 | if (code == FIX_TRUNC_EXPR) |
10248 | { | |
10249 | /* The signedness is determined from output operand. */ | |
b690cc0f RG |
10250 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
10251 | optab2 = optab_for_tree_code (c2, vectype_out, optab_default); | |
ebfd146a IR |
10252 | } |
10253 | else | |
10254 | { | |
10255 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
10256 | optab2 = optab_for_tree_code (c2, vectype, optab_default); | |
10257 | } | |
10258 | ||
10259 | if (!optab1 || !optab2) | |
10260 | return false; | |
10261 | ||
10262 | vec_mode = TYPE_MODE (vectype); | |
947131ba RS |
10263 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing |
10264 | || (icode2 = optab_handler (optab2, vec_mode)) == CODE_FOR_nothing) | |
ebfd146a IR |
10265 | return false; |
10266 | ||
4a00c761 JJ |
10267 | *code1 = c1; |
10268 | *code2 = c2; | |
10269 | ||
10270 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
10271 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff IE |
10272 | /* For scalar masks we may have different boolean |
10273 | vector types having the same QImode. Thus we | |
10274 | add additional check for elements number. */ | |
10275 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
928686b1 RS |
10276 | || known_eq (TYPE_VECTOR_SUBPARTS (vectype), |
10277 | TYPE_VECTOR_SUBPARTS (wide_vectype) * 2)); | |
4a00c761 | 10278 | |
b8698a0f | 10279 | /* Check if it's a multi-step conversion that can be done using intermediate |
ebfd146a | 10280 | types. */ |
ebfd146a | 10281 | |
4a00c761 JJ |
10282 | prev_type = vectype; |
10283 | prev_mode = vec_mode; | |
b8698a0f | 10284 | |
4a00c761 JJ |
10285 | if (!CONVERT_EXPR_CODE_P (code)) |
10286 | return false; | |
b8698a0f | 10287 | |
4a00c761 JJ |
10288 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
10289 | intermediate steps in promotion sequence. We try | |
10290 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do | |
10291 | not. */ | |
9771b263 | 10292 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
10293 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
10294 | { | |
10295 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
10296 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
10297 | { | |
7cfb4d93 | 10298 | intermediate_type = vect_halve_mask_nunits (prev_type); |
3ae0661a IE |
10299 | if (intermediate_mode != TYPE_MODE (intermediate_type)) |
10300 | return false; | |
10301 | } | |
10302 | else | |
10303 | intermediate_type | |
10304 | = lang_hooks.types.type_for_mode (intermediate_mode, | |
10305 | TYPE_UNSIGNED (prev_type)); | |
10306 | ||
4a00c761 JJ |
10307 | optab3 = optab_for_tree_code (c1, intermediate_type, optab_default); |
10308 | optab4 = optab_for_tree_code (c2, intermediate_type, optab_default); | |
10309 | ||
10310 | if (!optab3 || !optab4 | |
10311 | || (icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing | |
10312 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
10313 | || (icode2 = optab_handler (optab2, prev_mode)) == CODE_FOR_nothing | |
10314 | || insn_data[icode2].operand[0].mode != intermediate_mode | |
10315 | || ((icode1 = optab_handler (optab3, intermediate_mode)) | |
10316 | == CODE_FOR_nothing) | |
10317 | || ((icode2 = optab_handler (optab4, intermediate_mode)) | |
10318 | == CODE_FOR_nothing)) | |
10319 | break; | |
ebfd146a | 10320 | |
9771b263 | 10321 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
10322 | (*multi_step_cvt)++; |
10323 | ||
10324 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (wide_vectype) | |
10325 | && insn_data[icode2].operand[0].mode == TYPE_MODE (wide_vectype)) | |
5e8d6dff | 10326 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
928686b1 RS |
10327 | || known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type), |
10328 | TYPE_VECTOR_SUBPARTS (wide_vectype) * 2)); | |
4a00c761 JJ |
10329 | |
10330 | prev_type = intermediate_type; | |
10331 | prev_mode = intermediate_mode; | |
ebfd146a IR |
10332 | } |
10333 | ||
9771b263 | 10334 | interm_types->release (); |
4a00c761 | 10335 | return false; |
ebfd146a IR |
10336 | } |
10337 | ||
10338 | ||
10339 | /* Function supportable_narrowing_operation | |
10340 | ||
b8698a0f L |
10341 | Check whether an operation represented by the code CODE is a |
10342 | narrowing operation that is supported by the target platform in | |
b690cc0f RG |
10343 | vector form (i.e., when operating on arguments of type VECTYPE_IN |
10344 | and producing a result of type VECTYPE_OUT). | |
b8698a0f | 10345 | |
ebfd146a | 10346 | Narrowing operations we currently support are NOP (CONVERT) and |
ff802fa1 | 10347 | FIX_TRUNC. This function checks if these operations are supported by |
ebfd146a IR |
10348 | the target platform directly via vector tree-codes. |
10349 | ||
10350 | Output: | |
b8698a0f L |
10351 | - CODE1 is the code of a vector operation to be used when |
10352 | vectorizing the operation, if available. | |
ebfd146a IR |
10353 | - MULTI_STEP_CVT determines the number of required intermediate steps in |
10354 | case of multi-step conversion (like int->short->char - in that case | |
10355 | MULTI_STEP_CVT will be 1). | |
10356 | - INTERM_TYPES contains the intermediate type required to perform the | |
b8698a0f | 10357 | narrowing operation (short in the above example). */ |
ebfd146a IR |
10358 | |
10359 | bool | |
10360 | supportable_narrowing_operation (enum tree_code code, | |
b690cc0f | 10361 | tree vectype_out, tree vectype_in, |
ebfd146a | 10362 | enum tree_code *code1, int *multi_step_cvt, |
9771b263 | 10363 | vec<tree> *interm_types) |
ebfd146a | 10364 | { |
ef4bddc2 | 10365 | machine_mode vec_mode; |
ebfd146a IR |
10366 | enum insn_code icode1; |
10367 | optab optab1, interm_optab; | |
b690cc0f RG |
10368 | tree vectype = vectype_in; |
10369 | tree narrow_vectype = vectype_out; | |
ebfd146a | 10370 | enum tree_code c1; |
3ae0661a | 10371 | tree intermediate_type, prev_type; |
ef4bddc2 | 10372 | machine_mode intermediate_mode, prev_mode; |
ebfd146a | 10373 | int i; |
4a00c761 | 10374 | bool uns; |
ebfd146a | 10375 | |
4a00c761 | 10376 | *multi_step_cvt = 0; |
ebfd146a IR |
10377 | switch (code) |
10378 | { | |
10379 | CASE_CONVERT: | |
10380 | c1 = VEC_PACK_TRUNC_EXPR; | |
10381 | break; | |
10382 | ||
10383 | case FIX_TRUNC_EXPR: | |
10384 | c1 = VEC_PACK_FIX_TRUNC_EXPR; | |
10385 | break; | |
10386 | ||
10387 | case FLOAT_EXPR: | |
10388 | /* ??? Not yet implemented due to missing VEC_PACK_FLOAT_EXPR | |
10389 | tree code and optabs used for computing the operation. */ | |
10390 | return false; | |
10391 | ||
10392 | default: | |
10393 | gcc_unreachable (); | |
10394 | } | |
10395 | ||
10396 | if (code == FIX_TRUNC_EXPR) | |
10397 | /* The signedness is determined from output operand. */ | |
b690cc0f | 10398 | optab1 = optab_for_tree_code (c1, vectype_out, optab_default); |
ebfd146a IR |
10399 | else |
10400 | optab1 = optab_for_tree_code (c1, vectype, optab_default); | |
10401 | ||
10402 | if (!optab1) | |
10403 | return false; | |
10404 | ||
10405 | vec_mode = TYPE_MODE (vectype); | |
947131ba | 10406 | if ((icode1 = optab_handler (optab1, vec_mode)) == CODE_FOR_nothing) |
ebfd146a IR |
10407 | return false; |
10408 | ||
4a00c761 JJ |
10409 | *code1 = c1; |
10410 | ||
10411 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff IE |
10412 | /* For scalar masks we may have different boolean |
10413 | vector types having the same QImode. Thus we | |
10414 | add additional check for elements number. */ | |
10415 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) | |
928686b1 RS |
10416 | || known_eq (TYPE_VECTOR_SUBPARTS (vectype) * 2, |
10417 | TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 | 10418 | |
ebfd146a IR |
10419 | /* Check if it's a multi-step conversion that can be done using intermediate |
10420 | types. */ | |
4a00c761 | 10421 | prev_mode = vec_mode; |
3ae0661a | 10422 | prev_type = vectype; |
4a00c761 JJ |
10423 | if (code == FIX_TRUNC_EXPR) |
10424 | uns = TYPE_UNSIGNED (vectype_out); | |
10425 | else | |
10426 | uns = TYPE_UNSIGNED (vectype); | |
10427 | ||
10428 | /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer | |
10429 | conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more | |
10430 | costly than signed. */ | |
10431 | if (code == FIX_TRUNC_EXPR && uns) | |
10432 | { | |
10433 | enum insn_code icode2; | |
10434 | ||
10435 | intermediate_type | |
10436 | = lang_hooks.types.type_for_mode (TYPE_MODE (vectype_out), 0); | |
10437 | interm_optab | |
10438 | = optab_for_tree_code (c1, intermediate_type, optab_default); | |
2225b9f2 | 10439 | if (interm_optab != unknown_optab |
4a00c761 JJ |
10440 | && (icode2 = optab_handler (optab1, vec_mode)) != CODE_FOR_nothing |
10441 | && insn_data[icode1].operand[0].mode | |
10442 | == insn_data[icode2].operand[0].mode) | |
10443 | { | |
10444 | uns = false; | |
10445 | optab1 = interm_optab; | |
10446 | icode1 = icode2; | |
10447 | } | |
10448 | } | |
ebfd146a | 10449 | |
4a00c761 JJ |
10450 | /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS |
10451 | intermediate steps in promotion sequence. We try | |
10452 | MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */ | |
9771b263 | 10453 | interm_types->create (MAX_INTERM_CVT_STEPS); |
4a00c761 JJ |
10454 | for (i = 0; i < MAX_INTERM_CVT_STEPS; i++) |
10455 | { | |
10456 | intermediate_mode = insn_data[icode1].operand[0].mode; | |
3ae0661a IE |
10457 | if (VECTOR_BOOLEAN_TYPE_P (prev_type)) |
10458 | { | |
7cfb4d93 | 10459 | intermediate_type = vect_double_mask_nunits (prev_type); |
3ae0661a | 10460 | if (intermediate_mode != TYPE_MODE (intermediate_type)) |
7cfb4d93 | 10461 | return false; |
3ae0661a IE |
10462 | } |
10463 | else | |
10464 | intermediate_type | |
10465 | = lang_hooks.types.type_for_mode (intermediate_mode, uns); | |
4a00c761 JJ |
10466 | interm_optab |
10467 | = optab_for_tree_code (VEC_PACK_TRUNC_EXPR, intermediate_type, | |
10468 | optab_default); | |
10469 | if (!interm_optab | |
10470 | || ((icode1 = optab_handler (optab1, prev_mode)) == CODE_FOR_nothing) | |
10471 | || insn_data[icode1].operand[0].mode != intermediate_mode | |
10472 | || ((icode1 = optab_handler (interm_optab, intermediate_mode)) | |
10473 | == CODE_FOR_nothing)) | |
10474 | break; | |
10475 | ||
9771b263 | 10476 | interm_types->quick_push (intermediate_type); |
4a00c761 JJ |
10477 | (*multi_step_cvt)++; |
10478 | ||
10479 | if (insn_data[icode1].operand[0].mode == TYPE_MODE (narrow_vectype)) | |
5e8d6dff | 10480 | return (!VECTOR_BOOLEAN_TYPE_P (vectype) |
928686b1 RS |
10481 | || known_eq (TYPE_VECTOR_SUBPARTS (intermediate_type) * 2, |
10482 | TYPE_VECTOR_SUBPARTS (narrow_vectype))); | |
4a00c761 JJ |
10483 | |
10484 | prev_mode = intermediate_mode; | |
3ae0661a | 10485 | prev_type = intermediate_type; |
4a00c761 | 10486 | optab1 = interm_optab; |
ebfd146a IR |
10487 | } |
10488 | ||
9771b263 | 10489 | interm_types->release (); |
4a00c761 | 10490 | return false; |
ebfd146a | 10491 | } |
7cfb4d93 RS |
10492 | |
10493 | /* Generate and return a statement that sets vector mask MASK such that | |
10494 | MASK[I] is true iff J + START_INDEX < END_INDEX for all J <= I. */ | |
10495 | ||
10496 | gcall * | |
10497 | vect_gen_while (tree mask, tree start_index, tree end_index) | |
10498 | { | |
10499 | tree cmp_type = TREE_TYPE (start_index); | |
10500 | tree mask_type = TREE_TYPE (mask); | |
10501 | gcc_checking_assert (direct_internal_fn_supported_p (IFN_WHILE_ULT, | |
10502 | cmp_type, mask_type, | |
10503 | OPTIMIZE_FOR_SPEED)); | |
10504 | gcall *call = gimple_build_call_internal (IFN_WHILE_ULT, 3, | |
10505 | start_index, end_index, | |
10506 | build_zero_cst (mask_type)); | |
10507 | gimple_call_set_lhs (call, mask); | |
10508 | return call; | |
10509 | } | |
535e7c11 RS |
10510 | |
10511 | /* Generate a vector mask of type MASK_TYPE for which index I is false iff | |
10512 | J + START_INDEX < END_INDEX for all J <= I. Add the statements to SEQ. */ | |
10513 | ||
10514 | tree | |
10515 | vect_gen_while_not (gimple_seq *seq, tree mask_type, tree start_index, | |
10516 | tree end_index) | |
10517 | { | |
10518 | tree tmp = make_ssa_name (mask_type); | |
10519 | gcall *call = vect_gen_while (tmp, start_index, end_index); | |
10520 | gimple_seq_add_stmt (seq, call); | |
10521 | return gimple_build (seq, BIT_NOT_EXPR, mask_type, tmp); | |
10522 | } |